ALGORITHMS AND
PROTOCOLS FOR
WIRELESS SENSOR
NETWORKS
WILEY SERIES ON PARALLEL
AND DISTRIBUTED COMPUTING
Editor: Albert Y. Zomaya
A complete list of titles in this series appears at the end of this volume.
ALGORITHMS AND
PROTOCOLS FOR
WIRELESS SENSOR
NETWORKS
Edited by
Azzedine Boukerche, PhD
University of Ottawa
Ottawa, Canada
Copyright © 2009 by John Wiley & Sons, Inc. All rights reserved
Published by John Wiley & Sons, Inc., Hoboken, New Jersey
Published simultaneously in Canada
No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or
by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as
permitted under Section 107 or 108 of the 1976 United States Copyright Act, without either the prior
written permission of the Publisher, or authorization through payment of the appropriate per-copy fee to
the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, (978) 750-8400, fax
(978) 750-4470, or on the web at www.copyright.com. Requests to the Publisher for permission should
be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ
07030, (201) 748-6011, fax (201) 748-6008, or online at http://www.wiley.com/go/permission.
Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best efforts in
preparing this book, they make no representations or warranties with respect to the accuracy or
completeness of the contents of this book and specifically disclaim any implied warranties of
merchantability or fitness for a particular purpose. No warranty may be created or extended by sales
representatives or written sales materials. The advice and strategies contained herein may not be suitable
for your situation. You should consult with a professional where appropriate. Neither the publisher nor
author shall be liable for any loss of profit or any other commercial damages, including but not limited to
special, incidental, consequential, or other damages.
For general information on our other products and services or for technical support, please contact our
Customer Care Department within the United States at (800) 762-2974, outside the United States at (317)
572-3993 or fax (317) 572-4002.
Wiley also publishes its books in a variety of electronic formats. Some content that appears in print may
not be available in electronic formats. For more information about Wiley products, visit our web site at
www.wiley.com.
Library of Congress Cataloging-in-Publication Data
Algorithms and protocols for wireless sensor networks / edited by Azzedine Boukerche.
p. cm.
Includes bibliographical references and index.
ISBN 978-0-471-79813-2 (cloth)
1.
Sensor networks. 2.
Computer network protocols. 3.
Computer algorithms.
I. Boukerche, Azzedine.
TK7872.D48A422 2008
681’.2–dc22
2008016869
Printed in the United States of America
This book is dedicated to my parents and my family who have always been there with me.
Love you all.
Azzedine Boukerche
CONTENTS
Preface
About the Editor
Contributors
1. Algorithms for Wireless Sensor Networks: Present and Future
ix
xiii
xv
1
Azzedine Boukerche, Eduardo F. Nakamura, and Antonio A. F. Loureiro
2. Heterogeneous Wireless Sensor Networks
21
Violet R. Syrotiuk, Bing Li, and Angela M. Mielke
3. Epidemic Models, Algorithms, and Protocols in Wireless Sensor
and Ad Hoc Networks
51
Pradip De and Sajal K. Das
4. Modeling Sensor Networks
77
Stefan Schmid and Roger Wattenhofer
5. Spatiotemporal Correlation Theory for Wireless Sensor Networks
105
Özgür B. Akan
6. A Taxonomy of Routing Protocols in Sensor Networks
129
Azzedine Boukerche, Mohammad Z. Ahmad, Damla Turgut,
and Begumhan Turgut
7. Clustering in Wireless Sensor Networks: A Graph Theory Perspective 161
Nidal Nasser and Liliana M. Arboleda
8. Position-Based Routing for Sensor Networks: Approaches
and Obstacles
195
Marwan M. Fayed and Hussein T. Mouftah
9. Node Positioning for Increased Dependability of Wireless Sensor
Networks
225
Mohamed Younis and Kemal Akkaya
10. Mobility in Wireless Sensor Networks
267
Stefano Basagni, Alessio Carosi, and Chiara Petrioli
vii
viii
CONTENTS
11. Localization Systems for Wireless Sensor Networks
307
Azzedine Boukerche, Horacio A. B. F. Oliveira, Eduardo F. Nakamura,
and Antonio A. F. Loureiro
12. Location Discovery in Sensor Networks
341
Asis Nasipuri
13. QoS-Based Communication Protocols in Wireless Sensor Networks
365
Serdar Vural, Yuan Tian, and Eylem Ekici
14. Quality of Service in Wireless Sensor Networks
401
Gregory J. Pottie and Ameesh Pandya
15. Energy-Efficient Algorithms in Wireless Sensor Networks
437
Azzedine Boukerche and Sotiris Nikoletseas
16. Security Issues and Countermeasures in Wireless Sensor Networks
479
Tanveer Zia and Albert Y. Zomaya
17. A Taxonomy of Secure Time Synchronization Algorithms for
Wireless Sensor Networks
503
Azzedine Boukerche and Damla Turgut
18. Secure Localization Systems: Protocols and Techniques in
Wireless Sensor Networks
521
Azzedine Boukerche, Horacio A. B. F. Oliveira, Eduardo F. Nakamura,
and Antonio A. F. Loureiro
Index
535
PREFACE
With the recent technological advances in wireless communication and networking,
coupled with the availability of intelligent and low-cost actor and sensor devices with
powerful sensing, computation, and communication capabilities, wireless sensor networks (WSNs) are about to enter the mainstream. Today, one could easily envision a
wide range of real-world WSN-based applications from sensor-based environmental
monitoring, home automation, health care, security, and safety class of applications,
thereby promising to have a significant impact throughout our society. Wireless sensor
networks are comprised of a large number of sensor devices that can communicate
with each other via wireless channels, with limited energy and computing capabilities. However, due to the nature of wireless sensor networks, we are witnessing new
research challenges related to the design of algorithms and network protocols that will
enable the development of sensor-based applications. Most of the available literature
in this emerging technology concentrates on physical and networking aspects of the
subject. However, in most of the literature, a description of fundamental distributed
algorithms that support sensor and actor devices in a wireless environment is either
not included or briefly discussed. The efficient and robust realization of such large,
highly dynamic and complex networking environments is a challenging algorithmic
and technological task. Toward this end, this book identifies the research that needs to
be conducted on a number of levels to design and assess the deployment of wireless
sensor networks–in particular the design of algorithmic methods and distributed computing with sensing, processing, and communication capabilities. It is our belief that
this volume provides not only the necessary background and foundation in wireless
sensor networks but also an in-depth analysis of fundamental algorithms and protocols for the design and development of the next generations of heterogeneous wireless
networks in general and wireless sensor networks in particular. This book is divided
into 18 chapters and covers a variety of topics in the field of wireless sensor networks
that could be used as a textbook for graduate and/or advanced undergraduate studies,
as well as a reference for engineers and computer scientists interested in the field of
wireless sensor networks.
The rest of this book is organized as follows. In Chapter 1, we address the several
important algorithmic issues arising in wireless sensor networks and highlight the
main differences to classical distributed algorithms. Next, an algorithmic perspective
toward the design of wireless sensor networks is discussed followed by an overview
of well-known algorithms for basic services (that can be used by other algorithms in
WSNs), data communication, management functions, applications, and data fusion.
Chapter 2 introduces heterogeneous wireless sensor networks where more than one
ix
x
PREFACE
type of sensor node is integrated into a WSN. While many of the existing civilian
and military applications of heterogeneous wireless sensor networks (H-WSNs) do
not differ substantially from their homogeneous counterparts, there are compelling
reasons to incorporate heterogeneity into the network, such as improving the scalability of WSNs and addressing the problem of nonuniform energy drainage, among
others. Chapter 2 also discusses how these reasons are interrelated and how this new
dimension heterogeneity opens new challenges to the design of algorithms that run
on such wireless sensor networks.
In order to develop algorithms for sensor networks and in order to give mathematical correctness and performance proofs, models for various aspects of sensor networks
are needed. In the next three chapters, we focus upon the modeling, design, and analysis of algorithms and protocols for wireless sensor networks. Chapter 3 discusses
how biological inspired models, such epidemic models, can be used to design reliable
data dissemination algorithms in the context of wireless sensor networks. Recall that
reliable data dissemination to all sensor nodes is necessary for the propagation of
queries, code updates, and other sensitive WSN-related information. This is not a
trivial task because the number of nodes in a sensor network can be quite large and
the environment is quite dynamic (e.g., nodes can die or move to another location).
Chapter 4 provides an overview and discussion of well-known sensor network models
used today and shows how these models are related to each other. While the collaborative nature of the WSN brings significant advantages over traditional sensing, the
spatiotemporal correlation among the sensor observations is another significant and
unique characteristic of the WSN which can be exploited to drastically enhance the
overall sensor network performance. Chapter 5 presents the theoretical framework
to model the spatiotemporal correlation in sensor networks and describes in detail
how to exploit this correlation when designing reliable communication protocols for
WSN.
With the traditional TCP/IP models not suited to routing in wireless sensor networks, the network layer protocol has to be updated to be synchronized with the challenging constraints posed by WSNs. Hence, routing in these networks is a challenging
task and has thus been a primary focus with the wireless networking community. The
next chapters investigate the major issues to routing with the goals to devise new protocols to keep associated uncertainty under control. Chapter 6 highlights the properties
of a wireless sensor network from the networking point of view, and then it presents a
description of various well-known routing protocols for wireless sensor networks. The
common goals of designing a routing algorithm is not only to reduce control packet
overhead, maximize throughput, and minimize the end-to-end delay, but also to take
into consideration the energy consumption, especially in a sensor network comprised
of nodes that are considered lightweight with limited memory and battery power. In
order to achieve high energy efficiency and ensure long network lifetime for routing traffic control, as well as employ bandwidth re-use for data gathering and target
tracking, researchers have designed one-to-many, many-to-one, one-to-any, or oneto-all communications, routing, and clustering-based routing protocols. Chapter 7
presents different protocols developed to create clusters and select the best cluster
head using Graph Theory concepts. Chapter 8 discusses the merits and challenges of
PREFACE
xi
algorithms and protocols that provide point-to-point services through position-based
routing, where forwarding decisions are made by maximizing or minimizing some
function of node locations within a coordinate system. Sensors can generally be placed
in an area of interest either deterministically or randomly. However, controlled node
deployment is viable and often necessary when sensors are expensive or when their
operation is significantly affected by their position. Chapter 9 investigates the effect
node placement strategies on the dependability of WSNs, and it presents the various
sensor and base-station positioning protocols that have been developed to enhance
further the performance of WSNs and extend its network lifetime.
The next generation of wireless sensor networks are envisioned to support mobile
sensor devices and a variety of mobile robot sensor devices and a variety of wireless
multimedia sensor services. Chapter 10 presents several techniques for exploiting the
mobility of network components in large networks of resource constrained devices,
such as wireless sensor networks, and improving the performance of these networks
without significantly affecting data routing and end-to-end latency. A number of
mobility issues in WSNs as well as the pros and cons of providing mobility to the
normal nodes, relay nodes, and/or sink nodes are analyzed. Also in this chapter,
solutions that use mobility to alleviate the problem of energy depletion of nodes near
the sink are shown. However, this mobility as well as the random deployment of the
nodes in a WSN imposes another problem to the network: how to discover the current
physical position of the nodes. Chapters 11 and 12 focus on the different aspects of this
problem known as the localization problem. In Chapter 11, the localization systems
are divided into different components—distance estimation, position computation,
and localization algorithm—and several techniques employed by these components
are explained. On the other hand, Chapter 12 deals with more specific problems, such
as using the signals’ angle of arrival to estimate the position of the nodes.
Quality of service (QoS) provisioning in wireless sensor networks (WSNs) is an
important concept to enable mission-critical and real-time applications. In Chapter 13,
the necessity to support QoS in WSNs, QoS-based communication protocols, and
research directions to support QoS in WSNs is discussed. Chapter 14 presents some
background topics in network information theory relevant to the efficient collection,
compression, and reliable communication of sensor data. Then, it discusses how a
QoS perspective enables scalability in classical flat sensor networks. Finally, a number
of practical QoS approaches for high-fidelity data extraction in large-scale sensor
networks are explored. Chapter 15 focuses on several important aspects of energy
efficiency, like minimizing the total energy dissipation, minimizing the number of
transmissions, and balancing the energy load to prolong the system’s lifetime. Several
characteristic protocols and techniques in the recent literature that explicitly focus on
energy efficiency are presented. Such techniques include clustering and probabilistic
forwarding, adaptive transmission range management, and local optimization.
WSNs are supposed to be deployed in critical scenarios to be used in safety, emergency, and military applications. In these cases, security is a key technology in order
to make the gathered data a reliable information. Thus, we believe that a WSN book
would not be complete without a good review of the proposed techniques that aim to
provide the secure operation and communication in WSNs. Thus, the next chapters
xii
PREFACE
of this book investigate different aspects of providing security in WSNs. Chapter 16
focuses on general aspects of the problem, showing how WSNs are vulnerable to several attacks in the different network layers. Cryptography techniques for WSNs such
as cryptographic systems, authentication methods, and key distribution and management protocols are then studied and analyzed as a countermeasurement for a number
of the identified attacks. Also in this chapter, secure routing protocols that are resilient
to these attacks are discussed and explained. Besides securing the routing, it is also
important to secure other key protocols in WSNs such as the synchronization and
localization protocols. Chapter 17 provides a good overview of the proposed solutions for securing a time synchronization protocol to be used in critical applications
of WSNs. This chapter shows the importance of a secure synchronization system,
how current synchronization solutions are vulnerable to a number of attacks, and the
proposed techniques to secure these protocols. Finally, Chapter 18 takes the security
issue to the localization protocols. This chapter shows how the different components
of the localization systems–distance estimation, position computation, and localization algorithm–are vulnerable to a number of attacks and then shows the proposed
techniques and countermeasurements to secure these components and provide a secure localization system that are able to work in the presence of hostile nodes and
compromised environments.
It is our belief that this is the first book that covers the basic and fundamental algorithms and protocols for wireless sensor networks, making their design and analysis
accessible to all levels of readers.
Special thanks are due to all contributors for their support and patience, as well
as to the reviewers for their hard work and timely reports, which make this book
truly special. Last but not least, we wish to extend our thanks to Paul Petralia and
Whitney Lesch from John Wiley & Sons for their support, guidance, and certainly
their patience in finalizing this book.
Azzedine Boukerche
University of Ottawa
ABOUT THE EDITOR
Azzedine Boukerche is a Professor and holds a Canada Research Chair position at the
University of Ottawa. He is the Founding Director of Paradise Research Laboratory
at the University of Ottawa. Prior to this, he held a Faculty position at the University
of North Texas, and he was working as a Senior Scientist at the Simulation Sciences
Division, Metron Corporation, located in San Diego. He was also employed as a
faculty member at the School of Computer Science, McGill University, and he taught
at Polytechnic of Montreal. He spent a year at the JPL/NASA-California Institute
of Technology, where he contributed to a project centered around the specification
and verification of the software used to control interplanetary spacecraft operated
by JPL/NASA Laboratory. His current research interests include wireless ad hoc
and sensor networks, wireless networks, mobile and pervasive computing, wireless
multimedia, QoS service provisioning, large-scale distributed interactive simulation,
parallel discrete event simulation, and performance evaluation and modeling of largescale distributed and mobile systems. Dr. Boukerche has published several research
papers in these areas. He was the recipient of and/or nominated for the Best Research
Paper Award at IEEE/ACM PADS ’97, IEEE/ACM PADS ’99, IEEE ICC 2008, ACM
MSWiM 2001, and MobiWac’06, and he was the co-recipient of the 3rd National
Award for Telecommunication Software 1999 for his work on distributed security
systems on mobile phone operations.
Dr. A. Boukerche is a holder of an Ontario Early Research Excellence Award
(previously known as Premier of Ontario Research Excellence Award), an Ontario
Distinguished Researcher Award, and a Glinski Research Excellence Award. He is
a Co-Founder of QShine International Conference on Quality of Service for Wireless/Wired Heterogeneous Networks (QShine 2004) and has served as a General
Chair for the 8th ACM/IEEE Symposium on Modeling, Analysis, and Simulation
of Wireless and Mobile Systems, the 9th ACM/IEEE Symposium on Distributed
Simulation and Real-Time Application, and the 6th IEEE/ACM MASCOT ’98 Symposium; he has also served as the Vice General Chair for the 3rd IEEE International
Conference on Distributed Computing in Sensor Systems (DCOSS ’07), Program
Chair for IEEE Globecom 2007 and 2008 Ad Hoc, Sensor and Mesh Networking
Symposium, and a Program Co-Chair for ICPP 2008, the 2nd ACM Workshop on
QoS and Security for Wireless and Mobile Networks, ACM/IFIPS Europar 2002
Conference, IEEE/SCS Annual Simulation Symposium ’02, ACM WWW ’02, IEEE
MWCN 2002, IEEE/ACM MASCOTS ’02, IEEE Wireless Local Networks 03-04,
IEEE WMAN 04-05, and ACM MSWiM 98-99.
xiii
xiv
ABOUT THE EDITOR
Dr. A. Boukerche is an Associate Editor for ACM/Springer Wireless Networks,
IEEE Transactions on Vehicular Networks, IEEE Wireless Communication Magazine,
IEEE Transactions on Parallel and Distributed Systems, Elsevier’s Ad Hoc Networks,
Wiley International Journal of Wireless Communication and Mobile Computing,
Wiley’s Security and Communication Network Journal, Wiley’s Pervasive and Mobile Computing Journal, Elsevier’s Journal of Parallel and Distributed Computing,
and SCS Transactions on Simulation. He also serves as a Steering Committee Chair
for the ACM Modeling, Analysis and Simulation for Wireless and Mobile Systems
Symposium, the ACM Workshop on Performance Evaluation of Wireless Ad Hoc,
Sensor, and Ubiquitous Networks, and the IEEE/ACM Distributed Simulation and
Real-Time Applications Symposium (DS-RT).
CONTRIBUTORS
Mohammad Z. Ahmad, School of Electrical Engineering and Computer Science,
University of Central Florida, Orlando, FL 32816-2362
Özgür B. Akan, Next generation Wireless Communications Laboratory (NWCL),
Department of Electrical and Electronics Engineering, Middle East Technical
University, Ankara, Turkey 06531
Kemal Akkaya, Department of Computer Science, Southern Illinois University,
Carbondale, IL 62901
Liliana M. Arboleda, Department of Computing and Information Sciences,
University of Guelph, Guelph, Ontario N1G 2W1, Canada
Stefano Basagni, ECE Department, Northeastern University, Boston, MA 02115
Azzedine Boukerche, School of Information Technology and Engineering,
University of Ottawa, Ottawa, Ontario K1N 6N5, Canada
Alessio Carosi, Dipartimento di Informatica, Università di Roma “La Sapienza,”
Roma 00198, Italy
Sajal K. Das, Center for Research in Wireless Mobility and Networking
(CReWMaN), Department of Computer Science and Engineering, University of
Texas at Arlington, Arlington, TX 76019
Pradip De, Center for Research in Wireless Mobility and Networking
(CReWMaN), Department of Computer Science and Engineering, University of
Texas at Arlington, Arlington, TX 76019
Eylem Ekici, Department of Electrical and Computer Engineering, Ohio State
University, Columbus, OH 43210
Marwan M. Fayed, School of Information Technology and Information, University
of Ottawa, Ottawa, Ontario K1N 6N5, Canada
Bing Li, Department of Computer Science and Engineering, Arizona State
University, Tempe, AZ 85287-8809
Antonio A. F. Loureiro, Department of Computer Sciences, Federal University of
Minas Gerais, Belo Horizonte, Brazil, 31270-010
xv
xvi
CONTRIBUTORS
Angela M. Mielke, Distributed Sensor Networks Group, Los Alamos National
Laboratory, Los Alamos, NM 87545
Hussein T. Mouftah, School of Information Technology and Information, University of Ottawa, Ottawa, Ontario K1N 6N5, Canada
Eduardo F. Nakamura, Research and Technological Innovation Center (FUCAPI),
Brazil.
Asis Nasipuri, Department of Electrical and Computer Engineering, The University
of North Carolina at Charlotte, Charlotte, NC 28223
Nidal Nasser, Department of Computing and Information Sciences, University of
Guelph, Guelph, Ontario N1G 2W1, Canada
Sotiris Nikoletseas, Department of Computer Engineering and Informatics,
University of Patras, Patras, Greece; and Computer Technology Institute, (CTI),
Patras 26500, Greece
Horacio A. B. F. Oliveira, School of Information Technology and Engineering,
University of Ottawa, Ottawa, Ontario, Canada, K1N 6N5; Federal University
of Minas Gerais, Minas Gerais, Brazil, 31270-010; and Federal University of
Amazonas, Amazonas, Brazil, 69077-000
Ameesh Pandya, Department of Electrical Engineering, UCLA, Los Angeles, CA
90095
Chiara Petrioli, Dipartimento di Informatica, Università di Roma “La Sapienza,”
Roma 00198, Italy
Gregory J. Pottie, Department of Electrical Engineering, UCLA, Los Angeles, CA
90095
Stefan Schmid, Computer Engineering and Networks Laboratory (TIK), ETH
Zurich, CH-8092 Zurich, Switzerland
Violet R. Syrotiuk, Department of Computer Science and Engineering, Arizona
State University, Tempe, AZ 85287-8809
Yuan Tian, Department of Electrical and Computer Engineering, Ohio State
University, Columbus, OH 43210
Begümhan Turgut, Department of Computer Science, Rutgers University,
Piscataway, NJ 08854-8019
Damla Turgut, School of Electrical Engineering and Computer Science, University
of Central Florida, Orlando, FL 32816-2362
Serdar Vural, Department of Electrical and Computer Engineering, Ohio State
University, Columbus, OH 43210
Roger Wattenhofer, Computer Engineering and Networks Laboratory (TIK), ETH
Zurich, CH-8092 Zurich, Switzerland
CONTRIBUTORS
xvii
Mohamed Younis, Department of Computer Science and Electrical Engineering,
University of Maryland Baltimore County, Baltimore, MD 21250
Tanveer Zia, School of Information Technologies, The University of Sydney,
Sydney, NSW 2006, Australia
Albert Y. Zomaya, School of Information Technologies, The University of Sydney,
Sydney, NSW 2006, Australia
CHAPTER 1
Algorithms for Wireless Sensor
Networks: Present and Future
AZZEDINE BOUKERCHE
School of Information Technology and Engineering, University of Ottawa, Ottawa, Ontario K1N
6N5, Canada
EDUARDO F. NAKAMURA
Federal University of Minas Gerais, Brazil; and FUCAPI—Analysis, Research, and Technological
Innovation Center, Brazil
ANTONIO A. F. LOUREIRO
Department of Computer Sciences, Federal University of Minas Gerais, Belo Horizonte, Brazil
1.1 INTRODUCTION
Wireless sensor networks (WSNs) pose new research challenges related to the design
of algorithms, network protocols, and software that will enable the development of
applications based on sensor devices. Sensor networks are composed of cooperating sensor nodes that can perceive the environment to monitor physical phenomena
and events of interest. WSNs are envisioned to be applied in different applications,
including, among others, habitat, environmental, and industrial monitoring, which
have great potential benefits for the society as a whole. The WSN design often employs some approaches as energy-aware techniques, in-network processing, multihop
communication, and density control techniques to extend the network lifetime. In addition, WSNs should be resilient to failures due to different reasons such as physical
destruction of nodes or energy depletion. Fault tolerance mechanisms should take
advantage of nodal redundancy and distributed task processing. Several challenges
still need to be overcome to have ubiquitous deployment of sensor networks. These
challenges include dynamic topology, device heterogeneity, limited power capacity,
lack of quality of service, application support, manufacturing quality, and ecological
issues.
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
1
2
ALGORITHMS FOR WIRELESS SENSOR NETWORKS: PRESENT AND FUTURE
The capacity to transmit and receive data packets allows both information and
control to be shared among sensor nodes but also to perform cooperative tasks, all
based on different algorithms that are being specifically designed for such networks.
Some of the classes of algorithms for WSNs are briefly described in the following:
r Centralized algorithms execute on a central node and usually benefit from a
global network knowledge. This type of algorithm is not very common in WSNs
because the cost of acquiring a global network knowledge is usually unfeasible
in most WSNs.
r Distributed algorithms are related to different computational models. In a WSN,
the typical computational model is represented by a set of computational devices
(sensor nodes) that can communicate among themselves using a message-passing
mechanism. Thus, a distributed algorithm is an algorithm that executes on different sensor nodes and uses a message-passing technique.
r Localized algorithms comprise a class of algorithms in which a node makes
its decisions based on local and limited knowledge instead of a global network
knowledge. Thus “locality” usually refers to the node’s vicinity [1].
Algorithms for WSNs may also have some specific features such as selfconfiguration and self-organization, depending on the type of the target application.
Self-configuration means the capacity of an algorithm to adjust its operational parameters according to the design requirements. For instance, whenever a given energy
value is reached, a sensor node may reduce its transmission rate. Self-organization
means the capacity of an algorithm to autonomously adapt to changes resulted from
external interventions, such as topological changes (due to failures, mobility, or node
inclusion) or reaction to a detected event, without the influence of a centralized entity.
1.2 WIRELESS SENSOR NETWORKS: AN ALGORITHMIC
PERSPECTIVE
In the following, we present an overview of some algorithms for basic services (that
can be used by other algorithms), data communication, management functions, applications, and data fusion.
1.2.1 Basic Services
Some of the basic services that can be employed by other algorithms in wireless
sensor networks are localization, node placement, and density control.
Localization. The location problem consists in finding the geographic location of
the nodes in a WSN, which can be computed by a central unit [2] or by sensor nodes in a
distributed manner [3–8]. Essentially, the location discovery can be split in two stages:
distance estimation and location computation [4]. Usually, the distance between two
WIRELESS SENSOR NETWORKS: AN ALGORITHMIC PERSPECTIVE
3
A
c
b
a
B
C
(a)
(b)
(c)
Figure 1.1. Position estimation methods: (a) triangulation, (b) trilateration, and (c) multilateration. (Adapted from reference 10.)
nodes is estimated based on different methods, such as Received Signal Strength
Indicator (RSSI), Time of Arrival (ToA), and Time Difference of Arrival (TDoA) [4].
Once the distance is estimated, at least three methods can be used to compute the node
location: triangulation, trilateration, and multilateration [9], as depicted in Figure 1.1.
Another method to estimate the node location is called the Angle of Arrival (AoA),
which uses the angle in which the received signal arrives and the distance between
the sender and receiver.
Solutions for finding the nodes’ location are often based on localized algorithms in
the sense that every node is usually able to estimate its position. For instance, Sichitiu
and Ramadurai [11] use the Bayesian inference to process information from a mobile
beacon and determine the most likely geographical location (and region) of each
node, instead of finding a unique point for each node location. The Directed Position
Estimation (DPE) [8] is a recursive localization algorithm in which a node uses only
two references to estimate its location. This approach leads to a localization system
that can work in a low-density sensor network. Besides, the controlled way in which
the recursion occurs leads to a system with smaller and predictable errors. Liu et al.
[12] propose a robust and interactive Least-Squares method for node localization in
which, at each iteration, nodes are localized by using a least-squares-based algorithm
that explicitly considers noisy measurements.
Node Placement. In some applications, instead of throwing the sensor nodes on
the environment (e.g., by airplane), they can be strategically placed in the sensor field
according to a priori planning. In this approach, there is no need to discover the nodes’
location. However, good planning depends on the knowledge of the terrain and the
environmental particularities that might interfere in the operation of the sensor nodes
and the quality of the gathered data.
The node placement problem has been addressed using different approaches
[13–15]. However, current solutions are basically concerned with assuring spatial
coverage while minimizing the energy cost. The SPRING algorithm is a node placement algorithm that also performs information fusion. In SPRING it is possible to
migrate the fusion role.
Besides spatial coverage [13, 15], other aspects should be considered in a node
placement algorithm, such as node diversity [14] and the fusion performance. When
4
ALGORITHMS FOR WIRELESS SENSOR NETWORKS: PRESENT AND FUTURE
Figure 1.2. An example of node scheduling: Gray nodes are asleep and black nodes are awake.
nodes perform data fusion, an improper node placement may lead to the degradation
of information fusion as illustrated by Hegazy and Vachtsevanos [16].
Density Control. The main node scheduling objective is to save energy using a
density control algorithm [17–20]. Such algorithms manage the network density by
determining when each node will be operable (awake) and when it will be inoperable
(asleep). Figure 1.2 depicts an example of the result of a node scheduling algorithm
in which gray nodes are asleep because their sensing areas are already covered by
awaken nodes (in black).
Density control is an inherently localized algorithm where each node assesses its
vicinity to decide whether or not it will be turned on. Some of the node scheduling
algorithms, such as GAF [17], SPAN [19], and STEM [18], consider only the communication range to choose whether or not a node will be awake. Therefore, it is
possible that some regions remain uncovered, and the application may not detect an
event. Other solutions, such as PEAS [20], try to preserve the coverage. However,
none of the current node scheduling algorithms consider the information fusion accuracy. As a result, nodes that are important to information fusion might be turned
off. A key issue regarding density control algorithms is the integration with other
functions such as data routing. Siqueira et al. [21] propose two ways of integrating
density control and data routing: synchronizing both algorithms or redesigning an
integrated algorithm.
1.2.2 Data Communication
In wireless sensor networks, the problem of data communication is mainly related to
medium access control, routing, and transport protocols.
MAC Protocols. The link or medium access control (MAC) layer controls the
node access to the communication medium by means of techniques such as contention [22, 23] and time division [24, 25]. Basically, the MAC layer must manage
the communication channels available for the node, thereby avoiding collisions and
errors in the communication.
WIRELESS SENSOR NETWORKS: AN ALGORITHMIC PERSPECTIVE
5
Most solutions try to provide a reliable and energy-efficient solution. In this direction, Ci et al. [26] use prediction techniques to foresee the best frame size to reduce
the packet size and save energy. To avoid transmitting packets under unreliable conditions, Polastre et al. [23] apply filter techniques to estimate ambient noise and
determine whether the channel is clear for transmission. Liang and Ren [27] propose
a MAC protocol based on a fuzzy logic rescheduling scheme that improves existing
energy-efficient protocols. Their input variables are the ratios of nodes that (i) have an
overflowed buffer, (ii) have a high failing transmission rate, and (iii) are experiencing
an unsuccessful transmission.
Routing Protocols. Routing is the process of sending a data packet from a given
source to a given destination, possibly using intermediate nodes to reach the final
entity. This is the so-called unicast communication. In WSNs, data communication,
from the point of view of the communicating entities, can be divided into three cases:
from sensor nodes to a monitoring node, among neighbor nodes, and from a monitoring node to sensor nodes. Data communication from sensor nodes to a monitoring
node is used to send the sensed data collected by the sensors to a monitoring application. This class includes most of the routing protocols proposed in the literature [28].
Data communication among neighbor nodes often happens when some kind of cooperation among nodes is needed. Data communication from a monitoring node to a set
of sensor nodes is often used to disseminate a piece of information that is important
to those nodes. Based on an efficient dissemination algorithm, a monitoring node can
perform different activities, such as to change the operational mode of part or the
entire WSN, broadcast a new interest to the network, activate/deactivate one or more
sensor nodes, and send queries to the network.
The routing algorithms for wireless sensor networks can be broadly divided into
three types: flat-based routing, hierarchical-based routing, and adaptive-based routing. Flat-based routing assumes that all sensor nodes perform the same role. On
the other hand, nodes in hierarchical-based routing have different roles in the network, which can be static or dynamic. Adaptive routing changes its behavior according to different application and network conditions such as available energy
resources. These routing protocols can be further classified into multipath-based,
query-based, or negotiation-based routing techniques depending on the protocol
operation.
A natural routing scheme for flat networks is the formation of routing trees.
Krishnamachari et al. [29] provide analytical bounds on the energy costs and
savings that can be obtained with data aggregation using tree topologies. Zhou and
Krishnamachari [30] evaluate the tree topology with four different parent selection
strategies (earliest-first, randomized, nearest-first, and weighted-randomized) based
on the metrics, such as node degree, robustness, channel quality, data aggregation, and
latency. Tian and Georganas [31] identify drawbacks of pure single-path and multipath routing schemes in terms of packet delivery and energy consumption. The InFRA
algorithm [32] builds a routing tree by establishing a hybrid network organization in
which source nodes are organized into clusters and the cluster-to-sink communication
6
ALGORITHMS FOR WIRELESS SENSOR NETWORKS: PRESENT AND FUTURE
occurs in a multihop fashion. The resulting topology is a distributed heuristic to the
Steiner tree problem.
For the hierarchical topology, several algorithms are provided in the literature.
LEACH [33] is a cluster-based protocol that randomly rotates the cluster heads to
evenly distribute the energy load among the sensors in the network. PEGASIS [34]
is an improvement of LEACH in which sensors form chains, and each node communicates only with a close neighbor and takes turns to transmit messages to the sink
node.
The Directed Diffusion [35] is a pioneer protocol that tries to find the best paths
from sources to sink nodes that might receive data from multiple paths with different
data delivery frequencies. If the best path fails, another path with lower data delivery
frequency assures the data delivery. Ganesan et al. [36] propose a routing solution,
which evolved from Directed Diffusion, that tries to discover and maintain alternative
paths, connecting sources to sinks, to make the network more fault-tolerant.
Niculescu and Nath [37] propose the Trajectory-Based Forwarding (TBF) algorithm, a data dissemination technique in which packets are disseminated from a monitoring node to a set of nodes along a predefined curve. Machado et al. [38] extend
TBF with the information provided by the energy map [39] of a sensor network to
determine routes in a dynamic fashion.
In WSNs, routing protocols are closely related to information fusion because it
addresses the problem of delivering the sensed information to the sink node, and it is
natural to think of performing the fusion while the pieces of data become available.
However, the way information is fused depends on the network organization, which
directly affects how the role can be assigned. Hierarchical networks are organized into
clusters where each node responds only to its respective cluster-head, which might
perform special operations such as data fusion/aggregation. In flat networks, communication is performed hop-by-hop and every node may be functionally equivalent.
Transport Protocols. In general, transport protocols are concerned with the
provision of a reliable communication service for the application layer. This is
the main objective of the Pump Slowly, Fetch Quickly (PSFQ) protocol [40].
PSFQ is an adaptive protocol that makes local error correction using hop-by-hop
acknowledgement. In this case, the adaptation means that under low failure rates,
the communication is similar to a simple forward, and when failures are frequent,
it presents a store-and-forward scheme. Another transport protocol that aims to
provide a reliable communication is the Reliable Data Transport in Sensor Networks
(RMST) [41] that also implements a hop-by-hop acknowledgment. However, RMST
is designed to operate in conjunction with Directed Diffusion.
An interesting approach is introduced by the Event-to-Sink Reliable Transfer
(ESRT) protocol [42, 43]. This protocol is designed for event-based sensor networks,
and it changes the focus of traditional transport protocols. The authors state that for
WSNs a transport protocol should be reliable regarding the event detection task. ESRT
assumes that an event must be detected when the sink node receives a minimum number of event reports from sensor nodes. If this threshold is not achieved, the sink node
WIRELESS SENSOR NETWORKS: AN ALGORITHMIC PERSPECTIVE
7
does not recognize the event. Thus, ESRT adjusts the transmission rate of each node
in such a way that the desired threshold is achieved and the event is reliably detected.
1.2.3 Management Functions
In the following, we present some high-level management functions that can be used
by different monitoring applications in a WSN. We start by presenting a management
architecture, followed by a discussion of data storage, network health, coverage and
exposure, and security.
Architecture. A WSN management architecture can be used to reason about the
different dimensions present in the sensor network. In this direction, the MANNA
architecture [44] was proposed to provide a management solution to different WSN
applications. It provides a separation between both sets of functionalities (i.e., application and management), making integration of organizational, administrative, and
maintenance activities possible for this kind of network. The approach used in the
MANNA architecture works with each functional area, as well as each management
level, and proposes the new abstraction level of WSN functionalities (configuration,
sensing, processing, communication, and maintenance) presented earlier. As a result,
it provides a list of management services and functions that are independent of the
technology adopted.
Data Storage. Data storage is closely related to the routing (data retrieval) strategy.
In the Cougar database system [45], stored data are represented as relations whereas
sensor data are represented as time series. A query formulated over a sensor network
specifies a persistent view, which is valid during a given period [46]. Shenker et al. [47]
introduce the concept of data-centric storage, which is also explored by Ratnasamy
et al. [48] and Ghose et al. [49]. In this approach, relevant data is labeled (named) and
stored by the sensor nodes. Data with the same name are stored by the same sensor
node. Queries for data with a particular name are sent directly to the node storing that
named data, avoiding the flooding of interests or queries.
Network Health. An important issue underlying WSNs is the monitoring of the
network itself; that is, the sink node needs to be aware of the health of all the sensors.
Jaikaeo et al. [50] define diagnosis as the process of monitoring the state of a sensor
network and figuring out the problematic nodes. This is a management activity that
assesses the network health—that is, how well the network elements and the resources
are being applied.
Managing individual nodes in a large-scale WSN may result in a response implosion problem that happens when a high number of replies are triggered by
diagnostic queries. Jaikaeo et al. [50] suggest the use of three operations, built on
the top of the SINA architecture [51], to overcome the implosion problem: sampling,
self-orchestrated, and diffused computation. In a sampling operation, information
from each node is sent to the manager without intermediate processing. To avoid the
8
ALGORITHMS FOR WIRELESS SENSOR NETWORKS: PRESENT AND FUTURE
implosion problem, each node decides whether or not it will send its information
based on a probability assigned by the manager (based on the node density). In a
self-orchestrated operation, each node schedules its replies. This approach introduces
some delay, but reduces the collision chances. In a diffused computation, mobile
scripts are used (enabled by the SINA architecture) to assign diagnosis logic to sensor nodes so they know how to perform information fusion and route the result to
the manager. Although diffused computation optimizes bandwidth use, it introduces
greater delay and the resultant information is less accurate. The three operations provide different levels of granularity and delay; therefore they should be used in different
stages: Diffused computation and self-orchestrated operations should be continuously
performed to identify problems, and sampling should be used to identify problematic
elements.
Hsin and Liu [52] propose a two-phase timeout system to monitor the node liveliness. In the first phase, if a node A receives no message from a neighbor D in a
given period of time (monitoring time), A assumes that D is dead, entering in the second phase. Once in the second phase, during another period of time (query time), A
queries its neighbors about D; if any neighbor claims that D is alive, then A assumes
it was a false alarm and discards this event. Otherwise, if A does not hear anything
before the query time expires, it assumes that D is really dead, triggering an alarm.
This monitoring algorithm can be seen as a simple information fusion method for
liveliness detection where the operator (fuser) is a logical OR with n inputs such as
input i is true if neighbor i considers that D is alive and false otherwise.
Zhao et al. [53] propose a three-level health monitoring architecture for WSN.
The first level includes the digests that are aggregates of some network property,
like minimum residual energy. The second comprises the network scans, a sort of
feature map that represents abstracted views of resource utilization within a section
of the (or entire) network [54]. Finally, the third is composed by node dumps that
provide detailed node states over the network for diagnosis. In this architecture, digests
should be continuously computed in background and piggybacked in a neighborto-neighbor communication. Once an anomaly is detected in the digests, a network
scan may be collected to identify the problematic sections in the network. Finally,
dumps of problematic sections can be requested to identify what is the problem. The
information granularity increases from digests to dumps, and the finer the granularity, the greater the cost. Therefore, network scans and, especially, dumps should
be carefully used.
An energy map is the information about the amount of energy available at each
part of the network. Due to the importance of energy-efficiency solutions for WSNs,
the energy map can be useful to prolong the network lifetime and be applied to
different network activities in order to make a better use of the energy reserves. Thus,
the cost of obtaining the energy map can be amortized among different network
applications, and neither of them has to pay exclusively for this information itself.
The energy map can be constructed using a naive approach, in which each node sends
periodically only its available energy to the monitoring node. However, this approach
would spend so much energy, due to communication, that probably the utility of the
energy information would not compensate the amount of energy spent in this process.
WIRELESS SENSOR NETWORKS: AN ALGORITHMIC PERSPECTIVE
9
Zhao et al. [55] propose a more interesting solution that obtains the energy map using
an aggregation based approach. Mini et al. [39] propose another efficient solution,
based on a Markov Chain mechanism, to predict the energy consumption of a sensor
node in order to construct the energy map.
Coverage and Exposure. Coverage (spatial) comprises the problem of determining the area covered by the sensors in the network [13, 14, 56, 57]. Coverage allows
the identification of regions that can be properly monitored and regions that cannot.
This information associated with the energy map [54] can be used to schedule sensor nodes to optimize the network lifetime without compromising the quality of the
gathered information.
Azzedine Boukerche [57] defines coverage in terms of the best case (regions of high
observability) and the worst case (regions of low observability), and it is computed in
a centralized fashion by means of geometric structures (Delaunay triangulation and
Voronoi diagram) and algorithms for graph searching. Li et al. [56] extend this work
considering a sensing model in which the sensor accuracy is inversely proportional to
the distance to the sensed event, and they provide distributed algorithms to compute
the best case of coverage and the path of greater observability. Chakrabarty et al. [14]
compare coverage to the Art Gallery Problem (AGP), which consists in finding the
smallest number of guards to monitor the entire art gallery. Dhillon et al. [13] consider
coverage as the lowest detection probability of an event by any sensor. Exposure is
closely related with coverage and it specifies how well an object, moving arbitrarily,
can be observed by the WSN over a period of time [58].
Security. Security is an issue of major concern in WSNs, especially in surveillance
applications, with implication to other functions. For instance, despite the fact that
data fusion can reduce communication, fusing data packets makes security assurance
more complex. The reason is that intermediate nodes can modify, forge, or drop data
packets. In addition, source-to-sink data encryption may not be desirable because the
intermediate nodes need to understand the data to perform data fusion.
Hu and Evans [59] present a protocol to provide secure aggregation for flat WSNs
that is resilient to intruder devices and single device key compromises, but their
protocol may become vulnerable when a parent and a child node are compromised. The Energy-efficient and Secure Pattern-based Data Aggregation protocol
(ESPDA) [60] is a secure protocol for hierarchical sensor networks that does not
require the encrypted data to be decrypted by cluster heads to perform data aggregation. In ESPDA, the cluster head first requests nodes to send the corresponding
pattern code for the sensed data. If the same pattern code is sent to the cluster head by
different nodes, then only one of them is allowed to send its data. The pattern code
is generated based on a seed provided by the cluster head. No special fusion method
is actually applied in the ESPDA protocol, which simply avoids the transmission of
redundant data, so any information fusion must be performed by the sensor nodes,
not the cluster head. Secure Information Aggregation in Sensor Networks (SIA) [61]
presents a fuse–commit–prove approach in which fuser nodes need to prove that
they perform fusion tasks correctly. To avoid cheating by fuser nodes, SIA adopts
10
ALGORITHMS FOR WIRELESS SENSOR NETWORKS: PRESENT AND FUTURE
cryptographic techniques of commitments and provides random sampling mechanisms and interactive proofs to allow the user to verify the data given by fuser nodes,
even when the fuser nodes or some sensor nodes are corrupted.
1.2.4 Applications
Two of the most basic applications for wireless sensor networks are query processing,
and event and target tracking. The former is often used to answer queries posed by
users outside of the network, and the latter is used to know about events happening
inside the network, including specific targets. These two applications can actually
be seen as application protocols that might be present in different monitoring
applications.
Query Processing. Different solutions explore the query approach using innetwork processing to filter and/or aggregate the data during the routing process.
Directed Diffusion [35] introduces the concept of interests to specify which data will
be delivered through a publish/subscribe scheme, but no query language is specified.
Another possibility is to model the sensor network as a database so data access
is performed by declarative queries. The DataSpace Project [62] provides a means
of geographically querying, monitoring, and controlling the network devices that encapsulate data. DataSpace provides network primitives to assure that only relevant
devices are contacted when a query is evaluated. Sensor Information Networking Architecture (SINA) [51] is a cluster-based architecture that abstracts a WSN as a dense
collection of distributed objects where users access information through declarative
queries and execute tasks through programming scripts. The Cougar Project [45]
handles the network as a distributed database in which each piece of data is locally
stored in a sensor node and data are retrieved by performing aggregation along a query
tree. Temporal coherency-aware in-Network Aggregation (TiNA) [63] uses temporal
coherency tolerances to reduce the communication load and improve quality of data
when not all sensor readings can be propagated within a given time constraint. The
ACtive QUery forwarding In sensoR nEtworks (ACQUIRE) [64] system considers
the query as an active entity that is forwarded through the network searching for a
solution. In ACQUIRE, intermediate nodes, handling the active query, partially evaluate the queries by using information from nodes within d hops. Once the query is
fully evaluated, a response is sent toward the querying node. TinyDB [65] provides a
simple query language to specify the data of interest.
Event and Target Tracking. Event (target) tracking is one of the most popular
applications of sensor systems in general. The problem consists in predicting where an
event or target being detected is moving to. This is essentially a data fusion application.
Coates [66] uses filters for target tracking in cluster-based networks in which cluster
heads perform computations and share information, and the other cluster members
sense the environment. To track multiple targets, Sheng et al. [67] use filters that
run on uncorrelated sensor cliques that are dynamically organized based on target
trajectories. Vercauteren et al. [68] propose a collaborative solution for jointly tracking
CHALLENGE: SYNTHESIS PROCESS
11
several targets and classifying them according to their motion pattern. Schmitt et al.
[69] propose a collaborative algorithm to find the location of mobile robots in a known
environment and track moving objects.
1.2.5 Data Fusion
Data fusion algorithms [70] are orthogonal to the above-mentioned problems, in the
sense that these algorithms can be applied to any solution that needs to make inferences
or improve estimates.
Classical data fusion techniques have been used to assist solving many problems.
For instance, the Least-Squares method has been used to predict sensor data [71] and
find nodes’ locations [8, 12]; the moving average filter has been used to estimate link
connectivity statistics [72], estimate data traffic [73] and the number of events [74], and
track targets [75]; the Kalman filter has been applied to refine location and distance
estimates [6, 76], track different targets [77], predict the best frame size for MAC
protocols [26], and predict sensor data to reduce communication [78].
As discussed in the following, data fusion can have an important role when we
design an integrated solution for a wireless sensor network.
1.3 CHALLENGE: SYNTHESIS PROCESS
One of the most important challenges in the design of wireless sensor networks is to
deal with the dynamics of such networks. The physical world where the sensors are
embedded is dynamic. Over time, the operating conditions and the associate tasks to
be performed by the sensors can change. Some of the causes that might trigger these
changes are the events occurring in the network, amount of resources available at
nodes (particularly energy), and reconfiguration of nodes. Furthermore, it is important
that sensors adapt themselves to the environment since manual configuration may
be unfeasible or even impossible. In summary, the kind of distributed system we
are dealing with calls for an entire new class of algorithms for large-scale, highly
dynamic, and unattend WSN.
The complete design of a wireless sensor network, considering a particular application, should take into account many different aspects such as application goals, traffic
pattern, sensor node capability and availability, expected network lifetime, access to
the monitoring area, node replacement, environment characteristics, and cost. Given
a particular monitoring application, the network designer should clearly identify its
main goals and the corresponding QoS parameters. For instance, given a fire detection
application for a rain forest, we would like to guarantee that the network will operate
for the expected lifetime. However, as soon as a fire spot is detected, this information
should reach the sink node as fast and reliable as possible, probably not worrying
about the energy expenditure of the nodes involved in this communication.
Power-efficient communication paradigms for a given application should consider
both routing and media access algorithms. The routing algorithms must be tailored
for efficient network communication while maintaining connectivity when required
to source or relay packets. In this case, the research challenge of the routing problem
12
ALGORITHMS FOR WIRELESS SENSOR NETWORKS: PRESENT AND FUTURE
is to find a power-efficient method for scheduling the nodes such that a multihop path
may be used to relay the data. But, when we consider the particular aspects of the
monitoring application, we could apply, for instance, information fusion and density
control algorithms to reduce the amount of data packets to be relayed and sensor
nodes that need to be active, respectively.
As the sensor network starts to operate, it may be necessary to adjust the functionality of individual nodes. This refinement can take several different forms. Scalar
parameters, like duty cycle or sampling rates, may be adjusted using self-configuration
and self-organization algorithms. This process may occur in different ways along the
operation of the network lifetime.
Ideally, a WSN designer should come up with both the hardware and software
necessary to accomplish the aspects mentioned above. Unfortunately, it seems that
we are far from this scenario. We are still giving the first steps in the design process
of a wireless sensor network as we move toward to a more disciplined development.
Most of the studies found in the literature study particular problems for a WSN. That
is possibly the way we should go since we need to have more experience before we
can design a complete solution in a more systematic and automated way.
Figure 1.3 depicts a possible monitoring application for a rain forest. In this case,
we might be interested in detecting different events such as the presence of a rare
bird, a fire spot, and different environmental variables. The operation of the sensor
Figure 1.3. Monitoring application for a rain forest.
CHALLENGE: SYNTHESIS PROCESS
13
Figure 1.4. Synthesis process.
network can also be based on data received from a meteorological station, an unmanned airplane, or a satellite. Thus, given the different application requirements and
data sources, what are the best algorithms and sensor nodes that should be used to
accomplish the desired goals? This is a research challenge that we are starting to face
once more, and more real monitoring applications are being deployed. Notice that we
14
ALGORITHMS FOR WIRELESS SENSOR NETWORKS: PRESENT AND FUTURE
can even go one step further and build a specific hardware node that best fits to the
proposed solution, leading to a truly hardware–software codesign.
In order to achieve this proposed solution, we need a network synthesis process,
as depicted in Figure 1.4. This is similar to what happens currently in the design
of an integrated circuit (IC) that starts with its high-level specification and finishes
with its physical design. The synthesis process is guided by some aspects such as the
testability of the IC. It is important to design a more testable IC, since a chip is tested
not to check its logical correctness but to check its manufacturing process. In the case
of the WSN synthesis process, there are very interesting scientific challenges that we
need to overcome to have this automated development, as it happens in the synthesis
of an integrated circuits.
These challenges are related to the theory, techniques, methodologies, tools, and
processes. We need to propose new fundamental principles that will create a theory to
synthesize both the hardware and software of a wireless sensor network. This theory
will lead to techniques, methodologies, tools, and processes that will enable designers
to design new sensor networks for different monitoring applications in a systematic
way. In this vision, algorithms for wireless sensor networks have a fundamental role,
since they will be the outcome of this synthesis process.
BIBLIOGRAPHY
1. J. Feng, F. Koushanfar, and M. Potkonjak. Localized algorithms for sensor networks. In
I. Mahgoub and M. Ilyas, editors, Handbook of Sensor Networks, CRC Press, Boca Raton,
FL, 2004.
2. L. Doherty, K. S. J. Pister, and L. El Ghaoui. Convex position estimation in wireless sensor
networks. In INFOCOM [79], p. 1655–1663.
3. D. Niculescu and B. Nath. Ad hoc positioning system (APS). In 2001 IEEE Global Telecommunications Conference (GLOBECOM ’01), Vol. 5, San Antonio, TX, Novermber 2001,
IEEE, New York, pp. 2926–2931.
4. C. Savarese, J. M. Rabaey and J. Beutel. Locationing in distributed ad-hoc wireless sensor
networks. In Proceedings of the IEEE Signal Processing Society International Conference
on Acoustics, Speech, and Signal Processing 2001 (ICASSP ’01), Vol. 4, Salt Lake City,
UT, May 2001, IEEE, New York, pp. 2037–2040.
5. A. Savvides, H. Park and M. B. Srivastava. The bits and flops of the n-hop multilateration
primitive for node localization problems. In Proceedings of the 1st ACM International Workshop on Wireless Sensor Networks and Applications (WSNA’02), Atlanta, GA, September
2002, ACM, New York, pp. 112–121.
6. A. Savvides, C.-C. Han and M. B. Strivastava The n-hop multilateration primitive for node
localization. Mobile Networks and Applications, 8(4):443–451, 2003, ISSN 1383-469X.
doi: http://dx.doi.org/10.1023/A:1024544032357.
7. L. Hu and D. Evans. Localization for mobile sensor networks. In Proceedings of
the 10th Annual International Conference on Mobile Computing and Networking
(MobiCom ’04), Philadelphia, PA, 2004, ACM, New York, pp. 45–57, ISBN 1-58113868-7. doi: http://doi.acm.org/10.1145/1023720.1023726.
BIBLIOGRAPHY
15
8. H. A. B. F. de Oliveira, E. F. Nakamura, A. A. F. Loureiro, and A. Boukerche. Directed
position estimation: A recursive localization approach for wireless sensor networks. In
S. R. Thuel, Y. Yang, and E. K. Park, editors, Proceedings of the 14th IEEE International
Conference on Computer Communications and Networks (IC3N ’05), San Diego, CA.
October 2005, IEEE, pp. 557–562, ISBN 0-7803-9428-3.
9. J. D. Gibson, editor. The Mobile Communication Handbook, 2nd edition CRC Press, Boca
Raton, FL, 1999.
10. A. Savvides, C.-C. Han, and M. B. Strivastava. Dynamic fine-grained localization in ad-hoc
networks of sensors. In Mobicom [80], pp. 166–179, ISBN 1-58113-422-3.
11. M. L. Sichitiu and V. Ramadurai. Localization of wireless sensor networks with a
mobile beacon. In Proceedings of the 1st IEEE International Conference on Mobile
Ad Hoc and Sensor Systems (MASS 2004), Fort Lauderdale, FL, October 2004, IEEE,
pp. 174–183.
12. J. Liu, Y. Zhang, and F. Zhao. Robust distributed node localization with error management.
In Proceedings of the 7th ACM International Symposium on Mobile Ad Hoc Networking
and Computing (MobiHoc’06), Florence, Italy, 2006, ACM, New York, pp. 250–261, ISBN
1-59593-368-9. doi: http://doi.acm.org/10.1145/1132905.1132933.
13. S. S. Dhillon, K. Chakrabarty, and S. S. Iyengar. Sensor placement for grid coverage under imprecise detections. In Proceedings of 5th International Conference on Information
Fusion (Fusion 2002), Vol. 2, Annapolis, MD, July 2002, IEEE, New York, pp. 1581–
1587.
14. K. Chakrabarty, S. S. Iyengar, H. Qi, and E. Cho. Grid coverage for surveillance and target
location in distributed sensor networks. IEEE Transactions on Computers, 51(12):1448–
1453, 2002.
15. E. S. Biagioni and G. Sasaki. Wireless sensor placement for reliable and efficient data
collection. In Hawaii International Conference on Systems Sciences (HICSS 2003), Hawaii,
January 2003, IEEE, New York, pp. 127–136.
16. T. Hegazy and G. J. Vachtsevanos. Sensor placement for isotropic source localization. In
F. Zhao and L. J. Guibas, editors, IPSN ’03, Vol. 2634 of Lecture Notes in Computer Science,
Palo Alto, CA, April 2003, Springer, New York, pp. 432–441, ISBN 3-540-02111-6.
17. Y. Xu, J. Heidemann, and D. Estrin. Geography-informed energy conservation for ad-hoc
routing. In Mobicom [80], pp. 16–21, ISBN 1-58113-422-3.
18. C. Schurgers, V. Tsiatsis, S. Ganeriwal, and M. Srivastava. Optimizing sensor networks in
the energy-latency-density design space. IEEE Transactions on Mobile Computing, 1(1):
70–80, 2002.
19. B. Chen, K. Jamieson, H. Balakrishnan, and R. Morris. Span: An energy-efficient coordination algorithm for topology maintenance in ad hoc wireless networks. Wireless Networks,
8(5):481–494, 2002.
20. F. Ye, G. Zhong, J. Cheng, S. Lu, and L. Zhang. PEAS: A robust energy conserving protocol for long-lived sensor networks. In Proceedings of the 23rd International Conference on Distributed Computing Systems, Providence, RI, May 2003, IEEE, New York,
pp. 28–37.
21. I. G. Siqueira, C. M. S. Figueiredo, A. A. F. Loureiro, J. M. S. Nogueira, and L. B. Ruiz.
An integrated approach for density control and routing in wireless sensor networks. In
Proceedings of the IEEE 20th International Parallel and Distributed Processing Symposium
(IPDPS ’06). April 2006, IEEE, New York, CD-ROM.
16
ALGORITHMS FOR WIRELESS SENSOR NETWORKS: PRESENT AND FUTURE
22. A. Woo and D. E. Culler. A transmission control scheme for media access in sensor networks.
In Mobicom [80], pp. 221–235, ISBN 1-58113-422-3.
23. J. Polastre, J. Hill, and D. Culler. Versatile low power media access for wireless sensor networks. In J. A. Stankovic, A. Arora, and R. Govindan, editors, Proceedings
of the 2nd International Conference on Embedded Networked Sensor Systems (SenSys’04), Baltimore, MD, November 2004, ACM, New York, pp. 95–107, ISBN 1-58113879-2.
24. W. Ye, J. Heidemann, and D. Estrin. An energy-efficient MAC protocol for wireless
sensor networks. In INFOCOM, editor, INFOCOM 2002, New York, June 2002. IEEE,
New York, pp. 1567–1576, URL http://www.isi.edu/∼johnh/PAPERS/Ye02a.html.
25. V. Rajendran, K. Obraczka, and J. J. Garcia-Luna-Aceves. Energy-efficient, collision-free
medium access control for wireless sensor networks. In Akyildiz et al. [81], pp. 181–192,
ISBN 1-58113-707-9.
26. S. Ci, H. Sharif, and K. Nuli. A UKF-based link adaptation scheme to enhance energy
efficiency in wireless sensor networks. In Proceedings of the 15th IEEE International
Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC ’04), Vol. 4,
Barcelona, Spain, September 2004, IEEE, pp. 2483–2488.
27. Q. Liang and Q. Ren. An energy-efficient MAC protocol for wireless sensor networks. In 2005 IEEE Global Telecommunications Conference (GLOBECOM ’05), Vol. 1,
St. Louis, MO, November–December 2005, IEEE, New York, CD-ROM.
28. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cyirci. Wireless sensor networks: A
survey. Computer Networks, 38(4):393–422, 2002.
29. B. Krishnamachari, D. Estrin, and S. Wicker. The impact of data aggregation in wireless
sensor networks. In International Workshop of Distributed Event Based Systems (DEBS),
Vienna, Austria, July 2002, IEEE, pp. 575–578.
30. C. Zhou and B. Krishnamachari. Localized topology generation mechanisms for selfconfiguring sensor networks. In 2003 IEEE Global Telecommunications Conference
(GLOBECOM ’03), Vol. 22, San Francisco, CA, December 2003, IEEE, New York,
pp. 1269–1273.
31. D. Tian and N. D. Georganas. Energy efficient routing with guaranteed delivery in
wireless sensor networks. In IEEE Wireless Communications and Networking Conference (WCNC 2003), Vol. 3, New Orleans, LA, March 2003, IEEE, New York,
pp. 1923–1929.
32. E. F. Nakamura, H. A. B. F. de Oliveira, L. F. Pontello, and A. A. F.
Loureiro. On demand role assignment for event-detection in sensor networks. In P.
Bellavista, C.-M. Chen, A. Corradi, and M. Daneshmand, editors, Proceedings of the
11th IEEE International Symposium on Computers and Communication (ISCC ’06),
Cagliari, Italy, June 2006, IEEE, New York, pp. 941–947, ISBN 0-7695-2588-1. doi:
http://doi.ieeecomputersociety.org/10.1109/ISCC.2006.110.
33. W. Heinzelman, A. Chandrakasan, and H. Balakrishnan. Energy-efficient communication protocol for wireless microsensor networks. In Proceedings of the 33rd Hawaii
International Conference on System Sciences (HICSS ’00), Maui, HI, January 2000, IEEE,
New York, pp. 8020–8029, ISBN 0-7695-0493-0.
34. S. Lindsey, C. Raghavendra, and K. M. Sivalingam. Data gathering algorithms in sensor
networks using energy metrics. IEEE Transactions on Parallel and Distributed Systems,
13(9):924–935, 2002.
BIBLIOGRAPHY
17
35. C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed diffusion: A scalable and robust
communication paradigm for sensor networks. In Proceedings of the 6th Annual International Conference on Mobile Computing and Networking (MobiCom ’00), Boston, MA,
August 2000, ACM, New York, pp. 56–67.
36. D. Ganesan, R. Govindan, S. Shenker, and D. Estrin. Highly-resilient, energy-efficient
multipath routing in wireless sensor networks. ACM SIG-MOBILE Mobile Computing and
Communications Review, 5(4):11–25, 2001.
37. D. Niculescu and B. Nath. Trajectory-based forwarding and its applications. In Proceedings of the 9th Annual International Conference on Mobile Computing and Networking
(MobiCom ’03), 2003, pp. 260–272.
38. M. V. Machado, O. Goussevskaia, R. A. F. Mini, C. G. Rezende, A. A. F. Loureiro,
G. R. Mateus, and J. M. S. Nogueira. Event-to-sink reliable transport in wireless
sensor networks. IEEE Journal on Selected Areas in Communications, 23(12), 2005.
CD-ROM.
39. R. A. F. Mini, M. do V. Machado, A. A. F. Loureiro, and B. Nath. Prediction-based energy
map for wireless sensor networks. Ad Hoc Network Journal, 3:235–253, 2005.
40. C.-Y. Wan and A. Campbell. PSFQ: A reliable transport protocol for wireless sensor networks. In Proceedings of the 1st ACM International Workshop on Wireless Sensor Networks
and Applications (WSNA ’02), Atlanta, GA, September 2002, ACM, New York, pp. 1–11.
41. F. Stann and J. Heidemann. RMST: Reliable data transport in sensor networks. In Proceedings of the First IEEE International Workshop on Sensor Network Protocols and Applications (SNPA 2003), Anchorage, AK, May 2003, IEEE, pp. 102–112.
42. Y. Sankarasubramaniam, O. B. Akan, and I. F. Akyildiz. ESRT: Event-to-sink reliable transport in wireless sensor networks. In Proceedings of the 4th ACM International Symposium
on Mobile Ad Hoc Networking and Computing (MobiHoc ’03), Annapolis, MD, June 2003,
ACM, New York, pp. 177–188.
43. O. B. Akan and I. F. Akyildiz. Event-to-sink reliable transport in wireless sensor networks.
IEEE/ACM Transsactions on Networking, 13(5):1003–1016, 2005, ISSN 1063-6692.
44. L. B. Ruiz, J. M. S. Nogueira, and A. A. F. Loureiro. Manna: A management architecture
for wireless sensor networks. IEEE Communications Magazine, 41(2):116–125, 2005.
45. Y. Yao and J. Gehrke. The cougar approach to in-network query processing in sensor
networks. Sigmod Record, 31(3):9–18, 2002.
46. P. Bonnet, J. Gehrke, and P. Seshadri. Towards sensor database systems. In K.-L. Tan, M. J.
Franklin, and J. C. S. Lui, editors, Proceedings of the 2nd International Conference on
Mobile Data Management, Vol. 1987 of Lecture Notes in Computer Science, Hong Kong,
China, January 2001, Springer, pp. 3–14.
47. S. Shenker, S. Ratnasamy, B. Karp, R. Govindan, and D. Estrin. Data-centric storage in
sensornets. ACM SIGCOMM Computer Communication Review, 33(1):137–142, 2003.
48. S. Ratnasamy, B. Karp, S. Shenker, D. Estrin, R. Govindan, L. Yin, and F. Yu. Datacentric storage in sensornets with ght, a geographic hash table. Mobile Networks and
Applications, 8(4):427–442, 2003.
49. A. Ghose, J. Grossklags, and J. Chuang. Resilient data-centric storage in wireless ad-hoc
sensor networks. In M.-S. Chen, P. K. Chrysanthis, M. Sloman, and A. B. Zaslavsky, editors,
Proceedings of the 4th International Conference on Mobile Data Management, Vol. 2574
of Lecture Notes in Computer Science, Melbourne, Australia, January 2003, Springer,
New York, pp. 45–62.
18
ALGORITHMS FOR WIRELESS SENSOR NETWORKS: PRESENT AND FUTURE
50. C. Jaikaeo, C. Srisathapornphat, and C.-C. Shen. Diagnosis of sensor networks. In Proceedings of the 2001 IEEE International Conference on Communications (ICC’01), Vol. 1,
Helsinki, Finland, June 2001, IEEE, New York, pp. 1627–1632.
51. C.-C. Shen, C. Srisathapornphat, and C. Jaikaeo. Sensor information networking architecture and applications. IEEE Personal Communications Magazine, 8(4):52–59, 2001.
52. C.-f. Hsin and M. Liu. A distributed monitoring mechanism for wireless sensor networks. In
Proceedings of the ACM Workshop on Wireless Security (WiSe’02), Atlanta, GA, September
2002, ACM, New York, pp. 57–66.
53. J. Zhao, R. Govindan, and D. Estrin. Computing aggregates for monitoring wireless sensor networks. In Proceedings of the 1st IEEE International Workshop on Sensor Network
Protocols and Applications (SNPA 2003), Anchorage, AK, May 2003, IEEE, New York,
pp. 139–148.
54. J. Zhao, R. Govindan, and D. Estrin. Residual energy scans for monitoring wireless sensor
networks. In Proceedings of the IEEE Wireless Communications and Networking Conference (WCNC ’02), Vol. 1, Orlando, FL, USA, March 2002, IEEE, New York, pp. 356–362.
55. Y. J. Zhao, R. Govindan, and D. Estrin. Residual energy scans for monitoring wireless
sensor networks. In IEEE Wilress Communications and Networking Conference (WCNC
’02), Orlando, FL, March 2002, CD-ROM.
56. X.-Y. Li, P.-J. Wan, and O. Frieder. Coverage in wireless ad-hoc sensor networks. IEEE
Transactions on Computers, 52(6):753–763, 2003.
57. X. f. A. Boukerche. A coverage-preserving scheme for wireless sensor network with irregular sensing range. In INFOCOM [79], pp. 1303–1316.
58. S. Megerian, F. Koushanfar, G. Qu, G. Veltri, and M. Potkonjak. Exposure in wireless sensor
networks: Theory and practical solutions. Wireless Networks, 8(5):443–454, 2002.
59. L. Hu and D. Evans. Secure aggregation for wireless networks. In Workshop on Security
and Assurance in Ad Hoc Networks, Orlando, FL, USA, January 2003, IEEE, New York,
pp. 384–391.
60. H. Cam, S. Ozdemir, P. Nair, and D. Muthuavinashiappan. ESPDA: Energy-efficient and
secure pattern-based data aggregation for wireless sensor networks. In Proceedings of
the IEEE Sensors, Vol. 2, Toronto, Canada, October 2003, IEEE, New York, pp. 732–
736.
61. B. Przydatek, D. Song, and A. Perrig. SIA: Secure information aggregation in sensor networks. In Akyildiz et al. [81], pp. 255–265, ISBN 1-58113-707-9.
62. T. Imielinski and S. Goel. DataSpace: Querying and monitoring deeply networked collections in physical space. IEEE Personal Communications, 7(5):4–9, 2000.
63. M. A. Sharaf, J. Beaver, A. Labrinidis, and P. K. Chrysanthis. TiNA: A scheme for temporal
coherency-aware in-network aggregation. In Proceedings of the 3rd ACM International
Workshop on Data Engineering for Wireless and Mobile Access, San Diego, CA, September
2003, ACM, New York, pp. 69–76.
64. N. Sadagopan, B. Krishnamachari, and A. Helmy. Active query forwarding in sensor networks. Ad Hoc Networks, 3(1):91–113, 2005.
65. S. R. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong. TinyDB: An acqusitional
query processing system for sensor networks. ACM Transactions on Database Systems,
30(1):122–173, 2005, ISSN 0362-5915.
66. M. Coates. Distributed particle filters for sensor networks. In K. Ramchandran,
J. Sztipanovits, J. C. Hou, and T. N. Pappas, editors, IPSN’04, Berkeley, CA, April 2004,
ACM, pp. 99–107, ISBN 1-58113-846-6.
BIBLIOGRAPHY
19
67. X. Sheng, Y. H. Hu, and P. Ramanathan. Distributed particle filter with GMM approximation
for multiple targets localization and tracking in wireless sensor network. In IPSN, editor,
IPSN’05, Los Angeles, CA, USA, April 2005, IEEE, pp. 181–188, ISBN 0-7809-9201-9.
doi: http://dx.doi.org/10.1109/IPSN.2005.1440923.
68. T. Vercauteren, D. Guo, and X. Wang. Joint multiple target tracking and classification
in collaborative sensor networks. IEEE Journal on Selected Areas in Communications,
23(4):714–723, April 2005, ISSN 0733-8716. doi: 10.1109/JSAC.2005.843540.
69. T. Schmitt, R. Hanek, M. Beetz, S. Buck, and B. Radig. Cooperative probabilistic state
estimation for vision-based autonomous mobile robots. IEEE Transactions on Robotics
and Automation, 18(5):670–684, 2002.
70. A. Boukerche. In Handbook of Algorithms for Wireless and Mobile Networks and Computing, Chapman & Hall/CRC Press, Boca Raton, FL, 2005, pp. 841–864, ISBN 1-58488465-7.
71. S. Santini and K. Römer. An adaptive strategy for quality-based data reduction in wireless
sensor networks. In Proceedings of the 3rd International Conference on Networked Sensing
Systems (INSS 2006), Chicago, IL, June 2006, TRF, pp. 29–36, ISBN 0-9743611-3-5.
72. A. Woo, T. Tong, and D. Culler. Taming the underlying challenges of reliable multihop
routing in sensor networks. In Akyildiz et al. [81], pp. 14–27, ISBN 1-58113-707-9.
73. E. F. Nakamura, F. G. Nakamura, C. M. S. Figueiredo, and A. A. F. Loureiro. Using information fusion to assist data dissemination in wireless sensor networks. Telecommunication
Systems, 30(1–3):237–254, 2005, ISSN 1018-4864.
74. C. M. S. Figueiredo, E. F. Nakamura, and A. A. F. Loureiro. An event-detection estimation
model for hybrid adaptive routing in wireless sensor networks. In Proceedings of the 2007
IEEE International Conference on Communications (ICC ’07), Glasgow, Scotland, May
2007, IEEE, New York. IEEE CD-ROM.
75. C.-L. Yang, S. Bagchi, and W. J. Chappell. Location tracking with directional antennas in
wireless sensor networks. In 2005 IEEE MTT-S International Microwave Symposium Digest, Long Beach, CA, June 2005, IEEE, New York doi: 10.1109/MWSYM.2005.151640.
76. C. Hongyang, D. Ping, X. Yongjun, and L. Xiaowei. A robust location algorithm with biased
extended Kalman filtering of TDOA data for wireless sensor networks. In Proceedings of the
International Conference on Wireless Communications, Networking and Mobile Computing
(WCNM ’05), Vol. 2, Wuhan, China, September 2005, IEEE, New York, pp. 883–886. doi:
10.1109/WCNM.2005.1544192.
77. T. Li, A. Ekpenyong, and Y.-F. Huang. Source localization and tracking using distributed
asynchronous sensor. IEEE Transactions on Signal Processing, 54(10):3991–4003, 2006.
78. A. Jain, E. Y. Chang, and Y.-F. Wang. Adaptive stream resource management using Kalman
filters. In Proceedings of the 2004 ACM SIGMOD International Conference on Management
of Data (SIGMOD’04), Paris, France, 2004. ACM, New York, pp. 11–22, ISBN 1-58113859-8. doi: http://doi.acm.org/10.1145.1007568.1007573.
79. INFOCOM, editor. 20th Annual Joint Conference of the IEEE Computer and Communications Societies (INFOCOM 2001), Anchorage, AK, April 2001, IEEE, New York.
80. Mobicom, editor. Proceedings of the 7th Annual ACM/IEEE International Conference
on Mobile Computing and Networking (Mobicom ’01), Rome, Italy, July 2001, ACM,
New York, ISBN 1-58113-422-3.
81. I. F. Akyildiz, D. Estrin, D. E. Culler, and M. B. Srivastava, editors. Proceedings of the
1st International Conference on Embedded Networked Sensor Systems (SenSys ’03), Los
Angeles, CA, USA, November 2003, ACM, New York, ISBN 1-58113-707-9.
CHAPTER 2
Heterogeneous Wireless
Sensor Networks
VIOLET R. SYROTIUK and BING LI
Department of Computer Science and Engineering, Arizona State University, Tempe, AZ
85287-8809
ANGELA M. MIELKE
Distributed Sensor Networks Group, Los Alamos National Laboratory, Los Alamos, NM 87545
2.1 INTRODUCTION
Wireless sensor networks (WSNs) have emerged as an important new class of computation that embeds computing in the physical world. To date, most of the work
has focused on homogeneous WSNs, where all of the nodes in the network are of
the same type. However, the continued advances in miniaturization of processors and
in low-power communications combined with mass-produced sensors have enabled
the development of a wide variety of nodes. When more than one type of node is
integrated into a WSN, it is called heterogeneous. While many of the existing civilian
and military applications of heterogeneous wireless sensor networks (H-WSNs) do
not differ substantially from their homogeneous counterparts, there are compelling
reasons to incorporate heterogeneity into the network. These include:
r
r
r
r
r
r
Improving the scalability of WSNs.
Addressing the problem of nonuniform energy drainage.
Taking advantage of the multiple levels of fidelity available in different nodes.
Reducing energy requirements without sacrificing performance.
Balancing the cost and functionality of the network.
Supporting new and higher-bandwidth applications.
As we will see, many of these reasons are interrelated. However, before discussing the
new dimension that heterogeneity brings to the algorithms that run on such wireless
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
21
22
HETEROGENEOUS WIRELESS SENSOR NETWORKS
Figure 2.1. (a) Mica and (b) Stargate family of processors (not to scale).
sensor networks, we discuss the typical forms of and architectures for heterogeneous
wireless sensor networks.
2.1.1 Forms of Heterogeneity
A sensor node is made up of four basic components [1]: a processing unit, a transceiver
unit, a power unit, and a sensing unit. Heterogeneity may arise in each unit.
Nodes may vary significantly in their processing capability. For example, the Mica2
mote model MPR400CB [2] is based on the Atmel low-power ATmega128L microcontroller. It has only 128 Kbytes of program flash memory, 512 Kbytes of measurement (serial) flash memory, and 4 Kbytes of programmable read-only memory. In contrast, the Stargate [2] is a 400-MHz Intel PXA255 XScale processor with 64 Mbytes
of synchronous dynamic random access memory and 32 Mbytes of flash memory.
Figure 2.1 show the Mica and Stargate families of processors. Nodes with higher computational resources may perform more in-network processing, reducing the amount
and/or frequency of sensed information that needs to travel through the network.
Often, nodes that vary in transceiver unit also vary in their power unit. For example
the same Mica2 mote is a multichannel radio with four channels centered at 868 MHz.
It supports a data rate of 38.4 Kbaud drawing 27 mA to transmit at maximum power,
10 mA to receive, and less than 1 A to sleep. This mote is powered by two AA
batteries. The Stargate runs a version of IEEE 801.11a/b and can run off a lithium-ion
battery or an alternating-current power adaptor. Its power consumption is low, at less
than 500 mA. Typically, nodes with more powerful energy resources are used to form
a backbone of the network, taking the communication burden. In terms of energy
consumption, the wireless exchange of data between nodes strongly dominates other
node functions such as sensing and processing [3, 4].
A number of classes of sensors are available. These include light, temperature, relative humidity, barometric pressure, acceleration, seismic, acoustic, radar, magnetic,
camera, and global positioning system (GPS) among others. In each class, the sensors
vary greatly in fidelity and hence may vary significantly in accuracy and in reliability.
INTRODUCTION
23
Higher fidelity usually comes at a higher cost, so balancing the functionality and the
cost consequently impacts the network architecture.
Currently, almost all of the nodes in both the homogeneous and the heterogeneous WSNs considered are static. In the future, nodes that are mobile, such as those
considered in reference 5, will inevitably be integrated into the network.
2.1.2 Architectures for Heterogeneous Wireless Sensor Networks
Two classes of architecture have emerged for heterogeneous WSNs: staged and
hierarchical.
In a staged architecture the nodes are organized into a series of n tasks performed
step-by-step. Within each stage, nodes are typically homogeneous, while successive
stages are heterogeneous in their capabilities. As the nodes in stage i, 1 ≤ i ≤ n − 1,
complete their task, they trigger the nodes in the next stage i + 1 to carry out their
task. Often, decreasing numbers of nodes of increasing fidelity are used in successive
stages. Figure 2.2a illustrates this architecture of a heterogeneous WSN.
A hierarchical architecture is an organization of the nodes into a forest of trees.
There are two ways in which the forest is commonly formed for n distinct node types.
In the single-hop organization, each tree has as many levels n as node types. The root
node (level zero) of each tree in the forest usually corresponds to nodes of highest
fidelity. Nodes at level i, 1 ≤ i ≤ n − 1, in a tree typically correspond to nodes of
type i. Each leaf or intermediate node is connected directly (via an edge) to its parent.
In a multihop organization, the key difference is that leaf nodes may traverse a multihop
path to a node of higher fidelity. Therefore the number of levels in each tree does not
equal n. Furthermore, each tree may have a different height.
Figure 2.2b shows a hierarchical architecture made up of two node types. The
resulting forest has four trees, each of height one; each leaf node reaches its parent
in a singlehop. In contrast, Figure 2.2c also shows a hierarchical architecture for two
node types. Here the forest has four trees: one of height one, two of height two, and
one of height three. Hence, some of the low-fidelity nodes traverse a multihop path
to a node of higher fidelity.
While in each forest in Figures 2.2b and 2.2c the root of each tree forms a backbone
of the network, there is an important difference between them. In Figure 2.2b the
backbone is used to reach an information sink node, common in the architecture of
WSNs. In Figure 2.2c the network has no sink node. The high-fidelity nodes may
communicate amongst themselves and compute in a distributed manner.
Nodes physically organized into a hierarchy may be logically organized into stages
to accomplish a series of tasks. However, this need not be the case. The hierarchy
may exist solely to improve the scalability, functionality, or resource efficiency of the
network.
2.1.3 Chapter Organization
The rest of this chapter is organized as follows. In order to motivate some of the problems that arise in heterogeneous WSNs, Section 2.2 describes two testbeds. SensEye
HETEROGENEOUS WIRELESS SENSOR NETWORKS
Stage 1 Sensor Nodes
Increasing Node Numbers
(a)
Stage 2 Sensor Nodes
Increasing Node Fidelity
24
Stage n Sensor Nodes
Sink
Cluster-head
Sensor node
(b)
(c)
Figure 2.2. Architectures of heterogeneous WSNs: (a) Staged architecture (b) Single-hop
hierarchical architecture with sink (c) Multi-hop hierarchical architecture without Sink.
is a testbed for monitoring and surveillance in which four types of cameras are used
with three different hardware platforms, organized in a three-stage architecture. Radioactive source detection is the goal of the Los Alamos National Laboratory testbed.
It, too, is organized in a three-stage architecture. Section 2.3 examines the problems of scalability and nonuniform energy drainage in homogeneous WSNs and how
they are addressed in heterogeneous WSNs. Algorithms for topology formation and
routing are presented. Coverage is a fundamental problem in homogeneous WSNs.
The problems of differentiated and stochastic coverage in heterogeneous WSNs are
discussed in Section 2.4. Section 2.5 examines issues in the management of heterogeneous networks. Section 2.6 presents two new applications, live virtual reality and
HETEROGENEOUS WIRELESS SENSOR NETWORK TESTBEDS
25
do not disturb, that are enabled by heterogeneity. Section 2.7 provides a summary
of established research projects in heterogeneous WSNs. As well, a summary of
systems infrastructure (including system software, middleware, and simulators) under development for heterogeneous WSNs is provided. Finally, Section 2.8 provides
many potential directions of study in this emerging area of research in heterogeneous
wireless sensor networks.
2.2 HETEROGENEOUS WIRELESS SENSOR NETWORK TESTBEDS
2.2.1 A Heterogeneous Camera Sensor Network
SensEye is a heterogeneous camera sensor network motivated by two applications:
(a) the monitoring of rare species in remote forests and (b) surveillance in disaster
management [5]. Both applications share characteristics that involve three basic tasks:
1. Object Detection. The presence of a new object is detected in the monitored
environment. A good algorithm minimizes the latency in detection.
2. Object Recognition. Once a new object is detected, it is classified by type.
3. Object Tracking. If the object is of interest, then tracking is warranted. This
involves multiple tasks including computing the location and trajectory of
the object, the ability to track the object as it moves out of the visual range
of one camera sensor and into the range of another, and streaming video (or a
sequence of still images) of the object to a monitoring station.
In order to achieve low latency in detection without sacrificing energy efficiency, a
staged architecture is used. Low-fidelity cameras perform the simpler task of motion
detection, while higher-fidelity cameras are woken up on-demand for object recognition and tracking. Figure 2.3 shows the three stages of the SensEye architecture. The
imaging, processing, and networking capabilities improve with increasing stage.
The first stage of SensEye is made up of Mica2 motes [2] equipped with 900 MHz
radios and low-fidelity Cyclops [6] or CMUcam [7] camera sensors. The second
stage is made up of Stargate nodes [2] equipped with webcams. Each Stargate runs
Linux and is equipped with a webcam that can capture higher-fidelity images than
the cameras at stage one. Each stage two node has two radios: (1) an IEEE 802.11
radio that is used by the Stargates to communicate with each other and (2) a 900-MHz
radio that is used to communicate with the motes in the first stage. The third stage is
a sparse deployment of high-resolution pan–tilt–zoom (PTZ) cameras connected to
embedded personal computers. These cameras are used to fill any gaps in coverage
of the second-stage nodes and to provide additional redundancy for tasks such as
localization. There are no base stations in the architecture; nodes are assumed to
communicate in ad hoc mode in each stage and between stages.
The principles that guide the design of SensEye include:
r Each task is mapped to the least powerful stage with sufficient resources.
r Wake-up on demand, along with redundancy in coverage, is exploited.
26
HETEROGENEOUS WIRELESS SENSOR NETWORKS
Serial
Cable
Serial
Cable
Serial
Cable
Serial
Cable
Node
USB
CMUcam
Webcam
Stargate
Node
Node
Mini–ITX
Pan-Tilt-Zoom
Camera
CMUcam
Node
CMUcam
Ethernet
USB
Webcam
Stargate
Node
Node
CMUcam
Stage 1
Stage 2
Stage 3
Figure 2.3. Staged architecture of SensEye.
In general, object detection requires few resources and it is therefore performed at
stage one. The camera and the sensor node are duty-cycled, or woken up periodically,
to detect the presence of a new object. In addition, SensEye uses a randomized dutycycling algorithm where different cameras are woken up at different times to further
reduce object detection latency. A frame differencing algorithm is used to detect
objects. Each camera compares the newly acquired image to a background image
obtained when the system is calibrated; the pixel difference is used to indicate the
presence of a new object.
If a new object is detected in stage one, appropriate stage two nodes are woken up.
This involves computing the coordinates of the object and determining the stage two
nodes that have cameras pointing at its location; the details of object localization are
provided in reference 5. Each stage one node knows the visual range of each stage two
node in its vicinity and can therefore use the coordinates of the object to determine
the most appropriate stage two nodes. If no appropriate stage two node is identified,
a stage three camera is woken up, since it can use its pan and tilt capabilities to point
to the location of the object. Localization is feasible only when at least two stage one
nodes view the object. If only a single stage one node detects the object, then all stage
two nodes that have overlapping coverage with it are woken up.
The separation of object detection and recognition across stages introduces latency
between the execution of tasks. This latency includes the delay in receiving and
processing a wake-up packet as well as the delay in waking up a stage two node. The
wake-up process begins with the transmission of a wake-up packet to a stage two
node similar to “wake-on-wireless” energy saving strategy of Shih et al. [8]. Upon
receiving a wake-up packet, the stage two node transitions from a suspend to an awake
state. By running a minimum of device drivers, this transition time is kept small.
HETEROGENEOUS WIRELESS SENSOR NETWORK TESTBEDS
27
Accurate recognition of an object requires a higher-fidelity image and significantly
greater processing and memory resources than are available on a stage one node. As
a result, the recognition algorithm is executed using higher-fidelity webcams and
the more capable processors of stage two nodes. As a proof-of-concept, SensEye
implements two well-known recognition algorithms from the computer vision literature in the stage two nodes [5]. Object tracking in SensEye involves a combination of
detection, localization, and wake-up, in addition to recognition. The current system
can track objects moving slowly.
An experimental evaluation of SensEye shows that, compared to a flat network
architecture, an order of magnitude improvement in energy usage is obtained. Despite
the energy reduction, similar detection performance (only 6% more missed detections)
is obtained. Detection latency and energy usage at the stage one nodes is an order of
magnitude less than at the stage two nodes. The mean localization errors indicate that
detection can be performed by the lower-fidelity cameras of stage one while tracking
is best done using higher-fidelity cameras; see Kulkarni et al. [5] for the details of the
evaluation.
The SensEye heterogeneous camera sensor network testbed has demonstrated successfully the benefits of a staged architecture over a flat architecture with respect to
energy usage. Continuing research examines tradeoffs such as system cost and coverage. Design issues that impact performance, such as (a) the number of stages in the architecture and (b) the allocation of tasks to sensors, are also under study. The problem
of streaming video (or a sequence of still images) of the object to a monitoring station is
not addressed in this work; providing quality-of-service (QoS) support for such highbandwidth, real-time data is a challenging open problem in heterogeneous WSNs.
2.2.2 Detection of Radioactive Sources
A team of researchers at Los Alamos National Laboratory have spent the past several years focusing on the development of heterogeneous WSNs for event detection.
Typical deployments of sensor networks revolve around biological or environmental
monitoring applications where the emphasis is on collecting all of the data from a
sensor array to be sent back to a laboratory for detailed analysis. Applications of
interest to the Distributed Sensor Networks with Collective Computation (DSN-CC)
team have instead focused on the detection, classification, and tracking of radiological materials within the sensor network. These goals are very similar to those of the
SensEye system, requiring all processing to be performed within the network with no
use of base stations. However, this work relies on multiple sensor modalities instead
of a single sensor type for event detection.
The motivation for this research is to develop systems to guard against attacks from
radiological dispersal devices (RDDs) capable of contaminating an area or population
with fissile material. A potentially last line of defense for such attacks may reside in
systems placed along roadways that are able to detect such material in-transit and
alert the appropriate authorities before dispersal occurs [9].
One approach to such a threat employs portal monitoring equipment. Portals provide high fidelity results; however, they are large, conspicuous, and costly and require
28
HETEROGENEOUS WIRELESS SENSOR NETWORKS
Stage 3
Stage 2
Stage 1
First Stage Mica2
900MHz
Mica2
MTS310
First Stage Stargate
2.4GHz
Stargate
Vehicle
Highway
Vehicle
900MHz
Second Stage Stargate
2.4GHz
Stargate
Radiation Detector
Third Stage Stargate
2.4GHz
Stage 1
Stage 2
Stage 3
Stargate
Video Camera
Figure 2.4. Staged architecture for radioactive source detection.
considerable time and infrastructure to set up. Instead the DSN-CC project strives to
employ networks of small, low-cost, heterogeneous sensors in producing similar quality results in persistent applications where the system deployments may occur utilizing
fast, low-impact methods. Such a system must be heterogeneous in sensor type, as
well as in node backbone, to allow for data redundancy, in-network processing, and
hypothesis validation.
In the SensEye system a staged architecture is employed as a means of achieving
energy efficiency. While energy efficiency is an important factor for the development
of the DSN-CC system, the staged architecture is utilized instead as a means of gaining
confidence in a network-developed detection while decreasing system false alarms.
A single radiation detector provides specific detection and false alarm rates; coupling
a string of radiation detectors with seismic sensors, magnetometers, acoustic sensors,
atmospheric sensors, and video cameras increases dramatically the fidelity of the
decisions made within the network.
Figure 2.4 shows the three stages of the radiation detection application of the
DSN-CC system. Stage one consists of both Mica2 motes equipped with the MTS310
multisensor boards and Stargate nodes. Stage two is comprised of an array of radiation
detectors connected directly to the Stargate nodes, and stage three is a video camera
connected to a Stargate node. In all instances the Mica2 motes communicate through
an embedded 900-MHz radio while the Stargate nodes are equipped with both the
900-MHz radio and an IEEE 802.11 radio transmitting at the 2.4-GHz frequency.
Figure 2.5 shows the inexpensive Mica2 motes densely placed along the roadway
in stage one to allow for hardware failures through redundancy. The goal of stage one
is to detect the presence of a vehicle utilizing data from the acoustic and magnetometer
sensors for cueing stage two and three sensors. Vehicle classification may also occur
during this stage. The algorithms for vehicle detection and classification are embedded
within the network in the stage one Stargate nodes. The stage two nodes remain in a
background collection mode and stage three nodes remain inactive until a vehicle is
detected within the network by the stage one nodes. This helps to minimize the false
alarm rate of the radiation detection assets, as well as reducing the required energy
draw of the stage three nodes.
HETEROGENEOUS WIRELESS SENSOR NETWORK TESTBEDS
29
Figure 2.5. Stage one Mica2 vehicle detection network.
When a vehicle is detected within the sensing array, the stage one Stargate nodes
transmit cues to the second stage nodes. These radiation detection nodes change state
from background count collection to an active, timed-count collection. Each stage two
detector node takes a radiation count while the passing vehicle is directly in front of
it based upon its node location and the speed of the passing vehicle. These counts are
coherently added across the network and compared to the environmental background
readings [10]. If the vehicle is suspected of carrying radiological material, the stage
three video camera is cued to collect an image of the offending vehicle. This image,
along with the corroborating event information, is then relayed to a command-andcontrol console similar to Figure 2.6 for monitoring personnel for interdiction. In the
field, this command-and-control console is a tablet PC communicating to the network
via its IEEE 802.11 radio link.
Although commercially available hardware is limited in its performance and capabilities, ongoing field experiments provide an indication that heterogeneous systems
have the potential to provide low-cost, highly reliable solutions to many persistent
surveillance applications. Continuing research at the Los Alamos National Laboratory is focusing on (a) the operational issues of a network such as node and network
security, (b) validation studies of efficient communication protocol schemes, (c) development of additional embedded algorithms for further event classification and
30
HETEROGENEOUS WIRELESS SENSOR NETWORKS
Figure 2.6. Simple command-and-control console interface.
tracking, (d) the investigation of alternate backbone hardware, and (e) the investigation of long-range exfiltration schemes.
2.3 SCALABILITY AND SYSTEM LIFETIME
It is well known that homogeneous ad hoc networks, which include homogeneous
WSNs, suffer poor capacity. Gupta
√ and Kumar [11] were the first to show that
the throughput of a node is (1/ n log n), where n is the number of nodes in the
network—a very pessimistic result! In addition, as paths between nodes become
longer, the probability of packets being lost becomes higher. As the number of nodes
in a network grows, the successful end-to-end transmission rate drops significantly
[12]. Experimentation in simulation [13] and in testbeds [14] has confirmed that the
performance of homogeneous ad hoc networks does not scale with increasing n.
A primary difference between WSNs and ad hoc networks is that the traffic pattern is many-to-one, from the sensor nodes to the base station. Figure 2.7 shows a
homogeneous WSN with a base station at the center. A transmission from any sensor
node to the base station goes through one of the nodes within a distance of r of the
base station. These critical nodes have the highest burden of relaying traffic. As a
result, they are likely to exhaust their energy before other sensor nodes [15]. When
the critical nodes die, connectivity of the network is lost. Hence the energy drainage
rate of the critical nodes determine the system lifetime. Indeed, Du and Xiao [16]
found that when connectivity is lost, more than half of the nodes still have more than
50% of their energy left. This energy is wasted since communication with the sink is
no longer viable.
SCALABILITY AND SYSTEM LIFETIME
31
r
Figure 2.7. The energy of critical nodes (within the circle) drain nonuniformly.
Adding structure to the network can help improve scalability and also alleviate the
problem of nonuniform drainage of energy. Clustering is one way to add structure in a
homogeneous WSN. Heinzelman et al. propose LEACH, a clustered network in which
each cluster head aggregates data and transmits it directly to the base station [17]. The
cluster heads are periodically rotated for efficient load balancing and a consequent
lengthening of network lifetime. In order to minimize the total energy, the number of
cluster heads must scale as the square root of the total number of sensor nodes [17]. The
LRS [18] and power-aware chessboard-based adaptive routing (PCAR) [19] protocols
also aim to balance the energy consumption in a homogeneous sensor network. All of
these protocols suffer from overhead associated with frequent cluster-head rotation.
While the problem of routing in homogeneous WSNs has been considered in flat
architectures (see Directed Diffusion [20] as an example), when the architecture is
hierarchical, the routing protocol makes use of the hierarchy (see TTDD [21] and
LEACH [17] as examples). Hierarchy is shown to help a homogeneous WSN achieve
higher total throughput and increase the network lifetime.
In heterogeneous WSNs, it is often natural to organize the nodes into a hierarchy.
In this section we consider algorithms for heterogeneous WSNs to form hierarchical
topologies, and address the related problem of routing, in order to tackle the problems
of scalability and nonuniform energy drainage.
2.3.1 A Resource-Oriented Protocol: Topology Formation
and Routing
In WSN applications in which the sensor nodes are inherently heterogeneous in energy
resources, these differences should be considered in order to improve the network
capacity and extend the system lifetime.
32
HETEROGENEOUS WIRELESS SENSOR NETWORKS
Ma et al. [22, 23] consider a wireless in-home heterogeneous sensor network. This
may include devices embedded into everyday objects such as appliances, devices for
climate monitoring and environmental control, and medical devices integrated into
the home for monitoring medical conditions. Even mobile sensor nodes, such as those
carried by people or on mobile toys, are considered.
In-home sensor nodes are heterogeneous in their power units. Some nodes are
directly connected to the alternating current power supply and have, essentially, unlimited energy. Others are powered by batteries with varying capacities. To exploit
the heterogeneity in power units, a resource-oriented protocol (ROP) is proposed.
The goal is to achieve the longest system lifetime. To be precise, the system lifetime
for a sensor network is the shortest lifetime of any participating node in the network. The node lifetime for a sensor is the time at which the sensor exhausts all its
energy.
ROP exploits the existence of sensor nodes with unlimited resources. A topology
is formed to minimize the consumption of resources of energy-constrained sensor
nodes. Sensor nodes with unlimited energy serve as relays, since they can afford
higher transmission power and hence longer transmission range. This saves energy in
the energy-constrained sensor nodes and, as a result, extends the functional lifetime
of the network.
Figure 2.8 shows an example of ROP; node U has unlimited energy resources,
while the others are energy-constrained sensor nodes. Of these energy-constrained
nodes, nodes E1 and E2 are battery-powered local cluster heads with medium energy
capacity, and the rest of the nodes only have small energy capacity. When there is a
message that needs to be sent, for example, from a source node 1 to a destination node
7, a traditional multihop routing protocol might route the packets from node 1 through
nodes 2, 3, 4, 5, and 6 to 7. However, because of the existence of the unlimited energy
node U and the medium energy capacity node E1 , ROP would route the packets from
Node with unlimited energy resources
Energy-constrained cluster head
Energy-constrained node
U
E1
E2
4
5
8
3
1
2
6
7
Figure 2.8. Resource oriented routing versus multihop routing.
SCALABILITY AND SYSTEM LIFETIME
Time
Broadcast resource
characteristics
Small
Decide resource
levels
Medium
Aggregate data,
report to large
resource node
Receive Topology
Packet
Receive Topology
Packet
(a)
33
Resources:
Small
Medium
Large
Large
Broadcast its
neighbors
Contact other
large resource nodes
Broadcast Topology
(b)
Figure 2.9. Resource-oriented protocol for heterogeneous WSNs. (a) ROP topology formation
phase. (b) Sample ROP multihop hierarchical topology.
1 to 8, then to E1 and U, and then directly to 7. Although such a path is not energyefficient, it saves more energy in energy-constrained sensor nodes, in order to prolong
the operational lifetime of these nodes.
ROP has two phases: a topology formation phase and a topology update phase. The
topology formation phase, in turn, involves two steps. In the first step, each sensor
node reports its characteristics and available resources to all of its neighbors. The
local cluster-head aggregates these reports and sends it to the most powerful sensor
nodes. In the second step, these most powerful nodes decide the topology of the
network and broadcast routing information. On receipt of the topology packet, each
sensor node configures its route cache based on the topology decided. Figure 2.9a
shows the topology formation phase for a network with nodes at three resource levels:
small, medium, and large. This phase builds a multihop hierarchical topology with
the large resource nodes at level zero—that is, the root of each tree (see, for example,
Figure 2.9b).
In order to reduce the energy cost of the topology formation phase, some sensors
may be left isolated. The topology update phase takes care of this situation, and
it establishes routes to mobile nodes. ROP is one of the few heterogeneous WSN
protocols that incorporates mobile sensor nodes.
When a sensor wants to communicate with other sensors, it uses the route in its
cache. If the route is outdated, it sends a route request (RREQ) packet. The returned
route replaces the outdated route in its cache. If several routes are received, it chooses
the one with the largest resources; and if two routes have the same resources, the one
with fewer children is selected. The details of ROP and the reactive routing scheme
are described in Ma et al. [23].
The performance of ROP is evaluated in simulation. In ROP, energy efficiency
cannot always result in longer system lifetime. Rather, balancing resources among
sensors and saving energy for those more resource-constrained sensor nodes contributes to lengthening system lifetime.
2.3.2 Chessboard Clustering and Routing Protocol
Du and Xiao [16] propose a chessboard clustering and routing protocol for heterogeneous WSNs to overcome the performance bottleneck and poor scalability of
34
HETEROGENEOUS WIRELESS SENSOR NETWORKS
homogeneous WSNs and to address, at same time, the problem of nonuniform energy
consumption. A good observation is made that clustering alone does not solve the
problem of nonuniform energy drainage; indeed, the center node in Figure 2.7 can as
well be a cluster head rather than a base station.
Two types of nodes are assumed: a small number of high-end sensor nodes and
a large number of low-end sensor nodes. Each node is assumed to be aware of its
location. A cluster is formed around each high-end sensor node which serves as a
cluster head. Low-end sensor nodes perform the basic sensing as well as the relaying
of packets within the cluster. Given its powerful energy reserve and communication
ability, each high-end node performs data fusion within its cluster, and it transmits the
aggregated data to the sink via a single-hop link or a multihop path. In this way, the
network is divided into multiple regions, with each region assuming a smaller burden
of the communication due to the smaller number of sensor nodes within the cluster.
The network lifetime is therefore increased by transmitting fewer packets at lowend sensor nodes and utilizing the less power-constrained or non-power-constrained
nodes as much as possible.
Figure 2.10 shows the sensor field divided into equal-sized cells with adjacent cells
colored with different colors, resembling a chessboard. These nodes are assumed to
be uniformly and randomly distributed in the sensor field. Since each node knows its
location, it can determine if it is in a white cell or a black cell.
The basic idea is to use the underlying chessboard to define two clustered topologies, with only one clustering in use at a given time. In a white clustering, all highend sensor nodes in white cells are active while all high-end sensor nodes in black
cells are inactive. In a black clustering, all high-end sensor nodes in black cells are
active while all high-end sensor nodes in white cells are inactive. Low-end sensor
nodes are all active, forming multihop clusters around the currently active high-end
sensor nodes. The motivation for switching colors is as follows: sensor nodes that
are critical nodes in a white clustering are likely to become non-critical nodes in a
black clustering and vice versa. Since critical nodes consume more energy in packet
1
2
3
4
Figure 2.10. Chessboard clustering scheme.
SCALABILITY AND SYSTEM LIFETIME
2
3
2
1
Sensor Node
35
3
1
“Black” Cluster Head
“White” Cluster Head
Figure 2.11. A black (left) and white (right) clustering of the heterogeneous WSN.
forwarding than do other sensor nodes, switching the color of the clustering balances
the energy consumption among sensors, and prolongs the network lifetime.
Figure 2.11 shows an example of clustering based on black and on white cells,
respectively. In the black clustering, sensor node 3 is a critical node forwarding packets
on behalf of nodes 1 and 2. In the white clustering, nodes 1 and 2 become critical
forwarding packets on behalf of other nodes in the cluster; in particular, node 2 now
forwards the packets of node 3.
In order to form a black (white) clustering, each black (white) high-end sensor node
broadcasts a hello packet, containing its identifier and its location. Low-end sensor
nodes may receive hello packets from multiple black (white) high-end nodes. In a
two-dimensional sensor field, each low-end sensor node selects the closest high-end
sensor node as the cluster head; this leads to the formation of Voronoi cells where the
cluster heads correspond to the nuclei of the cells [16].
The decision to switch the color of the clustering is based on the energy levels of
the high-end nodes. Suppose the current clustering is black. Periodically, each black
high-end sensor node exchanges packets with its neighboring white high-end nodes.
The packets contain the energy remaining in the node. If the remaining energy of the
black high-end node drops below a threshold, its neighboring white high-end nodes
become active and initiate cluster formation. As the network runs, the black highend nodes drain their energy and become unavailable. Gradually, the white high-end
nodes become active.
Both intra- and intercluster routing protocols are proposed [13, 16]. Routing within
a cluster is achieved via a greedy geographic routing protocol. Each low-end sensor
node simply forwards a packet to the neighbor closest to the cluster head.
In order to support intercluster routing, after the clusters are formed, each cluster
head sends its location to the sink. The sink then broadcasts the locations of all clusters
heads. For a cluster head to communicate with the sink, it draws a line between
itself and the sink. The line intersects some number of Voronoi cells. The packet is
forwarded from the source cluster head to the sink through the cluster heads in these
relay cells. The chessboard routing protocol achieves a higher delivery ratio, lower
total energy consumption, smaller end-to-end delay, and better throughput than two
routing protocols for homogeneous WSNs. The details of the chessboard clustering
36
HETEROGENEOUS WIRELESS SENSOR NETWORKS
and routing protocols, as well as their performance evaluation in simulation, can be
found in references 13 and 16.
2.3.3 Analyses of System Lifetime in Heterogeneous WSNs
Mhatre and Rosenberg [24] present a cost-based comparative study of (a) homogeneous WSNs and (b) heterogeneous WSNs with two types of nodes. Their model
takes into account the cost of manufacturing the hardware as well as the battery energy of the sensor nodes. First, a single-hop clustered architecture is considered, with
LEACH [16] selected as the representative in a single-hop homogeneous WSN. For the
multihop homogeneous clustered architecture a multihop variant, called M-LEACH,
is proposed and analyzed. In comparing (a) the cost of the multihop homogeneous
clustered architecture with M-LEACH and (b) a multihop heterogeneous clustered
architecture, the homogeneous WSN can outperform the heterogeneous one if the
nonuniform energy drainage problem is not addressed.
Mhatre et al. continue their study of heterogeneous WSNs in reference 15. They
consider a WSN with nodes of two types distributed over a sensor field using twodimensional homogeneous Poisson point processes: (a) type 0 nodes with intensity
(average number per unit area) λ0 and battery energy E0 and (b) type 1 nodes with
intensity λ1 and battery energy E1 . The type 0 nodes do the sensing while the type 1
nodes act as the cluster heads. Nodes use multihop paths to communicate with their
closest cluster head. The optimum node intensities (λ0 , λ1 ) and node energies (E0 , E1 )
that guarantee a lifetime of at least T units, while ensuring both connectivity and
coverage of the surveillance area with high probability, are determined. The overall
cost of the network is minimized under these constraints. Here, the network lifetime
is defined as the number of successful data gathering trips (or cycles) that are possible
until connectivity and/or coverage are lost. Conditions for a sharp cutoff are taken into
account. This means that it is ensured that almost all the nodes run out of energy at
about the same time so that there is very little energy lost due to residual energy. The
results comparing a random deployment of nodes with a deployment in which nodes
are
√ placed deterministically along grid points show that λ1 scales approximately as
λ0 . The results can be extended to take into account unreliable nodes.
Duarte-Melo and Liu [25] examine the performance and the energy consumption
of a heterogeneous WSN providing periodic data from a sensor field to a remote
receiver. A flat homogeneous WSN is compared to one in which an overlay of fewer
more powerful sensor nodes is added. The energy consumption is formulated and the
estimated lifetime based on a clustering mechanism with varying parameters related
to the sensor field, such as size and distance, is studied. The optimal number of clusters
is quantified based on the model. Also, an allocation of energy between the two levels
of the architecture is discussed.
Li and Mohapatra [26] develop an analytical model for the problem of nonuniform
energy drainage. It is found that density does not affect the energy consumption rate
of a node. This confirms the fact that simply deploying more nodes in a network
cannot prolong its lifetime. Using the model, they investigate the effectiveness of some
existing approaches toward mitigating the nonuniform energy drainage problem in a
COVERAGE IN HETEROGENEOUS WSNs
37
formal manner. Using a hierarchical architecture alleviates, though does not eliminate,
the problem. Other approaches include investigating the impact of source bit rate and
the impact of traffic compression. Simulation is used to validate the analysis.
Ai et al. [27] develop an analytical model for energy dissipation for a homogeneous
WSN with a flat architecture and also for a heterogeneous WSN with a hierarchical
clustered architecture. The communication cost of multihop links increases with the
number of clusters, while the communication cost of forwarding messages on the
backbone increases with the number of clusters. Thus, there is an optimal number of
clusters that trade off the power consumption between multihop and single-hop links
to minimize the energy dissipation rate.
2.4 COVERAGE IN HETEROGENEOUS WSNs
One of the fundamental issues in homogeneous WSNs is the problem of coverage,
which reflects how well a sensor network is monitored by sensor nodes. Several forms
of coverage have been studied.
In order to achieve deterministic coverage, a static network of predefined shape
must be deployed. A grid-based sensor network is an example of a uniform deterministic deployment. In this case, the problem of coverage of the sensor field reduces
to the problem of coverage of one cell and its neighborhood [28, 29]. A weighted
deployment might be used in an art gallery where more valuable objects are equipped
with more sensors to maximize the coverage of the security system.
In many situations, deterministic deployment of the sensor nodes is neither practical nor feasible. Instead, sensor nodes may be randomly distributed in the sensor field
and stochastic coverage is considered [30–33]. The stochastic random distribution
model may be uniform, Gaussian, or any other distribution based on the application.
In this setting, Megerian et al. [29] study the worst-case and best-case coverage
problems. Informally, in the worst-case coverage problem, the goal is to find the
closest distance to sensor nodes that an agent traveling on any path in the sensor field
must encounter at least once. The idea is that the closest distance to sensor nodes
is one metric by which coverage of the field may be characterized. This scheme is
worst-case since the closest distances to sensor nodes is determined, even if the agent
tries to avoid them. At the other extreme is the best-case coverage problem, where the
goal is to find the farthest distance to sensor nodes that an agent traveling on any path
in the sensor field must have from the nodes even if it tries to stay as close to them as
possible. Provably optimal polynomial time algorithms for the best- and worst-case
coverage problems are provided [29].
Yan et al. [34] examine the problem of differentiated coverage (corresponding
to weighted deployments) in homogeneous WSNs. A protocol is designed in which
each node is able to dynamically decide a schedule for itself in order to guarantee a
degree of coverage. The schedule has an average energy consumption that is inversely
proportional to the node density.
We highlight the work on coverage in heterogeneous WSNs next.
38
HETEROGENEOUS WIRELESS SENSOR NETWORKS
2.4.1 Differentiated Coverage in Heterogeneous WSNs
In some applications, differentiated coverage is necessary. This is when a different
degree of coverage is applied to different parts of the network [34, 35]. For example,
some areas of a battlefield are of more interest than others.
Du and Lin [35] propose an algorithm for uniform coverage that can be extended
to provide differentiated coverage of a heterogeneous WSN. As in references 13
and 16, two types of nodes are assumed: a small number of high-end sensor nodes
and a large number of low-end sensor nodes. Each node is assumed to be aware of its
location.
A logical grid is assumed to overlay the sensor field with certain grid points requiring coverage k. In order to provide uniform coverage of the grid points, the goal
is to design a node scheduling algorithm that ensures that all grid points have the
required coverage while at the same time minimizes the total energy consumption
and balances node energy consumption.
The high-end sensor nodes know the locations of the low-end nodes and can
compute which low-end nodes cover a grid point. In Figure 2.12, the low-end sensor
nodes A, B, C, and D cover grid point 1. If k sensor nodes cover a grid point, then
an ideal schedule has each node awake for T/k time and asleep for T − T/k time,
in a round T . However, a low-end sensor node may need to cover other grid points.
Therefore the high-end sensor considers the assigned slots when each low-end sensor
is awake and assigns a time slot that has the maximal overlap with the existing awake
slots. For example, if node D already has a slot of [0, T/4] for covering grid point 1,
then the high-end node can assign an awake slot of [0, T/3] to D. Node D need only be
active during [0, T/3] in order to cover both grid points 1 and 2. If there is a conflict,
then a node may require the assignment of additional slots in which it is awake. After
determining the node schedule for all grid points, the high-end node broadcasts it to
all low-end sensor nodes. The schedule is updated periodically to ensure the coverage
algorithm is robust to sensor failure.
A
1
D
2
B
C
3
4
Figure 2.12. Coverage for grid points.
MANAGEMENT OF HETEROGENEOUS WSNs
39
The uniform coverage algorithm is extended to provide differentiated coverage. To
provide a coverage degree c of a certain grid point, the high-end node correspondingly
adjusts the time awake for each low-end sensor in the coverage area. For a grid point
covered by k sensor nodes, T/k slots in each round T are assigned to provide coverage
of degree 1. For coverage of degree c, the number of slots must be c · T/k. The
complete algorithm for differentiated coverage is provided in reference 35.
While the focus of Mhatre et al. [15] is on maximizing the lifetime of the network,
one of the constraints relates to coverage. Two types of node are deployed over a sensor
field for the purpose of surveillance. One type of node does the sensing while the other
type acts as cluster heads. An aircraft visits the area periodically and gathers data about
the activity in the field from the sensor nodes. The problem is treated assuming that
the base station (aircraft) receives updates from every cluster. However, if the base
station is interested in receiving updates from only a few clusters (an extrasensitive
region), then the analysis can be modified to accommodate this requirement. More
nodes are deployed over the regions of frequent updates, and these nodes are taken
into account in the overall network cost. The redundant nodes stay inactive while the
battery energy of other nodes lasts; they join the cluster when the other nodes start to
expire.
2.4.2 Stochastic Coverage in Heterogeneous WSNs
Lazos and Poovendran [31] study the following stochastic coverage problem in heterogeneous WSNs: Given a planar sensor field and n sensor nodes deployed according
to a known distribution, compute the fraction of the sensor field that is covered by at
least k sensor nodes, k ≥ 1. This may also be viewed as a problem in k-coverage [33].
The problem is formulated as a set intersection problem arising in integral geometry. Analytical expressions for stochastic coverage as then derived. The formulation
does not require the sensor nodes to have identical sensing capability, and it does not
restrict the distribution according to which the sensors are deployed. In addition, the
formulation is applicable to scenarios where the sensing area of each sensor node has
arbitrary shape. The validity of the derived expressions are verified by simulation.
2.5 MANAGEMENT OF HETEROGENEOUS WSNs
By definition, heterogeneous WSNs have more than one type of sensor, making their
management increasingly important. Management includes:
r
r
r
r
r
r
Coordinating and scheduling tasks for sensors.
Optimizing the use of capabilities and resources.
Managing the sensor data aggregation and correlation.
Assessing the situation.
Adapting the sensor network.
Reducing human involvement.
40
HETEROGENEOUS WIRELESS SENSOR NETWORKS
Visualization Layer
Query Results
Informantion
Tracking Algorithm
Data Fusion
Sensor
Configuration &
Management
Signal Processing
Command
Sensor Data
Physical Sensors
Figure 2.13. Position of management in the system model.
Vaidya et al. [36] propose a framework for sensor configuration and management
to take the responsibility of making decisions in order to coordinate the assignment
and scheduling of sensor nodes best suited for the application. The application
considered is tracking and movement of objects in a moderately occupied confined
space. Figure 2.13 shows how the management component is positioned in the
unified sensing system model.
A manager is designed to operate over a heterogeneous WSN that provides sensory
data from multiple types of sensor. The goal of the management system is to minimize
the energy consumption and the required bandwidth while preserving the quality of
tracking.
When tracking the movement of one object, the system uses a set of three sensor
nodes to determine the current location of the object and to predict the next set
of sensor nodes to use according to its velocity and direction. This allows the rest
of the sensor nodes in the network to go to sleep for the next detection round. For
multiple targets that are far away from each other, tracking is similar to a single object
moving. When multiple objects move very close to each other, there is ambiguity in
the data acquired from the sonar sensors. In this case, visual sensors come to the aid.
Figure 2.14 shows the complete flow chart for the sensor management system. With
the help of management system, a significant energy reduction is achieved compared
to a randomized activation scheme.
The challenge of correlating the data gathered by several sensor nodes listening to
live traffic is studied by Andersson et al. [37]. Correlating data from different types
of sensor brings a number of benefits. The first is a reduction of the number of alerts
that a user must address. The correlation engine should recognize when reports from
multiple sensors refer to the same incident. Correlation can enhance the detection
capability as a second benefit. In addition, correlation can exploit the complementary coverage from several sensors. Reports from several sensors employing diverse
NEW APPLICATIONS ENABLED BY HETEROGENEOUS WSNs
Line-Breaking sensor on?
41
No
Yes
Turn camera on, capture image, turn camera off.
Activate sonar sensors in area,
Analyze sonar data, turn sonar sensors off.
Are there sufficient data?
Yes
Compute next position of object
No
Wait until there
are sufficient
data
Assign probabilistic position
to tracked object
Turn camera on,
check identity of object
Turn next set of sonar sensors on
Figure 2.14. Sensor management flow chart.
analytical techniques may reinforce each other. A standard format is developed to
facilitate the interoperability of diverse sensors.
Management of the WSN may also help network scalability. In reference 38,
Gürgen et al. propose a hybrid approach offering scalable solutions that combine
the advantages of both centralized and distributed data stream management. Their
main concern is the querying and system management of large sets of sensors. In the
heterogeneous WSN considered, diverse types of sensor nodes are used, each with
a different data delivery rate. A three-stage architecture is proposed for distributed
evaluation of queries on the network: sensor nodes, gateway, and control site. In the
architecture, the load of query evaluation is distributed between stages. A mediatorwrapper [39] is applied at the gateway stage to serve as local query translator and
optimizer. The wrapper proposed is an approach for heterogeneous sensor data management and provides an integrated global view of the different types of sensor. With
the help of this three-stage management scheme, the query load is distributed; also, the
burden at each stage is decreased. Hence, the scalability of the network is improved
by manipulating the query in the heterogeneous WSN.
2.6 NEW APPLICATIONS ENABLED BY HETEROGENEOUS WSNs
Using a homogeneous WSN may not satisfy the requirements of an application [40].
The following two applications are examples requiring a heterogeneous WSN.
42
HETEROGENEOUS WIRELESS SENSOR NETWORKS
The ability to reliably deliver a large volume of data has opened a new range of
applications for heterogeneous WSNs. This includes audio/video surveillance and
the monitoring of telemetry data. Yuan et al. [40] consider a new live virtual reality
application. The idea is to provide a user with live real-time video of a monitored
sensor field together with the ability for the user to navigate virtually within the field.
One example is that of a heterogeneous WSN deployed for securing a building, where
a console operator can survey a building using a joystick for navigation.
Gnawali and Yarvis [41] propose a “do not disturb” application requiring a heterogeneous WSN. The purpose of this application is to alert people to keep the noise
down in an office environment when there are people working in a nearby area.
A “do not disturb” WSN consists of nodes with motion and sound sensors. Motion
detectors determine the occupancy status of a cubicle. Sound sensors in cubicles
measure the loudness of the sound heard in each cubicle. Sound sensors in the hallways
pick up the noise of impromptu hallway meetings. The “do not disturb” application
determines the source of the noise by data fusion; this requires CPU and memory
resources typically not available in sensor nodes. Figure 2.15a shows two types of
sensor node deployed in an office environment, as well as actuators to alert people
when it is too noisy; the network topology that corresponds to the deployment is
shown in Figure 2.15b.
The use of heterogeneous nodes allows a distributed architecture using ad hoc
deployment that is resilient to failure and has lower cost. In addition to the motion
and sound sensors, a sensor with more powerful processing ability is required. This
is where the motion and sound data are sent for processing and analysis, to determine
whether an actuator signal is needed. As a result, the successful transmission rate
increases, and delay decreases significantly. Moreover, as the network grows and
computational demand increases, more high-end sensor nodes can be added in the
network.
2.7 SUMMARY OF PROJECTS AND SYSTEMS INFRASTRUCTURE
2.7.1 Summary of Heterogeneous WSN Projects
A summary of heterogeneous WSN projects is given in Table 2.1.
2.7.2 Systems Infrastructure for Heterogeneous WSNs
A summary of systems infrastructure for heterogeneous WSNs is given in Table 2.2.
2.8 OPEN PROBLEMS
Research in heterogeneous WSNs is in its infancy and is therefore rich in open problems. Some of them include:
Inadequate Theory of Heterogeneous WSNs. Most of the models assume that
a heterogeneous WSN provides data that are clock-driven (or periodic).
OPEN PROBLEMS
Alert
Alert
43
Alert
Alert
Alert
Alert
Alert
Alert
Alert
Alert
Alert
Low-end motion and sound sensor
High-end sensor
Alert
Actuator
(a)
Alert
Alert
Alert
Alert
Alert
Alert
Alert
Alert
Alert
Alert
Alert
Low-end motion and sound sensor
High-end sensor
Alert
Actuator
(b)
Figure 2.15. (a) A “do not disturb” application deployed in an office area and (b) its corresponding topology.
Theory for query-driven and event-driven heterogeneous WSNs needs to be
explored. While Mhatre et al. [15] consider hardware cost, models that also
consider energy consumed in data processing (compression, fusion, etc.) are of
interest.
44
HETEROGENEOUS WIRELESS SENSOR NETWORKS
TABLE 2.1. Heterogeneous WSN Projects.
Project
Brief Description and Website
CENS,
University of
California,
Los Angeles
Center for Embedded Networked Sensing
A mission of CENS is to develop and demonstrate architectural
principles and methodologies for deeply embedded, massively
distributed, sensor-rich systems. Research areas that relate to
heterogeneous WSNs include the Multiscaled Sensing and
Actuation (MAS) project, the Tenet project, the EmStar family
of generalized deployment software tools, and individual
protocols such as the centralized (CentRoute) and distributed
(Hyper) routing protocols.
http://research.cens.ucla.edu/
CoSense,
Palo Alto Research
Center
Collaborative Sensemaking
Collaborative sensemaking of distributed sensor data for target
recognition and condition monitoring.
http://www2.parc.com/spl/projects/cosense/
DSN-CC,
Los Alamos National
Laboratory
Distributed Sensor Networks with Collective Computation
The goal of DSN-CC is to demonstrate in situ collective
computation abilities of heterogeneous sensor networks in
simulation and using inexpensive, readily available off-the-shelf
platforms. One application is a staged heterogeneous wireless
sensor network for the detection of radioactive sources.
http://www.lanl.gov/source/orgs/isr/dsn/background.shtml
GNOMES,
Rice University
Generalized Network of Miniature Environmental Sensors
GNOMES is a low-cost hardware and software testbed. It is
designed to explore the properties of heterogeneous wireless
sensor networks, to test theory in sensor networks architecture,
and to be deployed in practical application environments.
http://cmclab.rice.edu/projects/sensors/
HSN,
University of
California, Berkeley
Heterogeneous Sensor Networks
Heterogeneous sensor networks for automated target recognition
and tracking in urban terrain is the focus. Issues addressed
include: a new theory for distributed signal processing with
random spatiotemporal sampling of complex scenes, robust
design principles for sensor networks with both low- and
high-bandwidth sensors, and metrics for the design and
deployment of sensor networks and incorporating mobility into
sensor networks.
http://trust.eecs.berkeley.edu/hsn/
OPEN PROBLEMS
45
TABLE 2.1. (Continued)
Project
Brief Description and Website
Intel Research
Heterogeneous Sensor Networks
To address the scalability problem in WSNs, high-end nodes (such
as Intel XScale-based nodes) are overlaid on a sensor network.
Goals include identifying and utilizing heterogeneous
capabilities, such as links and services, for embedding local
processing, imposing a database model, and enhancing routing
protocols. Applications include preventive maintenance for
equipment in Intel’s fabs and sensor networks for theme parks.
http://www.intel.com/research/exploratory/heterogeneous.htm
Microsoft Research
Networked Embedded Computing Group
Microsoft is developing new service architectures, interoperation
protocols, and programming models that are resource-aware and
resource-efficient across heterogeneous devices that can range
from extremely limited sensor nodes to more powerful servers.
http://research.microsoft.com/nec/
SensEye,
University of
Massachusetts,
Amherst
A Multitier Multimodal Camera Sensor Network
Trends in technology have resulted in a spectrum of camera
sensors, wireless radios, and embedded sensor platforms.
SensEye is designed on the principle that multitier networks are
not only scalable, but also offer a number of advantages over
simpler, single-tier unimodal networks: lower cost, better
coverage, higher functionality, and better reliability.
http://sensors.cs.umass.edu/projects/senseye/
SensorNets,
Carnegie Mellon
University
Pervasive Infrastructure Sensor Networks
SensorNets creates a framework for applications of networks of
sensors in long-lived infrastructure systems such as buildings,
bridges, and highways—a heterogeneous collection of sensors
that must continue to operate even as parts of the infrastructure
are changed, upgraded, or remodeled. The project has four main
areas of thrust: devices, applications, systems, and data.
http://www.ices.cmu.edu/sensornets/
46
HETEROGENEOUS WIRELESS SENSOR NETWORKS
TABLE 2.2. Systems Infrastructure for Heterogeneous WSNs
Systems
Brief Project Description and Website
Aspen,
University of
Pennsylvania
Abstraction-based Sensor Programming at Penn
The Aspen project focuses on the challenges in developing a
programming environment and runtime system for complex
applications that may have heterogeneous types of sensor,
confidentiality requirements, different levels of connectivity,
and timing constraints. A programming model that handles
heterogeneous data stream types and sensor capabilities is under
development.
http://www.cis.upenn.edu/∼zives/aspen/
Avrora,
University of
California,
Los Angeles
Sensor Network Simulation
Avrora is an instruction-level sensor network simulator. Avrora
simulates a network of AVR/Mica2 motes. The goal is to
enhance Avrora with new capabilities for executing and
monitoring simulations of heterogeneous sensor networks.
Specifically, this includes supporting sensor code that is
dynamically updated, other sensor platforms, and source-level
monitoring of simulations.
http://research.cens.ucla.edu/projects/ 2006/Systems/Avrora
DSS,
Los Alamos National
Laboratory
Distributed Sensors Simulator
DSS is a simulation framework that assists in implementing
and debugging wireless distributed sensor networks. The user
provides data on node locations and characteristics, defines
event phenomena, and plugs in the applications each node runs.
DSS provides simulation of the wireless and environmental
channels and was specifically designed for investigations of
topological, phenomenological, networking, robustness,
and scaling issues in WSNs.
http://www.lanl.gov/source/orgs/isr/dsn/codes.shtml
EmStar,
University of
California,
Los Angeles
EmStar
EmStar is a family of tools, libraries, and services that provide an
environment to help enable the design, development,
and deployment of WSN applications. EmStar supports
heterogeneous deployments consisting of both mote-class and
microserver-class component systems.
http://research.cens.ucla.edu/
BIBLIOGRAPHY
47
Tradeoffs Between Local and Remote Processing. The tradeoffs regarding where
processing of the sensed data should be performed are not well understood.
What are the benefits of staged versus hierarchical architectures, and centralized
processing at a sink node versus distributed processing by the cluster heads?
Querying, In-Network Processing, Caching. A related question to the processing
tradeoff is how to support querying in a heterogeneous WSN, what to cache and
where to cache, and what kind of in-networking processing can be performed.
Event Detection. A large application of heterogeneous WSNs is in event detection.
How do we reliably detect events with a low false alarm rate?
Quality-of-Service Support. Heterogeneous WSNs bring the potential of of highbandwidth sources such as audio and video. Such data streams require qualityof-service support in order to meet delay, jitter, and related constraints.
Nonuniform Energy Drainage. While hierarchical architectures have alleviated the
problem of non-uniform energy drainage, the problem remains unsolved.
Mobility in Sensor Nodes. Eventually, mobile nodes will be integrated into heterogeneous WSNs. This will add another dimension of complexity to all of the
problems.
ACKNOWLEDGMENTS
The work of V. R. Syrotiuk and B. Li is supported, in part, by LANL contract 13638001-05 and NSF grant ANI-0240524. Any opinions, findings, conclusions, or recommendations expressed are those of the authors and do not necessarily reflect the views
of LANL or NSF.
The work of A. M. Mielke is supported by the U.S. Department of Energy/NNSA
and Los Alamos National Laboratory funds under Contract Number DE-AC5206NA25396 and is approved for public release under LA-UR-06-5787.
BIBLIOGRAPHY
1. I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. A survey on sensor networks.
IEEE Communications Magazine, 40(8):102–114, 2002.
2. Crossbow Technology, I. hppt://www.xbow.com/.
3. I. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks: A
survey. Computer Networks, 38(4):393–422, 2002.
4. Q. Gao, D. J. Holding, Y. Peng, and K. J. Blow. Energy efficiency design challenge in sensor
networks. In Proceedings of London Communications Symposium, September 2002.
5. P. Kulkarni, D. Ganesan, P. Shenoy, and Q. Lu. SensEye: A multi-tier camera sensor
network. In Proceedings of the 13th annual ACM International Conference on Multimedia
(MM ’05), November 2005, pp. 229–238.
6. M. Rahimi, R. Baer, O. I. Iroezi, J. C. Garcia, J. Warrior, D. Estrin, and M. Srivastava. Cyclops: In situ image sensing and interpretation in wireless sensor networks. In Proceedings
48
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
HETEROGENEOUS WIRELESS SENSOR NETWORKS
of the 3rd International Conference on Embedded Networked Sensor Systems (SenSys ’05,
(2005), pp. 192–204.
CMUcam. http://www-2.cs.cmu.edu/cmucam/cmucam2/index.html.
E. Shih, P. Bahl, and M. J. Sinclair. Wake on wireless: An event driven energy saving strategy
for battery operated devices. In Proceedings of the 8th Annual International Conference
on Mobile Computing and Networking (Mobicom ’02), 2002.
A. M. Mielke, S. M. Brennan, M. C. Smith, D. C. Torney, A. B. Maccabe, and J. F. Karlin.
Independent sensor networks. IEEE Instrumentation & Measurement Magazine, 8(2):33–
37, 2005.
S. M. Brennan, A. M. Mielke, D. C. Torney, and A. B. Maccabe. Radiation detection with
distributed sensor networks. IEEE Computer Magazine, 37(8):57–59, 2004.
P. Gupta, and P. R. Kumar. Capacity of wireless networks. IEEE Transactions on Information Theory, IT-46(2):388–404, 2000.
Intel heterogeneous sensor networks research group.
X. Du, and F. Lin. Improving routing in sensor networks with heterogeneous sensor nodes.
In Proceedings of the 61st IEEE Vehicular Technology Conference (VTC ’05-Spring),
May–June 2005, Vol. 4, pp. 2528–2532.
K. Xu, X. Hong, and M. Gerla. An ad hoc network with mobile backbones. In Proceedings
of the IEEE International Conference on Communications (ICC ’02), Vol. 5, April–May
2002, pp. 3138–3143.
V. Mhatre, C. Rosenberg, D. Kofman, R. Mazumdar, and N. Shroff. A minimum cost
heterogeneous sensor network with a lifetime constraint. IEEE Transactions on Mobile
Computing, 4(1):4–15, 2005.
X. Du, and Y. Xiao. Energy efficient chessboard clustering and routing in heterogeneous
sensor networks. International Journal of Wireless and Mobile Computing, 1(2):121–130,
2006.
W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan. Energy-efficient communication protocols for wireless microsensor networks. In Proceedings of the 33rd Hawaii
International Conference on System Sciences Vol. 2, January 2000.
S. Lindsey, C. S. Raghavendra, and K. Sivalingam. Data gathering in sensor networks using
the energy delay metric. In Proceedings of the 15th International Parallel and Distributed
Processing Symposium, 2001, pp. 2001–2008.
T. Lin, Y. Chen, and S. Lee. PCAR: A power-aware chessboard-based adaptive routing
protocol for wireless sensor networks. In Proceedings of the IEEE 6th Circuits and Systems
Symposium on Emerging Technologies: Frontiers of Mobile and Wireless Communication,
Vol. 1, 2004, pp. 145–148.
C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed diffusion: A scalable and robust communication paradigm for sensor networks. In Proceedings of the 6th Annual
ACM International Conference on Mobile Computing and Networking (Mobicom ’00),
August 2000, pp. 56–67.
F. Ye, H. Luo, J. Cheng, S. Lu, and L. Zhang. A two tier data dissemination model for
large-scale wireless sensor networks. In Proceedings of the 8th Annual ACM International
Conference on Mobile Computing and Networking (Mobicom ’02), 2002, pp. 148–159.
Y. Ma, S. Dalal, M. Alwan, and J. H. Aylor. ROP: A resource oriented protocol for heterogeneous sensor networks. In Proceeding of the 13th Virginia Tech/MPRG Symposium on
Wireless Personal Communication, June 2003, pp. 59–70.
BIBLIOGRAPHY
49
23. Y. Ma, and J. H. Aylor. System lifetime optimization for heterogeneous sensor networks
with a hub-spoke topology. IEEE Transactions on Mobile Computing, 3(3):286–294, 2004.
24. V. Mhatre, and C. Rosenberg. Homogeneous vs. heterogeneous clustered sensor networks:
A comparative study. In Proceedings of IEEE International Conference on Communications (ICC ’04), Vol. 6, 2004, pp. 3646–3651.
25. E. J. Duarte-Melo, and M. Liu. Analysis of energy consumption and lifetime of heterogeneous wireless sensor networks. In Proceedings of the IEEE Global Telecommunications
Conference (Globecom’02), November 2002, Vol. 1, pp. 21–25.
26. J. Li, and P. Mohapatra. An analytical model for the energy hole problem in many-toone sensor networks. In Proceedings of the 62nd IEEE Vehicular Technology Conference
(VTC ’05-Fall, Vol. 4, September 2005, pp. 2721–2725.
27. J. Ai, D. Turgut, and L. Bölön. A cluster-based energy balancing scheme in heterogeneous wireless sensor networks. In Proceedings of the 4th International Conference on
Networking (ICN ’05), April 2005, pp. 467–474.
28. B. Liu, and D. Towsley. A study of the coverage of large-scale sensor networks. In Proceedings of the IEEE International Conference on Mobile Ad-hoc and Sensor Systems,
October 2004, pp. 475–483.
29. S. Megerian, F. Koushanfar, M. Potkonjak, and M. B. Srivastava. Worst and best-case
coverage in sensor networks. IEEE Transactions on Mobile Computing, 4(1):84–92, 2005.
30. C. Huang, and Y. Tseng. The coverage problem in a wireless sensor network. In Proceedings
of the 2nd ACM International Conference on Wireless Sensor Networks and Applications,
2003, pp. 115–121.
31. L. Lazos, and R. Poovendran. Coverage in heterogeneous sensor networks. In Proceedings
of International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless
Networks, April 2006, pp. 1–10.
32. D. Miorandi, and E. Altman. Coverage and connectivity of ad hoc networks presence
of channel randomness. In Proceedings of the 24th Annual Joint Conference of the IEEE
Computer and Communications Societies (Infocom ’05), Vol. 1, March 2005, pp. 491–502.
33. G. Xing, X. Wang, Y. Zhang, C. Lu, R. Pless, and C. Gill. Integrated coverage and connectivity configuration for energy conservation in sensor networks. ACM Transactions on
Sensor Networks, 1(1):36–72.
34. T. Yan, T. He, and J. A. Stankovic. Differentiated surveillance for sensor networks. In
Proceedings of the 1st International Conference on Embedded Networked Sensor Systems
(SenSys ’03), 2003, pp. 51–62.
35. X. Du, and F. Lin. Maintaining differentiated coverage in heterogeneous sensor networks.
EURASIP Journal on Wireless Communications and Networking, 5(4):565–572, 2005.
36. D. Vaidya, J. Peng, L. Yang, and J. W. Rozenblit. A framework for sensor management in
wireless and heterogeneous sensor network. In Proceedings of the 12th IEEE International
Conference and Workshops on the Engineering of Computer-Based Systems (ECBS ’05),
April 2005, pp. 155–162.
37. D. Andersson, M. Fong, and A. Valdes. Heterogeneous sensor correlation: A case study of
live traffic analysis. In Proceedings of the 3rd Annual Information Assurance Workshop,
June 2002.
38. L. Gürgen, C. Labbe, V. Olive, and C. Roncancio. A scalable architecture for heterogeneous
sensor network. In Proceedings of the 16th International Workshop on Database and Expert
Systems Applications (DEXA ’05), 2005, pp. 1108–1112.
50
HETEROGENEOUS WIRELESS SENSOR NETWORKS
39. G. Wiederhold. Mediators in the architecture of future information systems. IEEE Computer Magazine, 25(3):38–49, 1992.
40. L. Yuan, C. Gui, C. Chuah, and P. Mohapatra. Applications and design of heterogeneous
and/or broadband sensor networks. In Proceedings of the 1st IEEE Workshop on Broadband
Advanced Sensor Networks (BaseNets ’04), October 2004.
41. O. Gnawali, and M. Yarvis. “Do Not Disturb” an application leveraging heterogeneous
sensor networks. Technical Report 03-808, University of Southern California, 2003.
CHAPTER 3
Epidemic Models, Algorithms,
and Protocols in Wireless Sensor
and Ad Hoc Networks
PRADIP DE and SAJAL K. DAS
Center for Research in Wireless Mobility and Networking (CReWMaN), Department of
Computer Science and Engineering, University of Texas at Arlington, Arlington, TX 76019
3.1 INTRODUCTION
Sensor networks are composed of a large number of sensing devices, which are
equipped with limited computing and radio communication capabilities. They have
diverse application areas, ranging from tracking and intrusion detection for security purposes to environment monitoring and traffic and location systems. However,
with the steady advancements in processor, memory, communication, and sensing
technology, along with a drive toward a smarter environment, there is an increased
interest in the development and deployment of wireless sensor networks to be used
for many interesting and new applications. These applications range from real-time
remote monitoring and control, military surveillance, and environmental monitoring
to healthcare management, construction safety, and so on.
Within the next few years, it is very likely that the number of deployed sensors
will see an exponential increase. Most of these networks will require applicationspecific functionalities and performance requirements [1]. However, the realization
of these networks poses a lot of challenges in system and network design, algorithm
and protocol design, and query language and database design. The primary issues
under focus which are critical to the proper functioning of wireless sensor networks
are energy consumption, connectivity, clustering techniques, data aggregation, and
so on. These issues stem mostly from the stringent resource constraints of the sensor
nodes. Therefore, in order to address these issues, we require efficient modeling
techniques and robust algorithms and protocols before actual implementation and
deployment is done.
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
51
52
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
Oftentimes, in the course of modeling complex entities and networks, we have
taken recourse to biologically inspired paradigms. Biologically inspired modeling
techniques are among the many mechanisms that have been adopted to accurately
model certain phenomena in wireless sensor networks. For example, data dissemination, routing algorithms, and broadcast protocols are among the few areas that have
been effectively modeled by epidemic theoretic concepts.
In this chapter, we address the modeling techniques, algorithms, and protocols
proposed in wireless sensor and ad hoc networks that are primarily based on Epidemic
Theoretic concepts and paradigms.
The rest of the chapter is organized as follows. In Section 3.2, we provide a general overview of Epidemic Theoretic concepts and analysis. In Section 3.3, we discuss
the data dissemination models in sensor networks and their use of epidemic theory.
Section 3.4 illustrates several reprogramming and code update protocols in sensor
networks that adopt epidemic theoretic principles. In Section 3.5, we look into epidemic protocols in ad hoc networks. Section 3.6 looks into some security aspects and
explains the propagation process modeling of malware in sensor networks. Finally,
we conclude the chapter in Section 3.7.
3.2 OVERVIEW OF EPIDEMIC THEORY
In order to appreciate the epidemiological models applied in wireless sensor networks,
we need to first understand the concept of epidemiology. In this section we provide a
terse description of the theory and its applications. Epidemic Theory [2] is the study
of the dynamics of how contagious diseases spread in a population, resulting in an
epidemic. Primarily, the theory mathematically models the propagation process of an
infection and measures its outcome in relation to a population at risk. The population
at risk basically comprises of the set of people who possess a susceptibility factor
with respect to the infection. This factor is dependent on several parameters such
as exposure, spreading rate, previous frequency of occurrence, and so on, which
define the potential of the disease causing the infection. Among the different models
characterizing the infection spread, two are quite popular. They are the Susceptible
Infected Susceptible (S-I-S) Model, Susceptible Infected Recovered (S-I-R) Model,
and so on. In the former, a susceptible individual acquires infection and then after
an infectious period (i.e., the time the infection persists) the individual becomes
susceptible again. On the other hand, in the latter, the individual recovers and becomes
immune to further infections.
An approach to model the propagation of an infection is to assume that the probability (per unit time) for a susceptible individual to acquire infection is equal to the
average rate at which new infective partners are acquired multiplied by the probability
of being infected by any one such partner. In the general deterministic S-I-R model,
if N(t), X(t), Y (t), and Z(t) denote the total population, the susceptibles, the infected,
and the recovered or immune individuals, respectively, at time t, we can say
N(t) = X(t) + Y (t) + Z(t)
(3.1)
OVERVIEW OF EPIDEMIC THEORY
53
If β denotes the infection rate and γ denotes the removal rate of infected individuals,
then assuming a homogeneous mixing model (i.e., each of the susceptibles can get in
contact with any of the infectives), it is simple to observe that in time t, there are
βxyt new infections and γyt removals. Therefore, the basic differential equations
that describe the rate of change of susceptibles, infectives, and recovered individuals
are given by
dX(t)
= −βXY,
dt
dY (t)
= βXY − γY,
dt
dZ(t)
= γY
dt
The above equations can be solved either approximately or precisely based on some
boundary conditions, such as, at the start of the epidemic, when t = 0, (X, Y, Z)
can take the values (x0 , y0 , 0). Note that, in particular, if y0 is very small, x0 is
approximately equal to N. It also follows that if the relative removal rate, µ = γ/β,
is greater than x0 , only then can an epidemic start to build up as this condition will
result in [dY (t)/dt]t=0 > 0, i.e. Y (t) will have a positive slope. Therefore, the relative
removal rate µ = x0 gives a threshold density of susceptibles.
On the other hand, the S-I-S model does not have the recovered subset Z(t), and
those who are infected fall back into the susceptible subset S(t) after their infectivity
duration.
An important aspect that is of particular interest in epidemiological studies is
the phenomenon of phase transition of the spreading process that is dependent on a
threshold value of the epidemic parameter; that is, if the epidemic parameter is above
the threshold, the infection will spread out and become persistent; on the contrary, if
the parameter is below the threshold, the infection will die out. Identification of this
threshold value is critical in the study of how an epidemic spreads and how it can be
controlled.
Apart from modeling technique based on the continuous differential rate equation, the study of epidemics has often been performed by treating the population as
a network graph, with the nodes representing each individual and the edges their
interaction. This form of analysis [3] has mainly been used in scenarios where the
end result of the epidemic spread is more important than the temporal dynamics of
the propagation. Several works have spawned from this formulation [3–8], where the
spread of diseases have been studied by modeling the social network as a scalefree topology. Several other works also exist that model the spread of computer
viruses [9, 10].
Epidemic Theory has found special attention in the design and modeling of several
phenomena and protocols in sensor networks wherever there is a scope of information
distribution on a large scale, preferably from a small number of sources to a large
number of recipients. Among the popular phenomena in sensor and ad hoc networks
54
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
where this theory has been adopted are data dissemination, broadcast protocols, and
routing. We will delve into some of these areas where Epidemic Theory has been used
to study and model several processes and functions of sensor networks.
3.3 DATA DISSEMINATION IN SENSOR NETWORKS:
MODEL AND PROTOCOLS
The problem of reliable data dissemination in the context of wireless sensor networks
is very critical. Reliable data dissemination to all nodes is absolutely necessary for
the propagation of queries, code updates, and other sensitive information in a wireless
sensor network. This is not a trivial task since the number of nodes in a sensor network
can be quite huge and the environment is dynamic (i.e., nodes can die or move), thus
making the topology change constantly.
Since data dissemination primarily deals with the transfer of messages from one
node to all nodes of a network, algorithms based on epidemiological formulations are a
perfect fit. Accordingly, these algorithms have been successfully used in disseminating
information in sensor networks and, depending on the application, the dissemination
can start at a single node, such as a base station, or at multiple sensor nodes. The
decentralized and distributed nature of wireless sensor networks fits the context of
epidemic algorithms aptly.
One of the prominent works of data dissemination in sensor networks is SPIN
[11]. An obvious problem with normal epidemic broadcast-based dissemination is
the inefficient use of bandwidth and other resources. Therefore, the basic epidemic
strategy needs to be optimized for sensor networks. In reference 11, the authors
proposed the concept of meta data or data descriptors to eliminate the chance of
redundant transmissions in sensor networks. Their work focuses on the efficient dissemination of individual sensor observations to all the sensors in a network. Their
main contribution was based on the basic deficiencies of classic flooding, namely,
Implosion, Overlap, and Resource Blindness. Implosion is sending data redundantly
to one’s neighbors regardless of whether they already received it. Coverage overlap of
nodes can make them gather the same data and flood it to common neighbors. Classic
flooding can be blind to the availability of resources when it is flooding data across
the network.
The use of metadata allows nodes to negotiate between themselves and prevent
redundantly transmitting the same information. Also, in SPIN, each node has a local
resource manager that keeps track of its resources and helps a node decide whether
to transmit or process data. SPIN first broadcasts metadata to its neighbors. Then,
if it receives a request for the data from any neighbor it sends the data to that
node.
There are four protocols in the SPIN family. The first two, SPIN-PP and SPIN-BC,
tackle the basic problem of data dissemination under ideal conditions. The other
two, SPIN-EC and SPIN-RL, are modified versions of the first two. SPIN-PP is
optimized for communicating in a point-to-point mode, where for each data
transmission between neighbors, a three-stage handshaking (ADV-REQ-DATA) is
DATA DISSEMINATION IN SENSOR NETWORKS: MODEL AND PROTOCOLS
55
Figure 3.1. SPIN-PP protocol. Node A sends advertisement messages (ADV ) to B. B responds
with a request (REQ) message. Then B starts to send ADV to its neighbors.
performed. As illustrated in Figure 3.1, a node sends an ADV message whenever it
has new data to advertise. Upon receiving an ADV message, the neighboring node
verifies whether it has already received or requested the advertised data. If not, it
responds by sending a REQ message for the missing data back to the sender. The initiator of the protocol responds to the REQ message with a DATA message containing
the missing data.
Although this protocol has been designed for a lossless environment, it can be
adapted for a lossy environment. Nodes can periodically send the ADV message to
counter lost ADV messages. For lost REQ and DATA messages, nodes can request
items that do not arrive within a fixed time period. SPIN-EC is a modification of
SPIN-PP so that when a node observes that it is approaching a low-energy threshold,
it reduces its participation in the protocol.
In SPIN-BC, which is a broadcast transmission protocol, each node transmits to
the broadcast address (Figure 3.2). Every node that is in the transmission range of the
sender processes the received message. This approach is justified because broadcast
and unicast transmissions use the same amount of network resources in a broadcast
network. The proliferation of redundant messages in the network can be curtailed by
SPIN-BC because a node A suppresses its own transmission whenever it observes
56
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
Figure 3.2. SPIN-BC protocol. (1) A sends ADV to all its neighbors. (2) C responds by
broadcasting a request, specifying as originator of the ADV. It also suppresses D’s REQ.
(3) After receiving the requested data, E’s REQ is also suppressed.
that another node B has transmitted the same message that A itself was supposed to
transmit.
We observe the epidemic nature of the dissemination of data in SPIN, especially in
SPIN-BC. Using the three-way handshake, a node that has the missing data passes it
on to a neighbor that does not have it, thereby infecting it in the process. The working
of the three-way handshaking protocol basically constitutes the contact and infecting
process of the SPIN protocol.
3.3.1 Infuse
For the reliable dissemination of data in sensor networks, the authors of Infuse [12]
proposed a TDMA-based data dissemination protocol for sensor networks. The primary purpose of the protocol was similar to that of Deluge [13]—that is, reliable dissemination of bulk data in a sensor network. We discuss Deluge later in this chapter.
In Infuse, the data dissemination protocol is based on a TDMA-based medium access
layer. Since TDMA ensures a deterministic slot when a sensor node should transmit
its packet, it offers a degree of reliability which is used by the data dissemination
strategy adopted in Infuse. The authors tackle the problem of random message losses
DATA DISSEMINATION IN SENSOR NETWORKS: MODEL AND PROTOCOLS
57
in the presence of channel errors by considering recovery algorithms based on sliding
window protocols, modified to use implicit acknowledgments.
In the ideal scenario without channel errors, the base station sends a special Start
Download message to the sensor which contains the number of subsequent packets to
follow in each TDMA slot. The sensor then reserves the necessary flash and downloads
the arriving packets.
For dealing with channel errors, Infuse uses an implicit acknowledgment technique. This happens because whenever a successor sensor forwards a data packet, the
predecessor node gets to hear it. This overhearing acts as an implicit acknowledgment
for the predecessor node. Furthermore, Infuse forwards a received packet in the next
TDMA slot, thus maintaining a pipeline effect of the transfer process which helps
in reducing the total latency of the dissemination process. The use of TDMA-based
data dissemination also allows Infuse to send the node to sleep except in its own
transmission slot, thereby making the Infuse protocol energy-efficient.
3.3.2 Firecracker
Routing a packet from one source to a single destination is fast because forwarding
nodes can retransmit without worrying about suppression or local density. At the
same time, routing cannot be used to disseminate data to all the nodes in a sensor
network because the nodes are not individually addressable. The Firecracker Protocol
[14] uses a combination of routing and broadcast principles to rapidly disseminate
data throughout a sensor network. As depicted in Figure 3.3, a data source first routes
the data to be distributed to distant points in the network. Once the data reaches
its destination, broadcast-based dissemination starts along the path like a string of
firecrackers. Firecracker is largely designed to disseminate small pieces of data that
would propagate fast, like small programs or configuration constants. While maintaining the energy efficiency of broadcasts, Firecracker can achieve dissemination
rates close to routing.
From an epidemic modeling standpoint, Firecracker is fundamentally an infection propagation strategy with a predetermined set of infective nodes defined by the
destination nodes of the routing protocol. Having strategically placed the infective
nodes at different points of the population, the protocol starts its final broadcast to
disseminate the information to the rest of the network. The dissemination strategy
Figure 3.3. The Firecracker disseminaton mechanism.
58
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
could be Trickle [15], and the routing strategy could be any suitable one used for
sensor networks.
To elucidate further, Firecracker is composed of three main parts: (a) Broadcast
Protocol, (b) Routing Protocol, and (c) Seed Selection. The broadcast protocol is very
important to the functioning of Firecracker. It not only must propagate data to nodes
that do not have the data, but also decide when to propagate. Moreover, the protocol
should minimize the cost of detection but propagate rapidly, and temporary network
disconnections should not prevent reception.
The basic purpose of the routing phase is to spread data to distant points in the
network so that the initial seeds are placed as deep as possible into the network. This
facilitates the following broadcast process to spread the data rapidly. In this regard,
a naming scheme that allows nodes to choose such points is helpful. Since wireless
data, even during the routing phase, is essentially broadcast in the neighborhood,
nodes along the route should be able to snoop on routed traffic to cache the data
as they pass by. Moreover, reliability and nonredundancy are more important than
minimum hop paths. Therefore, taking a long, winding path through different areas
of the network could benefit the subsequent broadcast protocol in quickly installing
the code in all the nodes.
The choice of the seed nodes is also equally important to the performance of
Firecracker. The farther the seeds are from the original source and the farther they are
spread apart from themselves, the faster the data would propagate to all the nodes of
the network.
In general, epidemic algorithms for data dissemination follow the model of nature
to spread information and define simple rules for information to flow between nodes
of a network. The authors in reference 16 have done a comparative study of epidemic
algorithms for data dissemination. Based on the style of communication between
neighboring nodes, they have classified epidemic algorithms for data dissemination
into three categories.
r Pull-Based. A node tries to extract new information from its neighbor.
r Push-Based. A node sends new information to a selected neighbor.
r Pull–Push-Based. A node asks its neighbors for new information as well as
sends new information to its neighbors.
They have studied the performance of these three classes of epidemic algorithms on
sensor networks. Their results show that both pull-based and push-based algorithms
perform better than the push–pull-based epidemic algorithms in terms of delivery rate
and scalability. The primary reason for this result is the restricted memory resource
of sensor devices.
In reference 17, the authors performed an experimental and empirical study of the
epidemic style algorithms in large-scale multihop wireless networks.
A smart tag-based data dissemination technique is explained by the authors
in reference 18, where mobile individuals, equipped with smart tags disseminate
data across disconnected static nodes spread across a wide area. When the mobile
CODE UPDATE PROTOCOLS IN SENSOR NETWORKS
59
individuals equipped with smart tags move into a sensor field, they get updated with
the latest information from the sensors. Later, when they move into another field, they
disseminate the newly acquired information. The concept of using carriers, who are
mobile, to carry data between connected components of the network has also been
used by the authors of the epidemic routing protocol [19]. However, they did it more
for the purpose of routing, whereas here the authors use smart tags to carry the sensed
information to another set of output devices like display units. The authors used
Bluetooth-enabled smart tags to illustrate the characteristics of their approach. As an
intuitive technique, their approach is suited for applications that are delay-tolerant.
3.4 CODE UPDATE PROTOCOLS IN SENSOR NETWORKS
Several protocols have been proposed for code update and propagation in sensor
networks. These protocols are mainly broadcast in nature, and tasks in sensor networks
are assigned through code updates, and all the nodes in the sensor network will have
the same code to execute. The propagation mechanism for the code update is basically
hop by hop to all nodes in the network. Needless to say, wireless sensor nodes have
limited energy, and therefore maintenance costs of the code updates must be low.
Another important requirement is rapid propagation of updates, because some tasks
may have to be activated as soon as possible and newly assigned tasks make the older
ones obsolete. Moreover, the update process should be scalable and should work in a
dynamically changing environment.
Being inherently broadcast in nature, these protocols and algorithms fundamentally
transmit code updates in a manner similar to an infection spread in a susceptible
population. In this subsection, we study some of these protocols and their mechanism.
3.4.1 Trickle
Trickle [15] is a broadcast algorithm for propagating and maintaining code updates
in a wireless sensor network. Conceptually, Trickle borrows from epidemiological
concepts and performs what the authors claim as polite gosip.
Sensor networks are generally deployed in remote areas and are expected to operate unattended for lengthy periods of time. Thus, there is every possibility that the
requirements and environments of a sensor network evolve. As a result, users need
to be able to introduce new code to retask the network. However, the large-scale and
embedded nature of the network requires these code updates to propagate through
the network. However, as is obvious, networking in sensor networks is very costly in
terms of energy consumption, and therefore an efficient and effective reprogramming
protocol is necessary.
An effective reprogramming protocol must transfer the code as fast as possible
because in the transition time when the code is propagating, the network is actually
in a useless state because the old and the new programs are concurrently running.
Propagation of code is costly, and nodes need to learn when they need to propagate
code. Nodes, therefore, periodically communicate to learn when there is new code.
60
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
To reduce energy costs, nodes transmit metadata to determine when code is needed.
However, the cost of periodically transmitting metadata consumes almost the same
amount of energy as actually transmitting the code itself. Therefore, there is a crucial
need for the reprogramming algorithm to be efficient in this aspect and effectively
determine when nodes should propagate code. Motivated by this requirement, the
authors in reference 15 have identified three main properties that a reprogramming
algorithm should have. They are as follows:
r Low Maintenance. When a network is in a stable state, metadata exchanges
should be infrequent, just enough to ensure that the network has a single program.
r Rapid Propagation. When the network discovers nodes that need update, it
should propagate the code as fast as possible and to every node of the network.
r Scalability. The algorithm should obviously be scalable and be robust against
any environmental changes and node failures.
Trickle tries to meet all these requirements. Its basic working principle is simple.
Every so often, a mote transmits code metadata if it has not heard a few other motes
transmit the same information. Trickle sends all messages to the local broadcast
address. When a neighbor receives a broadcast, either it is up to date, or it detects the
need for an update. Detection can be the result of either an out-of-date mote hearing
someone having a new code, or an updated mote hearing someone has old code. As
long as every mote communicates somehow, the need for an update is always detected.
It does not matter who transmits first, but as long as some nodes communicate with
each other at a nonzero rate, every node would be up to date. More formally, each node
maintains a counter c, a threshold k, and a timer t in the range of [0, τ]. k is a small,
fixed integer (e.g., 1 or 2) and τ is a time constant. When a node hears metadata identical to its own, it increments the counter c. At the timepoint t, which is uniformly
randomly chosen in the range of [0, τ], the mote broadcasts its metadata only if
c < k. When the interval of size τ completes, c is reset to zero and t is reset to
a new random value in the range [0, τ]. Thus, Trickle allows each node to broadcast its metadata at most once per period τ, thus maintaining the politeness of
its gossip. In each interval τ , the sum of receptions and sends of each mote is
k. The random selection of t uniformly distributes the choice of who broadcasts
in a given interval. This evenly spreads the transmission energy load across the
network.
In Figure 3.4, the solid line represents a transmission, the broken lines represent
reception, and the gray line means suppression of an advertisement. This mechanism
of Trickle not only allows us to scale to high network density, but also propagates
updates fast. It also distributes transmission load evenly as it spreads, and it is
simultaneously robust to transient disconnections. The experimental verification by
the authors shows that it imposes a maintenance overhead on the order of only a few
packets per hour per node.
The epidemiological essence in the working principle of Trickle is evident. The
objective is to propagate code as fast as possible to all nodes of the network. Thus,
CODE UPDATE PROTOCOLS IN SENSOR NETWORKS
61
Figure 3.4. Trickle metadata advertisement.
from an epidemic theoretic standpoint, the rate at which the metadata is exchanged
gives the rate at which the infection or propagation proceeds in the network. Since
after a node advertises metadata every node in the neighborhood gets updated with
the current code, Trickle succeeds in propagating the code update to all nodes in the
network. The propagation rate is dependent on the value of τ. With a large value of τ,
there is less communication overhead, but the code propagates slowly and conversely
in the case of a small τ.
3.4.2 Deluge
Another type of data dissemination protocol for supporting network programming
in sensor networks is Deluge (Figure 3.5) [13]. It is a reliable data dissemination
protocol for propagating large data objects from a few source nodes to many other
Figure 3.5. The Deluge state machine.
62
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
nodes in a wireless sensor network. Trickle’s key contribution is its polite gossip that
uses suppression and dynamic adjustment of the broadcast rate to limit transmissions
among neighboring nodes. It only provides a mechanism for a node to decide when to
propagate code. Deluge, on the other hand, though based on Trickle’s principles, has
the added feature of supporting the transfer of large data objects. It uses a three-phase
protocol similar to SPIN-RL [11].
Deluge, being an epidemic protocol, can disseminate large data objects as quickly
and reliably as possible. The basic local broadcast principle of Deluge is simple and
similar to Trickle, but it also addresses several subtle issues that improve its performance. The local suppression of redundant broadcasts makes it density-aware. Its
three-way handshaking mechanism ensures that there is a bidirectional link, thereby
making it a reliable data dissemination protocol. Moreover, by dynamically adjusting the rate of advertisements and emphasizing on the use of spatial multiplexing to allow parallel transfers, Deluge allows quick propagation of large blocks of
data.
Deluge divides the large data object into fixed-size pages for transfer. This enables
efficient incremental update and also limits the amount of state that should be reserved
at a time at the receiver. Each page is also divided into a fixed number of packets.
Because of the epidemic nature of the page propagation, Deluge offers CRC checks
at both the packet and page level to be safe from the negative effects of the epidemic
nature of data transfer. The protocol resides in one of the three states, namely, MAINTAIN, TX, and RX. In the MAINTAIN state, a node uses a summary advertisement
mechanism to ensure that all nodes in the communication range are up to date with
the current version of the object. In the RX state, the node is responsible for requesting all remaining packets of a page; and while in the TX state, it is responsible for
broadcasting all requested packets for a given page.
Another work, which was based on an epidemic style multihop reprogramming
service for sensor networks, was proposed in MNP [20]. One of the basic problems in
reprogramming and code update in a wireless network is the issue of message collision
and the hidden terminal problem. The authors counter this problem by proposing a
sender selection algorithm whereby it is guaranteed that in a neighborhood, there is
at most one sensor transmitting at a time. In the basic version of the sender selection
protocol, a node becomes a source node and starts advertising this fact only when it
acquires the new program code entirely.
Each source node maintains a variable that indicates the number of distinct requests
it has received so far, and it gets incremented each time a node receives a new download
request. Two messages are used for sender selection, namely, advertisement and
download request. The advertisement message contains information about the new
program and the source node. When a node j receives the advertisement request from
node i and it is in need of the new code, it sends a download request message to the
broadcast address so that any neighboring node k becomes aware that i is a potential
source.
In order to ensure that a node is aware of all the requesters who are likely to receive
the code, if it is chosen to transmit the code, the node sends a download request to all
senders of the advertisement messages. However, if node j loses to node k that has
EPIDEMIC MODELS IN AD HOC NETWORKS
63
more requesters, then whenever j attempts to advertise again, j must reset its request
counter value to zero and recalculate its requesters. After k finishes transmitting, it
sleeps for a while, so that other sources get a better chance to send. When it wakes up
and reenters the advertising state, its counter value is reset to zero, and a new round
of sender selection begins.
We observe that there is considerable similarity with MNP and the SPIN family of
protocols because both of them use a kind of three-way handshaking procedure for
disseminating the data.
For the transfer of large-sized data, MNP tries to incorporate pipelining into the
data dissemination process by breaking the large data into segments, each containing
a constant number of packets. Thus, the protocol now operates at the segment level
which helps in a node forwarding segments even if it has not received the whole data.
In this aspect it is very similar to Deluge [13].
MNP is equipped to address reliability issues like loss detection and recovery.
Each packet has a unique ID and each receiver is responsible for detecting its own
loss. Since the size of a segment is considerably small, a bitmap of the current segment is maintained in memory, where each bit corresponds to a packet. Using this,
a sensor node can receive packets in any order. This bitmap is called the Missing
Vector. A node also maintains a Forward Vector, which is a bitmap of the advertised segment. Whenever a node sends a download request, it puts its Missing Vector
in the request message. The advertising node marks its Forward Vector according
to the Missing Vector messages it receives. A node only sends the packets indicated in the Forward Vector. Upon receiving all the segments of a program, the node
reboots.
3.5 EPIDEMIC MODELS IN AD HOC NETWORKS
In ad hoc networks, the power supply of individual nodes, wireless bandwidths are
limited, and the channel conditions can vary significantly. Moreover, since nodes
can be mobile, routes may constantly change. Thus, to enable efficient communication, robust routing protocols must be developed. Several existing Mobile ad hoc
routing protocols [21, 22] have been developed that allow wireless nodes to communicate with one another without any preexisting network infrastructure. In this
section we look into some of the routing protocols that essentially have the flavor of
epidemiology.
3.5.1 Gossip
Although flooding has been used with some optimization to route packets in an ad
hoc network, many routing messages are propagated unnecessarily. The authors in
reference 23 have proposed a gossip-based approach where each node decides to
forward a message to another node based on some probability. They showed that this
technique could significantly reduce the number of routing messages sent. Gossip is
essentially an epidemic algorithm, where neighbors are chosen probabilistically to
64
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
propagate the information in the same way as an infection spreads in a susceptible
population.
In the gossip protocol a source sends the route request with probability 1. When
a node first receives a route request, with probability p it broadcasts the request to
its neighbors, and with probability 1−p it discards the request; if the node receives
the same route request again, it is discarded. Thus, a node broadcasts a given route
request at most once.
The problem with gossip is that if the source has very few neighbors, then the nodes
will not gossip and it would die out. Basically, from the epidemiological standpoint we
say that the phase transition did not happen and the propagation collapsed. In order to
circumvent this problem, the authors modify gossip so that each node forwards with
probability 1 for the first k hops before continuing to gossip with probability p. The
modified protocol is called the GOSSIP1(p, k) protocol.
The performance study of the gossip protocol in finite networks reveals several
important results. As expected, the location of the source node does not affect the
fraction of the source node receiving the messages. However, it does affect the number
of executions in which the gossip dies out. The number of executions in which the
gossip does not die out is higher for a more central node and is lower for a corner node.
The authors observe that lowering the probability significantly changes the fraction
of executions in which all nodes and no nodes get the message.
The authors suggest a few optimization techniques to the basic gossip protocol. In
many cases, a gossip protocol may be run in conjunction with other protocols. If the
other protocols maintain fairly accurate information regarding a node’s neighbors,
GOSSIP1 can make use of this information effectively, by a simple optimization. In a
random network, the number of neighbors of a node might not be very high. In such a
case, the gossip protocol might not propagate the information and die out. To overcome
such a situation, the authors proposed that the gossip probability at a node could be a
function of its degree, where nodes with lower degree gossip with higher probability.
The modified protocol has four parameters: p1 , k, p2 , and n. As in GOSSIP1, p1 is the
main gossip probability and k is the number of hops with which gossiping starts with
probability 1. The new features are p2 and n; the idea is that the neighbors of a node
with fewer than n neighbors gossip with probability p2 > p1 . Thus, if a node has
fewer than n neighbors, it would instruct its neighbors to broadcast with probability
p2 rather than p1 . The modified protocol is called GOSSIP2 (p1 ; k; p2 ; n). GOSSIP2
has significant impact in topologies that are random rather than regular.
However, GOSSIP1 and GOSSIP2 might suffer a premature death because the
probability is low. In order to detect whether the gossip is dying out, a node might
monitor the number of messages it is getting from its neighbors. If a node x has
n neighbors and the message does not die out, then it would expect that all of its
neighbors would get the message, and, if the gossip probability is p, it should get
roughly pn messages from its neighbors. If it gets significantly fewer than pn messages within a reasonable time interval, then this is a clue that the message is dying
out. The authors have proposed a modification to resolve this issue. If a node with n
neighbors receives a message and does not broadcast it, but then does not receive
the message from at least m neighbors within a reasonable timeout period, it
EPIDEMIC MODELS IN AD HOC NETWORKS
65
broadcasts the message to all its neighbors. If m is chosen too large, then there may
be too many messages. The experimental results show that the most significant performance improvement could be obtained with m = 1. Thus, in GOSSIP3 (p, k, m),
if a node originally did not broadcast a received message but did get the message
from at least m other nodes within some timeout period it will immediately broadcast
the message after the timeout period. In what follow, we will discuss several gossip
schemes.
Geographic Gossip for Efficient Aggregation. Gossip algorithms have also
been used for data aggregation in sensor networks. Their forte is their simplicity in approach. However, in their basic form they may waste significant energy
by essentially passing around redundant information. The authors in Geographic
Gossip [24] propose an alternative gossiping scheme that exploits geographic
information.
In a network of n sensors, a basic solution to the averaging problem (i.e., to compute
the average of all n sensor measurements) is based on the Gossip algorithms where
each node randomly picks a one-hop neighbor and exchanges their current values.
This is performed in an iterative fashion, and ultimately all nodes converge to the
global average in a distributed manner. The key issue here is the number of iterations
it takes for such a gossip algorithm to converge to a sufficiently accurate estimate.
Recent works [25–29] have dealt with variants of this problem. The convergence
time of this algorithm is closely linked with the mixing time of the Markov Chain
defined by a weighted random graph on the network. In reference 26, the authors
showed how to optimize the neighbor selection probabilities for each node in order to
find the fastest mixing Markov chain. However, for sensor network graphs, even an
optimized gossip algorithm can result in excess energy consumption. The authors of
Geographic Gossip exploit geographic information to build a completely randomized
and distributed algorithm that requires substantially less communication. The idea is
to include geographic routing to gossip with random nodes far away in the network.
Empowered with geographic knowledge, this protocol succeeds in quickly diffusing
information everywhere in the network and thus computes the average faster than the
standard nearest-neighbor gossip.
Smart Gossip. The authors of Smart Gossip [30] propose an adaptive form of
gossiping in sensor networks. They propose techniques by which a gossip-based
protocol can automatically and dynamically adapt to the network topology. Smart
Gossip copes well with wireless losses and unpredictable node failures that affect
network connectivity. The adaptivity of the gossiping strategy also extends itself
to provide reliability for disseminating messages. The authors argue that existing
gossip strategies are mostly static, since there is a fixed probability for transmitting
the received information. There are a few variants of the gossip protocol which are
adaptive. Haas et al. [23] proposed an adaptive form of gossip which chooses its
probability, based on the number of neighbors. The authors of Smart Gossip argue
that simply choosing gossip probabilities based on the number of neighbors is not
66
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
B
C
F
G
A
D
E
Figure 3.6. Example illustrating the argument for Smart Gossip.
correct. For example, in Figure 3.6, which is a subgraph of a random topology, node
C has a high degree and therefore its probability of gossiping would be low. This
could adversely affect the reception of the information at node F , which is solely
dependent on C for receiving messages originating at any node to the left of C.
From the point of view of gossip percolation, the authors extract out the notion of
dependence between a node X and a subset of its neighbors. These dependencies give
birth to parent–child relationships between neighbors based on the direction in which
the gossip can travel probabilistically. This dependency graph is just logical and also
probabilistic in nature. In other words, node X does not depend on any particular
parent, Y , to receive the gossip. Instead, it depends on a group of nodes, expecting at
least one member of this group to probabilistically deliver the gossip to it. The gossip
probabilities chosen at each node is therefore a function of the group size.
Based on this intuition, nodes promiscuously overhear broadcast messages and
extract information by applying simple rules and thereby deduce whether the sender
of the message is a parent, child or a sibling. A child node, on identifying its parent
set, calculates the probability with which it thinks its parents are required to send,
and it announces this probability by piggybacking it on every gossip it forwards. A
parent node overhears such announcements and assumes its gossip probability to be
the maximum of all the announced probabilities.
3.5.2 Epidemic Routing
The authors in reference 19 introduce epidemic routing for partially connected ad hoc
networks where pairwise exchanges of messages among mobile hosts in a random
manner ensure the eventual message delivery to the destination node. The prominent
goals of epidemic routing are to (i) maximize message delivery rate, (ii) minimize
message latency, and (iii) minimize the total resources consumed in message delivery.
Existing ad hoc routing protocols assume that there is a connected path from source
to destination. However, with the emergence of short-range wireless communication
environments (e.g., Bluetooth [31]) and the wide area over which such networks are
deployed, this assumption is not always a realistic one. Unfortunately, the current
ad hoc routing protocols are heavily dependent on consistent network connectivity
to deliver packets between the source and the destination and generally fail in the
presence of network partitions. At the same time, several applications based on a
EPIDEMIC MODELS IN AD HOC NETWORKS
67
mobile sensor network exist, where there are frequent and numerous formations of
network partitions.
In reference 19, the authors develop techniques for delivering application data with
a high probability even when there is never a fully connected path between the source
and destination. The main essence of their approach is to distribute data to connected
hosts of the network, whom they call carriers; and depending on node mobility, the
carriers can establish contact with other connected portions of the network. Through
such transitive transmission of data, messages have a high probability of eventually
reaching their destination. However, with basic random forwarding, the data might
be transmitted to a large number of carrier hosts other than the destination that is
not desirable. Since the overall goal of epidemic routing is not just to maximize
message delivery rate and minimize message delivery latency, but also to minimize
the aggregate system resources consumed in message delivery, the authors circumvent
this problem to a reasonable extent by placing an upper bound on the message hop
count and per-node buffer space (the amount of memory devoted to carrying other
host’s messages). Their results show that epidemic routing is able to successfully
deliver messages to the destination nodes where existing ad hoc routing protocols fail
because of limited node connectivity.
Although the authors of this work do not explicitly use the mathematical formulations of an epidemic model, they essentially follow the same principles of the model.
The contact rate between carriers and destination or intermediate connected nodes
is dependent on the mobility pattern of the carriers. However, since the notion is to
route and not broadcast the message, the authors successfully constrain resources at
nodes to restrict the number of messages a host is willing to carry on behalf of other
hosts.
3.5.3 Epidemiology and Mobile Ad Hoc Networks
Information diffusion in mobile ad hoc networks (MANET) has been an area where
epidemiological modeling concepts fit naturally. As mentioned earlier, several modeling formulations in epidemiology assume a homogeneously mixing population where
each infected individual has an equal probability of having contact with any susceptible individual. Scenarios that fit this assumption can borrow the differential-equationbased formulations popular in epidemiology. Information diffusion in MANETs fit
very closely in this model. Given the random mobility model, it’s a fair assumption
that the nodes can homogeneously mix. As a result, this phenomenon could be aptly
modeled based on the differential rate equation formulations. This has been done
by the authors in reference 32. Based on a simplistic S-I-S model, the authors have
simplistically modeled the spread of information in a MANET. They showed that the
information dissemination can be more or less accurately described by the infection
rate of the model. They derived expressions that show the change of infection rate
based on the node densities.
In reference 33, the authors address the issue of how to disseminate relevant
information to mobile agents within a geosensor network. In their work, the authors
propose an environment for simulating information dissemination strategies in
68
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
mobile ad hoc geosensor networks. A geosensor network is defined as a sensor network that monitors phenomena in geographic space [34]. In the context of geosensor
networks, the authors provide a decentralized location-based service that is able to
disseminate relevant geospatial information to spatially dispersed mobile users that
form a mobile ad hoc geosensor network. The authors explore the precise nature
of efficient information dissemination strategies based on localized communication
between agents in a geosensor network. Specifically, they are concerned with mobile location-aware agents who are able to sense information about their immediate
geospatial environment and communicate with other agents in their neighborhood.
The authors distinguish between three different strategies. The first strategy, Flooding, is where each geosensor node that encounters an event or receives a message
about an event passes on the information to every other node within its communication range. The second approach is referred to as an Epidemic, in which each
node only informs n other agents about the events. In the third approach which is
location-constrained, information is only passed on in proximity to the event, and then
discarded.
In reference 35, the authors propose a document oriented model for information
dissemination in mobile ad hoc networks. The problem of routing messages in disconnected or partially connected mobile ad hoc networks has been dealt by previous
works like references 19 and 36. The main contribution in reference 35 is the implementation of a service for document dissemination in ad hoc networks and then using
this service as a building block for application level services.
Any document that is sent in the network is cached as long as possible by as many
devices as possible, so that it can remain available for those devices that could not
receive it at the time it was sent originally. Other than providing a caching system
where documents can be maintained in mobile devices, their service also provides
facilities for document advertisement, document discovery, and document transport
between neighboring devices. A device can periodically advertise to its neighbors
about the documents stored in its cache. It can also search for specific documents in
its neighborhood and can either push documents toward or pull documents from its
neighbors.
In reference 37, the authors propose a middleware for a controlled epidemic style
dissemination for mobile ad hoc networks. Since traditional middleware primitives
offer very little information on dissemination mechanisms and epidemic algorithms
have hardly been used to control the spreading of information depending on the
desired reliability and network structure, the authors present a mobile ad hoc network
middleware that uses epidemic-style information dissemination techniques to tune
the reliability of the communication.
The authors argue that existing epidemic algorithms have little control on the
information dissemination process, and much of it is based on experimental results
and not on any analytical model. In other words, the information spread cannot be
accurately tuned in order to reach only a desired percentage of the hosts.
The authors, therefore, propose algorithms that rely on epidemic models and take
into account the underlying network structure. They design middleware interfaces
that allow programmers to set the reliability for unicasting and anycasting with a high
EPIDEMIC MODELS OF MALICIOUS CODE PROPAGATION
69
degree of accuracy. The middleware would have primitives for epidemic dissemination
and would take as control inputs the percentage of hosts to which the information
is disseminated. The authors use the infectivity, which is the probability of being
infected by a neighboring host, to control the reliability of the probabilistic unicast.
Thus, given an expected reliability value, the middleware is able to calculate the
infectivity accurately in order to obtain an infection rate proportional to the total
number of hosts in the network. For constructing the analytical model, the authors
adopt the simple S-I-S model of epidemiological spread to model the information
dissemination in a MANET. For the analytical model, the authors assume that there is
homogeneous mixing of the nodes and that the infectivity of a single host per message
is constant. Using the average node degree and the probability of infection, the authors
calculate the infection rate. Based on its calculation, the authors depict the epidemic
spread algorithm which is executed periodically.
3.6 EPIDEMIC MODELS OF MALICIOUS CODE PROPAGATION
Computer worms have recently emerged as one of the most critical threats against
information confidentiality, integrity, and service availability. Host machines in the
Internet have repeatedly revealed their susceptibility to malicious intrusions like
worms that have compromised millions of vulnerable hosts at an extremely fast pace
[38, 39]. Given that the threat of virus and worm propagation in wireless networks is
quite real, a few recent works have tried to focus on this idea and successfully utilized
the concept of epidemic theory to model the spread of worms in wireless ad hoc and
sensor networks.
In reference 40, the authors discuss the epidemic model of virus spreading in
mobile environments. Given the increasing rise in the usage of mobile devices, it is
a matter of time before viruses propagating over the air interface would be a major menace. Already, there are several viruses and worms that spread over the air.
For example, the Brador virus [41] infects Pocket PCs running Windows CE; and by
installing a backdoor, it allows a remote attacker access to the device. The Cabir worm
[42] infects cell phones running the Symbian operating system. Identifying these
examples, the authors of reference 40 stress several important factors like movement
of devices and the geographic locations while formulating epidemic models for virus
spread in mobile environments. In their model which they call probabilistic queuing,
the authors investigate the behavior of malicious codes that spread via proximity-based
point-to-point wireless links. They point out the drawbacks that existing epidemiological modeling of similar processes in mobile environments have, like ignoring
the node velocity and the nonhomogeneous connectivity distributions that often arise
in mobile environments. The Kephart–White (KW) Model [10] assumes a homogeneously connected topology, and the network is represented by a single parameter,
namely, the average node degree. However, mobile environments are too dynamic in
order for this model to fit. The KW model only considers mean connectivity, and it
discards useful information when the underlying distribution has significant variance,
which is normal in a mobile environment. Furthermore, the velocity is an important
70
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
factor that influences the way a virus can spread among mobile nodes. The authors
incorporate this parameter in their model, and they come up with a new epidemic
threshold value when the virus spread reaches endemic state. In the KW model, which
ignores node mobility, this threshold is crossed when the infecting rate is greater than
the curing rate. In their model, the authors incorporate the mobility model in their
derivation of a node’s degree distribution, which leads to a new value for the epidemic
threshold.
Several other works discuss various aspects of vulnerabilities of sensor and mobile
ad hoc networks in the light of epidemic theory.
3.6.1 TWPM
In reference 43, the authors develop a topologically aware worm propagation
model (TWPM) for wireless sensor networks. An important strategy effectively
employed by many recent worms (e.g., CodeRed v2) is localized scanning. The local scanning worms after compromising a host machine, instead of scanning a fixed
IP address space, scan neighboring hosts with a higher probability. This strategy
has proven to be quite effective since the presence of a single vulnerable host
implies that other hosts on the same network would also be vulnerable with a high
probability.
Since general routing strategies in sensor networks have each sensor maintain a
neighbor list, this procedure of localized scanning could be very effective for a virus
spreading in a sensor network. Moreover, since the worm under consideration employs
(next-hop) information from a sensor to infect other sensors, the authors refer to it as a
topologically aware worm. Based on the S-I-S model of epidemic spread, the authors
have constructed a differential-equation-based worm propagation model in sensor
networks. Apart from simultaneously capturing both time and space propagation
dynamics, TWPM incorporates physical, MAC and network layer considerations of
practical sensor networks.
Dividing the sensor network into equal-sized segments and using a constant rate
of infection, the authors have arrived at a closed-form expression for the number of
infectives at time t that also successfully captures the spatial information in terms of
the segment coordinates.
3.6.2 Compromise Propagation in Secure Sensor Networks
TWPM modeled the worm propagation process using a differential-equation-based
approach. However, generally in a static network (e.g., a sensor network), the
differential-equation-based approach is not feasible since it assumes a homogeneous
mixing of the susceptible nodes and the infected nodes. In such a scenario, a network
or graph-theoretic modeling technique is much more suitable to capture the propagation process. One such novel work has been done by the authors in reference 44,
where they model the process of how a compromised node in a sensor network gradually compromises other nodes and eventually compromises the whole network.
EPIDEMIC MODELS OF MALICIOUS CODE PROPAGATION
71
The authors assume the nodes in the sensor network to be uniformly randomly
deployed in an area and securely communicating among themselves. By secure communication, the authors assume a secret shared key-based communication paradigm.
They assume that a prior random key distribution technique has distributed secret
keys to each node, whereby they communicate with each other. Given such a securely
communicating sensor network, the authors study how an adversary who has captured
one or two sensor nodes and extracted their secret keys can possibly propagate the
compromise of nodes to the whole network.
When a node is captured and its keys are known by the attacker, secure communication can be established with neighboring nodes with which the captured node
shares keys. Being able to securely communicate with its neighbor, the node with
the malicious code can easily attain its susceptible neighbor’s trust and pass on the
malicious code to the latter. Once it has passed to the susceptible neighbor, the authors assume that the malicious code has the ability to acquire the secret keys of
the new node. This is when the new node also becomes infected and results in the
propagation of malicious code. This process continues until the whole network gets
compromised.
By constructing a random graph model of the key sharing overlay graph of the
sensor network, the authors present the compromise propagation model as a Poisson
process with a mean that is dependent on the infection probability and the infectivity
duration at each node. The propagation process was expressed by a transmissibility
factor of the infection, and it was basically analogous to the bond occupation probability on the graph representing the key sharing network. The size of the epidemic was
equivalent to the size of the giant component formed with edge existence probability
defined by the transmissibility of the compromise process.
The main focus of their work was to identify the phase transition points of
the process when it attains epidemic proportions. They studied the effects of the
compromise propagation under two scenarios, namely, without node recovery and
with node recovery. In the event of a compromise, the network may attempt to recover the particular node. Recovery might be realized in several possible ways. For
example, the keys of the nodes might be revoked and the node may be given a
fresh set of secret keys. In this context, key revocation, which refers to the task of
securely removing keys that are known to be compromised, has been investigated
as part of the key management schemes—for example, in reference 45. Moreover,
recovery can also be achieved by simply removing the compromised node from
the network—for example, by announcing a blacklist—or by simply reloading the
nodes programs. More sophisticated methods may include immunizing a node with
an appropriate antivirus patch that might render the node immune from the same
virus attack. Regardless, in their analysis, the authors studied virus spreading
under the two cases, respectively, depending on whether a node can be recovered
or not.
Since, contrary to the differential-rate-equation-based modeling methods, the
graph theoretic model does not capture the temporal effects of an epidemic, the
authors captured the temporal dynamics of the propagation process using simulation techniques.
72
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
3.7 CONCLUSION AND OPEN ISSUES
In this chapter we have delved into various epidemiological models and protocols
employed in wireless ad hoc and sensor networks. Starting from data dissemination
and gossip protocols to security issues in sensor networks such as propagation of
compromise of sensor nodes, we observe that there have been several works inspired
by this powerful concept of epidemic theory. The density and scale of a sensor network,
coupled with the objective of a one-to-many data transfer from a few nodes to the rest
of the network, unleash the efficiency with which this theory can effectively model and
provide solutions to several problems in ad hoc and sensor networks. In this chapter,
we have tried to touch most of the salient contributions that have adopted this theory
and provide a concise compilation under various categories of subclassifications.
Among the open issues related to the modeling of sensor networks, we find that not
many protocol models in sensor networks have been proposed with the nodes being
mobile. With mobile nodes, the dynamics of the network and its properties, like
connectivity, keep changing continuously, thus making it more difficult to capture in
an analytical closed-form model. For example, the analysis and performance study of
broadcast protocols in a mobile sensor network environment still requires considerable
research. Another important aspect that needs to be dealt with while modeling a sensor
network protocol is the node deployment scenario. Apart from the straightforward
uniform random deployment of the sensor nodes, there could be other distributions
for node deployment. A popular one among them is deploying nodes in packets such
that the resident points of nodes from each packet forms a two-dimensional Gaussian
distribution about the dropping point. Given such a position distribution of the sensor
network, epidemic modeling of dissemination or broadcast protocols would have to
deal with the change in degree distribution dependent on the location of a node in the
network.
3.8 EXERCISES
1. How is epidemic modeling in sensor networks different from that used for the
Internet?
2. What is the basic reproductive number in epidemiology? How is it different for
scale-free networks from other networks?
3. How does a homogeneously mixing population and a heterogeneously mixing one affect the mathematical formulation of the spreading process in
epidemiology?
4. How does mobility of the network nodes change the mathematical formulation
of an epidemic model of the network?
5. What are the important parameters that determine the choice between a continuous and a discrete time epidemic model of a network?
6. What are the important issues while epidemic modeling of malware spread in
a mobile environment? How do the location of a mobile device and the time of
the day affect the model?
BIBLIOGRAPHY
73
ACKNOWLEDGMENT
This work is supported by NSF ITR grant No. IIS-0326505 and Texas ARP grant
No. 14-748779.
BIBLIOGRAPHY
1. A. Woo, S. Madden, and R. Govindan. Networking support for query processing in sensor
networks. Communications of the ACM, 47(6):4752, 2004.
2. R. M. Anderson and R. M. May. Infectious Diseases of Human: Dynamics and Control,
Oxford University Press, Oxford, 1991.
3. M. E. J. Newman. Spread of epidemic disease on networks. Physics Reviews E, 66:
016128, 2002.
4. R. Pastor-Satorras and A. Vespignani. Epidemic dynamics and endemic states in complex
networks. Physics Reviews E, 63:066117, 2001.
5. R. M. May and A. L. Lloyd. Infection dynamics on scale-free networks. Physics Reviews
E, 64:066112, 2001.
6. I. de S. Pool and M. Kochen. Contacts and influence. Social Networks 1:148, 1978.
7. C. Moore and M. E. J. Newman. Epidemics and percolation in small-world networks.
Physics Reviews E 61:5678-5682, 2000.
8. P. Grassberger. On the critical behavior of the general epidemic process and dynamic
percolation. Mathematical Biosciences. 63:157, 1983.
9. C. Griffin and R. Brooks. A note on the spread of worms in scale-free networks. In IEEE
Transactions on Systems, Man and Cybernetics, 36(1):198–202, 2006.
10. J. Kephart and S. White. Directed-graph epidemiological models of computer viruses. In
Proceedings of the IEEE Computer Symposium on Research in Security and Privacy, May
1991, pp. 343–359.
11. J. Kulik, W. Rabiner, and H. Balakrishnan, Adaptive protocols for information dissemination in wireless sensor networks. In Proceedings of the Fifth ACM/IEEE Mobicom
Conference, Seattle, WA, August 1999. IEEE CD-ROM.
12. S. S. Kulkarni and M. Arumugam. INFUSE: A TDMA based data dissemination protocol
for sensor networks. Technical Report MSU-CSE-04-46, Department of Computer Science,
Michigan State University, November 2004.
13. J. W. Hui and D. Culler. The dynamic behavior of a data dissemination protocol for network
programming at scale. In 2nd International Conference on Embedded Networked Sensor
Systems, 2004. IEEE CD-ROM.
14. P. Levis and D. Culler. The firecracker protocol. In Proceedings of the 11th ACM SIGOPS
European Workshop, Leuven, Belgium, 2004. IEEE CD-ROM.
15. P. Levis, N. Patel, D. Culler, and S. Shenker. Trickle: A self-regulating algorithm for
code maintenance and propagation in wireless sensor networks. In First USENIX/ACM
Symposium on Network Systems Design and Implementation (NSDI), 2004. IEEE CDROM.
16. M. Akdere, C. C. Bilgin, O. Gerdaneri, I. Korpeoglu, O. Ulusoy, U. Cetintemel. A comparison of epidemic algorithms in wireless sensor networks. Computer Communication,
6:35–41, 2006.
74
EPIDEMIC MODELS, ALGORITHMS, AND PROTOCOLS IN WIRELESS SENSOR
17. D. Ganesan, B. Krishnamachari, A. Woo, D. Culler, D. Estrin, and S. Wicker. An empirical
study of epidemic algorithms in large scale multihop wireless networks. Intel Research,
Berkeley Technical Report IRB-TR-02-003, March 2002.
18. A. Beaufour, M. Leopold, and P. Bonnet. Smart-tag based data dissemination. In WSNA,
2002.
19. A. Vahdat and D. Becker. Epidemic routing for partially-connected ad hoc networks. Duke
Technical Report CS-2000-06, July 2000.
20. S. S. Kulkarni and L. Wang. MNP: Multihop network reprogramming service for sensor
networks. In Proceedings of the 25th International Conference on Distributed Computing
Systems ICDCS, June 2005, p. 716.
21. C. E. Perkins and E. M. Royer. Ad-hoc on-demand distance vector routing. In 2nd IEEE
Workshop on Mobile Computing Systems and Applications, February 1999, pp. 90–100.
22. D. B. Johnson and D. A. Maltz. Dynamic Source Routing in Ad Hoc Wireless Networks.
Kluwer Academic Publishers, Norwell, MA, 1996.
23. Z. J. Haas, J. Y. Halpern, and L. Li. Gossip-Based Ad Hoc Routing. In INFOCOM 2002.
IEEE CD-ROM.
24. A. G. Dimakis, A. D. Sarwate, and M. J. WainWright. Geographic gossip: Efficient aggregation for sensor networks. In Fifth International Symposium on Information Processing
in Sensor Networks (IPSN), 2006. IEEE CD-ROM.
25. S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah. Analysis and optimization of randomized
gossip algorithms. In Proceedings of the 43rd Conference on Decision and Control (CDC
2004), 2004. IEEE CD-ROM.
26. S. Boyd, A. Ghosh, B. Prabhakar, and D. Shah. Gossip algorithms: Design, analysis and
applications. In Proceedings of the 24th Conference of the IEEE Communications Society
(INFOCOM 2005), 2005. IEEE CD-ROM.
27. J.-Y. Chen and D. X. G. Pandurangan. Robust aggregates computation in wireless sensor networks: Distributed randomized algorithms and analysis. In Fourth International
Symposium on Information Processing in Sensor Networks (IPSN), 2005. IEEE CD-ROM.
28. R. Karp, C. Schindelhauer, S. Shenker, and B. Vocking. Randomized rumor spreading. In
Proceedings of the IEEE Conference of Foundations of Computer Science, (FOCS), 2000.
IEEE CD-ROM.
29. D. Kempe, A. Dobra, and J. Gehrke. Gossip-based computation of aggregate information.
In Proceedings of the IEEE Conference of Foundations of Computer Science (FOCS),
2003. IEEE CD-ROM.
30. P. Kyasanur, R. R. Choudhury, and I. Gupta. Smart gossip: Infusing adaptivity into gossiping protocols for sensor networks. In ICDCS 2006. IEEE CD-ROM.
31. J. Haartsen, M. Naghshineh, J. Inouye, O. J. Joeresson, and W. Allen. Bluetooth: Vision,
goals, and architecture. ACM Mobile Computing and Communications Review, 2(4):3845,
1998.
32. A. Khelil, C. Becker, J. Tian, and K. Rothermel. An epidemic model for information
diffusion in MANETs. In MSWiM, 2002. IEEE CD-ROM.
33. S. Nittel M. Duckham, and L. Kulik. Information dissemination in mobile ad hoc geosensor
networks. In Third International Conference on Geographic Information Science, 2004.
IEEE CD-ROM.
34. S. Nittel et al. Report from the first workshop on geo sensor networks. ACM SIGMOD
Record 33 2004.
BIBLIOGRAPHY
75
35. F. Guidec and H. Roussain. Asynchronous document dissemination in dynamic ad hoc
networks. In Second International Symposium on Parallel and Distributed Processing and
Applications (ISPA), 2004. IEEE CD-ROM.
36. A. Lindgren, A. Doria, and O. Schelen. Probabilistic routing in intermittently connected
networks. In Proceedings of the Fourth ACM International Symposium on Mobile Ad Hoc
Networking and Computing (MobiHoc 2003), June 2003. IEEE CD-ROM.
37. M. Musolesi and C. Mascolo. Controlled epidemic-style dissemination middleware for
mobile ad hoc networks. In Mobiquitous, 2006. IEEE CD-ROM.
38. S. Staniford, V. Paxson, and N. Weaver. How to own the Internet in your spare time. Usenix
Security Symposium, 2002.
39. C. Shannon and D. Moore. The spread of the witty worm. IEEE Security and Privacy, 2(4):
46–50, 2004.
40. J. W. Mickens and B. D. Noble. Modeling epidemic spreading in mobile environments. In
WiSE, 2005.
41. R. Wong and I. Yap. Security information. In Virus Encyclopedia: WINCE BRADOR.A,
Technical Details, 2004. Trend Micro Incorporated.
42. P. Ferrie, P. Szor, R. Stanev, and R. Mouritzen. Security Response: SymbOS.Cabir, 2004.
Symantec Corporation.
43. S. A. Khayam and H. Radha. A Topologically-aware worm propagation model for wireless sensor networks. In IEEE ICDCS International Workshop on Security in Distributed
Computing Systems SDCS, 2005. IEEE CD-ROM.
44. P. De, Y. Liu, and S. K. Das, Modeling node compromise spread in sensor networks using
epidemic theory. In World of Wireless, Mobile and Multimedia Networks, WoWMoM,
2006. IEEE CD-ROM.
45. H. Chan, V. D. Gligor, A. Perrig, G. Muralidharan. On the distribution and revocation of
cryptographic keys in sensor networks. In IEEE Transactions on Dependable and Secure
Computing 2005. IEEE CD-ROM.
CHAPTER 4
Modeling Sensor Networks
STEFAN SCHMID and ROGER WATTENHOFER
Computer Engineering and Networks Laboratory (TIK), ETH Zurich, CH-8092 Zurich,
Switzerland
4.1 INTRODUCTION AND MOTIVATION
In order to develop algorithms for sensor networks and in order to give mathematical
correctness and performance proofs, models for various aspects of sensor networks
are needed. This chapter presents and discusses currently used models for sensor
networks. Generally, finding good models is a challenging task. On the one hand, a
model should be as simple as possible such that the analysis of a given algorithm
remains tractable. On the other hand, however, a model must not be too simplistic
in the sense that it neglects important properties of the network. A great algorithm
in theory may be inefficient or even incorrect in practice if the analysis is based on
idealistic assumptions. For example, an algorithm that ignores interference may fail in
practice since communication happens over a shared medium. Many models for sensor
network have their origin in classic areas of theoretical computer science and applied
mathematics. Since the topology of a sensor network can be regarded as a graph,
the distributed algorithms community uses models from graph theory, representing
nodes by vertices and wireless links by edges. Another crucial ingredient of sensor
network models is geometry. Geometry comes into play as the distribution of sensor
nodes in space, as well as the propagation range of wireless links, usually adheres to
geometric constraints.
The chapter is organized as follows. In Section 4.2, the reader will become familiar with various models for the network’s connectivity. Connectivity models answer the question: Which nodes are “connected” to which other nodes and can
therefore directly communicate with each other. Section 4.3 then enhances these
connectivity models by adding interference aspects: Since sensor nodes communicate over a shared, wireless medium, a transmission may disturb a nearby concurrent
transmission. After having studied connectivity and interference issues, we look at
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
77
78
MODELING SENSOR NETWORKS
modeling questions related to algorithm design in Section 4.4. The reader is provided
with a survey of models that influence the feasibility and efficiency of certain operations on sensor networks. We draw some general conclusions in Section 4.5, and we
point out interesting areas for future research in Section 4.6.
4.2 MODELING THE SENSOR NODES’ CONNECTIVITY
A first and foremost modeling question concerns the connectivity of sensor nodes:
Given a set of nodes distributed in space, we need to specify which nodes can receive
a transmission of a node. Throughout this chapter, if a node u is within a node v’s
transmission range, we say that u is adjacent to v, or, equivalently, that u is a neighbor
of v. In the absence of interference (cf. Section 4.3), this relation is typically symmetric
(or undirected); that is, if a node u can hear a node v, also v can hear u. The connectivity
of a sensor network is described by a graph G = (V, E), where V (vertices) is the
set of sensor nodes, and E (edges) describes the adjacency relation between nodes.
That is, for two nodes u, v ∈ V , (u, v) ∈ E if v is adjacent to u. In an undirected graph,
it holds that if (u, v) ∈ E, then also (v, u) ∈ E; that is, edges can be represented by
sets {u, v} ∈ E rather than tuples.
The classic connectivity model is the so-called unit disk graph (UDG) [1]. The
name “unit disk graph” stems from the area of computational geometry; it is a special
case of the so-called intersection graph. In this model, nodes having omnidirectional
radio antennas—that is, antennas with constant gain in all directions—are assumed
to be deployed in a planar, unobstructed environment. Two nodes are adjacent if and
only if they are within each other’s transmission range (which is normalized to 1).
Model 4.2.1 (Unit Disk Graph (UDG)). Let V ⊂ R2 be a set of nodes in the twodimensional Euclidean plane. The Euclidean graph G = (V, E) is called unit disk
graph if any two nodes are adjacent if and only if their Euclidean distance is at most
1. That is, for arbitrary u, v ∈ V , it holds that {u, v} ∈ E ⇔ |u, v| ≤ 1. Figure 4.1
depicts an example of a UDG.
The UDG model is idealistic: In reality, radios are not omnidirectional, and even
small obstacles such as plants can change connectivity. Therefore, some researchers
1
V
W
u
Figure 4.1. Unit disk graph: Node u is adjacent to node v (distance ≤ 1) but not to node w
(distance > 1).
MODELING THE SENSOR NODES’ CONNECTIVITY
79
have proposed to study the other extreme and model the sensor network as a general
graph; that is, each node can be adjacent to every other node.
Model 4.2.2 (General Graph (GG)). The connectivity graph is a general undirected
graph G.
While a UDG is too optimistic, the GG is often too pessimistic, because the connectivity of most networks is not arbitrary but obeys certain geometric constraints.
Still, in some application scenarios it might be accurate to operate either on the UDG
or on the GG. Indeed, there are algorithms developed for the UDG which also perform
well in more general models. Moreover, some algorithms designed for the GG are
currently also the most efficient ones for UDGs (e.g., reference 2).
The research community has searched for connectivity models between the two
extremes UDG and GG. For example, the quasi unit disk graph model (QUDG)
[3, 4] is a generalization of the UDG that takes imperfections into account as they
may arise from non-omnidirectional antennas or small obstacles. These QUDGs are
related to so-called civilized graphs. The interested reader can find more information
in reference 5.
Model 4.2.3 (Quasi Unit Disk Graph (QUDG)). The nodes are in arbitrary positions in R2 . All pairs of nodes with Euclidean distance at most ρ for some given
ρ ∈ (0, 1] are adjacent. Pairs with a distance larger than 1 are never in each other’s
transmission range. Finally, pairs with a distance between ρ and 1 may or may not be
neighboring. An example is shown in Figure 4.2.
Note that, for ρ = 1, a QUDG is a UDG, and therefore the following theorem
holds.
Theorem 4.2.1. A UDG is a special case of a QUDG.
The QUDG model itself can be extended in several ways.
V5
V4
V2
V3
u
V1
P
1
Figure 4.2. Quasi unit disk graph from the perspective of node u: Node u is always adjacent
to node v1 (d(u, v1 ) ≤ ρ) but never to v5 (d(u, v5 ) > 1). All other nodes may or may not be in
u’s transmission range. In this example, node u is adjacent to v3 and v4 but not to v2 .
80
MODELING SENSOR NETWORKS
Model 4.2.4 (QUDG Variations). The QUDG as presented in Model 4.2.3 does not
specify precisely what happens if the distance is between ρ and 1. There are several
options. For example, one could imagine an adversary choosing for each node pair
whether they are in each other’s transmission range or not. Alternatively, there may
be a certain success probability of being adjacent: The corresponding probability
distribution could depend on the time and/or distance [6]. For example, the QUDG
could be used to study Rayleigh fading; that is, the radio signal intensity could vary
according to a Rayleigh distributed random variable. Also, a probabilistic on/off
model is reasonable, where in each round a link’s state changes from good to bad and
vice versa with a given probability.
Measurement studies suggest that in an unobstructed environment, and with many
nodes available, 1/ρ is modeled as a small constant [7]. Interestingly, many algorithms
can be transferred from the UDG to the QUDG at an additional cost of 1/ρ2 [4]. Note
that while for ρ ≈ .5 this factor is bearable, the algorithms are two orders of magnitude
worse if ρ ≈ .1. While the QUDG can be attractive to model nodes deployed in
fields with few obstacles, it does not make sense for inner-city or in-building networks
where obstructions cannot be ignored: Since a node may be able to communicate with
another node which is dozens of meters away, but not with a third node being just
around the corner, ρ would be close to 0.
However, even in such heterogeneous environments, the connectivity graph is still
far from being a general graph. Although nodes that are close but on different sides
of a wall may not be able to communicate, a node is typically highly connected to
the nodes which are in the same room, and thus many neighbors of a node are direct
neighbors themselves. In other words, even in regions with many obstacles, the total
number of neighbors of a node which are not adjacent is likely to be small. This
observation has motivated Model 4.2.6, see reference 8 for more details.
Model 4.2.5 (Bounded Independence Graph (BIG)). Let ϒ r (u) denote the set of
independent nodes that are at most r hops away from node u (i.e., nodes of u’s
r-neighborhood) in the connectivity graph G. Thus, a set S ⲴV of nodes is called
independent if all nodes in the set are pairwise not adjacent; that is, for all u, v ∈ S,
it holds that {u, v} ∈
/ E. Graph G has bounded independence if and only if for all
nodes u ∈ G, |ϒ r (u)| = O(poly(r)) (typically |ϒ r (u)| ∈ O(r c ) for a small constant
c ≥ 2).
The BIG model reflects reality quite well and is appropriate in many situations.
Figure 4.3 shows a sample scenario with a wall; in contrast to UDG and QUDG, the
BIG model captures this situation well.
Since the number of independent neighbors in a disk of radius r of a UDG is at most
O(r2 ), we have the following fact. The proof is simple (and similar to the upcoming
proof of Theorem 4.2.13) and left to the reader as an exercise.
Theorem 4.2.2. The UDG model is a special case of the BIG model. Similarly, if ρ
is constant, also a QUDG is a BIG.
MODELING THE SENSOR NODES’ CONNECTIVITY
81
W
u
V
Figure 4.3. Nodes u and v are separated by a wall. Nodes on the same side of the wall are
completely connected. However, due to the wall, although u can reach a distant node w, it
cannot hear the close node v. Such situations can be modeled by the BIG but not by the UDG
or the QUDG.
Observe that many models described so far can be generalized. For instance, the
UDG and QUDG models can be studied in three dimensions rather than in the plane,
or using different distance functions (norms). For more detailed information on the
concept of norms, the reader may want to consult any introductory book on linear
algebra.
Model 4.2.6 (Generalized (Q)UDG). One extension of the UDG and QUDG models is to consider nodes in R3 . Moreover, distances between nodes could be
modeled using the Manhattan norm (L1 norm). In the Manhattan norm, the distance between two points u = (x1 , y1 ) and v = (x2 , y2 ) in the plane is given by
d(u, v) = |x2 − x1
| + |y2 − y1 |, while in the Euclidean norm (L2 norm), the distance is d(u, v) = |x2 − x1 |2 + |y2 − y1 |2 . Alternatively, also the maximum norm
(L∞ norm) is popular, where d(u, v) = max |x2 − x1 |, |y2 − y1 |.
The UDG model has also been extended to more general metric spaces; for
example, in reference 9, it was extended to doubling metrics [10]. Note that
a metric space defines distances between all pairs of nodes while guaranteeing
non-negativity, identity of indiscernibles, symmetry, and triangle inequality. A doubling metric is simply a metric space with some additional constraints which are
described next.
Model 4.2.7 (Unit Ball Graph (UBG)). A doubling metric space is defined as follows: For a node u, let the ball Bu (r) denote the set of all nodes at a distance at most
r from u. It holds, for all nodes u and all r ≥ 0, that the ball
Bu (r) can be covered
by a constant number of balls of radius r/2; that is, Bu (r)⊂ i=1...c Bui (r/2), where
82
MODELING SENSOR NETWORKS
u2
r
u3
r/2
u
u1
Figure 4.4. The Euclidean plane forms a doubling metric. In this example, the nodes are
distributed in R2 , and three balls of radius r/2 are sufficient to cover all nodes in Bu (r); that is,
Bu (r) = Bu1 (r/2) ∪ Bu2 (r/2) ∪ Bu3 (r/2).
ui are arbitrary nodes and c is a (usually small) constant. In the UBG model, nodes
are assumed to form a doubling metric space. Two nodes u and v with d(u, v) ≤ 1 are
adjacent, whereas all other nodes are not.
The proof of the following theorem is left to the reader as an exercise.
Theorem 4.2.3. Nodes in a two-dimensional Euclidean plane (i.e., the metric space is
given by the Euclidean distances) form a doubling metric. A general graph, however,
does not.
Figure 4.4 shows an example for the Euclidean plane. In this setting, three balls of
radius r/2 are enough to cover all nodes in the ball of radius r around node u. To
see why a general graph may not form a doubling metric, consider a graph where all
nodes have distance 1 to all other nodes. Observe that it is possible to model a UDG
with a UBG by using the Euclidean distances of the UDG and connecting those node
pairs which have distance at most 1. Moreover, even a QUDG can be modeled with
a UBG. We have the following results.
Theorem 4.2.4. A UDG is a UBG.
Theorem 4.2.5. An undirected QUDG with constant ρ is a UBG.
Proof. The idea of the proof is as follows: First, we transform all distances in the
QUDG. We then show that during this transformation, all edges are maintained; that
is, the resulting graph is isomorphic to the QUDG. Moreover, it can be shown that
after the transformation, the graph also fulfills the requirements of a doubling metric
space.
83
MODELING THE SENSOR NODES’ CONNECTIVITY
We transform the distances between all pairs of nodes (u, v) in the QUDG as
follows. Let dQ (u, v) denote the distance from node u to node v in the QUDG, and let
dB (u, v) be the transformed distance in the UBG. Moreover, let ǫ > 0 be an arbitrary
small number.
dB (u, v) :=
dQ (u, v)/ρ
1
1+ǫ
dQ (u, v)
if dQ (u, v) ≤ ρ
if
ρ < dQ (u, v) ≤ 1 and v is adjacent to u
if ρ < dQ (u, v) ≤ 1 and v is not adjacent to u
if dQ (u, v) > 1
Observe that by this transformation, pairs of nodes that are adjacent in the QUDG are
assigned distances of at most 1 and are therefore also adjacent in the UBG. Similarly,
nodes that are not adjacent in the QUDG have a distance larger than 1 are therefore
not neighboring in the UBG either. Also observe that the transformation increases the
distance between two nodes by less than a constant factor of µ := (1 + ǫ)/ρ, but it
never decreases any distances. It remains to show that after the transformation, the
nodes indeed form a doubling metric space.
In order to form a metric space, the distances between the nodes are to fulfill the
following properties: (1) nonnegativity, (2) identity of indiscernibles, (3) symmetry,
and (4) triangle inequality. The nonnegativity and the identity of indiscernibles criteria are met trivially. The symmetry criterion, however, might not hold, because the
adjacency relation can be directed in a QUDG. Therefore, in the following, we consider undirected QUDGs only. Hence, since our distance transformation maintains
symmetry, Property 3 holds as well. It remains to discuss the triangle inequality.
Consider two arbitrary nodes u and v. Since in the QUDG, all distances are
Euclidean, it holds that
∀w : dQ (u, v) ≤ dQ (u, w) + dQ (w, v)
(4.1)
Let us now look at the following three cases in turn: (i) dQ (u, v) ≤ ρ, (ii) ρ <
dQ (u, v) ≤ 1, and (iii) 1 < dQ (u, v). In Case i, no node w with distance larger than ρ
from any of the two nodes u and v can challenge the triangle inequality. For all other
nodes w, however, it holds that dB (u, v) = dQ (u, v)/ρ ≤ (dQ (u, w) + dQ (w, v))/ρ =
dB (u, w) + dB (w, v). Here, the equalities hold by the definition of the transformation
function and the inequality is due to Eq. (4.1). Next, we tackle Case ii. Again, only
nodes w with dQ (u, w) ≤ ρ and dQ (w, v) ≤ ρ can challenge the inequality. However, we know that dQ (u, v) > ρ, and hence Eq. (4.1) yields dB (u, w) + dB (w, v) =
dQ (u, w)/ρ + dQ (w, v)/ρ > 1. Finally, the triangle inequality also holds in Case iii,
because the distance between u and v in the UBG is the same as in the QUDG, and
our transformation never decreases any distances.
We conclude the proof by showing that the metric space has a constant doubling dimension. Recall that all distances are only stretched by a constant factor
between 1 and µ in our transformation. Therefore, for all nodes u and arbitrary radii
84
MODELING SENSOR NETWORKS
QUDG
r, Bu
(r/µ)⊂BuUBG (r). Thus, at most a constant factor of O(log µ) times, more
balls are needed for the UBG than for the QUDG (the Euclidean plane) in the worst
case, and the claim follows.
The UBG itself has a polynomially bounded independence and is therefore a BIG.
Theorem 4.2.6. A UBG is a BIG.
Proof. Fix a node u. We have to prove that the total number of independent nodes in
Bu (r) grows polynomially in r. Observe that, due to the triangle inequality, in Bu (1/2)
there is at most one independent node. Thus, by the definition of a doubling metric,
there are at most c independent nodes in Bu (1), at most c2 in Bu (2), c3 in Bu (4),
and so on. Generally, there are at most clog r+1 independent nodes in Bu (r). Since
clog r ∈ O(r c ), the claim follows.
To conclude, we present two additional modeling aspects with which connectivity models are occasionally extended. The first aspect concerns the sensor nodes’
antennas.
Model 4.2.8 (Antennas). Besides omnidirectional antennas, there is a wide range of
more sophisticated antenna models. For example, a node can have a directional radio
antenna with more gain in certain directions.
Finally, as mentioned in the discussion of the QUDG, links are not always reliable:
Links may be up and down—for example, according to a probabilistic process.
Model 4.2.9 (Link Failures). Any graph-based model can be enhanced with probabilistic links.
4.3 INTERFERENCE ISSUES IN WIRELESS SENSOR NETWORKS
In wireless networks, the communication medium is shared and transmissions are
exposed to interference. Concretely, a node u may not be able to correctly receive a
message of an adjacent node v because there is a concurrent transmission nearby. In
some sense, an interference model explains how concurrent transmissions block each
other. Interference is a difficult phenomenon, with many hard-to-capture characteristics. A signal might, for example, interfere with itself due to multipath propagation
(e.g., a direct path canceling with a longer path reflecting on an object). A discussion of these effects is beyond the scope of this overview chapter. Instead we look at
models that capture reality from a worst-case perspective. The mother of all interference models is the so-called physical or SINR model [11–13], which is widely accepted
by information theorists. In this model, the successful reception of a message depends
on the received signal strength, the ambient noise level, and the interference caused
by simultaneously transmitting nodes.
INTERFERENCE ISSUES IN WIRELESS SENSOR NETWORKS
85
Model 4.3.1 (Signal-to-Interference Plus Noise (SINR)). Let Pr be the signal
power received by a node vr and let Ir denote the amount of interference generated by other nodes. Finally, let N be the ambient noise power level. Then, a node
Pr
vr receives a transmission if and only if N+I
≥ β. Thus, β is a small constant (der
pending on the hardware) and denotes the minimum signal to interference ratio that
is required for a message to be successfully received. The value of the received signal
power Pr is a decreasing function of the distance d(vs , vr ) between transmitter vs
and receiver vr . More specifically, the received signal power is modeled as decaying
with distance d(vs , vr ) as d(vs1,vr )α . The so-called path-loss exponent α is a constant
between 2 and 6 and depends on external conditions of the medium, as well as on the
exact sender–receiver distance. 1 Let Pi be the transmission power level of node vi .
A message transmitted from a node vs ∈ V is successfully received by a node vr if
Ps
d(vs , vr )α
N+
vi ∈V \{vs }
Pi
d(vi , vr )α
≥β
In other words, in the SINR model, a node correctly receives a transmission if the received signal power—which depends on the sending power and the distance between
sender and receiver—is large enough compared to the signal power of concurrent
(interfering) transmissions and the ambient noise level. Sometimes a variation of this
SINR model is used in literature. It has an additional requirement: For a successful reception, the received signal power must exceed a minimal threshold θ, that is,
Pr ≥ θ. In many situations, such a threshold can also be incorporated implicitly by
the ambient noise power level N. Moreover, researchers have also studied a probabilistic SINR model [14], where the gain of an antenna is described by a Gaussian
distribution—independently of the distance! Apart from the interference term, and
if all nodes send with the same transmission power level, the connectivity model of
SINR is exactly the UDG, with path-loss exponent α and minimum ratio β such that
the maximum distance for receiving a signal is 1. Hence, the SINR model can be
extended similarly to the UDG model. Now, observe that the SINR model does not
specify the signal power Ps used by a sender vs to transmit data to the receiver vr .
Three models are common:
Model 4.3.2 (Power Control). CONST: All nodes use the same constant transmission power. DIST: The power level depends on the distance d between sender and
receiver. Concretely, the transmission power is given by c · d α for some α ≥ 2 and
some constant c > 0. GEN: A general (or arbitrary) power level is assumed at the
1 In
free space, α roughly equals 2. In the so-called two-ray ground model, it is assumed that there are
two paths of the electromagnetic wave: a direct one and a ground reflected signal path; to describe this
situation, α = 4 is used. Finally, note that ever since Marconi’s first experiments, time has been devoted to
explain radio propagation phenomena, and there is a plethora of other proposals. For example, for small
urban cells, a photon propagation model has been suggested implying an exponentially growing path loss.
86
MODELING SENSOR NETWORKS
Figure 4.5. Sample network with heterogeneous transmission ranges. For instance, the node
on the far left saves energy and reduces interference by using only a small power level.
sender, which may change over time. Figure 4.5 depicts a network where each node
has a different power level.
Although the SINR model incorporates many important physical properties, it has
not received an appropriate amount of attention from the algorithms community [12].
This can be partially explained by the fact that the SINR model is complicated. For
instance, a lot of far-away transmissions sum up, and may interfere with a close-by
sender-receiver pair. In practice, however, these far-away transmissions often only
contribute to the ambient noise and need not be counted individually. Twiddling the
knobs of the model a bit more, we might not sum up all interfering transmissions, but
simply look at the worst—or, in the case of a CONST model; closest—disturbance:
A node receives a transmission if and only if the closest simultaneously transmitting
node is far enough.
Model 4.3.3 (Interfering Transmissions). SUM: All interfering transmissions are
taken into account. ONE: Only the worst (or closest) interfering transmission matters.
NULL: Pure connectivity models which do not consider interference aspects (cf.
Section 4.2).
ONE models are quite popular because of their simplicity. The UDI—an
interference-aware version of the UDG—is a prominent example (cf. Model 4.3.4).
Observe that, because of the constant transmission power, the power control type of
UDI is CONST. Figure 4.6 shows an example.
Model 4.3.4 (UDG with Distance Interference (UDI)). Nodes are situated arbitrarily in the plane. Two nodes can communicate directly if and only if their Euclidean
distance is at most 1, and if the receiver is not disturbed by a third node with Euclidean
distance less or equal a constant R ≥ 1.
Often the constant R of the UDI model is approximated in such a way that interference can be reduced to a parameter of the UDG. For instance, some MAC
INTERFERENCE ISSUES IN WIRELESS SENSOR NETWORKS
87
x
X
1
V
W
u
R
Figure 4.6. The UDI model has two radii: a transmission radius (length 1) and an interference
radius (length R ≥ 1). In this example, node v is not able to receive a transmission from node
u if node x concurrently transmits data to node w—even though v is not adjacent to x.
protocols (e.g., coloring algorithms [15]) have been proposed to reduce interference by
ensuring a certain hop distance between two senders. Concretely, it is assumed that
only the k-neighborhood of a receiver u can interfere with u. Clearly, this is a stark
simplification since in a UDG a (k + 1)-neighbor can be close to the receiver (see
Figure 4.7).
Model 4.3.5 (UDG with Hop Interference (UHI)). Nodes are located at arbitrary
positions in R2 . Two nodes are adjacent if and only if their Euclidean distance is at
most 1. Two nodes can communicate directly if and only if they are adjacent, and
Vk+1
Vk+2
1 +ε
V1
V2
Figure 4.7. Example where UHI fails: Nodes v1 and vk+2 are separated by a path of k + 1
hops, but are close (distance 1 + ǫ). Thus, concurrent transmissions of nodes v2 and vk+2 may
interfere at v1 in spite of their large hop distance.
88
MODELING SENSOR NETWORKS
if there is no concurrent sender in the k-hop neighborhood of the receiver (in the
UDG).
Observe that while the UHI model—for every k—sometimes overlooks interference terms which the UDI would take into account, the contrary does not hold.
Theorem 4.3.1. By choosing R = k, and since a hop has at most length 1, the UDI
model does not overlook any interference terms that UHI would have taken into
account. The contrary does not hold (cf. Figure 4.7).
Like UDI and UHI, also the protocol model (PM) is of type ONE (Model 4.3.3).
However, the senders in the PM model adapt their transmission power according to
DIST (Model 4.3.2)—that is, depending on the distance between sender and receiver.
Model 4.3.7 is a variation of the model introduced in reference 11.
Model 4.3.6 (Protocol Model (PM)). Let u1 , u2 , ..., uk , be the set of nodes transmitting simultaneously to receivers v1 , v2 , ..., vk , respectively. The transmission of
ui is successfully received by vi if for all j =
/ i, it holds that d(uj , vi ) > λ · d(uj , vj ),
where λ ≥ 1 is a given constant. That is, vi must not fall into a “guard zone” around
any sender uj which is a factor (1 + λ) larger than uj ’s transmission range.
Many interference models distinguish between senders and receivers assuming
that interference arises at senders and occurs at receivers. However, often receivers
acknowledge messages and are therefore also senders. If the original messages are
short (e.g., control messages), then the sender/receiver distinction may not make sense.
By this observation, some models (e.g., reference 16) simply consider the interference
of undirected links. Figure 4.8 depicts an example.
Model 4.3.7 (Direction). DIR: This class of interference models distinguishes between senders and receivers (interference disks around senders). UNDIR: Interference
originates from undirected links (interference “pretzels” around links).
Figure 4.8. DIR vs. UNDIR: On the left, only the sender transmits data (interference disks
around senders). On the right, there is no distinction between sender and receiver, and hence
interference arises from the entire link (“pretzels” around links).
INTERFERENCE ISSUES IN WIRELESS SENSOR NETWORKS
89
As in the case of connectivity models, the SINR, the UDI, and the UHI models
can be extended with directional antennas and link failures, and hence Models
4.2.14 and 4.2.15 also apply here. Moreover, also the idea of quasi unit disk graphs
(cf. Model 4.2.3) could be adopted. For example, the UDI can be “quasified” as follows: If two nodes are closer than a given threshold R1 , concurrent transmissions will
always interfere; if the distance is larger than a second threshold R2 , there will be
no interference. Finally, if the distance is between R1 and R2 , transmissions may or
may not interfere. However, these models are often too complicated to be handled
algorithmically. It is sometimes simpler to study general weighted interference graphs
instead. That is, similar to connectivity graphs, the interference model is based on
graphs; however, the edges are now weighted. Formally, in a weighted interference
graph H = G(V, E, w), V represents the set of sensor nodes, E represents the set of
edges, and w : E → R+ is a function assigning a positive value to each edge. The
weight denotes how large the interference between the corresponding nodes actually
is. As in the SINR model, a transmission is received correctly if the ratio between received signal power and the amount (either the sum or the maximal interfering signal
strength) of interfering traffic is smaller than a certain threshold.
Model 4.3.8 (General Weighted Graph (GWG)). A weighted interference graph
H is given. A receiver v successfully receives a message from a sender u, if and
only if the received signal strength (the weight of the link between u and v in H)
divided by the total interference (the sum or the maximum of the weights of the links
of concurrently transmitting nodes with a receiver v in H) is above the threshold given
by the signal-to-interference-plus-noise ratio.
The general weighted graph model is quite pessimistic, because it allows for nonnatural network topologies. Again—like in the BIG connectivity model—we need
a weighted graph model that captures the geometric constraints without making too
many simplifying assumptions. Again, one approach is to assume that the nodes form
a doubling metric (cf. UBG model of Section 4.2).
Model 4.3.9 (Doubling Metric (DM)). The DM model assumes that the nodes form
a doubling metric; that is, the set of nodes at a distance (which is now given by the
weights of the edges) of at most r from a node u can be covered by a constant number
of balls of radius r/2 around other nodes, for any r (cf. Model 4.2.9). Interference
can be incorporated in various ways. For example, the amount of interference at a
receiver u could depend on u’s distance (in the doubling metric space) to the closest
concurrently transmitting node (ONE model), or on the number of concurrent senders
(SUM model).
As a final remark, note that so far we have only presented binary interference
models: A message can be received either correctly or not at all. In practice, however,
also the transfer rate at which messages can be transmitted can depend on interference:
The larger the signal-to-noise ratio, the larger the available bandwidth. A WLAN
802.11, for example, exploits environments with less interference in order to transmit
90
MODELING SENSOR NETWORKS
Communication Model
Special Case of
Extensible with
Interference
Connectivity
GWG
ONE
SUM
GG
CONST
DIST
UBG
UDI
PM
QUDG
UHI
BIG
Other Metrics
DM
Antennas
Link Failures
SINR
“Quasi” Semantic
Other Metrics
Antennas
UDG
Figure 4.9. Overview of connectivity and interference models presented in Sections 4.2 and
4.3. The arrows show how the models are related.
more data per time unit. Of course, it would be possible to extend, for example, the
DM model to capture this aspect as well. However, since these issues are beyond the
scope of this chapter, we refer the reader to reference 17 for more details. To conclude,
Figure 4.9 summarizes the connectivity and the interference models.
4.4 ALGORITHM DESIGN
The main purpose of deploying sensor networks is the collection physical data such
as light intensity, sound, or temperature. In order to aggregate (e.g., compute the
minimum temperature, or the average, etc.) the data that are stored at the individual
nodes—and which are therefore distributed in space!— protocols or algorithms are
needed specifying how these operations are performed. For example, due to the limited
radio communication range, sensor nodes have to communicate (e.g., gather data) in
a multihop manner with each other—that is, the messages have to be relayed by
intermediate nodes—and hence a routing algorithm has to define which messages are
to be forwarded via which other nodes.
Algorithms for sensor networks come in different flavors. In the following, we
first describe the different types of algorithmic models appearing in literature today.
We then discuss modeling aspects that may influence an algorithm’s performance—
for instance, what kind of identifiers nodes have, or how the nodes are distributed in
space. Besides the classic evaluation criteria for algorithms—namely, time complexity
and space complexity—algorithms for sensor networks pose additional optimization
problems; for example, the number of messages that are sent should be small; or, in
order to maximize the lifetime of the network, the nodes’ energy consumption must
be minimized. In order to facilitate a better understanding of the different algorithm
types presented in the upcoming paragraphs, we will consider a sample problem: the
computation of dominating sets.
Definition 4.4.1 (Dominating Set Problem (DS)).. The computation of a dominating set (DS) is a fundamental operation in sensor networks. For instance, such a set
ALGORITHM DESIGN
91
Figure 4.10. A minimum dominating set with two dominators.
can be used to build node clusters. Moreover, it may serve as a basis for constructing
backbone networks that typically form a connected DS. A dominating set D⊂V of a
(undirected) network graph G = (V, E) is a set of nodes such that for all nodes u ∈ V
it holds that either u is in the dominating set itself (i.e., u ∈ D), or u is adjacent to
a node v in D (i.e., {u, v} ∈ E ∧ v ∈ D). It is often desirable to have small dominating sets. The minimum dominating set (MDS) is defined as the dominating set that
minimizes the number of dominators |D|. An example of an MDS is illustrated in
Figure 4.10. It can be shown that the MDS problem is NP-hard on general graphs and
that a logarithmic approximation is asymptotically optimal unless P ≈ NP [18]. For
simpler connectivity graphs such as the UDG graph, the approximation complexity
of the problem may be better; for example, there is a polynomial time approximation
scheme (PTAS) for UDG graphs!
The first category of algorithms we present here is similar to the classic (graph)
algorithms appearing in the field of theoretical computer science or applied mathematics. These global algorithms can operate directly on the entire network or graph and
can have complete information about the state of the system. For example, a system
designer planning a fixed sensor network can apply a global algorithm to determine
the optimal positions of the nodes in a given observation area.
Model 4.4.2 (Global Algorithms). A global algorithm can operate directly on the
entire network.
Kruskal’s algorithm for computing a minimum spanning tree [19] is an example of
a global algorithm: The algorithm receives the entire graph as input and can sort the
edges according to their weights. Kruskal’s algorithm thus has a complete visibility
of the entire graph and can perform arbitrary operations on it. No messages have to
be sent between nodes.
Example 4.4.3. Let us tackle our dominating set problem! When faced with the task
of designing an algorithm for a certain problem, it is often a good idea to start by
studying greedy algorithms—that is, algorithms that in every execution step “greedily” do the currently most promising thing. Interestingly, a greedy algorithm is often
92
MODELING SENSOR NETWORKS
optimal (see also Matroid theory [19]). So how can we greedily compute an MDS? A
straightforward approach is the following (see Algorithm 1): First, we initialize the
set of dominators with the empty set, that is, D := {}. We will call nodes in D black
(“dominators”), nodes that are covered by nodes in D gray (“dominated nodes”), and
all uncovered nodes white. Let w(v) be the number of white nodes adjacent to v,
including v itself. Then, in every step, we iterate over all nodes v (global operation!)
and compute the number w(v) of v’s white neighbors, remembering the node x having
the largest number. At the end of each step, we add node x to D. That is, we choose the
node to become a dominator that covers the most new nodes, greedily reducing the
number of the remaining nodes as much as possible. Obviously, the resulting dominators indeed form a DS. Moreover, it can be proven that the number of dominators is at
most a logarithmic factor larger than in the optimal case. This simple approximation
algorithm is therefore asymptotically optimal unless P ≈ NP!
ALGORITHM 1. Global and Greedy MDS Algorithm
1: D := {};
2: while ∃ white nodes do
3: x := {x|w(x) = maxv {w(v)}};
4: D := D ∪ x;
5: od;
However, unlike global algorithms, most algorithms for sensor networks are not
executed by a central designer, but rather by the sensor nodes themselves, for example,
during the system’s operation. A node a priori only knows its own state. In order to
learn more about the other nodes in the network, it is bound to communicate with
its neighbors by exchanging messages. Typically, when a node receives a message, it
performs some computation, and—depending on the computation’s results—sends a
new message to its neighbors. By this collaboration of the nodes, global operations
such as (multihop) routing between two nodes can be performed. Since the activity is
distributed among the nodes, these algorithms are called distributed algorithms [20].
Model 4.4.4 (Distributed Algorithms). In a distributed algorithm, every sensor
node runs its own algorithm. A priori, a node only knows the state of itself. In order to
learn more about the rest of the network, nodes repeatedly exchange messages with
adjacent nodes.
Thus, unlike in global algorithms, distributed algorithms do not see the entire graph
at the beginning. If more information about the neighborhood is needed, adjacent
nodes have to exchange messages. Distributed algorithms raise many interesting questions. For example: What can be computed in a distributed fashion, and what not?
In contrast to global algorithms, nodes have to be coordinated somehow, and—as
all nodes execute the same code—symmetries have to be broken. Besides an algorithm’s correctness, execution time to perform the task (time complexity), and
ALGORITHM DESIGN
93
memory requirements at the nodes (space complexity), a new criterion becomes
important, namely message complexity: Since distributed algorithms rely on message passing and since sending and receiving messages is an expensive operation
(e.g., energy consumption, queueing delay, congestion, etc.), a distributed algorithm
should minimize the total number of messages sent.
Example 4.4.5. In the following, we will see how dominating sets can be computed
with distributed algorithms. Consider the following idea: Each sensor node broadcasts
its own identifier plus the identifiers of all its neighbors to all other nodes in the
network. Hence, each node gets a picture of the entire connectivity graph. Then,
the node having the largest identifier computes the MDS using the global algorithm
described in Algorithm 1, and it broadcasts the corresponding solution back to all
nodes. Given this solution, each node can then decide whether it has to join the
dominating set or not. Note that this algorithm is indeed distributed and yields small
dominating sets: The size of the sets are again asymptotically optimal, unless P ≈ NP.
However, due to the broadcast operation, the message complexity is huge. What is
more, the algorithm does not scale to a large number of sensor nodes.
Many global algorithms can be turned into distributed algorithms by just collecting
the entire graph at each node and then computing the results locally. Of course, this is
inefficient and inappropriate in practice. Therefore, it is often better to study a more
restricted class of distributed algorithms, namely localized algorithms [21].
Model 4.4.6 (Localized Algorithms). A localized algorithm is a special case of a
distributed algorithm. At the beginning, a node has only information about its own
state. In order to learn more about the rest of the network, messages have to be
exchanged. In a k-localized algorithm, for some constant k, each node is allowed to
communicate at most k times with its neighbors. However, a node can decide to retard
its right to communicate; for example, a node can wait to send messages until all its
neighbors having larger identifiers have reached a certain state of their execution.
Localized algorithms are desirable in the sense that they only transmit a small
number of messages. Unfortunately, however, localized algorithms can be slow: A
node u might have to wait for a neighbor v to transmit all its messages, while node v
in turn has to wait for its neighbor w, and so on. As a matter of fact, there can be a
linear chain of causality, with only one node being active at any time. This yields a
worst-case execution time of (n), where n is the number of nodes.
Example 4.4.7. In order to compute the dominating sets of our sample problem in a
localized manner, a simple algorithm can be applied (cf. Algorithm 2). Each node v
waits until all its neighbors having a larger degree (or, in the case of the same degree; a
larger identifier) than v have decided whether to join the dominating set or not. Then, if
one of these nodes is a dominator, v decides not to join the dominating set. Otherwise,
v becomes a dominator. Thus, each node has to communicate at most twice with its
neighbors: once to find out their degree and once to tell them about its decision. This
94
MODELING SENSOR NETWORKS
1
n
2
1
n
3
n-1
2
3
n-1
4
12
5
11
6
7
10
9
8
4
12
5
11
6
10
9
8
7
Figure 4.11. Localized algorithms can have large execution times!
localized algorithm has therefore a low message complexity. However, the execution
time of this algorithm can be large. To see this, consider a cycle of nodes arranged
according to their identifiers as depicted in Figure 4.11. Since all nodes have the same
degree, the first node to become a dominator is node n. Only then, node n−1 can
make its decision: It does not join the dominating set. After that, node n−2 decides
to join the network, and so on. It’s only after a linear waiting time of O(n) time steps
that node 1 eventually can make its decision and terminate the algorithm!
ALGORITHM 2. Localized MDS Algorithm
1:
2:
3:
4:
5:
6:
7:
8:
9:
10:
11:
12:
13:
14:
(* Code executed by node v *)
send degree and ID to all neighbors;
receive messages from neighbors;
while (∃ undecided neighbor w with prio(w) > prio(v)) do
wait();
od;
(* Decision *)
if (∃ dominator in neighborhood) then
D := D;
send “I am dominated!” to neighbors;
else
D := D ∪ v;
send “I am a dominator!” to neighbors;
fi;
Researchers have proposed to study yet another kind of distributed algorithm that
overcomes the performance problems of localized algorithms, always terminating
after a constant number of communication rounds [2].
Model 4.4.8 (Local Algorithms). Again, at the beginning, each node only knows its
own state. In a k-local algorithm, for some constant k, each node can communicate at
most k times with its neighbors. However, in contrast to k-localized algorithms, nodes
cannot delay their decisions. In particular, all nodes process k synchronized phases,
ALGORITHM DESIGN
9
95
2
6
8
5
4
1
7
3
Figure 4.12. Illustration of local dominating set algorithm: Since all nodes with larger IDs
are connected to each other and cover all other neighbors of node 5, node 5 does not join the
dominating set—regardless of the decisions of other nodes.
and a node’s operations in phase i may only depend on the information received during
phases 1 to i − 1. The most efficient local algorithms are often randomized [22, 23],
that is, the number of rounds k can vary.
Observe that in a k-local algorithm, nodes can only gather information about nodes
in their k-neighborhood. In some local algorithms [22], the algorithm designer can
choose an arbitrarily small constant k (at the cost of a lesser approximation ratio). This
makes local algorithms particularly suited in scenarios where the nodes’ environment
changes frequently, because they are able to constantly adapt to the new circumstances.
Example 4.4.9. A dominating set can also be computed with a local algorithm: Each
node u asks its higher-priority neighbors (with respect to degrees and identifiers) about
their neighbors. If these higher-priority neighbors are connected and cover all of u’s
neighbors, then u does not join the dominating set and otherwise becomes a dominator. An example is illustrated in Figure 4.12. It can be shown that this algorithm even
results in a connected dominating set—that is, a dominating set where any two dominators are connected by a path that only consists of other dominators. Observe that the
algorithm is indeed “wait-free” or local, because a node can make its decisions only
based on the identifiers of its neighbors and independently of the neighbors’ decisions.
Two communication rounds are sufficient. From this point of view, this local algorithm
looks very appealing. Unfortunately, however,
in the worst case, its approximation
√
ratio of the optimal solution is as bad as ( n) already for simple connectivity graphs
such as the UDG.2 This indicates the existence of a challenging tradeoff: The smaller
the “horizon” of a local algorithm, the more difficult it is to find good approximations
of the optimal (global) solutions. In other words, there seems to be price of being
2 For
random UDGs, the performance is better: The algorithm achieves a constant approximation!
96
MODELING SENSOR NETWORKS
near-sighted: Similarly to online algorithms that cannot foresee the future and have
to be competitive to an optimal offline algorithm, local algorithms have a restricted
view of their neighborhood and are measured against the performance of a global
algorithm.
Note that due to the synchronous phases, local algorithms may make greater
demands on the media access sublayer than localized algorithms. In particular, in
unreliable wireless networks it seems to be costly to implement a media access
control scheme that allows for synchronous rounds, because messages will be lost
due to interference (conflicting concurrent transmissions) or mobility (even if the
nodes themselves are not mobile, the environment is typically dynamic, temporarily enabling/disabling links). A powerful concept for coping with failures is selfstabilization [24]. Fortunately, using a simple trick [25], every local algorithm is
immediately self-stabilizing. The trick works as follows (Section 4 of reference 25):
Every node keeps a log of every state transition it has taken until its current state;
generally, this boils down to memorizing the local variables of each step of the main
loop. If each node constantly sends its current log to all neighbor nodes, each node
can check and correct every transition it has made in the past. Assuming that all
inputs are correct (variable initialization and random seeds are stored in the imperishable program memory, and sensor information can be rechecked), every fault due
to memory or message corruption will be detected and corrected. For details we refer
to reference 25. Turning a k-local algorithm into a self-stabilizing algorithm with
reference 25 blows up messages by a factor k (in the worst case); on the other hand,
we immediately get an algorithm that works on a sensor network as the hardest wireless problems (messages lost due to interference and mobility) are covered by the
self-stabilization model. Also, in the case of an error (such as a lost message), only
the k-neighborhood of a node is affected.3
Having defined the most common types of algorithm, we now look at some algorithmic aspects in more detail. As mentioned, the message complexity—the total
number of messages sent by an algorithm—is a main evaluation criterion of distributed
algorithms. Because the number of messages typically depends on the amount of information that can be stored in a message, a model must specify the messages’ sizes. A
most popular model limits the message size to O(log n) bits, where n is the total number of nodes in the system. Hence, a message can store only a constant number of node
identifiers (e.g., the source and destination address of a routing packet). Moreover, it
is often assumed that if a node u sends a message to a neighbor v, all other neighbors of u will also receive the message (broadcast model). However, sometimes—for
example, for lower bound proofs [26]—models are considered where the message
size is unbounded, and where nodes can communicate with their neighbors individually (message-passing model). Algorithmic models also differ in their assumptions
about how nodes can access the wireless medium. The concrete MAC, however, can
3 In
principle, localized algorithms can also benefit from reference 25; however, errors are not restricted
to a k-neighborhood but may propagate through the entire network, resulting in a troublesome butterfly
effect.
ALGORITHM DESIGN
97
influence the number of retransmissions and hence an algorithm’s performance. Moreover, an algorithm may be able to coordinate the medium access itself.
Model 4.4.10 (Medium Access). Some researchers assume an ideal medium access
mechanism [27] where interference is impossible and where messages will always
be broadcast instantaneously to all neighbors (cf. models of Section 4.2). In addition,
adversarial models are used where an adversary schedules transmissions. Of course,
this model only makes sense if the adversary is restricted appropriately—that is, if
there are fairness guarantees. For example, the adversary might have to schedule
each node at least once every (n) rounds. One could also imagine an adversary that
delivers a message only to a subset of a node’s neighbors, because the other neighbors
experience collisions. Finally, completely unstructured radio networks [28] can be
considered where the algorithm designer has to implement her own medium access
scheme from scratch. These models can further be classified in terms of whether
collisions can be detected by a receiver or not.
As mentioned, a main objective of sensor networks is to collect physical data
distributed over a given region. To achieve this, typically one or more nodes observe
different sub-areas. Knowledge of the nodes’ distribution, however, can be important
for an algorithm designer. In a scenario where the nodes are dropped from an airplane,
one might expect that the nodes are roughly randomly distributed when they reach
ground.
Model 4.4.11 (Random Node Distribution). The simplest—and quite common—
way to model sensor networks is to assume a UDG in combination with a uniform node
distribution in the two-dimensional Euclidean plane. However, inspired by percolation
theory, also Poisson models have been proposed [29]; thus, the positions of the nodes
are distributed in R2 according to a homogeneous Poisson point process of constant
density λ per unit area.
While these random models may be fine to prove the performance of an algorithm, for
correctness and robustness issues, a more pessimistic model should be preferred—for
example, a worst-case distribution.
Model 4.4.12 (Worst-Case Node Distribution). Nodes are distributed arbitrarily in
the space given by the underlying graph (e.g., Euclidean plane, general graph, etc.).
Of course, there are again many models that lie between the two extremes. For example, random distributions with a density parameter varying over space could be
considered: One can imagine that there are several nodes per square meter in areas
that are “interesting” to observe, whereas in other “routing only” areas the nodes
are hundreds of meters apart. Finally, speaking of node distributions, there are also
models that do not allow nodes to be arbitrarily close or even assume the same position; for instance, there is such an assumption in the (1) model [30] or in so-called
civilized graphs [5]. Related to the distribution of nodes in space is also the issue of
98
MODELING SENSOR NETWORKS
the distribution of node identifiers. Because many algorithms are based on node IDs,
their performance can depend on how IDs are distributed among the nodes (and thus
also in space).
Model 4.4.13 (Node Identifiers). Typically, nodes can be assumed to have unique
identifiers. IDs could, for example, be generated during deployment using a random
number generator. Moreover, because RFID tags already have IDs, we believe that it
is reasonable to assume that sensor nodes obtain a unique ID during the production
process. Finally, also note that certain tasks cannot be solved by any distributed
algorithm if there are no identifiers, because there is no way to break symmetries
among the nodes. Similarly to the node distribution in space, the most common
models for ID distributions are random distributions and worst-case distributions.
Sometimes, it also matters from which range the identifiers are chosen. Again, many
variations are possible. For example, each of the n nodes can have a unique 128-bit
identifier (range 0, ..., 2128 − 1). Or, in a more restrictive case, the nodes may have
consecutive numbers (e.g., range 1, ..., n).
Alternatively—or in addition!—node IDs can contain location information—for
example, if the nodes are equipped with a Global Positioning System (GPS) or a
Galileo device. Location information can boost the performance of certain operations
[30]: for example, a routing algorithm can exploit geographic information to forward
the message to a neighbor which lies in the direction of the message’s destination
(greedy routing).
Model 4.4.14 (Location Information). Sensor nodes can have access to various
forms of (absolute or relative) geographic information about other nodes. For example, a node u might sense its distance to another node v, or sense in which direction
(angle of arrival) u lies, or even know v’s exact position.
Distributed algorithms for sensor networks are usually evaluated with respect to
their time complexity, their space complexity, and their message complexity. However, in order to be successful in a real sensor network, an algorithm has to pursue
additional objectives. For instance, if sensor nodes are deployed in large numbers,
recharging their batteries seems out of question, in particular in adversarial territory.
A node’s energy supply must suffice for the whole operational phase. Therefore, the
conservation of energy is of utmost importance. Basically, there are two approaches to
capture the energy consumption of a node. Historically, since during the transmission
of data much energy is consumed, a model has been studied which only takes the
transmission energy into account [31].
Model 4.4.15 (Transmission Energy). The energy consumed by a node is calculated
by the sum over all its transmissions. Thus, the energy needed to transmit one message
is of the form c · d α , where d is the distance between sender and receiver, α is the
path-loss exponent (usually α > 2), and c is a constant.
FINAL REMARKS
99
Although transmitting data is a costly operation, sensor nodes with short-range
radios available today spend as much energy receiving or waiting for data. Therefore,
techniques have been developed which allow nodes to change to a parsimonious sleep
mode [32]. During the time periods a node is sleeping, it cannot receive any data. The
idea is that if all nodes can somehow be synchronized to wake up at the same moment
of time to exchange data (e.g., every minute), much energy is saved. This motivates
the following model.
Model 4.4.16 (Sleeping Time). The energy consumed by a node is given by the
accumulated time in which it is not in sleep mode.
If there are no external disturbances, a node is assumed to live as long as it has some
energy left. The lifetime of the entire network is modeled in different ways.
Model 4.4.17 (Network Lifetime). In applications that depend on every single node,
the lifetime of a network can be defined as the time until the first node runs out of
battery power [33]. Alternatively, a network might be able to tolerate certain node
failures; for example, the network might live as long as all live nodes are still connected
to each other.
4.5 FINAL REMARKS
This chapter has given an overview and discussion of many sensor network models in
use today. It has been shown how the models are related to each other. Therefore, we
have assumed an algorithmic point of view and have concentrated on models of higher
levels of abstraction. Of course, it does not make sense to argue about which model is
“better” and which is “worse.” For example, a large warehouse has different physical
characteristics and signal propagation paths than an office building; or GPS might not
work indoors and hence algorithms based on coordinates are not be feasible; and so
on. A good model also depends on the question studied. A media access study might
need a detailed model capturing several low-level aspects; for example, it has to be
taken into account that a message might not be received correctly due to a nearby
concurrent transmission. Hence, it is crucial that the model appropriately incorporates
interference aspects. For a transport layer study, however, a much simpler model that
assumes random transmission errors might be sufficient. This chapter helps to compare
the different options.
Clearly, it is always desirable to have algorithms for sensor networks which can be
proved correct in the most general possible model that covers all possible characteristics of a real environment. Only then can we be sure that the algorithm will actually
work in practice. However, for efficiency considerations, a more idealistic model that
does not yield overly conservative results might be fine. Moreover, we believe that
when developing algorithms for sensor networks, it is often useful to study idealistic
models first, because these models are simpler and may provide helpful insights into
the given problem. After having found algorithms for these models, it is still possible
to tackle the more general cases.
100
MODELING SENSOR NETWORKS
4.6 FUTURE RESEARCH DIRECTIONS
Models for sensor networks have developed quickly and are now much more sophisticated than they were some years ago. However, we believe that the quest for new
models is still of prime importance. In particular, with the wider deployment of sensor
networks the experiences with complex issues such as connectivity and interference
will increase; consequently, models will evolve as well. Interestingly, many problems
are still unexplored for the models presented in this chapter. For many models, there
exist no algorithms with provable performance for fundamental operations such as
the computation of dominating sets. For example: How well can the minimum dominating set be approximated in bounded independence connectivity graphs (BIG)?
Such questions are exciting because they are interdisciplinary and require knowledge
of various mathematical fields. We want to encourage the more advanced readers to
address these problems!
4.7 EXERCISES
The following exercises are based on this chapter only, but sometimes require some
mathematical background.
1. Name some scenarios where the QUDG model is not appropriate. Which alternative model might capture the situation better? Under what circumstances
may it still be useful to study the QUDG?
2. Prove that both the UDG and the QUDG have a bounded independence (i.e.,
that they are a BIG). Hint: The proof is similar to the proof of Theorem 4.2.13.
3. (a) Prove that the two-dimensional Euclidean plane has a constant doubling
dimension.
(b) Formally prove that not every general graph has a doubling dimension.
4. A basic operation in sensor networks is the distributed computation of a (minimum) connected dominating set (CDS). A CDS is a dominating set with the
additional requirement that any two dominators are connected to each other by
a path that only consists of other dominators. For instance, such a CDS can be
useful to establish a routing backbone network.
(a) Can you come up with a local algorithm which computes a CDS in a UDG if
all nodes are equipped with a GPS device (i.e., if they know their position)?
(b) Does your algorithm yield the optimal solution, or just an approximation?
Can you prove its quality?
(c) How efficient is your algorithm, in terms of number of messages transmitted
and in terms of communication rounds needed?
5. The connected dominating set problem is well-studied.
(a) Surf the web to find the currently best-known algorithm for the UDG.
BIBLIOGRAPHY
101
(b) Can you find an algorithm that works on general connectivity graphs and
that does not require nodes to have position information?
(c) Compare the complexity and efficiency of the best-known UDG algorithm
to the best-known GG algorithm: What is the “price” of the more pessimistic
model?
6. Is the UBG model equivalent to the BIG model?
BIBLIOGRAPHY
1. B. N. Clark, C. J. Colbourn, and D. S. Johnson. Unit disk graphs. Discrete Mathematics,
86:165–177, 1990.
2. F. Kuhn and R. Wattenhofer. Constant-time distributed dominating set approximation.
In Proceedings of the 22nd Annual Symposium on Principles of Distributed Computing
(PODC), 2003, pp. 25–32.
3. L. Barrière, P. Fraigniaud, and L. Narayanan. Robust position-based routing in wireless ad
hoc networks with unstable transmission ranges. In Proceedings of the 5th International
Workshop on Discrete Algorithms and Methods for Mobile Computing and Communications (DIALM), 2001, pp. 19–27.
4. F. Kuhn, R. Wattenhofer, and A. Zollinger. Ad-hoc networks beyond unit disk graphs.
In Proceedings of the 1st ACM Joint Workshop on Foundations of Mobile Computing
(DIALM-POMC), 2003.
5. S. O. Krumke, M. V. Marathe, and S. S. Ravi. Models and approximation algorithms for
channel assignment in radio networks. Wireless Networks, 7(6):575–584, 2001.
6. I. Stojmenović, A. Nayak, and J. Kuruvila. Design guidelines for routing protcols in ad
hoc and sensor networks with a realistic physical layer. IEEE Communications Magazine,
6:36–42, 2005.
7. D. Ganesan, D. Estrin, A. Woo, D. Culler, B. Krishnamachari, and S. Wicker. Complex
behavior at scale: An experimental study of low-power wireless sensor networks. Technical
Report, UCLA Computer Science Technical Report UCLA/CSD-TR 02-0013, 2002.
8. F. Kuhn, T. Nieberg, T. Moscibroda, and R. Wattenhofer. Local approximation schemes
for ad-hoc and sensor networks. In Proceedings of the Joint Workshop on Foundations of
Mobile Computing (DIALM-POMC), 2005, pp. 97–103.
9. F. Kuhn, T. Moscibroda, and R. Wattenhofer. On the locality of bounded growth. In Proceedings of the 24th ACM Symposium on Principles of Distributed Computing (PODC),
2005, pp. 60–68.
10. A. Gupta, R. Krauthgamer, and J. R. Lee. Bounded geometries, fractals, and low-distortion
embeddings. In Proceedings of the 44th Annual IEEE Symposium on Foundations of Computer Science (FOCS), 2003, p. 534.
11. P. Gupta and P. R. Kumar. The capacity of wireless networks. Transactions of Information
Theory, 46(2):388–404, 2000.
12. T. Moscibroda and R. Wattenhofer. The complexity of connectivity in wireless networks.
In Proceedings of the IEEE Infocom, 2006. CD-ROM.
13. T. Rappaport. Wireless communications: Principles and practices,” Prentice Hall, Englewood Cliffs, NJ, 1996.
102
MODELING SENSOR NETWORKS
14. D. Bliò and G. Proietti. On the complexity of minimizing interference in adhoc and sensor
networks. Technical Report, Dipartimento di Informatica TRCS 009/1006, 2006.
15. T. Moscibroda and R. Wattenhofer. Coloring unstructured radio networks. In Proceedings of
the 17th Annual ACM Symposium on Parallelism in Algorithms and Architectures (SPAA),
2005, pp. 39–48.
16. F. Meyer auf de Heide, C. Schindelhauer, K. Volbert, and M. Grünewald. Energy, congestion
and dilation in radio networks. In Proceedings of the 14th Annual ACM Symposium on
Parallel Algorithms and Architectures (SPAA), 2002, pp. 230–237.
17. A. El Fawal, J.-Y. Le Boudec, Ruben Merz, B. Radunovic, J. Widmer, and G. M. Maggio.
Trade-off analysis of PHY-aware MAC in low-rate low-power UWB networks. IEEE Communications Magazine, 43(12):147, 2005.
18. R. Raz and S. Safra. A sub-constant error-probability low-degree test, and sub-constant
error-probability PCP characterization of NP. In Proceedings of the 29th Annual ACM
Symposium on Theory of Computing (STOC), 1997, pp. 475–484.
19. T. H. Cormen, C. E. Leiserson, R. L. Rivest, and C. Stein, Introduction to Algorithms, MIT
Press, Cambridge, MA, 2001.
20. D. Peleg, Distributed Computing: A Locality-sensitive Approach, Society for Industrial
and Applied Mathematics, Philadelphia, PA, 2000.
21. Y. Wang, X.-Yang Li, P.-J. Wan, and O. Frieder. Sparse power efficient topology for wireless
networks. Journal of Parallel and Distributed Computing, 2002.
22. F. Kuhn, T. Moscibroda, and R. Wattenhofer. The price of being near-sighted. In Proceedings of the ACM-SIAM Symposium on Discrete Algorithms (SODA), 2006.
23. M. Luby. “A simple parallel algorithm for the maximal independent set problem. SIAM
Journal of Computing, 15(4):1036–1053, 1986.
24. E. W. Dijkstra. Self-stabilizing systems in spite of distributed control. Communications of
the ACM, 17(11):643–644, 1974.
25. B. Awerbuch and G. Varghese. Distributed program checking: A paradigm for building
self-stabilizing distributed protocols. In Proceedings of the 32nd Annual IEEE Symposium
on Foundations of Computer Science (FOCS), 1991. CD-ROM.
26. F. Kuhn, T. Moscibroda, and R. Wattenhofer. What cannot be computed locally!. In Proceedings of the 23rd ACM Symposium on the Principles of Distributed Computing (PODC),
2004. CD-ROM.
27. P.-J. Wan, K. M. Alzoubi, and O. Frieder. Distributed construction of connected dominating
set in wireless ad hoc networks. Mobile Networks and Applications, 9(2):141–149, 2004.
28. T. Moscibroda, P. von Rickenbach, and R. Wattenhofer. Analyzing the energy-latency
trade-off during the deployment of sensor networks. In Proceedings of the IEEE Infocom,
2006. CD-ROM.
29. O. Dousse, P. Thiran, and M. Hasler, Connectivity in ad-hoc and hybrid networks. In
Proceedings of the IEEE Infocom, 2002. CD-ROM.
30. F. Kuhn, R. Wattenhofer, and A. Zollinger. Worst-case optimal and average-case efficient
geometric ad-hoc routing. In Proceedings of the 4th ACM International Symposium on
Mobile Ad Hoc Networking & Computing (MobiHoc), 2003, pp. 267–278.
31. A. E. F. Clementi, G. Huiban, and P. Penna. On the approximation ratio of the MST-based
heuristic for the energy-efficient broadcast problem in static ad-hoc radio networks. In
BIBLIOGRAPHY
103
Proceedings of the 17th International Symposium on Parallel and Distributed Processing
(IPDPS), Washington, DC, 2003. CD-ROM.
32. V. Raghunathan, C. Schurgers, S. Park, and M. Srivastava. Energy aware wireless microsensor networks. IEEE Signal Processing Magazine, 19(2):40–50, 2002.
33. J.-H. Chang and L. Tassiulas. Energy conserving routing in wireless ad-hoc networks. In
Proceedings of the IEEE Infocom, 2000, pp. 22–31.
CHAPTER 5
Spatiotemporal Correlation Theory
for Wireless Sensor Networks
ÖZGÜR B. AKAN
Department of Electrical and Electronics Engineering, Middle East Technical University,
Ankara, Turkey 06531
5.1 INTRODUCTION
Wireless sensor networks (WSNs) are generally composed of densely deployed sensor nodes that collaboratively observe and communicate their readings of a certain
physical phenomenon [1]. In general, the main objective of the WSN is to reliably
detect/estimate event features from the collective information provided by sensor
nodes. Therefore, the energy and hence processing constraints of small wireless sensor nodes are overcome by this collective sensing notion that is realized via their
networked deployment.
To this end, accurate and efficient operation of any WSN deployment requires
that we maintain sufficient network and sensing coverage in the deployment field.
To assure network and sensing coverage, WSN applications require sensor nodes
to be densely deployed in the field. Dense deployment of sensor nodes makes the
sensor observations highly correlated in space domain. Similarly, in periodic sensing
applications, all consecutive sensor readings are temporally correlated.
While the collaborative nature of the WSN brings significant advantages over
traditional sensing including greater accuracy, larger coverage area, and extraction
of localized features, the spatiotemporal correlation among the sensor observations
is another significant and unique characteristic of the WSN which can be exploited
to drastically enhance the overall network performance. In general, and based on the
application, the physical phenomenon to be observed can be modeled as point source
(e.g., target detection/tracking) or field source (e.g., monitoring of magnetic field and
seismic activities) [2]. Events generating a signal that originates from a single point in
the field can be modeled as a point source. The cases where the physical phenomenon
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
105
106
SPATIOTEMPORAL CORRELATION THEORY FOR WIRELESS SENSOR NETWORKS
is dispersed over the field can be modeled as field source. The characteristics of the
correlation in the WSN can be summarized as follows [3]:
r Spatial Correlation. Typical WSN applications require spatially dense sensor
deployment in order to achieve satisfactory coverage [4]. As a result, multiple sensors record information about a single event in the sensor field. Due to
high density in the network topology, spatially proximal sensor readings are
correlated with the degree of correlation increasing with decreasing internode
separation.
r Temporal Correlation. Some of the WSN applications such as event tracking may
require sensor nodes to periodically perform observation and transmission of the
sensed event features. The nature of the energy-radiating physical phenomenon
constitutes the temporal correlation between each consecutive observation of
a sensor node [5]. The degree of correlation between consecutive sensor measurements may vary according to the temporal variation characteristics of the
phenomenon.
The existence of the above-mentioned spatial and temporal correlations bring significant potential advantages for the development of efficient communication protocols well-suited for the WSN paradigm. For example, intuitively, due to the spatial
correlation, data from spatially separated sensors is more useful to the sink than highly
correlated data from nodes in proximity. Therefore, it may not be necessary for every
sensor node to transmit its data to the sink. Instead, a smaller number of sensor readings may suffice to communicate the information on the sensed phenomenon to the
sink within a certain reliability/fidelity level. Similarly, for a certain target tracking
application, the measurement reporting frequency, at which the sensor nodes transmit
their observations, can be adjusted such that a temporally correlated phenomenon
signal is captured at the sink within a certain distortion level and with minimum
energy expenditure.
There has been some research effort to study the correlation in WSNs [2, 3, 6–11]
most of which mainly investigate the information and coding theoretical aspects of
the correlation. Clearly, it is of great importance to capture the spatiotemporal correlation characteristics to be able to design energy-efficient communication protocols
that can exploit the potential advantages of correlation in WSNs. In this chapter, a
theoretical analysis of spatiotemporal correlation in WSNs is presented. The objective
of this analysis is to capture the spatiotemporal correlation characteristics of sensor
networks to reveal its potential advantages and possible approaches to exploit correlation in designing efficient communication techniques for WSNs. More specifically, in
Section 5.2, the theoretical framework is developed to model the spatiotemporal correlation in sensor networks. Based on the developed correlation model, the spatial
and temporal correlations are separately captured in Section 5.3.1 and 5.3.2, respectively. In order to understand the joint effects of spatial and temporal correlation, spatiotemporal correlation characteristics of point and field sources are then investigated
in Section 5.3.3. In Section 5.4, based on this analysis, possible approaches are also
COMMUNICATION ARCHITECTURE AND CORRELATION MODEL
107
discussed to exploit spatiotemporal correlation for efficient communication in WSNs.
Finally, the concluding remarks, open research problems, and potential directions are
discussed in Section 5.5.
5.2 COMMUNICATION ARCHITECTURE AND CORRELATION
MODEL FOR WIRELESS SENSOR NETWORKS
In a sensor field, each sensor observes the noisy version of a physical phenomenon.
The sink is interested in observing the physical phenomenon using the observations
from sensor nodes with the highest accuracy. The physical phenomenon in interest
can be modeled as a spatiotemporal process s(t, x, y) as a function of time t and spatial
coordinates (x, y).
Depending on the specific sensor application, the physical phenomenon may be
a spatiotemporal process generated by a point source in case of applications such as
object tracking. In this case, the sink is interested in reconstructing the source signal
at a specific location (x0 , y0 ) based on sensor observations. In other applications, the
spatiotemporal process may be a combination of multiple point sources where the sink
is interested in reconstructing the signal in multiple locations or over an event area.
Although the reconstruction is application specific, the properties of the observations
can be modeled based on the spatiotemporal process s(t, x, y).
The model for the information gathered by N sensors in an event area is illustrated
in Figure 5.1. The sink is interested in estimating the event source, S, according to the
observations of the sensor nodes, ni , in the event area. Each sensor node ni observes
N1
X1[n]
S1[n]
E
Y1[n]
N2
S 2[n]
S
SM[n]
X2[n]
E
Y2 [n]
NM
X M[n]
E
YM[n]
Wireless
Sensor
Network
NN
X N [n]
S N [n]
E
YN [n]
Figure 5.1. Correlation model and architecture.
D
^
S
108
SPATIOTEMPORAL CORRELATION THEORY FOR WIRELESS SENSOR NETWORKS
Xi [n], the noisy version of the event information, Si [n], which is spatially correlated
to the event source, S. In order to communicate this observation to the sink, each node
has to encode its observation. The encoded information, Yi [n], is then sent to the sink
through the sensor network. The sink, at the other end, decodes this information to
get the estimate, Ŝ, of the event source S. The encoders and the decoders are labeled
as E and D, respectively, in Figure 5.1. Using this model, we will exploit various
aspects of correlation among sensor readings in terms of both time and space.
Each observed sample, Xi [n], of sensor ni at time n is represented as
Xi [n] = Si [n] + Ni [n]
(5.1)
where the subscript i denotes the spatial location of node ni [i.e., (xi , yi )], Si [n] is the
realization of the space–time process s(t, x, y) at time t = tn 1 and (x, y) = (xi , yi ),
and Ni [n] is the observation noise. {Ni [n]}n is a sequence of i.i.d Gaussian random
2 . We further assume that the noise each sensor
variables of zero mean and variance σN
node encounters is independent of each other; that is, Ni [n] and Nj [n] are independent
for i =
/ j and ∀n.
As shown in Figure 5.1, each observation Xi [n] is then encoded into Yi [n] by the
source-coding at the sensor node as
Yi [n] = fi (Xi [n])
(5.2)
and then sent through the network to the sink. The sink decodes the received data to
reconstruct an estimation Ŝ of the source S
Ŝ = g(Y1 [n1 ], ..., Y1 [nτ ]; ...; YN [n1 ], ..., YN [nτ ])
(5.3)
based on the data received from N nodes in the event area over a time period τ =
tnτ − tn1 . The sink is interested in reconstructing the source S according to a distortion
constraint
D = E d(S, Ŝ)
(5.4)
In the next sections, the general distortion function in (5.4) will be used to independently obtain the distortion functions for spatial and temporal correlation in the
WSN, which will be further extended to capture the joint spatiotemporal correlation
analysis of point and field sources as well.
1 Note
that we use a discrete-time model since each node is assumed to sample the physical phenomenon
synchronously after the initial wake-up.
SPATIOTEMPORAL CORRELATION IN WIRELESS SENSOR NETWORKS
109
5.3 SPATIOTEMPORAL CORRELATION IN WIRELESS SENSOR
NETWORKS
5.3.1 Spatial Correlation in Wireless Sensor Networks
In this section, based on the communication architecture and the theoretical correlation
model presented in Section 5.2, the spatial correlation between observations of each
sensor node is modeled. The information gathered by N sensors in an event area
can be modeled as shown in Figure 5.1. The sink is assumed to be interested in a
point source S. Since we only consider the spatial correlation between nodes, in this
analysis we assume that the samples are temporally independent. Hence, by dropping
the time index n, (5.1) can be restated as
i = 1, ..., N
Xi = Si + Ni ,
(5.5)
The sink is interested in reconstructing the source S according to observations of
nodes ni which observe the spatially correlated version of S at (xi ,yi ), that is, Si . The
physical phenomenon is modeled as joint Gaussian random variables (JGRVs) at each
observation point as
E{Si } = 0,
i = 1, ..., N
var{Si } = σS2 ,
i = 1, ..., N
cov{Si , Sj } = σS2 corr{Si , Sj }
corr{Si , Sj } = ρi,j = Kϑ (di,j ) =
E[Si Sj ]
σS2
where di,j = si − sj denotes the distance between nodes ni and nj located at
coordinates si and sj , respectively and Kϑ (·) is the correlation model. The covariance
function is assumed to be nonnegative and to decrease monotonically with the distance
d = si − sj , with limiting values of 1 at d = 0 and of 0 at d = ∞. Generally,
covariance models can be classified into four groups [12]:
r Spherical:
KϑS (d)
=
1−
0
3 d
2 θ1
+
1 d 3
2 θ2
if 0 ≤ d ≤ θ1
if d > θ1
;
θ1 > 0
In this model, two observations taken more than θ1 distance apart are uncorrelated.
r Power Exponential:
θ2
KϑPE (d) = e(−d/θ1 ) ;
θ1 > 0, θ2 ∈ (0, 2]
110
SPATIOTEMPORAL CORRELATION THEORY FOR WIRELESS SENSOR NETWORKS
For θ2 = 1 the model becomes exponential, while for θ2 = 2 it becomes squared
exponential.
r Rational Quadratic:
RQ
Kϑ (d)
=
1+
d 2
θ1
−θ2
;
θ1 > 0, θ2 > 0
r Matérn:
KϑM (d) =
d
d θ2
1
;
K
θ
2
2θ2 −1 Ŵ(θ2 ) θ1
θ1
θ1 > 0, θ2 > 0
where Kθ2 (·) is the modified Bessel function of second kind and order θ2 .
The correlation model can be chosen according to the properties of the physical
phenomenon the sink is interested in. Since we are interested in S, which is also a
JGRV, we use a special notation with
var{S} = σS2
corr{S, Si } = ρs,i = Kϑ (ds,i ) =
E[SSi ]
σS2
where ds,i denotes the distance between the source S and the node ni . The observation
noise Ni of each node ni is modeled as i.i.d. Gaussian random variable with zero mean
2 , that is, N ∼ N(0, σ 2 ).
and variance σN
i
N
As each sensor node ni observes an event information Xi , this information is
encoded and then sent to the sink through the WSN. In traditional point-to-point
communication, the optimum performance is obtained by compressing the information according to the source statistics and then adding redundant information to
accommodate the errors introduced in the wireless channel. This technique is known
as the separation principle. In WSNs, where multiple nodes try to send information
about the same event, however, it is known that joint source-channel coding outperforms separate coding [10, 13]. In addition, for Gaussian sources, if the source is
Gaussian and the cost on the channel is the encoding power, then uncoded transmission is optimal for point-to-point transmission [14]. Furthermore, for sensor networks
with finite number of nodes, uncoded transmission outperforms any approach based
on the separation paradigm leading to the optimal solution for infinite number of
nodes [10]. Hence, we adopt uncoded transmission for the sensor observations in this
model. Each node ni sends to the sink, a scaled version, Yi , of the observed sample
Xi according to encoding power constraint PE .
Yi =
PE
X,
2 i
σS2 + σN
i = 1, ..., N
(5.6)
111
SPATIOTEMPORAL CORRELATION IN WIRELESS SENSOR NETWORKS
2 are the variances of the event information S and the observation
where σS2 and σN
i
noise Ni , respectively.
The sink needs to calculate the estimation of each event information, Si , in order to
estimate the event source S. Since uncoded transmission is used, it is well known that
minimum mean square error (MMSE) estimation is the optimum decoding technique
[15]. Hence, the estimation, Zi , of the event information Si is simply the MMSE
estimation of Yi , which is given by
Zi =
E[Si Yi ]
Yi
E[Yi2 ]
(5.7)
Note that the estimated values of Zi ’s are spatially correlated since the actual event
information Si ’s are spatially correlated. This spatial correlation results in redundancy
in each event information sent to the sink. Although the sink is interested in estimating
the event source, S, with a distortion constraint, intuitively, this constraint can still be
met by using a smaller number of sensor nodes rather than all the nodes in the event
area. In order to investigate the distortion achieved with a smaller number of nodes
sending information, we assume that only M out of N packets are received by the
sink, where N is the total number of sensor nodes in the event area. Since the sink
decodes each Yi using the MMSE estimator, the event source can simply be computed
by taking the average of all the event information received at the sink. Then, Ŝ, the
estimate of S, is given as
Ŝ(M) =
1
M
M
(5.8)
Zi
i=1
The distortion achieved by using M packets to estimate the event S is given as
D(M) = E[(S − Ŝ(M))2 ]
(5.9)
where we use the mean-squared error as the distortion metric. Using (5.5) and (5.6)
in (5.7), the estimate Zi of each event information Si can be written as
Zi =
denoting α =
E[Si Yi ]
E[Yi2 ]
σS2
P
(Si + Ni )
2
+ σN
P
2 ,
σS2 +σN
E[Si Yi ] = ασS2
2
2
2
E[Yi ] = α σS2 + σN
(5.10)
112
SPATIOTEMPORAL CORRELATION THEORY FOR WIRELESS SENSOR NETWORKS
Then, (5.10) is restated as
Zi =
σS2
σS2
(Si + Ni )
2
+ σN
(5.11)
Using (5.11) and (5.8) in (5.9), the distortion function D(M) is found to be [3]
D(M) = σS2 −
+
M
σS4
2
ρ(s,i) − 1
2)
M(σS2 + σN
i=1
σS6
2 )2
M 2 (σS2 + σN
M
M
ρ(i,j)
(5.12)
/ i
i=1 j =
D(M) shows the distortion achieved at the sink as a function of number of nodes M
that send information to the sink and correlation coefficients ρ(i,j) and ρ(s,i) between
nodes ni and nj and between the event source S and node ni , respectively. Based on
the distortion function, we discuss possible approaches that can be used in the design
of efficient medium access control in sensor networks in Section 5.4.1.
5.3.2 Temporal Correlation in Wireless Sensor Networks
As mentioned in Section 5.1, the energy-radiating physical phenomenon constitutes
the temporal correlation between each consecutive observation of a sensor node [5].
For the periodic sensing applications such as event tracking, all consecutively taken
sensor observations are temporally correlated to a certain degree. In this section, we
establish the theoretical analysis for this temporal correlation, which will be further
elaborated in the context of correlation-based reliable event transport approach discussed in Section 5.4.2.
Here, we consider the temporal correlation between the sensor observations and
hence we omit the spatial variation in this analysis. We are interested in estimating
the signal s(t) in a decision interval of τ. In our theoretical analysis, we model an
event-to-sink distortion metric, where all the information coming from the sensor
nodes in the event area is considered as if it is generated by a single source node
during the decision interval τ.
Assume that the sensed information from the sensors are sent to the sink using
a reporting frequency of f . In this case, we seek to control the reporting frequency
f such that a desired distortion level is not exceeded in the estimation of the event
features at the sink. The event signal s(t) is assumed to be a Gaussian random process
with N(0, σs2 ). The sink is interested in finding the expectation of the signal s(t) over
the decision interval τ, that is, S(τ). Assuming that the observed signal s(t) is widesense stationary (WSS), the expectation of the signal over the decision interval τ can
be calculated by the time average of the observed signal [16], that is,
SPATIOTEMPORAL CORRELATION IN WIRELESS SENSOR NETWORKS
S(τ) =
t0 +τ
1
τ
s(t) dt
113
(5.13)
t0
where t0 is the time the sensor node wakes up for the sampling of the signal. With a
change of variables, S(τ) can be shown as
S(τ) =
1
τ
τ
0
s(t0 + Ŵ) dŴ
(5.14)
We define the value of the signal at each sampling interval as
n
S[n] = s t0 +
f
(5.15)
where f is the sampling frequency and S[n] are JGRV with N(0, σs2 ).2 For the derivation of the distortion function, the following definitions are needed:
E{S[n]} = 0
E{(S[n])2 } = σS2
E{S[n]S[m]} = σS2 ρ̂S (n, m)
E{s(t)s(t + δ)} = σS2 ρS (δ)
where ρ̂S (n, m) = ρS (|m − n|/f ) is the covariance function that depends on the
time difference between signal samples. Among the covariance models introduced in
Section 5.3.1, we use the power exponential model in the derivation since the physical event information (such as an electromagnetic wave), is modeled to have an
exponential autocorrelation function [17]. Hence, the covariance function becomes
ρS (δ) = e−|δ|/θ1
(5.16)
Each sensor node observes the noisy version of the signal given as
X[n] = S[n] + N[n]
(5.17)
and the transmitted signal is expressed by
Y [n] =
2 Note
σS2
PE
X[n]
2
+ σN
that the samples of a Gaussian random process are jointly Gaussian [16].
(5.18)
114
SPATIOTEMPORAL CORRELATION THEORY FOR WIRELESS SENSOR NETWORKS
based on the discussion in Section 5.3.1. Using the MMSE estimator at the sink, each
sample is estimated as
E S[n]Y [n]
Y [n]
Z[n] =
E Y 2 [n]
(5.19)
Hence, each estimated sample from the sensor nodes can be represented as
Z[n] =
σS2
S[n]
+
N[n]
2
σS2 + σN
(5.20)
After collecting all the samples of the signal in the decision interval τ, the sink
estimates the expectation of the signal over the last decision interval by
Ŝ(τ) =
1
τf
τf
(5.21)
Z[k]
k=1
where τf is the total number of sensor samples taken within a decision interval with
duration of τ. As a result, the distortion achieved by using τf samples to estimate the
event is given as
D=E
S(τ) − Ŝ(τ)
2
(5.22)
Using the definitions above and substituting (5.14), (5.20), and (5.21) into (5.22), the
distortion function can easily shown to be [3]
D(f ) = σS2 +
−
σS4
σ6
+ 2 2 2S 2 2
2
2
τf (σS + σN ) τ f (σS + σN )
2σS4 θ1
2)
τ 2 f (σS2 + σN
τf
k=1
2−e
− fθk
1
−e
τf
e
−( |k−l|
f )/θ1
k=1 l =
/ k
−(τ− fk )/θ1
(5.23)
It is observed from (5.23) that the distortion in the estimation decreases with
increasing f . Note that a distortion level D for the estimation of event features from
the sensor observations corresponds to a certain reliability level of the event-to-sink
communication in the WSN.
SPATIOTEMPORAL CORRELATION IN WIRELESS SENSOR NETWORKS
115
5.3.3 Joint Spatiotemporal Correlation in Wireless Sensor Networks
In the previous sections, the spatial and temporal correlations were separately captured. In order to understand the joint effects of spatial and temporal correlation,
spatiotemporal correlation characteristics of point and field sources are investigated
next.
Spatiotemporal Characteristics of Point Sources. In many WSN applications such as target detection and fire detection, the goal is to estimate the properties
of an event generated by a single point source, through collective observations of
sensor nodes. In this section, we first introduce our model for the point source and
formulate its spatiotemporal characteristics. Next, we derive the distortion function
for the estimation of the point source.
Here, we are interested in observing the joint behavior of spatial and temporal correlation. Therefore, in order to capture the spatiotemporal correlation characteristics,
we follow a different approach than in Section 5.3. Here, the point source is assumed
to generate a continuous signal that is modeled by a random process fS (s, t), where s
denotes the outcome and t denotes time. For ease of illustration, we use fS (t) in the
remainder of the chapter. We model the point source, fS (t), as a Gaussian random
process such that fS (t) is first-order stationary; that is, µS (t) = µS and has a variance
σS2 . Without loss of generality, we assume µS = 0.
For ease of illustration, we assume that the coordinate axis is centered at the point
source. As a result, the received signal, f (x, y, t), at time t at a location (x, y) can be
modeled as
√
x2 + y2 − xθ2 +y2
s
f (x, y, t) = fS t −
(5.24)
e
v
which is the delayed and attenuated version of the signal fS (t). In this model, we
assume that the event signal travels with the speed, v, and is attenuated based on an
exponential law, where θs is the attenuation constant. Note that the function f (x, y, t)
is also a Gaussian random process and that the samples taken by the sensors are jointly
Gaussian random variables (JGRVs). Since µS = 0, the mean of the received signal
is given by µE = 0.3 The variance of the received signal is also given as follows:
√
2
2
2
(5.25)
σE2 (x, y) = E f 2 (x, y, t) = σS e− x +y /θs
An interesting result from (5.25) is that the variance of the signal observed at
location (x, y) depends on the distance between the observation location and the
point source. As in Figure 5.1, the received signal at time tk by a sensor ni at location
(xi , yi ) is given by
Si [k] = f (xi , yi , tk )
3 The
subscripts S and E that are used here represent the source and event, respectively.
(5.26)
116
SPATIOTEMPORAL CORRELATION THEORY FOR WIRELESS SENSOR NETWORKS
Assuming wide-sense stationarity, the spatiotemporal correlation function for two
samples of a point source taken at locations (xi , yi ) and (xj , yj ) and at times tk and
tl , respectively, is given by
ρp (i, j, k, l) =
E Si [k] Sj [l]
σE (xi , yi ) σE (xj , yj )
= ρS (
t)
(5.27)
where t = |tk − tl − (di − dj )/v|, di = xi2 + yi2 is the distance of the sensor ni to
the point source, and ρS ( t ) = E[fS (t) fS (t + t )]/σS2 is the correlation function of
the point source which is given by ρS ( t ) = e− t /θt , where θt is a constant governing the degree of correlation. Note that the spatiotemporal correlation between two
samples, ρp (i, j, k, l), depends mainly on the difference between sample times tk and
tl since generally v ≫ (di − dj ).
In WSNs, we are interested in estimating the signal generated by the point source
using the samples collected by the sensor nodes. The expectation of the generated
signal, fS (t), over an interval τ is given by
S(τ) =
1
τ
τ
fS (t) dt
(5.28)
0
Each sensor node, ni , receives the attenuated and delayed version of the generated
signal fS (t), that is, Si [k]. Due to the impurities in the sensor circuitries, the sampled
signal is the noisy version of this received signal which is given by
Xi [k] = Si [k] + Ni [k]
(5.29)
where the subscript i denotes the location of the node ni , that is, (xi , yi ), k denotes
the sample index which corresponds to time t = tk , Xi [k] is the noisy version of the
2 ).
actual sample Si [k], and Ni [k] is the observation noise, that is, Ni [k] ∼ N(0, σN
Si [k] is given by (5.24) and (5.26).
The observed information, Xi [k], is then encoded and sent to the sink through
the WSN. It has been shown that joint source-channel coding outperforms separate
coding. Moreover, as discussed in Section 5.3, for WSNs with a finite number of nodes,
uncoded transmission outperforms any approach based on the separation paradigm
leading to the optimal solution for infinite number of nodes [3]. In the light of these
results, we assume that uncoded transmission is deployed in each node. Hence, the
transmitted observation, Yi [k], is given by
Yi [k] =
PE
X [k],
2 i
σS2 + σN
i = 1, ..., N
(5.30)
SPATIOTEMPORAL CORRELATION IN WIRELESS SENSOR NETWORKS
117
2 are the variances of the event information S [k] and the observation
where σS2 and σN
i
noise Ni [k], respectively.
The transmitted information is decoded at the sink. Since uncoded transmission is
used, it is well known that minimum mean square error (MMSE) estimation is the optimum decoding technique [3]. Hence the estimation, Zi [k], of the event information
Si [k] is simply the MMSE estimation of Yi [k], which is given by
Zi [k] =
σE2 (xi , yi )
S
[k]
+
N
[k]
i
i
2
σE2 (xi , yi ) + σN
(5.31)
The sink is interested in estimating the expected value of the event during a decision
interval τ that is given by (5.28). Assuming each sensor node sends information at a
rate of f samples/sec, this estimation can simply be found by
Ŝ(τ, f, M) =
1
τfM
M
τf
Zi [k]
(5.32)
i=1 k=1
where M is the number of sensor nodes that send samples of the observed point source.
M nodes are chosen among the nodes in the network to represent the point source and,
hence, are referred to as representative nodes. Consequently, the distortion achieved
by this estimation is given by [2]
Dp (τ, f, M) = E (S(τ) − Ŝ(τ, f, M))2
(5.33)
where the subscript p denotes the point source. Using (5.24), (5.25), (5.28), (5.31),
and (5.32), (5.33) can be expressed as
M
τf
σS4 e−3di /θs
2
Dp (τ, f, M) =
− 2
2
τ fM
σ 2 e−2di /θs + σN
i=1 k=1 S
θt 2 − e− tk +di /c − e− τ−tk −di /c /θt
σS2
+
+
2
σN
τfM 2
M
σS4 e−2di /θs
i=1
1
τ2f 2M2
2
σS2 e−di /θs + σN
M
M
τf
τf
2
α ρ(i, j, k, l)
i=1 j=1 k=1 l=1
where
α=
σS8 e−2(di +dj )/θs
2 −d /θ
2
2
σS e j s + σN
σS2 e−di /θs + σN
(5.34)
118
SPATIOTEMPORAL CORRELATION THEORY FOR WIRELESS SENSOR NETWORKS
√
and di = (xi + yi ), and ρ(i, j, k, l) is the spatiotemporal correlation function given
in (5.27).
Spatiotemporal Characteristics of Field Sources. In some WSN applications
such as temperature monitoring and seismic monitoring, the physical phenomenon is
dispersed over the sensor field and, hence, can be modeled as a field source. Thus,
here we explore the spatiotemporal characteristics of observing such a phenomenon
in WSNs.
As in Section 5.3.3, the event signal f (x, y, t) is assumed to be a Gaussian random
process with N(0, σs2 ). The sink is interested in estimating the signal f (x0 , y0 , t) over
the decision interval τ at location (x0 , y0 ). Assuming that the observed signal f (x, y, t)
is wide-sense stationary (WSS), the expectation of the signal over the decision interval
τ [i.e., S(τ)] can be calculated by the time average of the observed signal as
S(τ) =
1
τ
τ
f (x0 , y0 , t) dt
(5.35)
0
where (x0 , y0 ) is the event location. The signal Si [k] received at time tk by a sensor
node at location (xi , yi ) is defined as in (5.26), and the Si [k]’s are JGRV with N(0, σs2 ).
The covariance of two samples, Si [k] and Sj [l], is given by
cov{Si [k], Sj [l]} = σS2 ρs (i, j) ρt (δ)
(5.36)
where
ρs (i, j) = e−di,j /θs
and
ρt (δ) = e−|δ|/θt
(5.37)
are spatial and temporal correlation functions, respectively, δ = (k − l)/f , f is the
sampling rate, di,j = (xi − xj )2 + (yi − yj )2 is the distance between two nodes ni
and nj , and θs and θt are spatial and temporal correlation coefficients, respectively.
Following the discussion and derivations in Section 5.3.3, the noisy version of
the signal, Xi [k], and the transmitted signal, Yi [k], are given by (5.29) and (5.30),
respectively. The estimation Zi [k] can be found as
Zi [k] =
σS2
S
[k]
+
N
[k]
i
i
2
σS2 + σN
(5.38)
After collecting the samples of the signal in the decision interval τ from M nodes,
the sink estimates the expectation of the signal over the last decision interval as given
in (5.32). As a result, the distortion achieved by this estimation is given as in (5.33).
Using the definitions above and substituting (5.35), (5.38), and (5.32) into (5.33), the
distortion function can be derived as [2]
COROLLARIES AND EXPLOITING CORRELATION IN WIRELESS SENSOR NETWORKS
Df (τ, f, M) = σS2 −
τf
k=1
+
2σS2
2)
τ 2 fM (σS2 + σN
119
M
ρs (i, s)
i=1
− t− fk /θt
θt 2 − e−k/(f θt ) − e
2
σS4 σN
σS6
+
2 )2
2 ))2
τfM (σS2 + σN
(τfM (σS2 + σN
M
M
τf
τf
i=1 j=1 k=1 l=1
ρs (i, j)ρt (|k − l|/f )
(5.39)
In order to provide further insight into the spatiotemporal correlation characteristics and distortion analysis derived in this section, next we discuss possible
approaches that can be used in the design of efficient communication techniques
exploiting the spatiotemporal correlation observed in the wireless sensor networks.
5.4 COROLLARIES AND EXPLOITING CORRELATION IN WIRELESS
SENSOR NETWORKS
Spatiotemporal correlation, in addition to the collaborative nature of the WSN, bring
significant potential advantages for the development of efficient communication
protocols well-suited for the WSN paradigm. In this section we discuss possible
approaches exploiting spatiotemporal correlation to achieve energy-efficient medium
access and reliable event transport in WSN, respectively.
5.4.1 Spatial Correlation and Medium Access Control
The shared wireless channel between sensor nodes and energy considerations of
the WSN make the medium access control (MAC) a crucial part of the wireless
sensor networking. Furthermore, the scarce energy sources of sensor nodes necessitate
energy-aware MAC protocols. Hence, MAC protocols for WSN should be tailored to
the physical properties of the sensed phenomenon and the specific network properties
so that the access to the channel is coordinated with minimum collisions without
affecting the connectivity throughout the network.
In WSNs, many individual nodes deployed in large areas sense events and
send corresponding information about these events to the sink. As discussed in
Section 5.3.1, due to the physical properties of the event, this information may be
highly correlated according to the spatial distribution of the sensor nodes. Intuitively,
data from spatially separated sensors are more useful to the sink than highly correlated
data from closely located sensors. Hence, it may not be necessary for every sensor
120
SPATIOTEMPORAL CORRELATION THEORY FOR WIRELESS SENSOR NETWORKS
node to transmit its data to the sink. Instead, a smaller number of sensor measurements
might be adequate to communicate the event features to the sink within a certain reliability constraint. As a result, the MAC protocol can reduce the energy consumption
of the network by exploiting spatial correlation in the WSNs without compromising
on the access latency as well as the distortion achieved.
In order to gain more insight to these intuitions, a case study is performed using the distortion function (5.12). In a 500-by-500 grid, 50 sensor nodes are randomly deployed. The Power Exponential model is used with θ2 = 1 and θ1 =
{10, 50, 100, 500, 1000, 5000, 1000} as the covariance model for the covariance function, Kϑ (·) in (5.36). The parameter θ1 controls the relation between the distance of
the nodes and the correlation coefficient. For each value of θ1 , the distortion function
(5.12) is calculated varying the number of sensor nodes sending information. Starting
from 50 nodes, we decrease the number of nodes that send event information to the
sink. We refer to these nodes as the representative nodes.
Representative nodes are selected randomly among the 50 nodes for each trial,
and the distortion function is calculated according to the locations of these nodes.
The average distortion calculated from these simulations and the distribution of the
distortion for each number of representative nodes is shown in Figure 5.2.
As shown in Figure 5.2, the achieved distortion stays relatively constant when
the number of representative nodes is decreased from 50 to 15. This behavior is
due to the highly redundant data sent by the sensor nodes that are close to each
other. In addition, with increasing θ1 , the observed event distortion decreases because
close nodes become less correlated with increasing θ1 . Based on the results shown in
Figure 5.2 and the distortion function (5.12), the following discussions about the
observed distortion at the sink can be made:
13
10
50
100
500
1000
5000
10000
Observed Event Distortion
12
11
10
9
8
7
6
5
4
3
2
0
5
10
15
20
25
30
35
40
45
Number of Representative Nodes
50
Figure 5.2. Observed event distortion for different θ1 values according to changing number of
representative nodes.
COROLLARIES AND EXPLOITING CORRELATION IN WIRELESS SENSOR NETWORKS
121
Remark 1. The minimum distortion is achieved when all the nodes in the event area
send information to the sink. However, the achieved distortion at the sink can be
preserved even though the number of the representative nodes decreases. As a result,
significant energy saving is possible by allowing a smaller number of nodes to send
information to the sink about an event.
Remark 2. Based on (5.12), there are two factors affecting the distortion other than
the number of representative nodes.
1. The correlation coefficient, ρ(s,i) , between a node ni sending information and
the event source S affects the distortion function negatively. The distortion
increases as the distance between the event source S and the node ni increases.
Intuitively, if a representative node is chosen apart from the source, it observes
relatively inaccurate data resulting in higher distortion at the sink.
2. The correlation coefficient, ρ(i,j) , between each representative node ni and nj
affects the distortion positively. As the distance between nodes increases, distortion decreases. Since nodes that are further apart observe less correlated data,
the distortion is decreased if these nodes are chosen as the representative nodes.
Consequently, due to the spatial correlation between sensor observations, significant energy saving can be achieved by choosing representative nodes among the
nodes in the event area without degrading the achieved distortion at the sink. It is
clear that smaller number of nodes transmitting information reduces contention in
the wireless medium, resulting in decreased energy consumption. Energy consumed
from both transmission of packets and collision penalties can be reduced drastically
if the spatial correlation is exploited. As a result, it is important to find the minimum
number of representative nodes that achieve the distortion constraint given by the
sensor application. This minimum number can be given as
M ∗ = arg (min {D(M) < Dmax })
M
where Dmax is the maximum distortion allowed by the sensor application.
It is important to note that the minimum number of representative nodes, M ∗ ,
depends on the locations of the representative nodes. It follows from our previous discussions that, for a fixed number of representative nodes, the minimum distortion can
be achieved by choosing these nodes such that (i) they are located as close to the event
source as possible and (ii) they are located as farther apart from each other as possible.
As an example, as illustrated in Figure 5.3, choosing representative nodes such
that they are spread over the event area results in a decrease in distortion, due to less
redundant data sent by these nodes. Note that such a formation also improves the
medium access performance during the transmission of the information. Since the
representative nodes are not located close to each other, the probability of collision
in the wireless medium decreases. As a result, exploiting spatial correlation not only
improves the distortion but also utilizes the wireless channel due to the spatial reuse
property of the wireless medium.
122
SPATIOTEMPORAL CORRELATION THEORY FOR WIRELESS SENSOR NETWORKS
Sink
Sink
Event Area
Figure 5.3. Spatial re-usage in sensor networks [18].
In a recent work [18], a correlation-based MAC protocol has been proposed which
exploits the spatial correlation between closely located sensor nodes that regulate
medium access and prevent redundant transmissions from closely located sensors.
Based on the spatial correlation among sensor nodes, the MAC protocol collaboratively regulates medium access so that redundant transmissions from correlation
neighbors are suppressed. In addition, necessary mechanisms for the efficient transmission of the information from the sensor nodes to the sink have also been incorporated into the proposed correlation-based MAC solution. The experimental results in
reference 18 reveal that significant performance gains and energy savings are obtained
by exploiting spatial correlation in sensor networks.
5.4.2 Spatiotemporal Correlation and Reliable Event
Communication
In order to realize the potential gains of the WSN, it is imperative that desired event
features are reliably communicated to the sink. Unlike traditional communication networks, the sensor network paradigm necessitates that the event features are estimated
within a certain distortion bound (i.e., required reliability level) at the sink as discussed
in Section 5.2. Reliable event detection at the sink is based on collective information
provided by source nodes and not on any individual report. Hence, conventional endto-end reliability definitions and solutions are inapplicable in the WSN regime and
would only lead to overutilization of scarce sensor resources. On the other hand, the
absence of reliable transport altogether can seriously impair event detection, which
is the main objective of WSN deployment. Hence, the WSN paradigm necessitates a
collective event-to-sink reliability notion rather than the traditional end-to-end notion
[19]. Such event-to-sink reliable transport notion based on collective identification of
spatially and temporally correlated data flows from the event to the sink is illustrated
in Figure 5.4 and depends on the following definitions:
COROLLARIES AND EXPLOITING CORRELATION IN WIRELESS SENSOR NETWORKS
Event radius
123
Sink
Figure 5.4. Typical sensor network topology with event and sink. The sink is only interested
in collective information of sensor nodes within the event radius and not in their individual
data.
Definition 1. The observed event distortion Di is the distortion achieved [i.e., as in
(5.23)] when the sink performs estimation of the signal S being tracked in decision
interval i.
Definition 2. The desired event distortion D∗ is the maximum distortion allowed to
assure reliable event detection in the estimation performed by the sink. This upper
bound for the distortion level is determined by the application and based on the
physical characteristics of the signal S being tracked.
Based on the packets generated by the sensor nodes in the event area, the sink
estimates the event features to determine the necessary action and observes Di at
each decision interval i. Note that a distortion level D for the estimation of event
features from the sensor observations corresponds to the reliability level of the eventto-sink communication in the WSN. If observed event distortion is less than the
distortion bound, i.e., Di < D∗ , then the event is deemed to be reliably detected.
Else, appropriate action needs to be taken to assure the desired reliability level in the
event-to-sink communication.
The main rationale behind such event-to-sink reliability notion is that the data
generated by the sensors are temporally correlated, which tolerates individual packets to be lost to the extent where the desired event distortion D∗ is not exceeded.
Let f be the reporting frequency of a sensor node defined as the number of samples taken and hence packets sent out per unit time by that node for a sensed phenomenon. This reporting frequency can be attributed to increase in sampling rate as in
Section 5.3.2. Hence, the reporting frequency f controls the amount of traffic injected
to the sensor field while regulating the number of temporally correlated samples taken
from the phenomenon. This, in turn, affects the observed event distortion—that is,
event detection reliability. Thus, the reliable event transport problem in WSN is to
determine the reporting rate (f ) of source nodes so that the maximum event estimation distortion bound D∗ is not exceeded; that is, required event detection reliability
is achieved at the sink, with minimum resource utilization.
124
SPATIOTEMPORAL CORRELATION THEORY FOR WIRELESS SENSOR NETWORKS
5.5
10
50
100
500
1000
5000
10000
Observed Event Distortion
5
4.5
4
3.5
3
2.5
2
1.5
10 −1
100
101
Reporting Frequency
10 2
103
(s −1)
Figure 5.5. Observed event distortion for varying normalized reporting frequency.
The determination of an appropriate reporting frequency f in order to assure the
desired event distortion with minimum energy expenditure and without causing congestion is a challenging issue. As derived in Section 5.3.2, the distortion Di observed
in the estimation of the signal S being tracked depends on the reporting frequency f
used by the sensor nodes sending their readings to the sink in the decision interval i.
A case study with the same network configuration and parameters in Section 5.4.1 is
also performed to observe the variation of the observed event distortion at the sink
for varying reporting frequency f —that is, distortion function D(f ) in (5.23). It is
observed from (5.23) and Figure 5.5 that the observed event distortion at the sink
decreases with increasing f . This is because the number of samples received in a
decision interval i increases with increasing f conveying more information to the
sink from the event area. Note that after a certain reporting frequency f , the observed
event distortion cannot be further reduced. Therefore, a significant energy saving can
be achieved by selecting small enough f which achieves desired event distortion D∗
and does not lead to an overutilization of the scarce sensor resources.
On the other hand, any f chosen arbitrarily small to achieve a certain distortion
bound D∗ using (5.23) may not necessarily achieve the desired distortion level and
hence assure the event transport reliability. This is mainly because all of the sensor
samples generated with this chosen reporting frequency may not be received because
of packet losses in the sensor network due to link errors and network disconnectivity.
Similarly, very high values of f do not bring any additional gain in terms of observed
event distortion as shown in Figure 5.5; on the contrary, they may endanger the event
transport reliability by leading to congestion in the sensor network. Let fmax be the
maximum reporting frequency which the network capacity can accommodate. Thus,
f > fmax leads to congestion and hence packet losses, resulting in an increase in the
observed event distortion.
CONCLUSION AND OPEN RESEARCH ISSUES
125
Therefore, the main objective would be to operate the network at optimal operating
region—that is, achieve required event reliability level (desired distortion level in the
estimation) with minimum energy expenditure. To achieve this objective and address
the reliable event transport problem, an event-to-sink reliable transport (ESRT) protocol has been developed in reference 19 based on the event-to-sink reliability notion
for WSN. The objective of this scheme is to achieve reliable event transport with
minimum energy expenditure and congestion control by exploiting the correlation
and the collaborative nature of the WSN. To help accomplish this, the protocol uses
a congestion control mechanism that serves the dual purpose of reliable detection
and energy conservation. For example, in the states where the observed reliability
is greater than that required (i.e., very low observed event distortion) and there is
no congestion, the protocol conservatively reduces the reporting frequency f to conserve energy, while not compromising on the event estimation distortion. On the other
hand, the protocol pursues more aggressive update policies in the network states with
congestion and low event reliability—that is, high observed event distortion [19]. As
a result, the spatiotemporal correlation conveyed in the physical characteristics of the
phenomenon and deployment of the sensor network can be exploited in addressing the
energy-efficient reliable event communication problem in wireless sensor networks.
5.5 CONCLUSION AND OPEN RESEARCH ISSUES
In addition to its collaborative nature, the existence of spatiotemporal correlation
among the sensor observations is a significant and unique characteristic of the WSN.
In this chapter, a theoretical analysis of spatiotemporal correlation characteristics
in WSN is presented. It has been shown via mathematical analysis, their results,
case studies, and discussions that correlation in WSNs can be exploited to significantly improve the energy-efficiency in WSNs. Therefore, this theoretical framework
provides tools for finding the feasible operating region in terms of spatial and temporal
resolution for a specific distortion constraint considering spatiotemporal correlation,
signal properties, and network variables in WSNs.
Although there exists a considerable amount of existing research and studies on
the correlation characteristics of sensor networks, there are many open research issues and new directions on this topic. One important research direction would be to
obtain real sensor data and capture the spatiotemporal correlation behavior observed
in different practical sensor network applications in order to enhance the correlation models developed so far. This will improve the accuracy of the correlation and
distortion analysis derived in the current literature.
On the other hand, the effects of the network parameters such as topology, node
distribution, radio, and sensing ranges of sensor nodes, heterogeneity of events occurring in the network need to be carefully investigated in order to reveal other important
characteristics of spatiotemporal correlation in WSNs.
The correlation in WSNs can be considered in developing new energy-efficient networking protocols specifically tailored for the WSN paradigm. These protocols utilizing the correlation to conserve energy resources may drastically enhance the overall
network performance. Furthermore, spatiotemporal correlation could be a baseline for
126
SPATIOTEMPORAL CORRELATION THEORY FOR WIRELESS SENSOR NETWORKS
cross-layer design of energy-efficient communication techniques for wireless sensor
networks.
Moreover, the spatiotemporal correlation characteristics can also be exploited for
efficient distributed source coding and information processing techniques. To this
end, new spatiotemporal correlation modeling analysis can be developed for wireless
multimedia sensor networks [20] that involve the communication of event data in the
form of multimedia such as still image, video, and audio. Based on the spatiotemporal correlation models that can capture the unique communication paradigm of
multimedia over WSNs, energy-efficient multimedia processing and communication
algorithms can be devised for wireless multimedia sensor networks as well.
5.6 EXERCISES
1. Explain different types of correlation observed in wireless sensor networks. Provide a practical example for each case and discuss the main reasons for the observed correlation in the examples you provide.
2. What is the relation between the correlation and the distortion observed in the
estimation of event features in a given sensor network? Propose another simple
(yet practical) definition for distortion [as in (5.9)] which could be used as a
reliability indicator in wireless sensor networks.
3. Derive the distortion function in (5.12) using the Spherical Covariance Model
instead of the Power Exponential Model. Obtain the distortion as a function of
number of nodes M that send information to the sink and correlation coefficients
and comment on the results comparing them with Figure 5.2.
4. In the analysis of temporal correlation in WSNs, it is assumed that the observed
signal s(t) is wide-sense stationary (WSS). What is the effect of this assumption?
How could this assumption be relaxed without losing the validity of the results
of the analysis?
5. Note that the distortion analysis in this chapter does not explicitly incorporate the
effects of channel and network capacity on the successful reception performance
of the transmitted samples. Propose a modification to the model presented in
Section 5.2 in order to include the effects of channel error rate and network
congestion on the derived distortion functions.
6. Based on the spatiotemporal correlation characteristics and the distortion functions derived in this chapter, propose a new routing protocol for WSNs which can
find the minimum energy and minimum distortion paths from the event field to
the sink. Clearly state the assumptions you make and outline the algorithm you
propose by explaining the rational behind your idea.
BIBLIOGRAPHY
1. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks:
A survey. Computer Networks (Elsevier) Journal, 38(4):393–422, 2002.
BIBLIOGRAPHY
127
2. M. C. Vuran and O. B. Akan. Spatiotemporal characteristics of point and field sources in
wireless sensor networks. In Proceedings of IEEE ICC 2006, Istanbul, Turkey, June 2006.
IEEE CD-ROM.
3. M. C. Vuran, O. B. Akan, and I. F. Akyildiz. Spatiotemporal correlation: Theory and applications for wireless sensor networks. Computer Networks Journal (Elsevier), 45(3):245–
261, 2004.
4. S. Meguerdichian, F. Koushanfar, M. Potkonjak, and M. B. Srivastava. Coverage problems
in wireless ad-hoc sensor networks. In Proceedings of IEEE INFOCOM 2001, Anchorage,
Alaska, April 2001.
5. J. Kusuma, L. Doherty, and K. Ramchandran. Distributed compression for sensor networks.
In Proceedings of IEEE Image Processing 2001, Vol. 1, October 2001, pp. 82–85.
6. S. Bandyopadhyay and E. J. Coyle. Spatiotemporal sampling rates and energy efficiency
in wireless sensor networks. In Proceedings of the IEEE INFOCOM 2004, Vol. 3, 2004,
pp. 1728–1739.
7. R. Cristescu, B. Beferull-Lozano, and M. Vetterli. On network correlated data gathering.
In Proceedings of IEEE INFOCOM 2004, Vol. 4, March 2004, pp. 2571–2582.
8. R. Cristescu and B. Beferull-Lozano. Lossy network correlated data gathering with highresolution coding. In Proceedings of IPSN 2005, April 2005, pp. 218–224.
9. R. Cristescu and M. Vetterli. On the optimal density for real-time data gathering of
spatiotemporal processes in sensor networks. In Proceedings of IPSN 2005, April 2005.
10. M. Gastpar and M. Vetterli. Source-channel communication in Sensor Networks. In Proceedings of the 2nd International Workshop on Information Processing in Sensor Networks
(IPSN’03), 2003.
11. M. Gastpar and M. Vetterli. Power, spatiotemporal bandwidth, and distortion in large sensor
networks. IEEE Journal on Selected Areas in Communications, 23(4):745–754, 2005.
12. J. O. Berger, V. de Oliviera, and B. Sanso. Objective bayesian analysis of spatially correlated
data. Journal of the American Statistical Association 96, 1361–1374, 2001.
13. S. S. Pradhan, R. Puri, and K. Ramchandran. n-Channel symmetric multiple descriptions,
Part I: (n,k) source-channel erasure codes. accepted for publication in IEEE Transactions
on Information Theory, 2003.
14. T. J. Goblick. Theoretical Limitations on the Transmission of Data from Analog Sources.
IEEE Transformation on Information Theory, IT-11(4):558–567, 1965.
15. V. Poor. An Introduction to Signal Detection and Estimation, 2nd edition, Springer-Verlag,
Berlin, 1994.
16. S. Haykin. Communication systems. 3rd edition, John Wiley & Sons, New York, 1994.
17. G. L. Stuber, Principles of Mobile Communication, Kluwer Academic Publishers, Norwell,
MA, 2001.
18. M. C. Vuran and I. F. Akyildiz. Spatial correlation-based collaborative medium access
control in wireless sensor networks. IEEE/ACM Transactions on Networking, 14(2):
316–329, 2006.
19. O. B. Akan and I. F. Akyildiz. Event-to-sink reliable transport in wireless sensor networks.
IEEE/ACM Transactions on Networking, 13(5):1003–1016, 2005.
20. E. Gurses and O. B. Akan. Multimedia Communication in Wireless Sensor Networks.
Annals of Telecommunications, 60(7–8):799–827, 2005.
CHAPTER 6
A Taxonomy of Routing Protocols
in Sensor Networks
AZZEDINE BOUKERCHE
School of Information Technology and Engineering, University of Ottawa, Ottawa, Ontario K1N
6N5, Canada
MOHAMMAD Z. AHMAD and DAMLA TURGUT
School of Electrical Engineering and Computer Science, University of Central Florida, Orlando,
FL 32816-2362
BEGUMHAN TURGUT
Department of Computer Science, Rutgers University, Piscataway, NJ
6.1 INTRODUCTION
Sensors, in the sense of devices that perform the measurement of certain quantities
and transform them into a computer readable digital format, have been around for at
least a few decades. These sensors were either connected to a data collection device
or connected directly to computers, using traditional wired communication such as
a serial interface. The development of highly integrated computer devices, wireless
radios, and miniaturization allowed the development of wireless sensor nodes, which
are miniature computers with integrated sensing and wireless communication capabilities [1, 2]. The continuing miniaturization effort allowed the development of nodes
with a physical size of several millimeters. Although many of these devices can act
as a general-purpose computer, with the ability to perform computation as well as
sensing, there are obvious limitations. First, the limited memory and computational
power does not allow us to run a full-featured operating system. Most of the time, the
networking stack needs to be simplified as well. The wireless transmission range is
limited by the small size of the antennas. The power resources of the sensor nodes
are limited by the physical size of the batteries; moreover, some of the proposed
deployment models do not allow the sensor nodes to recharge their batteries.
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
129
130
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
Let us now consider the deployment models of the wireless sensor nodes. Classical
sensors are typically used in pre-engineered deployment, being placed at a carefully
chosen locations. For instance, the space shuttle uses several dozen temperature sensors, carefully positioned in the structure of the spacecraft and reporting through
a wired connection to a central computer. Naturally, wireless sensor nodes can be
deployed similarly, simply by replacing the wired connection with a wireless one.
However, we can also choose a radically different deployment method: Blanket the
desired area with a large number of nodes. Instead of careful positioning, we need
to worry only about making sure that every area of interest is covered by one node
(or preferably, several). The deployed nodes might not have the transmission range
to reach the central computer, but they can transmit the collected information on a
hop-by-hop basis to collection points called sinks. We call the resulting structure a
“wireless sensor network (WSN).” In our example application, a WSN can provide
several advantages: Due to the large number of sensors, it can collect more data than
sensors with pre-engineered deployment. As several sensors cover the same area, they
provide fault tolerance. In addition, having computational as well as sensing capabilities, the wireless sensor network can provide preliminary processing of the collected
data concomitantly with the sensing and forwarding.
Let us now investigate the properties of a wireless sensor network from the networking point of view. First of all, there is no infrastructure available. Because the sinks are
accessible only to a limited subset of nodes, the sensor nodes need to participate in the
forwarding of the packets. Due to the random deployment, the routing architecture
cannot be pre-established; the network needs to be set up through self-configuration.
In these respects, wireless sensor networks are similar to ad hoc wireless networks.
There are, however, several important differences. The sensor nodes have significantly lower communication and computation capabilities than do the full-featured
computers participating in ad hoc networks. The problem of energy resources is especially difficult. Due to their deployment model, the energy source of the sensor node
is considered nonrenewable (although some sensor nodes might be able to scavenge
resources from their environment). Routing protocols deployed in sensor networks
need to consider the problem of efficient use of power resources.
An additional difference between ad hoc and sensor networks refers to the uniqueness of the nodes. Ad hoc nodes have a hard-wired unique MAC address, which
forms the basis of node identification on the higher levels of the networking stack.
The cheap, disposable sensor nodes usually come without any pre-wired identifiers;
they acquire a unique identity only after deployment, by virtue of their position in the
environment.
In addition to these, several other factors such as the large number of nodes in
sensor networks, the high failure rates of the sensor nodes, and the frequent use of
broadcasting in sensor networks as opposed to the typically unicast communication
in ad hoc networks [3] require new types of MAC [4, 5] and routing protocols,
specifically targeted toward the requirements of WSNs.
In this chapter, we succintly present the major applications of the WSNs, describe
some of the design issues associated with routing algorithms for WSN, and finally
present a survey of the state of the art in WSN routing protocols.
DESIGN ISSUES
131
6.2 APPLICATIONS
Sensor networks can be deployed in a wide variety of applications. One of the main
classification criteria is whether the sensor nodes are mobile or immobile. The data
collection might be either continuous or periodic; the latter can lead to bursty traffic
patterns. Naturally, every application requires a specific set of sensor types. Some of
the most popular sensor types are: light, sound, magnetic field, accelerator, temperature, humidity, chemical composition such as soil makeup, mechanical stress levels
on an object, and many others [6]. Some of the primary application domains for sensor
networks are the following:
Environmental. Environmental sensors can be used to detect and track natural disasters such as forest fires or floods. They can also be used to track the movement
of birds and other animals.
Military. The sensor networks will be an integral part of the future C4ISRT systems (command, control, communications, computing, intelligence, surveillance, reconnaissance, and targeting.) They can, for instance, be used to track
the movement of the enemy in the battlefield. The main advantage of WSNs is
that they can be deployed and operated remotely, without putting human lives
at risk. Naturally, military deployments bring their own challenges of security
and confidentiality.
Health. Sensor networks can be used in hospitals and clinics for patient monitoring
and tracking of various systems and humans. Sensors can be also used to track
and monitor the drug doses prescribed to patients and prevent situations where
the drugs are administered to the wrong patient. Sensors can be deployed for
telemonitoring of patients, a promising new direction for at-home monitoring
and care for the elderly.
Home. The various home appliances can be sensor enabled and interconnected
with each other and a central control system of the home. These sensor-enabled
sensor homes might not only offer additional conveniences, but will also be
safer and more energy-efficient.
6.3 DESIGN ISSUES
The challenges posed by the deployment of sensor networks is a superset of those
found in wireless ad hoc networks. Sensor nodes communicate over wireless, lossy
lines with no infrastructure. An additional challenge is related to the limited, usually
nonrenewable energy supply of the sensor nodes. In order to maximize the lifetime of
the network, the protocols need to be designed from the beginning with the objective
of efficient management of the energy resources [3]. Let us now discuss the individual
design issues in greater detail.
Fault Tolerance. Sensor nodes are vulnerable and frequently deployed in dangerous environment. Nodes can fail due to hardware problems or physical damage
132
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
or by exhausting their energy supply. We expect the node failures to be much
higher than the one normally considered in wired or infrastructure-based wireless networks. The protocols deployed in a sensor network should be able to
detect these failures as soon as possible and be robust enough to handle a relatively large number of failures while maintaining the overall functionality of the
network. This is especially relevant to the routing protocol design, which has to
ensure that alternate paths are available for rerouting of the packets. Different
deployment environments pose different fault tolerance requirements.
Scalability. Sensor networks vary in scale from several nodes to potentially
several hundred thousand. In addition, the deployment density is also variable.
For collecting high-resolution data, the node density might reach the level
where a node has several thousand neighbors in their transmission range. The
protocols deployed in sensor networks need to be scalable to these levels and
be able to maintain adequate performance.
Production Costs. Because many deployment models consider the sensor nodes
to be disposable devices, sensor networks can compete with traditional
information gathering approaches only if the individual sensor nodes can be
produced very cheaply. The target price envisioned for a sensor node should
ideally be less than $1.
Hardware Constraints. At minimum, every sensor node needs to have a sensing
unit, a processing unit, a transmission unit, and a power supply. Optionally,
the nodes may have several built-in sensors or additional devices such as
a localization system to enable location-aware routing. However, every
additional functionality comes with additional cost and increases the power
consumption and physical size of the node. Thus, additional functionality
needs to be always balanced against cost and low-power requirements.
Transmission Media. The communication between the nodes is normally implemented using radio communication over the popular ISM bands. However,
some sensor networks use optical or infrared communication, with the latter
having the advantage of being robust and virtually interference free.
Power Consumption. As we have already seen, many of the challenges of
sensor networks revolve around the limited power resources. The size of the
nodes limits the size of the battery. The software and hardware design needs
to carefully consider the issues of efficient energy use. For instance, data
compression might reduce the amount of energy used for radio transmission,
but uses additional energy for computation and/or filtering. The energy policy
also depends on the application; in some applications, it might be acceptable to
turn off a subset of nodes in order to conserve energy while other applications
require all nodes operating simultaneously.
6.4 SENSOR NETWORKS ROUTING PROTOCOLS
Routing has been carried out in sensor networks with the emphasis on conserving energy. Power efficiency is the most important design metric because it is an
SENSOR NETWORKS ROUTING PROTOCOLS
133
Routing protocols in sensor networks
Attributebased
Flat
Geographical
Hierarchical
Multi-path
QoS-based
Directed Diffusion
GRAdient
SPEED
LEACH
M-MPR
SER
EAD
SAR
Rao et. al
PEGASIS
Ganesan et. al
Arrive
Youssef et. al
MCFA
Seada et. al
TEEN/APTEEN
ReInForM
Subramanian et. al
Al-Karaki et. al
Rumor
Figure 6.1. Categories of sensor routing protocols.
important design goal of any sensor network. Routing algorithms are also datacentric and employ attribute-based addressing strategies along with location awareness. These can be used with various clustering and hierarchical approaches to make
an efficient routing algorithm for sensor networks. A robust and scalable stategy is
required in designing a routing protocol that is also energy-efficient with minimal
control overhead. Different routing protocols in sensor networks can be divided into
six categories based on their underlying architectural framework. These different
categories are as follows (see Figure 6.1).
r
r
r
r
r
r
Attribute-based (Section 6.4.1)
Flat (Section 6.4.2)
Geographical (Section 6.4.3)
Hierarchical (Section 6.4.4)
Multipath (Section 6.4.5)
QoS-based (Section 6.4.6)
Some prior studies discussing the various sensor network routing protocols have
been published [3, 7–9]. We list the classifications of sensor network routing protocols
mostly based on the type of deployment of these networks. This classification helps
the advent of newer ideas with special focus on the application being serviced by
the sensor network and hence develops efficient algorithms to facilitate better routing
techniques in these networks.
6.4.1 Attribute-Based Protocols
Attribute-based routing protocols in sensor networks concentrate on routing data
packets based on the content of the packets and are not device-specific. Hence, these
are also known as data-centric routing approches. Since each node within the network
is engaged in the routing mechanism, these nodes can make any decision and apply
any routing rules to packets—that is, either forward or drop the packets. In this class
of routing algorithms, contents of the transmitted data are evaluated at each hop in
the network.
134
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
Event
Event
Source
Event
Source
Interests
Source
Gradients
Sink
Sink
Sink
Figure 6.2. Directed diffusion [10].
Directed Diffusion [10]. Directed diffusion (Figure 6.2) is a data-centric approach,
which means that all communication is for named data. It can be categorized under both attributed-based routing and flat routing protocols. Data generated by the
application-aware sensor nodes is named using attribute–value pairs. A node requests
data by sending interests for named data. A sensing task is disseminated via sequence of local interactions throughout the network as an interest for named data.
Nodes diffusing the interest set up their own caches and gradients within the network to which the data delivery is carried out. During the data transmission, reinforcement and negative reinforcement techniques are used to converge to efficient
distribution. Intermediate nodes fuse interests and aggregate, correlate, or cache
data. Algorithm 1 presents a pseudocode of the directed diffusion algorithm.
ALGORITHM 1. Pseudocode Describing the Directed Diffusion Algorithm
Setup phase:
1.
The base station broadcasts its set of interests
2.
Do
3.
ForEach network node N receiving an interest from node M
4.
N forwards the received interest to its neighbors
(other than M)
5.
N sets up a gradient with M
6.
EndFor
7.
Until all gradients are set up
8.
Check for loops in the paths and remove them
Operating phase:
1. ForEach node
2.
Colect sensor data.
3.
Receive messages containing sensor data readings.
4.
Aggregate, correlate or fuse data (if necessary)
5.
If data maches an interest
6.
Forward the data according to the gradient associated
with the interest
7.
EndIf
8. EndFor
SENSOR NETWORKS ROUTING PROTOCOLS
135
Energy-Aware Data-centric Routing (EAD) [11]. This protocol proposes to
build a virtual backbone which contains all the active sensors in the network to facilitate energy aware routing. A routing heuristic allows us to build a broadcast tree
rooted at a gateway to enable the data-centric approach. The tree spans all sensors
within the network and has a large number of leaves that save power by turning off
their radios. The active sensors continue working as relays for traffic generated in the
network.
In data-centric routing, backbone senders are in charge of data processing and
information dissemination throughout the network. At each individual sensor, the
local raw data is initially combined/aggregated with data from other sensors located
farther away from the sink. This aggregated data are then sent to a sensor closer to
the sink or to the sink itself. EAD consists of two main components: the neighboring
broadcast scheduling and the distributed competition among neighbors. These components ensure that the final tree has many leaves, and sensors with relatively higher
residual power also have a greater chance of being part of the virtual backbone. EAD
basically makes certain the formation of a specially rooted broadcast tree designed
for data-centric routing. This protocol is suitable for applications requiring frequent
queries and events.
Constrained Shortest-Path Energy-Aware Routing [12]. The distance from
a source to a destination can be used as a metric for energy consumption and estimation
of the propagation delay between them. It is shown that by changing the transmission
power level and thereby changing the network topology graph, the algorithm can be
optimized to ensure higher throughput and energy efficiency while maintaining lower
end-to-end delay.
The nodes are grouped into clusters with each cluster having a clusterhead or gateway node. Routing decisions are determined and maintained at the clusterhead. Such
a centralized approach is more efficient than a distributed approach since it entails
less control packet overhead maintainance. The network operates in two main cycles:
data and routing. The data cycle consists of the nodes sending data to the gateway
nodes, whereas during the routing cycle the routing state of each node is determined
by the clusterhead and the routing information is sent to all the nodes accordingly.
The constrained shortest-path algorithm uses the distance between any two nodes
to determine transmission power required to send packets from one to another. The
transmission energy varies inversely with d n , where d is the distance between the
transmitter and receiver and n is a value based on the system and application in
use. The connectivity between nodes in a cluster can be maintained by this energy
parameter, making the network topology dynamic. If no constraints are enforced for
the transmission energy, then each node uses its maximum energy to transmit directly
to the destination. This is certainly not maintainable in the long run; therefore, a
constraint has to be put in place. Rerouting is carried out by the clusterhead if (i) the
sensors within the clusters are reorganized, (ii) the battery level of the active nodes
falls under a certain threshold, or (iii) there are some changes required in the energy
model of the nodes.
136
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
Rumor [13]. Rumor allows the routing of queries to nodes that have observed an
event of interest. As a result, retrieval of data is based on events and not on an
addressing scheme. An event is an activity related to the phenomena being sensed
(e.g., increased movement in an area being monitored). Events are assumed to be
localized phenomena that occur in fixed regions of space. A query is issued by the
sink node for one of two reasons: as an order to collect more data or as a request for
information. Once a query arrives at its destination, data are issued to the originator of
the query. Depending on the amount of data (whether it is more or less) being issued
to the originator of the query, shorter paths from the source to the sink are discovered.
Any node may generate a query for a particular event. If it knows the route to the
event, it transmits the query. Otherwise, the query is sent in a random direction, and
this continues until the query reaches a node that has a route to the event.
Each node has its neighbor list and an events table with forwarding information
to all the events it knows. After a node witnesses an event, an agent may be created,
which is a long-lived packet and travels around the network. Each agent contains an
events table, including the routing information for all events it knows. Since an event
happens in a zone, composed of several or many nodes, it is possible that multiple
agents are created from the zone and moving in the network. When an agent travels
in the network, its routing table is updated if there exist a shorter path to an event
within the routing table of the node it is visiting. In a similar way, the routing table
of the currently visited node is updated if its route to an event is more costly than the
agent’s route. See Figure 6.3.
Rumor routing uses agents that have a limited life determined by a TTL field; these
agents create paths in the direction of any events they may come across. If an agent
crosses a path to an event that it has not yet come across in the network, it creates a
path that leads to both events.
If flooding was to happen on a regular basis, network resources would be consumed
quickly, thus Rumor routing was created to be an alternative to flooding queries and
events. When a query is generated, it is sent randomly through the network until it
finds the event path instead of flooding it. When the query finds the event path, it is
A
text
Event
Node with path to event
Event 1
Node with path to event 1
Node
Query source = query path to event
Event 2
Node with path to event 2
A Agent
Figure 6.3. Rumor [13].
Node with path to events 1 and 2
SENSOR NETWORKS ROUTING PROTOCOLS
137
routed directly to the event. Only if the path cannot be found, it will be flooded as a
last resort.
6.4.2 Flat Protocols
In flat sensor networks, there is a large number of nodes that collaborate together to
sense the environment. These nodes are similar to each other in all respects; and due
to their sheer number, they cannot be assigned specific global IDs. Routing protocols
servicing such flat networks fall under this category of sensor networks.
Gradient Broadcast (GRAB) [14]. This protocol is designed for efficient data
forwarding in large-scale dense sensor networks where particular objects or events
called stimuli are monitored by all the sensor nodes in the neighborhood. An earlier
version of this work has been reported in reference 15. These sensors pick up the same
stimuli simultaneously but collectively elect a leader that creates a sensing report on
behalf of the entire group. The election is carried out by the signal strength of the
field created by the stimulus. All the nodes picking up the stimulus broadcast their
respective signal strengths with a random delay to avoid collisions. Whenever a node
hears this signal from another node, the node checks the strength of the signal and
compares it with its own. If it is lower than the received signal, it stops; otherwise it
rebroadcasts its own signal strength. This helps in rolling the messages to the center
of the stimuli (CoS), where ultimately the leader is chosen. This is the node with the
largest signal strength, and then it generates the report. This leader is named as the
data source. On the other hand, each node always maintains a cost value, a metric
representing the cost of sending data from the node to the destination sink. It is the
job of the sink to build and maintain this cost field associated with each node, but the
nodes also know their respective costs. Nodes located closer to the sink obviously
have lower cost values than ones that are located farther away. The main difference
of GRAB from other protocols is that instead of the sender deciding the path for the
data packet, it is the intermediate nodes who decide whether or not they should be
forwarding the packet to the sink. This decision is made by the receiver nodes simply
by comparing their respective cost values with that of the sender; and if it is lower,
then they accept forwarding the packet. Due to this inherent mechanism, data will
always be forwarded on the shorter path—that is, the path of descending cost to the
sink.
However, some small issues come up if there are multiple paths to the destination.
To solve some of these problems, the source also assigns a credit value along with
the report it generates and sends out for transmission. With this credit value, a report
may be sent out on a mesh of interleaved paths where none of the paths have a total
cost greater than the cost of the source in addition to its credit. This simply makes use
of the multiple paths if available and hence introduces a certain degree of robustness
into the protocol. This degree of overhead and robustness is the “width” of the mesh,
which is ultimately dependent on the credit of the report. This credit value also helps
in controlling the robustness degree, which may be useful for different scenarios
138
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
and events occurring within the sensor network. No state information is required to
be maintained at the individual nodes, thereby eliminating overhead due to network
complexity and path maintenance.
Sequential Assignment Routing (SAR) [16–18]. This is the very first sensor
routing algorithm to include QoS concepts. It is mainly a table-driven multipath
approach toward routing and concentrates on energy efficiency and fault tolerance.
Path selection is made by considering the energy resources of the nodes on the path
along with the priority levels of each packet. The packet priority is calculated using
an innovative technique. There is an energy cost and delay value associated with each
link, which provides a resistance to packet flow. It is against this metric that a packet
will have credits so that it achieves a certain priority in using these paths. A weighted
QoS metric, which is the product of the additive QoS metric mentioned earlier and
a certain weight coefficient, measures the QoS provided to each packet relative to
the priority level of the packet [16]. The main objective of the SAR algorithm is to
minimize this average weighted QoS metric.
The SAR algorithm creates trees that have roots at the one-hop neighbors of the sink
by taking the priority level of each packet, the QoS metric, and the energy resources
for the nodes on each path. With these trees, multiple paths are created and maintained
from source to destination. When a path fails, the recovery procedure is carried out
by forcing updates of the routing table between upstream and downstream nodes.
Hence, there is always an available path for sending the packet, and any path failure
always leads to a local path restoration being carried, which mainly means updating
the respective routing tables for the nodes.
The major disadvantage of this protocol is that each node has to maintain entire
routing tables of not only one path between a set of sources and destinations, but
many available paths. This leads to greater overhead and memory requirements and
also leads to more energy waste, especially in case of dense networks.
Minimum Cost Forwarding Algorithm (MCFA) [15]. This paper presents a
cost-field approach to minimum cost forwarding in sensor networks. The authors
propose a “cost field,” which is defined as “the minimum cost from that node to the
sink on the optimal path.” Since the direction of the routing is always known, the
sensor node is not required to have unique ID or maintain a routing table, but rather
each node maintains the minimum cost estimate from itself to the sink. Each sensor
node broadcasts the message to be forwarded to its neighbors. When the neighbor
node receives the message, it verifies if it is on the minimum cost path toward the sink.
If so, it broadcasts the message to its neighbors. This forwarding scheme continues
until the message reaches to the sink.
6.4.3 Geographical Routing
Geographical routing uses location devices to estimate the location of a node prior
to forwarding the packets to the destination region. In wireless sensor networks,
SENSOR NETWORKS ROUTING PROTOCOLS
139
geographical forwarding (GF) is a widely used approach due to its low overhead
and localized interactions. In GF, location information is exchanged with the neighbors, and the neighbor closest to the destination is selected as a result of forwarding
decisions.
Speed [19]. This is the stateless protocol for soft real-time communication in sensor
networks. Each node maintains information about all of its neighbors and uses a
geographic forwarding technique to find the paths. The SPEED protocol tries to
ensure a certain speed for each packet sent such that the application has an idea
of the average end-to-end delay between the nodes. If the routing protocol makes
the estimated speed of the packet available, the application can calculate this delay
beforehand, simply dividing the distance between the source and the sink by the
estimated speed. The protocol also provides congestion avoidance mechanism when
needed.
The routing component, called the stateless geographic nondeterministic forwarding (SNFG), in SPEED uses four modules listed below to function effectively.
1. Beacon Exchange: Similar to any other geographic routing algorithm, location
information is updated periodically by nodes using specific beacon packets
broadcast. Two other types of on-demand beacons in SPEED include delay
estimation and backpressure beacons. These are mainly used to quickly identify
traffic changes in the network.
2. Delay Estimation: This is carried out at each node by calculating the time elapsed
since the acknowledgment receipt of the last data packet sent. By observing
these delay values, the SNFG selects the best node conforming to the speed
requirement of the application.
3. Neighborhood Feedback Loop: If no node with an appropriate delay value is
obtained, this module is consulted for the relay ratio of the node. This ratio is
simply the miss ratio of all the neighbors of the node—that is, the nodes that
were unable to provide the desired speed for the application. A random number
between zero and one is generated; and if the relay ratio is less than the random
number generated, the packet is dropped.
4. Backpressure Rerouting: This module helps to reduce congestion by sending
messages to the source nodes such that alternate routes can be discovered.
Due to the stateless architecture of SPEED, it requires minimal memory and minimal MAC layer support. It also provides traffic load balancing and also enables QoS
routing and congestion management in the sensor network applications.
Geographic Routing with No Location Information [20]. This is a geographic
routing algorithm that does not use any location information. The previous geographic
protocols presented always assume the prescence of location identification techniques; however, this requirement cannot always hold. This scheme assigns virtual
140
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
coordinates in such a scenario in which any other geographic routing protocols can
be deployed without any location information.
The protocol has three stages. The initial assumptions are gradually removed as
the protocol proceed further. It is essential to point out that the virtual coordinates
assigned to the nodes do not have to accurately match to the geography of the places,
but they should reflect the underlying connectivity. The local connectivity information
is used to set these coordinates. The three stages of the protocol are described below
in detail.
1. Perimeter Nodes Know Location: Initially, it is assumed that all perimeter nodes
know their locations and the nonperimeter nodes are assigned virtual coordinates. The analogy used, which is borrowed from graph theoretic approaches,
states that “each neighbor relation is represented by a force that pulls the neighbors together.” Thus, the force in the x and y direction is proportional to the
difference in x and y coordinates of the nodes, respectively. Iteratively, each
node updates its virtual coordinates by calculating the average of x and y coordinates of the neighbor set of each node. As the number of iterations increase,
the nodes tend to move toward the perimeter nodes closest to them, and ultimately the algorithm converges to a steady state where the nodes are spread
throughout the region.
2. Perimeter Nodes Are Known: The perimeter nodes are aware that they are on
the perimeter; but unlike the first step, they do not know their location. Each
perimeter node broadcasts a HELLO packet to the entire network to discover the
hop distance to every other perimeter node. These distance pairs are stored in a
perimeter vector in which each perimeter node in turn broadcasts its perimeter
vector to the entire network. As a result, every perimeter node becomes aware
of its hop distance from other perimeter nodes. A triangulation algorithm is executed to enable the nodes to compute the coordinates of all the perimeter nodes
in the network. The triangulation algorithm could fail under some conditions.
As a part of the solution to this problem, two nodes are designated as bootstrap
beacons. These nodes flood the network with HELLO messages. The perimeter
nodes include these beacons in their triangulation algorithm to compute the
coordinates of all the perimeter nodes.
3. No Location Information: The nodes use the following criteria as stated in
reference 20 to decide if they are perimeter nodes: If a node is the farthest
away, among all its two-hop neighbors from the first bootstrap node, then the
node decides that it is on the perimeter.
Energy-Efficient Forwarding Strategies for Geographic Routing in Lossy
Wireless Sensor Networks [21]. The main aspect of geographic forwarding includes the greedy forwarding of message packets by a node, usually to the neighbor
located closest to the destination. This strategy works only under the following assumptions: (i) sufficiently dense network; (ii) accurate localization; and (iii) high link
reliability irrespective of the distance within the physical radio ranges. It has been
SENSOR NETWORKS ROUTING PROTOCOLS
141
observed that even though the first two assumptions may hold for most types of systems, the final assumption regarding high reliable links will not be fully satisfied in
a realistic scenario. This is because wireless links are inherently unreliable in nature
and are subject to various environmental conditions causing the weak links to drop
a high number of packets. These packets are then retransmitted, which leads to high
energy consumption. In order to alleviate the problem, the neighbors are classified
based on link reliability. Not only could some neighbor links be weaker than the others, their loss characteristics could also be different. Hence, a blacklisting/neighbor
selection scheme is devised to avoid the weak and unreliable links. The distancehop energy tradeoff performance metric is proposed for geographic forwarding. If
the forwarding scheme aims at minimizing hops as in greedy forwarding, a significant energy consumption can occur due to transmitting and/or retransmitting the
packets on long and possibly unreliable weak links. On the other hand, if smaller
distances are covered across stronger links, more hops would be visited again, causing increased energy usage. The optimal choice is generally in the transitional region
between these two strategies. However, let us note that not too many links should be
blacklisted since this would cause greater route disconnections and lower delivery
rates.
Geographic Routing with Limited Information in Sensor Networks
[22]. This protocol shows that even for instances where there is erroneous or limited
location information due to faulty GPS, the order of routing delays is within a constant
factor of straight-line greedy routing strategies. The limited information means that
nodes only know the relative quadrant or zone where the destination node is situated.
Nodes can be forwarded in the wrong direction if the GPS at nodes are faulty or
biased even when the nodes have the correct destination coordinates. The four basic
scenario models are discussed below.
1. Location Errors (Imprecise GPS): An angular error is introduced within greedy
straight-line routing so that the packet is forwarded anywhere in the zone within
the angle.
2. Limited Destination Information: It is assumed that the node has only a coarse
estimate of the destination location such as the quadrant or half-plane information.
3. Small Fraction of Nodes with Routing Information: In this scenario, only a
small fraction of nodes know about the destination quadrant. If a node that has
no information about the destination has a packet to send, it simply selects a
neighbor randomly to forward the packet.
4. Throughput-Capacity Networks: The packet is forwarded in the right direction
but not in a straight line along the shortest path. It uses specific progressive
routing strategies to selectively forward the packets toward the destination.
This leads to spatial “hot spots” within the network because many paths may
intersect due to suboptimal selection of routes.
142
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
6.4.4 Hierarchical Protocols
As the name suggests, this class of routing techniques create a virtual hierarchy among
the nodes of the sensor network. Such a hierarchy may be based on zones which
comprise of divisions of the entire network area, node functionality or node location.
Irregardless of the system creating the network hierarchy, the routing protocol must
be designed to make maximum use of this hierarchical pattern.
Low-Energy Adaptive Clustering Hierarchy (LEACH) [23]. LEACH is based
on a simple clustering mechanism by which energy can be conserved since cluster
heads are selected for data transmission instead of other nodes in the network (Algorithm 2). By the received signal strengths, local cluster heads are selected to serve as
the routers to the data sinks. A sensor node sends its data to the local cluster head in
turn transmitted to the nearest cluster head on the way to the sink. Since the cluster
heads are only responsible for bulk of the data transfer, the overhead is minimized;
however, if the cluster heads are chosen beforehand and remain fixed throughout the
network lifetime, they will easily die out, thereby ending the lifetime of the member
nodes of the particular cluster as well. To solve this problem, LEACH performs a
periodic randomized rotation of the cluster head to enable all the nodes within the
cluster to take on a collective responsibility in order not to drain the battery of a single
node. The optimal number of cluster heads is considered 5 percent of the total number
of nodes.
LEACH also performs local data fusion and aggregation to compress the data
received from each cluster. Sensor nodes are selected as cluster heads by the node
choosing a random number between zero and one. The node is selected as a cluster
head for the current round if the number is less than the following threshold values:
T (n) =
p
1 − p ∗ (r mod p1 )
if
n∈G
where p is the desired percentage of cluster heads, r is the current round, and G is
the set of nodes that have not been cluster heads in the last p1 rounds.
Once all the nodes are organized into clusters, the cluster head will create a schedule
for the nodes in its cluster which enables the radio components of each cluster node
to be turned off for most of the time. Each node transmits its data to the cluster head
according to its schedule. On completion, the cluster head aggregates and sends all
the data to the sink.
LEACH achieves a significant reduction in energy dissipation when compared
with direct communication and other minimum energy routing protocols. Properties
of LEACH includes (i) dynamic clustering to increase network lifetime; (ii) singlehop routing from node to cluster head, hence saving energy; (iii) distributiveness;
(iv) additional overhead due to cluster head changes and calculations leading to energy
inefficiency for dynamic clustering in large networks.
SENSOR NETWORKS ROUTING PROTOCOLS
143
ALGORITHM 2. Pseudo-code Describing the Operation of the LEACH Protocol
Setup Phase:
In this phase clusters are created---cluster heads (CHs) are
chosen
1.
2.
3.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
ForEach (node N)
N selects a random number r between 0 and 1
If (r < Threshold value)
N becomes a CH
N broadcasts a message advertising its CH status
Else
N becomes a regular node
N listens to the advertising messages of the CHs
N chooses the CH with the strongest signal as its
cluster head
N informs the selected CH and becomes a member of
its cluster
EndIf
ForEach (clusterhead CH)
CH creates a TDMA schedule for each node to transmit
data
CH communicates the TDMA schedule to each node in the
cluster
EndFor
Steady State Phase:
1.
ForEach (regular node N)
2.
N collects sensed data
3.
N transmits the sensed data to the CH in the
corresponding TDMA time slot
4.
EndFor
5.
ForEach (cluster head CH)
6.
CH receives data from the nodes of the cluster
7.
CH aggregates the data
8.
CH transmits the data to the base station
9.
EndFor
Power-Efficient Gathering in Sensor Information Systems (PEGASIS)
[24]. PEGASIS forms chains of the sensor nodes instead of forming multiple clusters
as performed in LEACH protocol. Each node in the chain can transmit and receive
data from its neighbors. In the entire chain, one node is selected to transmit all the
data received to the sink or base station. The chain construction follows a greedy
approach. The problem of building a chain to minimize the total length is similar to
the traveling salesman problem.
144
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
The elected local leader in the chain waits for data from its closest neighbors.
These neighbors first receive data from their own respective closest neighbors and
aggregate the data before transmitting to the leader. The leader then sends the data
received from its closest neighbors to the sink.
Even though PEGASIS is similar to LEACH, it differs in the following ways. It uses
multihop routing while only one node is selected to transmit to the base station. The
reduction in overhead due to dynamic clustering as in LEACH leads to a performance
gain of almost 100 to 300 percent in PEGASIS. This overhead is reduced when
the following occur: (i) Transmission distances of non-leader nodes are minimized.
(ii) One transmission is made to the sink per round by aggregating all the data. This
reduces local energy consumption but introduces a large delay for nodes farther away
from the leader node of that chain. It also results in a single point of failure by the
bottleneck created at the chain leader. (iii) The number of transmission among the
nodes are reduced leading to overall energy efficiency.
Threshold-Sensitive Energy-Efficient Sensor Network Protocol (TEEN)
[25]. The TEEN protocol is designed to respond to sudden changes in the sensed
attributes and uses a hierarchical model along with a data-centric mechanism. Clusters
are formed in a hierarchical fashion at different levels with elected clusterheads serving
as communication links between each other and the data sink.
Initially, the clusters are formed after which each clusterhead broadcasts two
threshold values to all the nodes. These are hard and soft thresholds for the sensed
attributes. The hard threshold is the minimum possible value of an attribute based
on which the sensor will be transmitting data to the sink. When the sensed value
of the attribute is greater than this threshold, the data are sent to the clusterhead.
This enables the nodes to transmit only relevant data. Once a value above the hard
threshold is sensed, the node checks if the difference in the current and earlier values is greater than the soft threshold; if so, the new data are transmitted. Hard and
soft threshold values can be adjusted per requirements allowing to control the packet
transmissions.
This protocol is succeeded by the adaptive threshold-sensitive energy-efficient
sensor network protocol (APTEEN) [26], which aims at capturing periodic data
collections and respond to time-critical events. While the architecture remains the
same as TEEN, APTEEN supports three types of data query: (1) historical, to analyze
and monitor past data values and take decisions based on these recorded values;
(2) one-time, to take a snap view of the current network situation and visualize it at a
particular time instant; (3) persistent, to monitor the network over a continuous time
interval especially during an event taking place.
Data Aggregation—Exact and Approximate Algorithms [27]. The aim is to
develop better data aggregation and in-network data processing schemes for energy
savings in sensor networks. These lead to lesser packet transmissions and reduce
redundancy, thereby helping in increasing the network lifetime.
SENSOR NETWORKS ROUTING PROTOCOLS
145
The protocol employs a hierarchical model that uses data aggregation and innetwork processing at two different levels of the network hierarchy. A set of nodes,
called the local aggregators (LA), are elected. These form a backbone routing architecture on which the first instance of data aggregation and routing is performed.
From the set of LAs, another set of aggregators, called master aggregators (MA),
are selected to form the second level of nodes to carry out the second level of data
aggregation. Since choosing the optimal set of MAs is an NP hard problem, a separate
integer linear program (ILP) is developed to identify the optimal MA set. The goal
of maximizing network lifetime must be considered during the selection process of
MAs.
The time required to solve the ILP increases exponentially with the number of
LAs. Three near-optimal approximation algorithms, which serve dual purposes of
selecting MAs and employing routing to the external traffic sink, are proposed. The
differences from conventional data aggregation approaches are as follows: (i) Data
aggregation is performed at optimal network points rather than at arbitrary points in
other approaches. The selected points maximize the network lifetime. (ii) Routing
and data aggregation are carried out simultaneously with the proposed algorithms.
(iii) There are no additional overheads due to dynamic clustering and topologies.
The architecture is simple and uses a fixed hierarchy of nodes with no additional
complications.
6.4.5 Multipath Routing
Multipath routing techniques compute multiple paths from source to destination to
effectively route around failed nodes or invalid links. In single-path routing protocols,
if a link fails, additional control packets have to be generated and broadcast to discover
newer paths. Multipath routing avoids such pitfalls, and there is always a secondary
route ready in the event of a route failure.
Meshed Multipath Routing (M-MPR) [28]. Two ways of affecting disjoint multipath routing (MPR) include (i) Disjoint (or split) MPR (D-MPR) with selective
forwarding (SF) in which case each packet is sent along different disjoint routes and
the decision of path selection is made by the source on packet-by-packet basis and
(ii) D-MPR with packet replication (PR) (or limited flooding) where multiple copies
of a data packet are transmitted simultaneously along multiple disjoint routes from a
source to a destination.
A meshed multipath is set up in three steps: (i) acquiring neighborhood information,
(ii) route discovery, and (iii) route reply as described below.
r Acquiring Neighborhood Information: Each active node broadcasts its ID, residual battery power, and location information to local neighbors. For each active neighbor i, a node maintains the following information in its database:
IDi , locationi , residual poweri . Since the sensor nodes are considered stationary, period update on neighborhood status is not needed unless the node is
146
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
S
D
Figure 6.4. M-MMPR: Source-to-destination meshed multipath [28].
entering to sleep mode or has just woken up. In this case, the nodes status is
locally broadcast based on which of the neighborhood tables of nearby nodes
are updated.
r Route Discovery: Each node attempts to form a meshed multipath based on the
neighborhood database and location information of the controller node. So far,
the intermediate node is allowed to accept multiple discovery packets. During
source-to-destination route discovery process, at most two copies of a discovery
packet are accepted by an intermediate node, and the first arrived packet is
forwarded to maximum two downstream neighbors nodes to ensure the reduction
of the receiver complexity and power consumption of a node as can be seen
from Figure 6.4. A maximum two forwarding node is chosen since this allows
an alternate route with minimum possible extra control overhead.
The route packet has the following fields: source ID, source location,
intermediate node ID, next node ID1, next node ID2, destination ID,
destination location, and TTL, where the IDs of forwarding nodes
(next node IDi, i = 1, 2), intermediate node ID, and TTL values are
updated at each intermediate stage.
Each intermediate node maintains the following information in its
routing database: previous node IDi, previous node IDn, next node ID1,
next node ID2. Since several nodes are targeting the same destination, an intermediate node can have more than two “previous node” entries in its routing
table although there will be no more than two “next nodes” as can be seen from
Figure 6.5. The list of “previous node” is bounded since the number of local
c
d
b
y
x
D
Figure 6.5. M-MPR. Meshed topology formed by many-sources-to-a-destination routes [28].
SENSOR NETWORKS ROUTING PROTOCOLS
147
neighbors are finite and no entry is created in the routing table for discovery
packets coming from an upstream neighbor which is already listed in the list.
If an intermediate node that has already forwarded a discovery packet receives
another discovery packet, it updates the “previous node” list in its routing table
and drops the packet.
Entry in the routing table at each node is maintained as a soft state that is deleted
after a timeout unless a reply is received from a controller node. Since most
sensor applications are data-centric, delay differences (jitter) between packet
arrivals is not a big concern. No other resource reservation apart from storing
and maintaining upstream and downstream nodes information is made during
this phase; therefore, the route discovery phase can be considered as a topology
construction process.
r Route Reply: Route reply message identifies the nodes comprised the meshed
path. The controller node, upon receiving the discovery packets from a single
source, selects the first two and sends a route reply following the original links
by the route discovery packets in reverse direction with the following fields:
source ID, source location, intermediate node ID, previous node ID1,
previous node ID2. Each intermediate node changes the states of its corresponding entries from soft to permanent for the duration of its active
participation, updates the fields of the reply packet other than the source information, and forwards the reply packet to its upstream node. When forwarding
the route reply message, the node does not require the knowledge of source
information.
In case of the discovery packets arriving to controller node from several sensor
nodes, multicast reply is used. If an intermediate node is out of service or goes
to sleep mode, the upstream nodes select necessary neighbors to sustain connectivity. Intermittent “link breakage” will not trigger reconfiguration of meshed
multipath, instead it is handled using selective forwarding. In the constructed
meshed topology, the number of downstream links is no more than two, whereas
the number of upstream nodes can be more. As can be observed from Figure 6.5,
node n has three upstream nodes (a, b, and c) and two downstream nodes (x
and y).
After the meshed multipath is constructed, the information packet is forwarded from the source to the destination using either the packet replication
(PR) or selective forwarding (SF) variants. The methods are explained below in a
nutshell.
r In PR, a source packet is copied from along all possible paths to its destination.
To reduce power consumption due to transmission of multiple copies of the same
packet, there is a provision of discarding the packets if a node receives more than
one correct copy of the packet from one of the upstream nodes to be forwarded
to the downstream node.
148
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
r In the SF variant of M-MPR, if more than one downstream node is available at
either the source or an intermediate node, the information packet is forwarded
to only one of the nodes based on local conditions.
In addition to a fault tolerance objective, the selective forwarding approach along
meshed multipath offers more efficiency than PR in terms of resource utilization
and congestion avoidance. The main difference between PR and SF is that in PR,
the packet is intended for multiple neighbors in which each of them will receive
and forward the packet whereas in SF, only one receiver will receive and forward
the packet. Due to broadcast nature, M-MPR requires less transmission energy than
D-MPR; MPR provides more flexibility in selective forwarding decisions than
D-MPR, resulting in more successful packet delivery rate.
ALGORITHM 3. Pseudocode Description of the Meshed Multipath Routing Algorithm
Setup phase:
1.
2.
3.
4.
ForEach (node N)
N broadcasts node information
N listens to broadcast packets of neighboring nodes
N adds the received information to the neighborhood
database
5. EndFor
6. ForEach (node N)
7.
N forwards a route discovery packets to a maximum of
2 hops
8.
N listens to route discovery packets and sends route
reply packets
9.
N listens to route reply packets and sets up the
meshed multipaths.
10. EndFor
Operating phase:
1.
2.
3.
4.
5.
6.
6.
7.
8.
ForEach (node N)
Select forwarding technique T.
If T is packet replication
N forwards data to all possible paths to the
destination
ElseIf T is selective forwarding
N chooses a path to destination based on some criteria
N forwards the data on the selected path
EndIf
EndFor
SENSOR NETWORKS ROUTING PROTOCOLS
149
Highly Resilient, Energy-Efficient Multipath Routing in Wireless Sensor
Networks [29]. Energy-efficient multipath routing defines a set of localized algorithms to construct multiple paths from the source to the destination (Algorithm
3). These multiple paths may come in handy if one particular route fails during the
network operation. Multipaths are constructed between nodes using two different approaches. The first is the original node disjoint paths where none of the paths created
intersect with each other, meaning that they are completely disjoint. This ensures
that for k paths constructed, no k node failures can eliminate all the paths. The second technique defines the braided paths in which there are generally no completely
disjoint paths, but instead numerous partially disjoint alternative paths.
Conventional approaches to energy-efficient and robust routing have resulted in
periodic flooding of the network, which in turn lead to energy inefficiency. By constructing alternate paths initially, there may be no need for periodic flooding of the data
packets. In the event of the primary path failure along with all the other alternate paths
simulatenously, there is no other option left but to flood the network to ensure that the
data reach the destination. However, such a condition should not occur very frequently.
The two mechanisms for identifying disjoint paths and braided paths are as follows:
r Disjoint Multipaths: These can be constructed by identifying a few alternate
paths that are node disjoint with the primary path. Here, initially the primary
path is selected between the source and the sink. This is followed by selecting
the first alternate path as the best path, which is node disjoint with the primary
path. Other secondary paths are selected based on the first disjoint path, and
this process continues until the number of alternate paths reach the maximum
specified limit. These disjoint paths are resilient, but generally energy-inefficient.
See Figure 6.6 [29].
r Braided Multipaths: In this technique, the alternate paths created are not completely node disjoint from the primary path, but instead have a few nodes in
common. Selecting a braided path is similar to selecting disjoint paths; however,
each node on the primary path receiving a reinforcement propagates it to its most
preferred neighbors, who in turn carry out the same operation. This ultimately
leads to multiple alternate paths, some of which can be parts of the primary path
and are not completely node disjoint. See Figure 6.7 [29].
In this system, two different modes are considered: isolated and patterned node
failures. In the first case, each node has an individual probability of failure whereas
in the second case, all nodes within a particular radius fail simultaneously.
ReInForM [30]. The reliable information forwarding using multiple paths (ReInForM) protocol brings to the forefront issues relating to information aware data delivery in sensor networks. It may happen that essential data are sent along a lossy
and unreliable route. Let us look at an example in which a sensor network senses
temperature at some points in the forest. On a given day, a sensor may sense a temperature of 60◦ F at a particular point that is normal. On the other hand, at that very
150
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
Source
Sink
Sink
Source
Primary-path
reinforcement
Low-rate samples
(a)
(b)
Source
Sink
Source
Sink
B
P
alternate-path
reinforcement
negative
reinforcement
A
P1
(c)
X
(d)
Q
Source
Sink
B
P
A
P1
(e)
Figure 6.6. Construction of localized disjoint paths. Diagram (a) presents the low-rate date
sample. Primary path P and the alternate path negative reinforcement are shown in diagrams
(b) and (c), respectively. The alternate path P1 and the idealized algorithm are given in diagrams
(d) and (e) [29].
moment, another sensor placed some distance away may be sensing a temperature
of 1000◦ F. This packet containing the abnormal temperature is the more important
packet and should be sent to the sink by more reliable links to ensure delivery with
the lowest latency. Basically, the proposed mechanisms support delivery of data at
a(i)
a(i)
a(i-1)
a(i-1)
Source
Sink
n(k+1)
n(k)
n(k-1)
n(k-2)
a(i-2)
a(i-3)
(a)
n(k-3)
n(k+1)
n(k)
n(k-1)
n(k-2)
n(k-3)
a(i-2)
(b)
Figure 6.7. Braided multipaths. Diagrams (a) and (b) present idealized and localized braided
multipaths [29].
SENSOR NETWORKS ROUTING PROTOCOLS
151
any desired reliability. However, with the increase in reliability, one expects an increase in the overhead required to transfer the packet. Multiple paths from source
to destination are discovered while reliability is ensured by redundant copies of a
packet. The desired reliability, the local channel error conditions, and the neighborhood information at each node determines the degree of redundancy needed. Another
key feature is that data caching is not required at intermediate nodes, thereby saving
memory.
In multihop sensor networks where the source is located far away from the sink
with high number of channel error occurences, a simple packet forwarding strategy
may become unreliable. To improve the reliability, two solutions have been proposed:
(i) No acknowledgments—sending data from source to sink along the shortest path
with no acknowledgments. Multiple copies of the packet are sent along the same path
to increase the probability of packet delivery. (ii) With acknowledgments—sink sends
back an acknowledgment to the source after receiving the data packet. The simulation
results show that the acknowledgement-based scheme has similar overhead to the nonacknowledgment-based scheme. Hence, the non-acknowledgment scheme is used in
the protocol description.
ReInForM is based on the Dynamic Packet State (DPS) [31] approach used for
data networks. The source uses local knowledge of the network such as channel
error, hop count to sink, and so on, and sends multiple copies of the data packets through multiple paths. With every hop, network conditions are recorded in
the packet header and forwarded to the next node. Every intermediate node takes
forwarding decisions based on this stored information. This strategy enables the forwarding node to route the packets according to local network conditions, but without maintaining the state information on its own. It also helps in making localized
decisions on packet forwarding since network characteristics at an earlier point on
the path may be much different from those at the current point. This mechanism
adapts to channel errors and topological deviations while maintaining a reasonable
overhead.
6.4.6 QoS-Based Protocols
Routing is generally carried out based on metrics, prominent among which are
control packet overhead and energy efficiency. However, additional quality of service
(QoS) parameters may also be considered to facilitate more efficient routing in
sensor networks.
Stream-Enabled Routing (SER) [32]. SER protocol allows sources to choose
the routes based on the instruction (or task) provided by the sinks. An instruction
is a predefined identifier value. Therefore, only an identifier is sent rather than the
attribute list, resulting in memory conservation. It takes into account the available
energy of the sensor nodes, QoS requirements of the instruction, memory limitation
of nodes, and the localized effect of dense nodes. Sinks can give new instructions to the
152
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
sources without establishing another path. SER uses four types of messages: (i) scout
message (S-message), (ii) information message (I-message), (iii) neighbor–
neighbor message (N-message), and (iv) update message (U-message).
An S-message is used during the source discovery to determine sources that
will process the instruction (or task) specified in the S-message. Sources decide
the type and level of the routes needed by the instruction. There are four types of
routes, each with two levels (i.e., level 1 and level 2). The µ value is the radius of the
level 2 routes.
Stream is identified by both the type and the level of route. Each level 2 stream
includes level 1 stream. In level 2, the size of the radius µ of the stream can be
determined based on QoS specified in the instruction. Combination of types and levels
creates different kinds of QoS for a stream. After streams are chosen, the source sends
an N-message to establish the streams back to the sink. The repairs of streams are
accomplished via N- and S-messages. Once the streams are accomplished, data
travels from sources to the sink through either level 1 or level 2 stream with an
I-message. The sink can update the instruction (or task) at the sources through
either level 1 or level 2 stream using a U-message. Both sources and sink can
terminate the streams using U-message.
SER has seven phases:
1. Source Discovery. A sink broadcasts an S-message to find routes from sink to
source. S-message contains the following fields: TID, NAP, LID, NH, AE.
Table 6.1 represents the parameters used throughout the algorithm description.
The average energy of a node is given by the following formula:
AE =
NHi−1 × AEi−1 + Ei
NHi−1 + 1
The TID field is further composed of LI, MT, INS, and TLOC fields. When
a sensor node receives an S-message, it determines if the instruction, (INS),
is intended for the node. If the INS in the S-message is not intended
for the node, the node stores the fields of S-message in a connection-tree
(C-tree). A C-tree is a logical tree that represents possible connections through
the node. It maintains the node’s neighbors that can participate in a routing back
to sink.
A sensor node in an established route knows the LID values of both uplink
and downlink nodes Initially, DLID and ULID values are not set, and DSP
and NS values are set to OFF. Updated values of AE, NH, and LID fields of
an S-message are broadcast to the neighbors.
If a sensor node receives the same S-message from its neighbors, it dismisses it. The sources store an S-message in a task-tree (T-tree). T-tree has
x DLID values since source can select up to x LIDs to route I-message
back to the sink based on QoS requirement. The max value of x is the number
of neighbor nodes. Each DLID value corresponds to a DSP indicator. In the
T-tree, the leaf nodes has no ULID and NS indicator since the sources are the
SENSOR NETWORKS ROUTING PROTOCOLS
153
TABLE 6.1. SER Protocol Parameters
TID
Task ID
NAP
LID
NH
AE
LI
MT
Network access point (represents a unique sink)
Local ID
Number of hops from the sink
Average energy of a route
Length indicator
Message type, (MT = 0 (S-message); MT = 1 (I-message);
MT = 2 (U-message); MT = 3 (N-message)
Instruction
Targeted location
Downlink sensor problem
Node selected
LID of downlink sensor node
LID of uplink sensor node
Selected ID
Message
Flow indicator (message going uphill or downhill)
Current number of hops
Description
New INS
INS
TLOC
DSP
NS
DLID
ULID
SLID
MES
FI
CNH
Payload
NINS
destination of an S-message. A source can receive an x S-message since it has
x neighbors. The route associated with the first received S-message is considered the shortest route. Sources select a neighbor node to send an I-message
back to sink based on the QoS requirement of INS.
2. Route Selection. Once the sources receive the S-message, they determine
the QoS requirement of task in the S-message. There are four types of stream
for communication between sources and sinks, and each stream can be either
level 1 or level 2:
r Type 1: Time critical but not data critical
r Type 2: Data critical but not time critical
r Type 3: Not time and data critical
r Type 4: Not time and data critical
After the sources select the neighbor nodes, the sources broadcast an
N-message to their neighbors indicating the level and size of the stream.
N-message contains TID, NAP, LID, SLID, and MES fields. If the stream
is level 1, µ = 0 (width of the stream). At level 1, messages are routed back to
the sink via hop-by-hop communication. Messages are sent to only one node.
The Level 2 stream contains the level 1 stream, which serves as a backbone in
setting up the level 2 stream. The value of µ is the number of hops away from
the nodes in the level 1 stream. Messages can flow downhill to the sink or uphill
154
3.
4.
5.
6.
7.
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
to the sources by flooding through only the nodes that are part of the stream.
An I-message flows downhill from sources to sink by using the NH value
stored in each node in the C-tree. The nodes near the sources have higher NH
values. U-message flows uphill from sink to sources by using the negative of
the NH value. The nodes near the sources have higher negative NH values.
Route Establishment. An N-message is used by a sensor node to inform
neighbors about its local information. The source sends an N-message to
establish a stream back to the sink. Sensor nodes that are not part of a stream
delete all data associated with the N-message from the C-tree. If intermediate
nodes between the sources and sinks have not received an N-message in
response to an S-message in a set time interval, the sensor node deletes the
C-tree branch that is associated with the S-message. After the N-message
arrives to the sink, the minimum delay or maximum average energy stream
is established. Sources can start sending I-messages to the sink. An Imessage contains TID, FI, CNH, and Payload fields.
I-Message Transmission. The neighbor nodes can determine if they need
to route the I-message by the TID since each neighbor nodes maintain a
C-tree. When a source broadcasts an I-message, it sends a CNH field with
the value from a T-tree. Intermediate nodes between sources and sink use a
C-tree. FI and CNH fields are only used when the stream is level 2. Each
node only rebroadcasts once to avoid a node from broadcasting the same
message over again. After an I-message is received, the sensor nodes turn off
the receiver for some amount of time if the sleep mode operation is on such that
the node can avoid listening neighbors broadcasting the same I-message.
The C-tree indicates which instructions the sensor nodes need to route.
Route Reconnection. If a sensor node is low on energy or there is too much
noise around when transmitting at level 1, it can broadcast an N-message
by setting up a reconnect message indicator. Once the neighbors receive an
N-message, they check their C-tree to decide if there are possible alternate
routes. An N-message will be broadcasted until the alternate route is found.
Sudden death of route: If the stream suddenly terminates, the sink cannot
get the I-messages. The sink sends out a new S-message with higher
QoS requirement version of the same instruction (higher QoS INS value). New
streams can be found to avoid broken paths. Multiple streams of level 2 can be
set up between source and sink to improve robustness of I-message routing.
Instruction Update. A U-message allows a sink to update its instruction
to the sources. The U-message from the sink to the sources flows uphill
while it flows downhill from sources to the sink when streams are level 2. A
U-message contains TID, FI, CNH, and NINS fields.
Task Termination. A task at the sources are terminated in two ways:
(a) Sources have finished the task associated with the instruction given by sink.
A U-message with the task completed instruction indicator is broadcast
by sources.
SENSOR NETWORKS ROUTING PROTOCOLS
155
(b) A sink decides to terminate the instruction. A U-message with the task
termination instruction indicator is set by the sink.
The streams are torn down by removing C-tree branches at the intermediate nodes
and removing a T-tree at the sources.
Algorithm for Robust Routing in Volatile Environments (ARRIVE)
[33]. This is a probabilistic algorithm and makes packet forwarding decisions based
on localized information, and it has a tree-like topology rooted at the sink of the
network. The forward approach is employed to achieve end-to-end reliability. The
packet loss is avoided by sending multiple packets of the single event. Basically, the
three sources of packet loss expected are (i) isolated link, (ii) patterned node failures,
and (iii) malicious or misbehaving nodes. Figure 6.8 presents the overview of the
ARRIVE algorithm.
Waiting for
packet
No
packet
received
passively
process
packet
No
is packet
for me?
Yes
Yes
Neighbor
check threshold
conditions
where to
send
packet?
check threshold
conditions
Parent
check threshold
conditions
send packet
Figure 6.8. Overview of Arrive protocol [33].
156
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
r Event: Identified by [SourceID, EventID]
r Level: Each node has unique level indicating distance from source to sink
(in terms of hops)
r Parents: Nodes one level closer to the sink
r Neighbors: Nodes on the same level and be able hear each other
r Push: Push packet to one of the neighbors
r Forward: Forward packet to one of the parents
r Forwarding Probability: Included in the packet header and used to probabilistically select whether to push or forward
r Reputation History: Each node keeps this information for each of its parents
and neighbors
r Convergence: Prevents multiple packets of the same event being sent to same
source of failure
ARRIVE achieves diversity in paths in two ways: (i) Upon receiving a packet, the
next hop is selected probabilistically based on link reliability and node reputation and
(ii) when more than two or more packets of the same event are processed, these packets
are ensured to follow different outgoing links. Each keeps the following information:
level, neighbors list, parents list, reputation history of neighbors and parents, and
convergence history of specific events.
Several assumptions are made about the network and the algorithm. The network,
which is almost considered a static network, is assumed to be dense enough for
sufficient multiplicity of paths to exist between sources and sink for the algorithm
to perform well. Sensors used are considered to have a low per-node cost. Routes
used by the packets are unlikely to be optimal due to the probabilistic nature of the
algorithm. Messages flow only from nodes to one and only sink.
The breadth first search rooted at the sink is used to initialize level, parents, and
neighbors state information at each node. When a node hears a packet, it checks to
see if the packet is addressed for it. If so, threshold processing takes place. Nodes are
filtered by their reputation and convergence history of the neighbors and parents. A
decision needs to be made to either choose to forward the packet to a parent or push
it to one of its neighbors with the probability value found in the packet header. This
is randomly determined by the forwarding probability function Pr(f ). Each node is
weighed by their reputation. The destination is randomly selected from the rest of the
nodes (since bad reputation nodes are eliminated). If the the packet is forwarded to
one of the parents, Pr(f ) is not changed; however, its value is increased.
6.5 CONCLUSIONS
In this chapter, we have discussed various routing protocols in sensor networks. The
common goals of designing a routing algorithm is not only to reduce control packet
overhead, maximize throughput, and minimize the end-to-end delay, but also to take
TABLE 6.2. Comparison of Sensor Network Routing Protocols
Protocol
Directed diffusion
EAD
RUMOR
Youssef et al.
LEACH
PEGASIS
TEEN
APTEEN
Al-Karaki et al.
M-MPR
Ganesan et al.
ReInForM
SPEED
Seada et al.
Subramanian et al.
Rao et al.
SER
ARRIVE
SAR
GRAdient
MCFA
Classification
Mobility
Attribute-based
Attribute-based
Attribute-based
Attribute-based
Hierarchical
Hierarchical
Hierarchical
Hierarchical
Hierarchical
Multipath
Multipath
Multipath
Geographical
Geographical
Geographical
Geographical
QoS
QoS
Flat
Flat
Flat
Restricted
Virtual backbone
Very restricted
Cluster based
Fixed base station
Fixed base station
Fixed base station
Fixed base station
No
No
No
Limited
No
No
No
Yes
No
No
No
No
No
Power
usage
Limited
Limited
—
Low
Highest
Highest
Highest
Highest
—
—
High
—
—
High
—
—
—
—
—
High
Low
Negotiation
based
Data
Aggregation
Yes
No
No
No
No
No
No
No
Yes
No
No
No
No
—
—
No
Yes
Yes
Yes
No
No
Yes
Yes
Yes
Yes
Yes
No
Yes
Yes
Yes
No
No
Yes
No
Yes
No
—
No
Yes
Yes
No
No
Scalability
Restricted
Restricted
High
Limited
High
High
High
High
High
Restricted
High
Restricted
Restricted
Restricted
Restricted
High
Restricted
Restricted
Restricted
High
High
157
158
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
into consideration the energy consumption, especially in a sensor network comprised
of nodes that are considered lightweight with limited memory and battery power.
We have divided the sensor routing protocols into six categories: (i) attribute-based,
(ii) flat, (iii) geographical, (iv) hierarchical, (v) multipath, and (vi) QoS-based. We
have then compared the protocols belonging to the same category based on various
characteristics.
6.6 EXERCISES
1. List the characteristics of the directed diffusion protocol which makes it considered both a flat and attribute-based sensor routing protocol (see Table 6.2).
2. For which type of sensor applications is the EAD protocol most suited? Detail
some of the factors by which EAD helps in providing better services to these
applications.
3. On what basis are the local cluster heads selected in the LEACH protocol?
Describe the cluster-head selection procedure in detail. Why are cluster heads
not selected at the initial network deployment time itself?
4. How do PEGASIS and LEACH protocols differ?
5. In the ReInForM protocol, determine the reasons why the No ACK scheme
has a similar control packet overhead to the ACK scheme, since at first glance
it would logically seem that the scheme without the ACK packets would result
in much less control overhead.
6. The SPEED protocol is characterized by minimal memory requirements. What
are the probable reasons for such a characteristic?
7. The ARRIVE protocol uses a tree-based routing approach. Discuss various
advantages derived by using such a routing scheme.
BIBLIOGRAPHY
1. D. E. Culler, D. Estrin, and M. B. Srivastava. Guest editors’ introduction: Overview of
sensor networks. IEEE Computer, 37(8):41–49, 2004.
2. G. J. Pottie and W. J. Kaiser. Wireless integrated network sensors. Communications of
ACM, 43(5):51–58, 2000.
3. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks:
A survey. Computer Networks, 38(4):393–422, 2002.
4. W. Ye, J. S. Heidemann, and D. Estrin. Medium access control with coordinated adaptive
sleeping for wireless sensor networks. IEEE/ACM Transactions on Networking, 12(3):493–
506, 2004.
5. V. Rajendran, K. Obraczka, and J. Garcia-Luna-Aceves. Energy-efficient collision-free
medium access control for wireless sensor networks. In Proceedings of SenSys, 2003,
pp. 181–192.
BIBLIOGRAPHY
159
6. D. Estrin, R. Govindan, J. S. Heidemann, and S. Kumar. Next century challenges: Scalable
coordination in sensor networks. In Proceedings of MOBICOM, 1999, pp. 263–270.
7. S. Tilak, N. B. Abu-Ghazaleh, and W. R. Heinzelman. A taxonomy of wireless micro-sensor
network models. ACM Mobile Computing and Communications Review, 6(2):28–36, 2002.
8. J. N. Al-Karaki and A. E. Kamal. Routing techniques in sensor networks: A survey. IEEE
Communications, 11(6):6–28, 2004.
9. K. Akkaya and M. Younis. A survey on routing protocols for wireless sensor networks. Ad
Hoc Networks Journal, 3:325–349, 2005.
10. C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed diffusion: A scalable and robust communication paradigm for sensor networks. In Proceedings of MOBICOM, 2000,
pp. 56–67.
11. A. Boukerche, X. Cheng, and J. Linus. Energy-aware data-centric routing in microsensor networks. In MSWIM ’03: Proceedings of the 6th ACM International Workshop on
Modeling Analysis and Simulation of Wireless and Mobile Systems, 2003, pp. 42–49.
12. M. A. Youssef, M. F. Younis, and K. Arisha. A constrained shortest-path energyaware routing algorithm for wireless sensor networks. In Proceedings of WCNC, 2002,
pp. 794–799.
13. D. Braginsky and D. Estrin. Rumor routing algorithm for sensor networks. In Proceedings
of WSNA, 2002, pp. 22–31.
14. F. Ye, G. Zhong, S. Lu, and L. Zhang. Gradient broadcast: A robust data delivery protocol
for large scale sensor networks. Wireless Networks, 11(3):285–298, 2005.
15. F. Ye, A. Chen, S. Liu, and L. Zhang. A scalable solution to minimum cost forwarding in
large sensor networks. In Proceedings of the Tenth International Conference on Computer
Communications and Networks (ICCCN), 2001, pp. 304–309.
16. K. Sohrabi, J. Gao, V. Ailawadhi, and G. J. Pottie. Protocols for self-organization of a
wireless sensor networks. IEEE Personal Communications Magazine, 7(5):16–27, 2005.
17. K. Sohrabi. On low power wireless sensor networks. Ph.D. thesis, Electrical and Computer
Engineering Department, UCLA, June 2000.
18. J. L. Gao. Energy efficient routing for wireless sensor networks. Ph.D. thesis, Electrical
and Computer Engineering Department, UCLA, June 2000.
19. T. He, J. A. Stankovic, C. Lu, and T. F. Abdelzaher. SPEED: A stateless protocol for
real-time communication in sensor networks. In Proceedings of ICDCS, 2003, pp. 46–58.
20. A. Rao, C. Papadimitriou, S. Shenker, and I. Stoica. Geographic routing without location
information. In Proceedings of MOBICOM, 2003, pp. 96–108.
21. K. Seada, M. Zuniga, A. Helmy, and B. Krishnamachari. Energy-efficient forwarding
strategies for geographic routing in lossy wireless sensor networks. In Proceedings of
SenSys, 2004, pp. 108–121.
22. S. Subramanian and S. Shakkottai. Geographic routing with limited information in sensor
networks. In Proceedings of IPSN, 2005, pp. 269–276.
23. W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan. Energy-efficient communication protocol for wireless microsensor networks. In Proceedings of HICSS, 2000.
24. S. Lindsey and C. Raghavendra. PEGASIS: Power-efficient gathering in sensor information
systems. In Proceedings of IEEE Aerospace Conference, 2002, pp. 1125–1130.
25. A. Manjeshwar and D. P. Agrawal. TEEN: A routing protocol for enhanced efficiency in
wireless sensor networks. In Proceedings of IPDPS, 2001, pp. 2009–2015.
160
A TAXONOMY OF ROUTING PROTOCOLS IN SENSOR NETWORKS
26. A. Manjeshwar and D. P. Agrawal. APTEEN: A hybrid protocol for efficient routing and
comprehensive information retrieval in wireless sensor networks. In Proceedings of IPDPS,
2002, pp. 195–202.
27. J. N. Al-Karaki, R. Ul-Mustafa, and A. E. Kamal. Data aggregation in wireless sensor
networks - exact and approximate algorithms. In Proceedings of IEEE Workshop on High
Performance Switching and Routing (HPSR), 2004, pp. 241–245.
28. S. De, C. Qiao, and H. Wu. Meshed multipath routing with selective forwarding: An
efficient strategy in wireless sensor networks. Computer Networks, 43(4):481–497, 2003.
29. D. Ganesan, R. Govindan, S. Shenker, and D. Estrin. Highly-resilient, energy-efficient
multipath routing in wireless sensor networks. ACM Mobile Computing Communication
Review, 5(4):10–24, 2001.
30. B. Deb, S. Bhatnagar, and B. Nath. ReInForM: Reliable information forwarding using
multiple paths in sensor networks. In Proceedings of LCN, 2003, pp. 406–415.
31. I. Stoica, S. Shenker, and H. Zhang. Stateless fair queueing: Achieving approximately
fair bandwidth allocations in high speed networks. In Proceedings of SIGCOMM, 1998,
pp. 118–130.
32. W. Su and I. F. Akyildiz. A stream enabled routing (SER) protocol for sensor networks. In
Proceedings of Med-hoc-Net, 2002.
33. C. Karlof, Y. Li, and J. Polastre. ARRIVE: Algorithm for robust routing in volatile environments. Technical Report UCB/CSD-03-1233, Computer Science Division, UC Berkley,
May 2003.
CHAPTER 7
Clustering in Wireless
Sensor Networks: A Graph
Theory Perspective
NIDAL NASSER and LILIANA M. ARBOLEDA
Department of Computing and Information Sciences, University of Guelph, Guelph, Ontario
N1G 2W1, Canada
7.1 INTRODUCTION
Wireless Sensor Networks (WSNs) consist of tiny sensing devices that are spread
over a large geographic area and can be used to collect and process environmental
data such as temperature, humidity, light conditions, seismic activities, images of the
environment, and so on. These data can be used to detect certain events and to trigger
activities. For example, sensors distributed over large woodland could automatically
raise an alarm if a fire has broken out somewhere, or sensors distributed over a large
farmland could trigger irrigation if the ground of a field is not moist enough.
A WSN is usually comprised of a large number of sensors that are physically
small, communicate wirelessly among each other, and are deployed without prior
knowledge of the network topology. Due to the limitation of their physical size,
the sensors tend to have storage space, energy supply, and communication bandwidth so limited that every possible means of reducing the usage of these resources
is aggressively sought. The use of the WSN potential will provide efficient and
costs-effective solutions for several problems. However, it is necessary to implement mechanisms or procedures to deal with the sensor constraints. The hierarchical
organization of the sensors, grouping them and assigning those specific tasks into
the group before transferring the information to higher levels, is one the mechanisms proposed to deal with the sensors limitations and is commonly referred to as
clustering.
The creation of clusters in a WSN field is generally done by taking into account
the proximity between the sensors, measured through the radio-frequency signal they
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
161
162
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
emit. Each cluster has a cluster head, which is the node that directly communicate
with the sink (base station) for the user data collection. By assuming roles within
a cluster hierarchy, the nodes in a WSN can control the activities they perform and
therefore reduce their energy consumption. However, the election of when to act
as a simple data provider (saving energy) and when to act as a gateway (cluster
head) between the nodes and the base station is a difficult problem. To make this
decision, it is necessary to take into account several aspects including power level
signal, transmission schedules, nodes localization, and networking function. Clustering helps in solving some of the sensors’ constraints by reducing the cost of transmitting data to base stations, reducing the power consumption in the devices, facilitating
the gathering of sensed data, maximizing the routing process execution, and allowing
scalability.
One of the techniques used to create the clusters and represent the topology of
the WSN involves the use of Graph Theory concepts to represent the topology of the
sensors in the field. Graph theory can be used to create the sensor clusters and can
help in identifying the cluster head. Sensor network can be represented by a graph G =
(V, E), where the vertices (V ) represent the sensors and the set of links (E) represents
the connections between vertices if the they are within the transmission range of each
other, as shown in Figure 7.1.
In this chapter we will study different existing approaches to (a) create clusters
and select the cluster head using Graph Theory concepts and (b) identify if they
can be used in either static WSN or mobile WSN. In mobile WSN, the dynamic
changes in the topology of the clusters creates additional challenges. In this situation it is necessary to implement mechanisms to control the changes in the cluster
graph.
Sensors Field
{G=(V,E)}
14
2
1
15
13
3
4
17
12
Transmission Link 18
{ (u,v) ε E}
11
5
6
16
10
Sensors
{u,v ε V}
7
19
9
8
20
Figure 7.1. Graph representation of a WSN.
FUNDAMENTAL CORRESPONDING BETWEEN WIRELESS SENSOR NETWORKS
163
7.2 FUNDAMENTAL CORRESPONDING BETWEEN
WIRELESS SENSOR NETWORKS AND GRAPH THEORY
A wireless sensor network is a set of sensors deployed in a sensor field to monitor
specific characteristics of the environment, to measure those characteristics, and to
collect the data related to those phenomena. The sensors are small devices with limited
resources: limited battery power, low memory, little computing capability, very low
data rates, low bandwidth processing, variable link quality, and so on. However,
despite their constraints, when the sensors are deployed in large numbers, they can
provide us with a very real picture of the field being sensed. WSN can provide an area
coverage that was not possible with other wired and wireless networks. They can be
deployed in different environments and can be permanently attended or can be left
unattended once they have been deployed in the field.
The use of the WSN potential will provide efficient and cost-effective solutions for
many problems. However, it is necessary to implement mechanisms or procedures to
deal with the sensor constraints. The use of clustering techniques has been proposed
to help solve some of those constraints, by allowing the organization of the sensors in
a hierarchical manner, grouping them into clusters and assigning a specific task to the
sensor in the clusters, before moving the information to higher levels. The concept
of clustering is very useful in different contexts of WSN. Clustering is a fundamental
mechanism to design scalable sensor network protocols. In general terms, clustering is
the classification of similar objects into different groups or subsets. The formed subsets
in some sense belong together, because they share one or more similar characteristics
or behaviors. Examples of such common characteristics could be: proximity according
to some defined distance measure, similar behaviors, common data patterns, and so
on. In the most general problem the number of clusters or groups is unknown, as are
the properties that make them similar.
Clustering techniques have been proposed in wireless networks in order to achieve
high energy efficiency and assure long network lifetime, for bandwidth reuse, for
data gathering [1] and target tracking [2], one-to-many, many-to-one, one-to-any, or
one-to-all communications, routing [3–6], and so on. Clustering is particularly useful
for applications that require scalability to hundreds or thousands of nodes. Scalability
in this context implies the need for load balancing, efficient resource utilization, and
data aggregation [7]. Also, many routing protocols can use clustering to create a
hierarchical structure and minimize the path cost when communicating with the base
station. In many sensors network applications where data collection and processing
can be done in situ, this hierarchical approach is a promising method for efficiently
organizing the network. Also, many signal processing algorithms used for extraction
of final information from the data gathered by the sensors are well-suited for local
processing of data within the clusters.
Graph Theory concepts can be used to describe, analyze and represent a WSN in
a very clear way. Several of these concepts refer to structures and algorithms that had
been previously used to address other aspects like topology management, localization
techniques and routing, not only in WSN but also in other types of wireless and wired
networks [8–11]. For example, the OSPF (Open Shorted Path First) routing algorithm
164
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
used for routing in wired LANs implements the SPF (Shorted Path First) algorithm
or Dijkstra’s algorithm, which solves the single-source shortest path problem for a
directed graph with nonnegative edge weights.
Graph data structures and algorithms can easily represent the network with stationary wireless sensor. The construction of basic graph structures like trees, cliques
or dominating sets, and the use of graph algorithms like Breadth First Search (BFS)
or Depth First Search (DFS) will help in the construction of clusters to improve the
communication in the WSN. However, in mobile WSN, the dynamic change of the
network topology due to the sensor movement creates additional challenges when
forming the clusters. In this situation it is necessary to implement additional mechanisms to control the changes in the cluster graph.
In this section, we give the basic concepts and definitions that provide a detailed
overview of the corresponding relation between clustering in WSN and graph theory,
focusing on concepts and definitions that are important for the understanding of the
material to follow.
7.2.1 Wireless Sensor Networks and Graph Theory Concepts
A cluster in WNS consists of three main different elements: sensor nodes (SNs), base
station (BS), and cluster heads (CH); see Figure 7.2. The SNs are the set of sensors
present in the network, arranged to sense the environment and collect the data. The
main task of an SN in a sensor field is to detect events, perform quick local data
processing, and then transmit the data. But the greatest constraint it has is the power
consumption, which usually is caused when the sensor is observing its surroundings,
SN
CH
SN
CH
SN
SN
SN
CH
CH
SN
SN
SN
SN
SN
BS
CH
CH
SN
SN
SN
Figure 7.2. Elements in a clustered WSN.
FUNDAMENTAL CORRESPONDING BETWEEN WIRELESS SENSOR NETWORKS
165
and communicating (sending and receiving) data. The BS is the data processing point
for the data received from the sensor nodes, and where the data are accessed by the
end-user. It is generally considered fixed and at a far distance from the sensor nodes.
The CH acts as a gateway between the SNs and the BS. The function of the cluster
head is to perform common functions for all the nodes in the cluster, like aggregating
the data before sending it to the BS. In some way, the CH is the sink for the cluster
nodes, and the BS is the sink for the cluster heads. This structure formed between the
sensor nodes, the sink, and the base station can be replicated as many times as it is
needed, creating the different layers of the hierarchical WSN.
The SNs and the communication links between them can be represented by an
undirected graph G = (V, E), where each vertex v ∈ V (the set of vertices in the
graph) represents a sensor node with a unique ID. An edge (u, v) ∈ E (the set of
edges in the graph) represents a communication link if the corresponding nodes u and
v are within the transmission range of each other.
The graph is formed by defining the neighborhood of each node. The neighborhood
of each one of the nodes in the network and the k-neighborhood of the nodes can be
defined as follows:
Definition 1. Node’s Neighborhood. The neighborhood N(v) is the set of nodes
(neighbors) that reside within the circular transmission range of node v, which means
the vertices adjacent to v. If v is included into the neighborhood, it is called closed
neighborhood of v and it is represented N[v].
Definition 2. k-Neighborhood. The k-neighborhood of v, Nk (v), is the set of nodes
with distance at most k from v.
Nk (v) = {u|u ∈ V ∧ d(u, v) ≤ k}
Following the creation of the graph, by defining the adjacent nodes in the network
and the communication links between nodes, it is possible to determine which nodes
are reachable from a specific node and it allows us to calculate the correspondent hop
distance (see Definition 3) between any source and target nodes. It also allows us to
determine the kth power graph of a node (see Definition 4), to limit the number of
nodes that will be considered among the node’s transmission radius when creating
the clusters.
Definition 3. Hop Distance. The shortest path between two nodes u and v is the path
with the minimum number of hops between them. The distance d(u, v) is the number
of hops in the shortest path between u and v.
Definition 4. kth Power Graph. The kth power graph of G, Gk = (V, Ek ), is the
graph between the nodes in V and an edge between every pair of nodes u, v ∈ V ,
such that d(u, v) ≤ k in G.
166
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
Knowing the existent links between the nodes, the next process involved in the
communication is determining the routes available to send information through
the WSN. One of the most common mechanisms used for message delivery inside
the network is based on a spanning trees structure, defined below, that allows the use
of algorithms like Depth First Search (DFS) and Breadth First Search (BFS) to send
messages along the created graph in linear time. We use T to represent the spanning
tree of Gc , rooted at y. The ith level of the tree is the set of nodes with hop distance
equal to i from y. The depth of the tree, depth(T ), is the index of the farthest level in
the tree.
Definition 5. Spanning Tree. A spanning tree is a connected and undirected graph,
with no cycles. Each tree has with n vertices. Each pair of vertices has exactly one
path connecting them, creating n − 1 edges in the tree.
Definition 6. Depth First Search (DFS). Depth First Search (DFS) is an uninformed
tree search that works by expanding the first child node that appears on the graph and
goes deeper each time until a goal node is found, or until it hits a node that has no
children. After this the search backtracks, returning to the most recent node it hadn’t
finished exploring.
Definition 7. Breadth First Search (BFS). Breadth First Search (BFS) is a graph
search algorithm that begins at the root node and explores all the neighboring nodes.
After this, it explores the unexplored neighbor nodes of each of the previously found
nodes, repeating this procedure until it finds the goal node.
Cluster Representation. There are many different graph concepts used for the
creation of clusters in a WSN. Among them is the following definition for cluster:
Definition 8. Cluster. A cluster is any subset of nodes C⊂V . y ∈ V is the cluster
head and Gc = (C, Ec ) is the cluster graph.
Ec = {(u, v)|u, v ∈ C ∧ (u, v) ∈ E}
If Gc is connected, then the cluster is connected. dc (u, v) is the shortest path inside
the cluster, and the cluster radius is the maximal distance between y and any other
node v ∈ C.
maxv∈C dc (y, v)
After determining the nodes’ neighbors and the distance between them, it is possible to use additional graph concepts like the node’s weight to use it as one of the
parameters in the definition and functionality of the clusters.
FUNDAMENTAL CORRESPONDING BETWEEN WIRELESS SENSOR NETWORKS
167
Definition 9. Graph Weight. The nodes in the network graph can have a positive
weight wv . The total weight of a cluster is given by
Wsum (C) =
wv
v∈C
Dominating Sets and Covers. Another group of graph definitions which are
very useful when trying to model the clusters is the concept of sets. Set definition
in the context of WSN are listed below. By being able of defining these sets in the
network, the nodes can calculate their real coverage and establish the stability of
the communication paths between them. Generally, the clusters are defined using
the vertex cover of the graph (see Definition 10) and the nodes that belong to that
vertex cover are selected as the cluster heads. The remaining nodes should calculate
the stability of their communication link to the neighboring CHs and join the cluster
corresponding to the link with the best connection stability.
Definition 10. Vertex Cover. A vertex cover is a subset S ⊂ V , such that for all edges
in E, S ∩ e =
/ 0, which means that every edge has at least one edge in S.
Definition 11. Independent Set. The independent set ISG is the subset S ⊂ V , such
that there is no edge between any pair of nodes in S. The maximum independent set
in G is equal to |V |, the size of the minimum vertex cover of G.
Definition 12. Dominating Set. The dominating set DSG is the subset S ⊂ V such
that every node is in S or if it is in V − S and has at least one neighbor (adjacent
vertex) in S. A vertex of S is said to dominate itself and all adjacent vertices. An edge
is dominated if either of its endpoints is in S. The nondominated edges are called free
edges.
Definition 13. Minimum Dominating Set. The Minimum Dominating Set (MDS )
problem is the problem of finding a dominating set of minimum size. This is an
NP-complete problem.
Definition 14. Independent Dominating Set. The independent dominating set
IDS G is the set S ⊂ V that is both dominating and independent.
Definition 15. Connected Dominating Set. The connected dominating set CDS G
is a dominating set whose induced subgraph S ′ is connected.
Definition 16. Weakly Induced Subgraph. For any subset S ⊂ V , the subgraph
weakly induced by S, Sw , is the graph (N[S], E ∩ (N[S] × S). This means that Sw
contains the vertices of S, their neighbors, and all edges with at least one endpoint in
S. A vertex subset S is a weakly connected dominating set if S is dominating and
Sw is connected.
168
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
7.3 GRAPH-BASED APPROACHES FOR CLUSTERING IN WSN
In this section, we present a classification of proposed clustering techniques based on
the application of Graph Theory concepts. Although many of the references are related
with general wireless and ad hoc networks and not specifically WSN, we include them
either because such approaches can be applied to WSN, especially static WSN, or
because we need to clarify why they are not suitable for WSN. We summarize the
main characteristics for each one of the algorithms and protocols, with emphasis on
describing the ones suitable for WSN.
Once it has been decided that the creation of clusters inside the WSN is an appropriate solution to support other network and application functions, it is necessary
to clearly define the characteristics desired in the clusters to obtain a “well-formed”
structure. Some of those characteristics are as follows:
1. Every node should be in exactly one cluster. The objective of this is to maximize
the average cluster sizes while maintaining full coverage.
2. Guarantee the total coverage of the network.
3. Minimize the number of CH to provide an efficient network coverage while minimizing the cluster overlap. A minimum cluster overlap reduces the amount of
channel contention between clusters and improves the efficiency of algorithms
that execute at the level of the CH.
4. Create a highly uniform, balanced clustering.
As an example of the importance of highly uniform clustering with low overlap,
consider the clustered broadcast protocol described by Ni et al. [12]. In this protocol,
the broadcast message is relayed from CH to CH, which then broadcast the message
to their associated nodes in each cluster, called followers. In a clustering with few CH
and large cluster sizes, the clusters have minimal overlap and provide the best coverage
of the network with the fewest clusters. In this configuration, the number of repeated
broadcast transmissions over any area will be small, thus reducing the amount of
transmission collisions and channel contention. This allows for faster, more efficient
and more reliable communications. On the other hand, a poor clustering containing a
lot of cluster overlap and a large number of CHs lose most of the benefits of clustering
because transmissions will be repeated in areas of overlap with significant channel
contention
In the following subsections we make a classification of the graph theory based
clustering protocols for static and mobile WSN. This classification contains two main
groups as shown in Figure 7.3: static-based WSN algorithms and mobile-based WSN
algorithms.
7.3.1 Centralized Algorithms and Self-Elective Protocols
Many centralized algorithms often use graph theoretic properties for clustering. These
algorithms deal with the topology of the entire network as a whole, creating structures
GRAPH-BASED APPROACHES FOR CLUSTERING IN WSN
169
Graph Theory Approaches Used for Clustering in WSN
Mobile Sensors
Stationary Sensors
Centralized
algorithms &
self elective
protocols
. Node ID and
node degree
[19]
. Spanning tree
[33, 4]
Localized
protocols &
emergent
algorithms
. ACE [11]
Distributed
dominating setbased
algorithms
. Span [16]
. DFS [17]
. IDS [5]
. LRG [24, 22]
Combined &
additional
approaches
. Zonal
algorithm [13,
18]
. Vertex Cover
[6]
. Max-Min DCluster [2]
. GHS [18]
. Greedy set
cover algorithm
[12]
. TASC [35]
. Balanced
Clusters [20]
Zone-based
clustering
. Zonal weaklyconnected
clustering
algorithm [13,
15]
Peer-to-peer
generalized
clustering
model
. CDC [32]
CNet(G)
clustering
. Eulerian Cycle
[34]
Figure 7.3. Graph theory approaches used for clustering in WSN.
that often present an increased vulnerability to node failure in certain key parts of the
network, usually near the root of the tree (if they use a spanning tree) or near the BS.
Additionally, these protocols may require significant communications or computation
overhead for very large networks.
It is necessary to clarify that not all the clustering algorithms and protocols that
have been proposed using graph theory concepts are suitable for WSN. For example,
in reference 13, the authors propose a technique where each cluster forms a clique.
However, they do not select a CH that makes the protocol unfeasible for WSN.
Additionally, their “createClusters()” function has a relatively large overhead of
O(d 3 ), where d is the density of the network (number of nodes per area unit).
A group of clustering protocols called self-elective protocols use the Node ID and
Node Degree to select the CH. In reference 14, Gerla and Tsai proposed two weightbased clustering algorithms, where each vertex v selects the node with optimal weight
within N(v) as CH. In the first algorithm, the optimal vertex is the one with lowest
node ID as shown in Figure 7.4. The neighborhood of the node selected as CH is the
cluster. A node that can hear two or more CHs is a “gateway” or border node.
In the second clustering algorithm, the highest-degree node in a neighborhood is
selected as the optimal node to be the CH and the neighbors are covered by it (see
Figure 7.5). Although the algorithm is expected to perform well on many randomly
defined graphs, it may not produce any CH for graphs that do not have any node with
the highest number of neighbors (like interval graphs). Thus, the algorithm must be
completed by adding nontrivial tie resolution rules.
170
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
1
12
14
13
15
2
3
11
16
10
17
4
6
20
19
18
7
5
Cluster
9
8
Cluster Head
Figure 7.4. Node ID clustering.
In references 15 and 16 the authors propose algorithms that control the size of
each cluster and the number of hierarchical levels. However, these algorithms are not
suitable for WSN, since they add a significant computation overhead to the nodes in
the networks and incorrectly assume that the topology can be controlled and deployed
in a regular organized basis. Additionally, they are unable to prevent two nodes that
are just over one cluster radius apart from simultaneously electing themselves as CH,
causing a large overlap in the network’s clusters. This problem occurs sufficiently
frequently to make the resultant cluster packing inefficient.
N(1)=1
N(12)=3
N(11)=3
N(3)=3
12
(-1 outside (-1 outside
the cluster) the cluster)
1
N(2)=3
2
3
N(4)=3
(-1 outside
the cluster) 4
11
N(13)=2
16
10
N(16)=3
N(10)=4
N(19)=2
(-1 outside
the cluster)
19
6
20
N(6)=4
N(20)=2
18
N(18)=2
N(14)=3
14
13
N(7)=1
N(5)=4
7
5
N(9)=2
Cluster
Border Nodes
Cluster Head
8
9
N(8)=2
Figure 7.5. Node degree clustering.
15 N(15)=2
17 N(17)=3
GRAPH-BASED APPROACHES FOR CLUSTERING IN WSN
171
Zhang and Arora [17] present a centralized scheme to produce an approximate hexagonal close packing. However, they assume that each node knows their
precise location, which may be difficult to address in WSN. The clustering algorithms that use a centralized approach are generally based on the use of Spanning
Trees to associate the vertices and to have several parent nodes in charge of the interactions with their child nodes and with their own parents in the superior level. A
generalized clustering technique creating a tree-based construction for network partitioning is presented in reference 18. Thaler and Ravishankar propose to construct a
top-down hierarchy based on an initial root node. However, the level of broadcasting
messages used to create the tree is too large.
Another tree-based clustering protocol is proposed by Banerjee and Khuller in
reference 19. They use a Breadth First Search (BFS) algorithm to partition the network.
However, a drawback for using their algorithm in WSN is that only one node initiates
the clustering process, which can provoke the loss of an entire subtree if one of
the higher-level nodes in the tree suffers a failure. Additionally, this protocol still
requires O(n) time in linear networks, where n is the entire size of the network, which
is outperformed by other protocols as we will show ahead.
7.3.2 Localized Protocols and Emergent Algorithms
In contrast with centralized algorithms, localized algorithms are characterized by
reducing the amount of central coordination necessary and only require each node
to interact with its local neighbors. Emergent algorithms are a class of localized
algorithms. The emerging localized algorithms have the additional characteristic that
the individual agents (i.e., sensor nodes) only encode simple local behaviors and do
not explicitly coordinate on the global scale. A localized protocol for a sensor network
is a protocol in which each sensor node only communicates with a small set of other
sensor nodes within close proximity in order to achieve a desired global objective.
Locality reduces the chances for protocol failure due to transmission errors and node
failure.
In reference 20, the authors present an emergent algorithm called ACE (Algorithm
for Cluster Establishment). The goal of ACE is to select the smallest set of cluster
heads such that all nodes in the network belong to a cluster. The problem is similar
to the minimum dominating set problem in graph theory. In ACE, the sensors are
considered stationary and with uniformly random coordinates in the sensor field.
The proposed algorithm is completed in constant time regardless of the size of the
network, and it consists of two logical parts: The first part controls how clusters can
spawn (by having a node elect itself to be leader), and the second part controls how
clusters “migrate” dynamically to reduce overlap. The sensors continually compute
the clusters, without the necessity of a specific event. This makes the protocol suitable
for proactive networks. ACE results in a highly uniform cluster formation that can
achieve a packing efficiency close to hexagonal close-packing, which minimizes the
overlap between uniform circular clusters while ensuring full coverage. ACE is scaleindependent (it is completed in constant time regardless of the size of the network)
and operates without requiring geographic knowledge of node positions or any kind
of distance or direction estimation between nodes.
172
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
14
2
1
15
13
3
12
6
17
4
16
12
18
11
5
19
6
18
10
7
19
9
8
16
11
10
7
15
13
3
17
4
5
14
2
1
9
20
(a) N (1) = {2, 3, 4}
N (15) = {12, 13, 14, 16, 17, 18}
8
(b)
20
Cluster Heads:
4, 7, 12, 16, 20
N (20) = {9, 19}
Figure 7.6. ACE cluster formation. (a) Initial state. (b) CHs dominating set.
The nodes initiate actions at random intervals to avoid collisions. Each time that
an action can be initiated for a node is called a node’s iteration. The communication
between nodes is local and do not require time synchronization. ACE distinguishes
between three possible node states: unclustered (still during the cluster formation process), follower (a node that belongs to a cluster with a designated CH), or cluster head.
The iteration interval, or time interval between iterations among nodes, is uniformly
random distributed. All nodes in the sensor field are initially in the unclustered state
as shown in Figure 7.6a. When a node is unclustered at the beginning of its iteration,
it assesses its surroundings and counts the number of loyal followers it would receive
if it declared itself a CH of a new cluster. A loyal follower is a node that can belong
only to the cluster that would be formed by the current node sensing its signal.
If the number of loyal followers for the node is greater than or equal to its spawning
threshold function, the node will span a new cluster. The “Spanning Threshold Funct
tion” (fmin ) is given by the formula: fmin = (e−k1 ( cI ) − k2 )d, where l is the number of
loyal followers in the network, c is the desired number of algorithm iterations, I is the
expected length of the iteration interval, t is the time passed since the protocol began,
d is the estimated average degree of a node in the network, and k1 and k2 are constants
to determine the shape of the exponential graph. fmin is an exponentially decreasing
function, where the nodes at the beginning of the algorithm have less probability of
becoming a cluster head, since the number of loyal followers is usually going to be
less than fmin , but the probability grows as the cluster formation process progresses.
Based on the fmin function, the node will decide if it is becoming a cluster head or
not. The set of CHs is a minimum dominating set in G (see Figure 7.6b).
If at the beginning of the iteration interval a CH already exists, this will send a
POLL message to all its followers to determine which of them is the best candidate
to become the new leader of the cluster. The best candidate to become CH is the node
that has the greatest number of possible followers while minimizing the amount of
overlap with existing clusters. Once the best candidate is determined by the current
GRAPH-BASED APPROACHES FOR CLUSTERING IN WSN
173
cluster head, it will promote the best candidate as the new cluster head and abdicate its position as the old cluster head. In some cases, the best candidate can be the
same node that was already acting as CH. When a node becomes a CH, it conserves the previous ID if it was promoted by a previous CH or generates a random cluster ID if it declared itself as CH. After generating the cluster ID, the CH
broadcasts a RECRUIT message to its neighbors, who will become followers of
the new cluster. A node can be a follower of more than one cluster while the protocol is running (it picks a single cluster for membership only at the end of the
protocol).
When the cluster head promotes a new node to CH, the position of the cluster will
appear to migrate in the direction of the new CH because some of the former followers
of the old CH are no longer part of the cluster, while some new nodes near the new
CH become new followers of the cluster. Nodes that are not within one-hop radius
of any CH can pick a clustered neighbor to act as their bridge to the CH, becoming
two-hop followers. In the worst-case scenario, the loss of a CH in ACE would leave
at most one cluster of the nodes unclustered.
Listing 1 shows and outline of the algorithm used by the nodes during ACE’s
first iteration, according to the pseudocode presented in reference 10. The number of
iterations that the algorithm should execute is a nonfixed parameter that should be
determined by the system using the algorithm, in order to determine a good tradeoff
between communication overhead and cluster size. This algorithm only covers the
aspects related to the clusters formation and does not include aspects related to data
transmission after that. According to the simulations presented in the paper, the
authors determined that 3 iterations were good for its execution, but they do not
guarantee that this number is going to be the correct one for all executions of the
protocol.
LISTING 1. Pseudocode for the ACE Algorithm
public class ACENode {
int ID, clusterID, bestLeader;
String state;
//time since the node started the protocol
double time;
int expectedIterationLenght;
int numLoyalFollowers, bestFollowerCount;
public ACENode( ){
//All nodes initiate in an UnclusteredState
state = ‘‘Unclustered’’;
}
public int ScaleOneIteration( ){
if (time > 3*expectedIterationLenght) {
(Continued)
174
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
(LISTING 1. Continued)
if (state==‘‘ClusterHead’’){
return 1; //DONE }
if (state==‘‘Clustered’’){
//wait for cluster-head to terminate
waitClusterHead();
//pick one as cluster-head
pickNewClusterHead();
return 1; //DONE }
if (state==‘‘Unclustered’’){
//pick a random clustered node to act as
proxy
//after it terminates
pickRandomProxy();
//wait for random proxy to terminate
waitRandomProxy();
return 1; //DONE }
}else{
if (state==‘‘Unclustered’’){
numLoyalFollowers = countLoyalFollowers();
if (numLoyalFollowers >= fmin(time)){
clusterID = generateNewRandomID();
locallyBroadcast(‘‘Recruit’’,ID,
clusterID); }
}else{
if (state==‘‘ClusterHead’’){
bestLeader = ID;
bestFollowerCount = numLoyalFollowers;
//all nodes where the node is a potential
new cluster-head
ArrayList potCH = definePotentialClusterHeads();
Iterator iter = potCH.iterator();
ACENode potential=null;
while (iter.hasNext()){
potential = (ACENode) iter.next();
followerCount = PollForNumLoyalFollowers
(potential,clusterID);
if (followerCount > bestFollowerCount){
bestLeader=potential.ID;
bestFollowerCount=
followerCount; }
}
if (bestLeader!= ID) {
send(bestLeader,‘‘Promote’’,clusterID);
//waits for bestLeader to broadcast
GRAPH-BASED APPROACHES FOR CLUSTERING IN WSN
175
it’s ‘‘Recruit’’
//message
while(!recruitMessageReceived());
locallyBroadcast(‘‘Abdicate’’, ID,
clusterID); }
}
}
}
}
}
Because of being an strictly localized algorithm, ACE uses a fixed amount of time
O(d) for its execution, regardless of the total amount of nodes in the network. The
estimated average degree d of a node in the network is calculated prior to the network
deployment.
7.3.3 Distributed Dominating Set-Based Algorithms
One of the problems that arise in a number of distributed network applications is how
to locate a small number of centers in the network such that every node is close to
at least one center. Clearly, this is the same objective targeted by clustering, and one
of the main objectives when doing this is to define the cluster concentration points
while creating the minimum number of clusters in the network and guaranteeing the
total coverage of the network.
Graph Theory approaches this problem by calculating a dominating set DSG in
the graph. Computing a DSG is also a way of increasing the network’s lifetime. The
idea is to build a network backbone (the DS) consisting of nodes with higher energy.
In reference 21, the nodes in the dominating set (called coordinators) change from
time to time in order to maximize their lifetime. An approach similar to the ones
presented in reference 14 is proposed in reference 22 using both the node ID and the
node degree of the vertices to select the CH. After that, they use clustering algorithms
designed based on a Depth First Search (DFS) traversal of nodes in the network.
These algorithms are initialized by any node and are fully distributed. They create
and maintain k-clusters, in which any node is at distance at most k hops from the CH.
In the clustering process, each node is either a CH, a covered nodes or an undecided
node. In the basic k-cluster algorithm, each undecided visited node in DFS declares
itself a CH and covers all its k-hop neighbors. In the highest connectivity k-cluster
algorithms, each undecided visited node checks all its undecided k-hop neighbors
and chooses one with the largest connectivity to be its CH. Chen and Stojmenovic
consider the mobility of nodes and describe algorithms for modifying cluster structure
in the presence of topological changes. The efficiency of the clustering algorithms
is tested by measuring the average number of created clusters, the number of border
nodes, and the cluster size in random unit graphs. Basagni [23], also uses the IDSG
approximation but generalizes the idea of using “weight,” such that any meaningful
parameter can be used in order to best exploit the network properties.
176
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
The fast distributed approximation algorithms proposed in reference 24 use a
synchronous model of computation using a variant of DSG called the k-dominating
set where every node is within a distance k of some server or directory copy—in
this case, of some CH. The proposed algorithm is called Local Randomized Greedy
Algorithm (LRG) and is similar to the greedy algorithm for Minimum Set Covering
Computation presented in reference 25, but instead of only considering the span of the
nodes, LRG considers how the nodes covered by an individual node are also covered
by other nodes.
The algorithm creates the clusters in three steps. First each node v calculates its
span d(v); this is how many of the current adjacent uncovered nodes it will cover if it
becomes a CH. This span d(v) plus the node ID is sent to all nodes within a distance
equal to 2 around u. Each neighbor computes the maximum and forwards it one more
step. With this information, every node can determine whether it is a candidate or not.
In the third step, called the selection step, a node u calculates its support s(u), which
is the number of candidates that cover u. The node u sends its s(u) to the candidates in
its N(u). Each node v adds itself to the dominating set lexicographically higher than
that of any node within its 2-hop neighbors. One drawback of this approach for WSN
is that once a node and it neighbors are covered, the node does not participate further
in the cluster formation process, which will cause a higher rate of energy consumption
in the node if it is a CH, leading it to its death because of battery depletion.
7.3.4 Combined and Additional Approaches
In this section we present some other clustering mechanisms that combine some of
the previously explained approaches and therefore cannot be classified in only one of
the categories. We include the Zonal Algorithm, the Max-Min D-Cluster formation,
and the Topology Adaptive Spatial Clustering algorithm (TASC).
The Zonal Algorithm proposed in reference 26 aims to find weakly connected
dominating sets in the network. The algorithm consist of three phases: (1) An input
graph representing the network is partitioned into regions of approximately size x. (2)
The distributed algorithm for weakly connected sets is run in each region. (3) Some
additional border vertices are added.
In phase 1, graph partitioning using minimum spanning forests, a sender v of
degree d(v) can broadcast to its neighbors in constant time with O(d(v)) messages. A
vertex can also send a message to any adjacent vertex in constant time. For finding a
Minimum Spanning Tree (MST), all edge weights are distinct, breaking ties using the
vertex IDs of the endpoints. The MST is unique for a given graph with distinct edge
weights. The algorithm maintains a spanning forest. Initially, the spanning forest is
a collection of trees of single vertices. At each step the algorithm merges two trees
by including an edge in the spanning forest. During the process of the algorithm, an
edge can be in any of the three states: tree edge, rejected edge, or candidate edge.
All edges are candidate edges at the beginning of the algorithm. When an edge is
included in the spanning forest, it becomes a tree edge. If the addition of a particular
edge would create a cycle in the spanning forest, the edge is called a rejected edge
and will not be considered further. In each iteration, the algorithm looks for the
GRAPH-BASED APPROACHES FOR CLUSTERING IN WSN
177
u
Figure 7.7. GHS (Gallager, Humblet, and Spira) algorithm for finding an MST (Minimum
Spanning Tree).
candidate edge with minimum weight, and the algorithm changes it to a tree edge
merging two trees into one. During the algorithm, the tree edges and all the vertices
form a spanning forest. The algorithm terminates when the forest becomes a single
spanning tree. This partitioning process is based on the GHS algorithm proposed in
reference 27, however, any distributed algorithm for constructing spanning trees can
be used. Figure 7.7 shows an example of the zonal division obtained after applying
the GHS algorithm.
In phase 2, for computing the weakly connected DSG of the regions, they use a
color-based algorithm. This is a distributed implementation of the centralized greedy
algorithm for finding small weakly connected DSG in graphs. The root of the spanning
tree is used as an arbitrator. The root starts an iteration by broadcasting a request
through the tree to search for the vertex with the maximum improvement value. Once
the root has determined the global maximum, it sends a message to that vertex and
colors it. Phase 3 is used to fix the cluster borders by constructing a bipartite graph.
Two regions Ri and Rj joined by a dominated edge can comprise a single region
with a dominating set Si ∪ Sj and do not need to have their shared border fixed. If
neighboring regions Ri and Rj are not joined by a shared dominated edge, the region
with the lower subscripts adds a new vertex from the RRji border into the dominating
set. Figure 7.8 shows an example where R3 and R4 share a dominated edge and would
be fixed into one only region R3 .
In reference 28, the author presents a polynomial-time approximation algorithm
to minimize the number of clusters in the network. First, it uses the concept of vertex
cover to create the clusters and then creates a tree inside each cluster to improve
the transmission processes in the network. This approach uses the community access
network concept to connect static wireless networks with a wired backbone network.
178
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
R4
R5
R3
w
u
y
v
R2
R1
Figure 7.8. Border fixing in the zonal clustering algorithm.
It divides the network into clusters and selects a single access point at each cluster.
This access point has a direct link to the wired network and acts as a gateway between
the wireless nodes and the wired network.
For its execution, the proposed algorithm assumes that the upstream and downstream bandwidth requirements of every node in the network are the same. It also
assumes that there is a maximal weight W ∈ R+ for the graph. The algorithm performs an initial operation to create the clusters and in the process takes into account
the following constraints: (a) The resulting clusters should be disjoint, which means
that every node is included in a single tree; (b) they should contain all network nodes;
(c) each cluster must induce a connected graph with fixed bound R, which is the radius
for multihop routing with bounded propagation delay and therefore is the maximum
value for the depth of every tree; and (d) a spanning tree T is created inside each
cluster to simplify the routing process and maximize the network utilization. The
delivery trees inside the clusters must cover all the nodes in the network. Each node
knows its parent and its descendant nodes in T. The total weight of all the nodes in a
tree is at most W. A graph decomposition that satisfies the previous requirements is
called a feasible partition, and every tree in the partition is called a feasible delivery
tree.
In the first stage of the clustering algorithm, it looks for the minimal number
of cluster heads such that they cover their R-neighbor nodes. The author proposes
three different ways to select the CHs: (1) to use the greedy set cover algorithm
proposed in reference 29, (2) to use a shifting strategy approach, which is based on
the “divide-and-conquer” method and performs a series of shifting operations to
subdivide the networks in a progressive manner, finding a near-optimal solution but
with the expense of a high running time, or (3) to use a dominating independent set
approach, where each iteration of the algorithm selects the node with the maximal
number of uncovered nodes in the R-neighborhood of each node v. The second stage
of the algorithm performs a cluster refinement by associating each node with its
GRAPH-BASED APPROACHES FOR CLUSTERING IN WSN
179
closest yi ∈ Y (the CHs set), and it breaks ties by selecting the CH with the lowest
ID. The third stage creates the tree inside the cluster and executes a post-order tour
of the tree to calculate its total weight. Finally, the algorithm reduces the relay load
of the cluster nodes by selecting as an access point a node that is closer to the tree
center. The root of every tree T is the access point to the wired network. This root
node coordinates all the transmissions inside the clusters by using slotted channels
and polling mechanisms. The delivery mechanism attempts to maximize the clusters’
throughput.
According to the simulations presented in reference 28, this clustering technique
yields results close to optimal. Since the weight partitioning algorithm allows the
average cluster size to be as small as W/3, it produces clusters that are small in comparison to other clustering techniques. However, there is a drawback in this proposed
technique: The access point constitutes a single point of failure in each cluster. If the
access node is a cute node (if its removal leaves the graph disconnected), the system
attempts to attach the disconnected nodes to an adjacent cluster. Otherwise, the algorithm should send the proper recovery messages to modify T for reconnecting the
detaching nodes, which can be very costly for a WSN. In the Max–Min D-Cluster
formation proposed in reference 30, the CHs are selected such that they form a d-hop
DSG , where d > 1 distance hops form the vertices in the DSG away from any other
vertex in C. Since d is an input value to the heuristic, it enables control over the
density of cluster heads in the network.
The Topology Adaptive Spatial Clustering (TASC) algorithm presented in
reference 31 is a distributed algorithm that partitions the network into a set of
locally isotropic, nonoverlapping clusters without prior knowledge of the number
of clusters, cluster size, and node coordinates.
The spatial grouping of nodes with respect to regions of close proximity and similar deployment density does the following: (a) It promotes efficient data aggregation,
(b) it allows the use of different data compression rates in each cluster, (c) it improves the overall compression rate in the whole network, and (d) as a consequence,
it also helps reducing the propagation of redundant data inside the network. Spatial
clustering would also assist transmission power control, since intracluster communication requires less transmission power in dense clusters. TASC is useful for proactive
networks, where the clusters provide the information needed for the cluster formation
process. The nodes in the sensor field are assumed to be stationary. Each node can
measure distances to its 1-hop neighbors and has knowledge of its 2-hop neighborhood, as shown in Figure 7.9.
TASC operates on a combination of node weights and a dynamic density reachability criterion. The main objective of TASC is to cluster nonuniform sensor networks.
In this type of network, the node density variations are globally large but there exist
subgroups of nodes such that density variations are locally small. The idea of TASC
is to partition those networks in a way such that relative node density variation in individual clusters is smaller than relative node density variation in the whole network,
obtaining a distribution similar to the one shown in Figure 7.10.
For the cluster formation, two different parameters must be previously specified: required minimum cluster size and density reachability parameter. The latter
180
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
Node i
N (i)
S = {v ε N(i)
density in terms
of distance is
similar}
Figure 7.9. TASC neighbor’s information.
parameter allows each node to further limit the number of nodes that it can potentially nominate as CH by considering only density reachable nodes as nomination candidates. This effect pulls cluster leaders toward the most dense groups
in the cluster, but nomination among density reachable candidates is still based on
weights.
Listing 2 shows an outline of the algorithm used by the nodes during TASC’s
execution, according to the pseudocode presented in reference 31. Initially, each
node in the WSN computes its own weight based on shortest Euclidean paths in its
Figure 7.10. TASC cluster’s nodes distribution.
GRAPH-BASED APPROACHES FOR CLUSTERING IN WSN
181
2-hop environment. In a nonuniform deployment WSN, the node that tends to be the
midmost related to all shortest communication paths (in terms of hops) gets the largest
weight. Instead of incrementing the weight by one each time a node is used in a path,
we increment the weight as a function of the distance a node contributes to the path.
LISTING 2. Pseudocode for the TASC Algorithm
public class TASC {
int ID;
double weight;
TASC nominee, leader;
int clusterSize;
ArrayList clusterMembers;
/*
* 2HNeighborhood = Euclidean paths in the 2-hop
neighborhood of the node
* internode = Inter-node distance measurements
* minClustSize = Minimum cluster size
* Dr = Density reachability parameter
*/
public TASC(2HNeighborhood,internode,minClustSize,Dr){
//The node computes its own weight based on
//2HNeighborhood
weight = computeWeight(2HNeighborhood);
broadcastToNeighborhood(weight);
//all weights received
If (receiveWeights()){
//finds the heaviest density reachable node
nominee = findHDRNode(Dr);
broadcastToNeighborhood(nominee);
}
//all nominations have been received
If (receiveNominations()){
//Selects the closest nominee
leader = findClosestNominee();
broadcastToNeighborhood(leader.ID, this.ID);
}
//this node is leader
(Continued)
182
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
(LISTING 2. Continued)
If (leader.ID == this.ID){
//Wait until election timeout
while (!electionTimeOut);
broadcastToNeighborhood(clusterMembers,
clusterSize);
}
//cluster size is received
If ((clusterSize = receiveClusterSize())!=0){
If (clusterSize < minClustSize){
//Selects the closest neighbor for which
//clustersize >= minimum cluster size
leader=selectClosestCluster
(2HNeighborhood);
//joins previous node’s cluster;
this.joinCluster(leader);
}
this.broadcastToNeighborhood(leader.ID,
clustersize);
}
}
}
The weight value calculated by the node is broadcasted to its 2-hop neighborhood.
Since this is done by all nodes, each node also receives the weights of its 2-hop neighbors. By comparing the weights, each node nominates the node having biggest weight
in the density-reachable subset of its 2-hop neighbors and broadcasts its nominee to
its 2-hop neighborhood. After receiving all nominees in its 2-hop neighborhood, each
node elects the closest nominee to its leader. Each node that ends up in a cluster where
the total number of nodes is smaller than a prespecified minimum cluster size joins
the closest cluster, where the number of nodes exceeds the required minimum cluster
size.
After creating the clusters, TASC uses an “all-pairs-shortest path routing.” Instead
of trying to forward traffic to the neighboring node that is closest to the destination,
TASC routing is based on distance measurements to extract information about the
network topology. More specifically, node weight is a measurement of two key quantities: (1) the frequency a node is found on the shortest path between pairs of nodes,
and (2) the distance contribution of the edges of that node with respect to the total
length of the path.
As mentioned at the beginning of the section, creating balanced clusters is one
of the main objectives for WSN. In reference 32, the authors propose an algorithm
for clustering the sensor nodes such that each cluster (and its corresponding CH) is
balanced and the total distance between sensor nodes and CH is minimized. Balancing
the clusters is needed for evenly distributing the load on all master nodes. If there is
CLUSTER’S CONSTRUCTION AND MAINTENANCE IN MOBILE ENVIRONMENTS
183
no balance constraint, the authors propose the utilization of Voronoi Diagrams [7],
by constructing it based on the number of CH deployed previously in the sensor field.
However, each CH can only manage a certain number of communication channels.
In this case, trying to solve the k-clustering problem optimally helps to maintain the
network’s operation. This k-clustering problem attempts to group the sensor nodes
such that each cluster is balanced and has exactly one CH. To solve the problem, the
authors transform it into a min-cost flow instance by adding a source node s and a
sink node t to G, both with infinite capacity. Each edge has a weight equal to the
message transmission energy dissipation between the two end vertices. There are n
directed edges from s to all vertices corresponding to sensors nodes. Similarly, there
are k directed edges from vertices corresponding to CH to t. All edges incident to s
or t have weight 0. Finally, nodes corresponding to sensors have capacity 1, while
nodes corresponding to CH have capacity nk . Each flow solution is corresponding to a
k-clustering solution. Constructing G and the corresponding k-clustering solution can
be done in O(n.k) time. Hence, the k-clustering problem can be optimally solved in
O((n + k)3 ) time. However, the major drawback of this proposal is that it assumes that
the nodes are not randomly deployed in the sensor field. This is not a valid assumption
for the majority of WSN application.
7.4 CLUSTER’S CONSTRUCTION AND MAINTENANCE
IN MOBILE ENVIRONMENTS
Maintaining clusters as the topology changes is an issue that arises in mobile wireless
sensor networks. A local change in the weight may result in a different vertex becoming a cluster head. As an example, if we are using IDSG , when two CH become
adjacent, one of them has to abdicate. This and other event occurrences can generate a
total reordering of the nodes in the WSN and create the necessity to have maintenance
operations, which in turn may cause other changes to propagate to the network.
Despite the evident impact of mobility issues in WSN, very few clustering proposals take it into account. Most proposals consider the nodes as stationary since
designing mechanisms to solve the issue is a very challenging task. During the development of this survey, we only found three research papers involving mobility in the
clustering techniques proposed for WSN using Graph Theory. References 33–35 are
the subject of our next analysis in this section. The purpose is to analyze how those
proposals approach the creation and maintenance of clusters in a mobile environment.
7.4.1 Zone-Based Clustering
Chen and Liestman [33], present a modification to their “Zonal weakly connected
clustering algorithm,” presented in reference 26, including cluster maintenance in the
presence of network topology changes. In this new proposal, the cluster maintenance is
divided into two layers: intrazonal (inside the zones) and interzonal (between zones,
at the borders). One of the assumptions made to manage mobility is that all the
changes occur sequentially and that the network can be restructured before the next
184
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
topology change occurs. As a result of the nodes motion, the original weakly connected
graph can become disconnected. This creates induced disconnected subgraphs that
are controlled using some timeout mechanism that allows the formation of a new
zone inside the network. If when the timeout interval is completed no other zones
have annexed the nodes to its cluster, a new cluster is created.
The cluster formation is performed in three phases. Phase 1 divides the network
into zones, creating a set of trees. Phase 2 executes a greedy algorithm to find a small
weakly connected dominating set within each zone. Phase 3 fixes the border vertex
of each zone. A border vertex of a zone is a vertex that is adjacent to vertices of
other zones. In this last phase, if in the border between two zones a dominating edge
exists between the border vertices, there is no need of new vertices, but if there is no
dominating edge, it is necessary to add new vertices to “fix” the border. All the cluster
formation process is supposed to be executed at the beginning of the network’s life.
After this process, the following operations are maintenance operations. There are
four different events to be considered during this maintenance stage: (1) edge-down,
(2) vertex-down, (3) edge-up, and (4) vertex-up.
In the edge-down event, if the lost edge is a free, nondominated edge, the graph
remains the same and no action is taken, but if the edge (u, v) is a dominating edge, its
loss must be reported by the vertices incident upon it. u and v broadcast an edge-loss
message to determine if the graph has been broken into two pieces. If so, the root node
in the cluster performs the breach suturing procedure to add a dominated vertex to
the zone and fix the breach. The dominated vertex added is the one with the greatest
improvement value, the node that is going to allow the current zone to connect with
the greatest number of other zones. In the example shown in Figure 7.11 a, the initial
clustering scheme has the edges (2,13) and (4,5). If (2,13) is down, nothing happens
because it is not a dominating edge. If (4,5) is down, the root has to perform the breach
suturing to create the new edge (3,5) and must calculate the new dominating set in G
(Figure 7.11b).
14
14
15
15
1
2
2
17
1
18
11
19
9
Z2
20
8
(a)
Z3
5
10
7
18
11
Z3
5
16
12
3
3
Z1
Z
16
12
4
4
17
6
13
1
13
6
10
7
19
9
Z2
8
20
(b)
Figure 7.11. Breach suturing procedure in an edge-down event.
CLUSTER’S CONSTRUCTION AND MAINTENANCE IN MOBILE ENVIRONMENTS
185
The vertex-down event is treated very similar to the edge-down event, since a
detecting vertex cannot easily distinguish between the loss of an edge and the loss of
a neighboring vertex. If the vertex that detects the loss still has another dominated edge
incident to the lost vertex’s zone, it does nothing; otherwise, it will send a message
to its root for it to fix the border. However, a special operation must be performed in
case the vertex down is the root vertex for the zone. In this case, another vertex must
take on the role of the root. To achieve this, every root has a neighbor as a backup.
If the loss is discovered by the backup, it assumes the role of the root. However, it
may be the case that what is lost is the edge between the backup and the root, not
the nodes. In this case, when the duplicated roots detect the redundant condition, the
backup must give up on its new role and the original root joins the regions of itself
and its backup neighbor.
The edge-up event is caused by one node moving closer toward another. In this
case, if the new edge is a free edge, no changes need to be made, but if the new edge
is a dominated edge, some vertices in the DSG can become redundant and can be
eliminated from the weakly dominating subgraph.
When a new vertex is added in an vertex-up event, it is dominated if it has a
neighbor in the dominating set; otherwise, it changes its status to be dominating.
As can be observed in the previous operations, it is necessary to add or remove
vertices from the zones every time the network topology changes. To accomplish this,
a bipartite graph is formed between the dominating vertices of the zones.
Although Chen and Liestman [33] show in their results that the weakly connected
dominating set remains roughly the same size over time, it is not clear from the
paper how they add or remove vertices to the zones. If these operations involve the
additional deployment of the nodes in the field, the application of this algorithm in
many sensor network applications will not be feasible, since they consider a unique
node deployment at the beginning of the WSN operation.
7.4.2 Peer-to-Peer Generalized Clustering Model
Ramaswamy et al. [34], present CDC: Connectivity-based decentralized node clustering scheme, a clustering scheme where every node only requires local knowledge
about its neighbors, providing the ability to cluster the network automatically or to
discover clusters around a given set of nodes. CDC is able to handle nodes’ dynamics
without resorting the whole network at each entry and exit.
CDC is based on the idea of flow in the network, where every edge is considered
a road between two points. Every time a message is passed through an edge, it decrements its TTL (Time to Live) counter until it becomes negligible and is discarded,
but it also accumulates a communication weight. Nodes with greater communication
weight are those that lie inside a cluster. To calculate the accumulated communication
weight for each node, a group of messages is sent from some specific nodes called
the originators.
To select the originators, two properties are used: (1) The set of originators should
be spread out in all regions of the graph, and (2) a node vl is considered to be a
good originator if it acquires more weight due to messages initiated by it than the
186
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
weight acquired by messages initiated by any other originator: TotalWeight(Vl , Vl ) ≥
TotalWeight(Vl , Vi )∀Vi ∈ V . To evaluate the first property, each node checks if it has
already received any clustering messages from other nodes in its vicinity. If so, it
means that there are other neighbor nodes that have already chosen to be originators
and, hence, the node opts not to become an originator. Otherwise, the node moves
to evaluate the second property. This property, however, can only be evaluated if the
CDC algorithm has already been executed, but to execute the CDC algorithm we need
to select the originators. To solve this situation, the authors use the THP (two-hop
return probability). THP, calculated as
TwoHopProb(Vl ) =
Vi ∈N(Vl )
1
Degree(Vl ) × Degree(Vi )
determines the probability of returning to a node Vl in the graph in two hops if random
messages were sent starting at Vl . If the THP is higher than a predefined threshold,
then the node chooses to be an originator.
The selected originators use the same TTL value for their messages. Messages
are sent to the network, and each node that receives the messages calculates (after an
stabilization time) which originator has a higher weight and joins it.
To manage with the dynamics of nodes, two mechanisms are considered: node
entry and node exit. In the node entry mechanism, the entering node calculates
its attraction to the vertices on an specific cluster. The Neighbor-Attraction function
toward another node is defined as
NbrAttractionVN+1 (Vj ) =
1
Degree(Vj )
where VN+1 is the entering node and Vj ∈ N(VN+1 ). The node joins the cluster where
the sum of the attractions to the nodes in that cluster is the highest. In the node exit
mechanism, the neighbors of the exiting node have two possibilities. If they are not in
the same cluster as the exiting node, they only update their connectivity information.
If they are in the same cluster as the exiting node, they have to perform the Node
attraction computation; to define it after the node leaves, they are going to stay in
the same cluster or they are going to “move” to a new one. If the leaving node is an
originator, the neighbors have to be informed and they elect the neighbor with the
highest number of in-cluster edges as the new originator. Special cases may occur
when a node is entering the network while one of its neighbors is leaving it, or vice
versa. In these cases, the computation of the attraction has to be done more than once
and increases the computation overhead of the clustering algorithm.
Although the authors state that this model is suitable for any P2P network, including
sensor networks, we consider that this proposal is not suitable for WSN. Without
considering the “special cases,” the amount of information that the nodes have to
handle is large and the computation overhead would cause a rapid depletion of the
sensor’s energy.
CLUSTER’S CONSTRUCTION AND MAINTENANCE IN MOBILE ENVIRONMENTS
187
7.4.3 Cluster-Based Graph Network
In the schema for the construction and maintenance of a cluster-based architecture
for a sensor network proposed in reference 35, the network is treated as a bidirectional
graph G and CNet(G) is a cluster-based network of G. The CNet(G) is a spanning tree
of G. The nodes in the WSN are grouped into disjoint clusters, where the backbone
is a tree consisting of the cluster heads and gateway nodes. A cluster of G is a star
subgraph of G, where the cluster head has an edge to each other cluster member and
no edges exist between two cluster members in the cluster.
To minimize the number of clusters, two cluster heads cannot be neighbors with
each other. Thus, they form a maximal independent set in G. To guarantee the communication between cluster heads, any two cluster heads are joined through one special
cluster member called a gateway node, which is an intersection of neighbors in G of
two cluster heads. The cluster members that do not act as gateways are called pure
cluster members. A transmission between a cluster head and its members is called
local transmission, and a transmission between cluster heads is called backbone transmission.
Figure 7.12 shows an example of a clustering scheme, where nodes 3, 7, 10, and
16 are CH, nodes 6 and 13 are gateway nodes, and the remaining nodes are pure
cluster members. CH and gateway nodes form the network backbone. Bold edges and
the vertices they connect form the CNet(G) in G.
The broadcasting technique used here is called Eulerian broadcasting procedure.
In this technique, a message called token starts from the source node, visits every
node in depth-first order, and turns to the source node, forming an Eulerian cycle.
When a node v gets the token, it sends the token with the message and its ID to one
of its neighbors which has not been visited by the token yet. If v has no unvisited
11
12
10
8
2
13
1
16
9
6
3
14
15
5
4
7
Cluster
Cluster Head
Gateway Node
Figure 7.12. CNet(G) clustering scheme.
Edge in Cnet(G)
Backbone tree
BT(G)
188
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
neighbors, it returns the token to the node from which it got the token for the first
time. Two operations are defined to treat the movement of nodes: node-move-in and
node-move-out. These operations are supposed to be atomic, and they can not be
performed simultaneously.
In this clustering architecture, a new node u announces itself by sending a message
to join the existing CNet(G) and the network reorganizes itself. When u performs the
node-move-in operation, it checks the status of its neighbors. If there exist cluster
heads in N(u), u selects one to be its head and itself becomes a pure cluster member.
Otherwise, if there are gateway nodes in its neighbors, u selects one of them as its
gateway and becomes a cluster head of a new cluster. If there are no cluster heads
and no gateway nodes in N(u), u becomes a cluster head and sets one neighboring
pure cluster member to be the gateway of itself. After this process, the nodes in
N(u) update their information. According to
the authors, a CNet(G) can be formed
statically and dynamically in O(n) and O( ni=1 Ti ), respectively, where T (i) is the
number of rounds that a node-move-in operation requires for an i-node graph.
When a node u wishes to leave the network, it must perform the node-move-out
operation. If u is a pure cluster member, it sends an “I’m leaving” message and simply
leaves from the network. Otherwise, the node-move-out algorithm works as follows:
If u is a cluster head or a gateway node, CNet(G) is divided into two subtrees. One is
the tree T with u as the root (not including the root in CNet(G)) and one is the tree H
with the root of CNet(G) as the root. u is removed from T and the other nodes of T
are added to H by using the node-move-in operation on each node, so that the resulted
tree is CNet(G). If u is the root of the whole network and N(u) =
/ ∅, a cluster head
that is connected with u by a gateway is elected and is set as the new root of CNet(G);
if N(u) = ∅ and u is the only cluster head in CNet(G), one of the nodes in N(u)
becomes the new root and calls the Eulerian procedure to create a new CNet(G).
Uchida et al. also propose a second version of the node-move-in and node-moveout operations, considering networks where the first operation is performed more
frequently than the second one. In this case, the list of neighbors for every node is
not maintained. For the node-move-in operation, one leader is first elected from the
neighbors of the new node u. If S = {the leader (not a cluster head)} ∪ {neighbor
cluster heads of u}, every node v ∈ S sends its ID at some round t. If u receives
an ID at round t, it means |S| = 1 and S contains only the leader and there are no
cluster-head neighbors of u. If u receives no ID at round t, it means |S| ≥ 2 and there
exists at least one cluster head from which u can elect its cluster head. Otherwise,
it can check for gateways in order to select one of them as its gateway when it
becomes a CH. For the node-move-out operation, u recognizes its neighbors (except
its parent and its children) by calling the Eulerian procedure described above, and
after that it performs the same operations described in the former node-move-out
operation.
Although this clustering architecture is the one we consider the most appropriate
between the tree approaches considered in this section, there are still some issues
remaining—for example, if one sensor is aware when it is leaving the network in
case of energy depletion or malfunction, instead of just moving out of the cluster
range.
EXERCISES
189
7.5 SUMMARY AND OPEN RESEARCH PROBLEMS
In this chapter we study the fundamental concepts of clustering process in WSN
from the Graph Theory perspective. To this end, we survey current existing works
in the literature which use graph theory concepts to perform clusters in WSN. Most
of the algorithms and protocols presented in this chapter meet the design criteria
mentioned in Section 7.3 for clustering in WSN; however, the inclusion of mobility
as a new criterion for the clustering creation and maintenance adds new challenges
for these clustering techniques. Research for the maintenance of the cluster organization under this new scenario is seldom seen. To cope with the typical dynamics of
mobile WSN, some of the existing clustering protocols, like the ones presented in
Section 7.4, include new operations that allow new nodes to be incorporated into appropriate existing clusters and also allow the properly management of the departure of
nodes from the clusters. However, those mechanisms still have to be further modified
to handle situations where the nodes cannot give any warning to the networks about
their abandon of the network, like in the case of energy depletion.
Likewise, there are still several aspects to improve the performance of clustering
mechanisms in WSN. One of those aspects is related to exploiting the redundancy and
spatial correlation of the nodes in the sensor field. This can allow the implementation
of fault tolerance techniques during the data transmission phase of the clustering
algorithms, taking into account that sensors will measure similar data as long as they
remain located in a relatively close position among them. In the same way, this aspect
can help reduce the energy consumption in the nodes, thus making their lifetime
longer.
The inclusion of security techniques also affects the functioning of the clustering methods, when the information exchanged between the nodes is critical for the
network operation. Integrating dynamic clustering with secure information qualitydriven clustering and routing protocols will allow a more efficient communication
inside the network and between the cluster heads and the base station.
In the mobility scenarios, there is still some work to do to (a) determine the best
mechanisms to trigger the update routines in the WSN, (b) look for a better stability
in the network, and (c) maintain a consistent state of the initially created clusters.
Issues such as how to adapt the clusters very quickly to the continuous topology
changes or how to make the network react when those changes take effect during an
update period are still to be considered. Also, the inclusion of mobility patterns in the
simulation will allow us to create test environments more similar to the real world and
to investigate how the complex set of interactions occur between the nodes deployed
in the field.
7.6 EXERCISES
1. Why do you think creating the minimum number of clusters (and also with
smaller sizes) in the network is one of the main objectives of clustering techniques?
190
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
2. In the Local Randomized Greedy Algorithm (LRG) proposed in reference 24,
the authors define the clusters using the calculation of the set covering in the
graph, taking into account the span of the nodes and how the nodes covered by
an individual node are also covered by other nodes. How does this last aspect
improve the performance of the clustering process?
3. How do spatial correlation and data correlation affect clustering mechanisms in
WSN? Is it feasible to exploit redundancy by taking these aspects into account
when developing clustering mechanisms?
4. In topology adaptive spatial clustering (TASC), the authors use the 2-hop
neighborhood of the nodes. Why use the 2-hop instead of reducing it to
only the 1-hop neighborhood or extending it to the n-hop environment of the
nodes?
5. How would you modify the node-move-out operation in the cluster-based graph
technique proposed in Section 7.4.3, so that the nodes that remain in the network
could react even if the moving-out node does not announce that it is leaving the
network?
6. Clustering techniques should remain as a network layer problem or should be
integrated with the lower and upper layers (MAC, Application)? For example,
should they allow the management of administrative policy constraints during
the clusters creation?
7. What is the relation between mobility and topology changes and energy constraints in WSN?
8. The Algorithm for Cluster Establishment (ACE) presented in Section 3.2 uses
a localized emergent algorithm to create the clusters in the network. What are
the advantages or disadvantages of this kind of algorithm, compared to those
that rely on global interaction or information?
9. How does the in-network processing required for the creation of clusters affect
the security mechanisms in WSN?
10. How often should the topology change mechanisms be triggered in Mobile
WSN to ensure the consistency of the clusters and maintain the network
functions?
11. An extension to the mobility management in WSN considers the problem of
“group mobility,” where the nodes do not move using an individual pattern but
following a group pattern. How do you think this affects the proposals presented
in Section 7.4?
BIBLIOGRAPHY
1. M. Ye, C. Li, G. Chen, and J. Wu. EECS: An energy efficient clustering scheme in wireless sensor networks. In National Laboratory of Novel Software Technology, Nanjing
University—China, Department of Computer Science and Engineering, Florida Atlantic
University, USA, 2005.
BIBLIOGRAPHY
191
2. W.-P. Chen, J. C. Hou, and L. Sha. Dynamic clustering for acoustic target tracking in
wireless sensor networks. In Proceedings of the 11th IEEE International Conference on
Network Protocols (ICNP 2003), November 2003, pp. 284–294.
3. J. N. Al-Karaki and A. E. Kamal. Routing techniques in wireless sensor networks: A survey.
Department of Electrical and Computer Engineering, Iowa State University, December
2004.
4. W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan. Energy-efficient communication protocol for wireless microsensor networks. In Proceedings of the 33rd Hawaii
International Conference on System Sciences, Vol. 8, January 2000, pp. 3005–3014.
5. C. Liu, K. Wu, and J. Pei. A dynamic clustering and scheduling approach to energy saving
in data collection from wireless sensor networks. In Proceedings of the Second Annual
IEEE Communications Society Conference on Sensor and Ad Hoc Communications and
Networks (SECON 2005), Santa Clara, California, September 2005.
6. A. Manjeshwar and D. P. Agrawal. TEEN: A routing protocol for enhanced efficiency
in wireless sensor networks. In Proceedings of the 15th International and Distributed
Processing Symposium, April 2001, pp. 2009–2015.
7. F. Aurenhammer. Voronoi diagrams—A survey of a fundamental geometric data structure.
ACM Computing Surveys, 23(3):345–405, 1991.
8. A. Boukerche. Performance evaluation of on-demand routing protocols for ad-hoc wireless
networks. ACM/Kluwer Mobile Networks and Applications, 9:333–342, 2004.
9. A. Boukerche and S. K. Das. Congestion control performance of R-DSDV protocol for
multihop wireless ad hoc networks. ACM/Kluwer Wireless Networks, 9:261–270, 2003.
10. A. Boukerche and S. Rogers. GPS query optimization in mobile and wireless ad hoc
networks. In Proceeedings of the 6th IEEE Symposium on Computers and Communications,
July 2001, pp. 198–203.
11. A. Boukerche and S. Vaidya. A performance evaluation of a dynamic source routing discovery optimization using GPS system. Kluwer Telecomunication Systems, Vol. 22:337–354,
2003.
12. S.-Y. Ni, Y.-C. Tseng, Y.-S. Chen, and J.-P. Sheu. The broadcast storm problem in a mobile
ad hoc network. In Proceedings of the Annual ACM/IEEE International Conference on
Mobile Computing and Networking (MOBICOM 1999), August 1999, pp. 151–162.
13. P. Krishna, N. H. Vaidya, M. Chatterjee, and D. K. Pradhan. A cluster-based approach
for routing in dynamic networks. ACM SIGCOMM Computer Communication Review,
27(2):49–65, 1997.
14. M. Gerla and J. T.-C. Tsai. Multicluster, mobile, multimedia radio network. ACM/Kluwer
Journal of Wireless Networks, 1(3):255–265, 1995.
15. R. Ramanathan and M. Steenstrup. Hierarchically-organized, multihop mobile wireless
networks for quality-of-service support. ACM/Baltzer Mobile Networks and Applications,
3(1):101–119, 1998.
16. R. Krishnan, R. Ramanathan, and M. Steenstrup. Optimization algorithms for large selfStructuring networks. In Proceedings of the Conference on Computer Communications,
Eighteenth Annual Joint Conference of the IEEE Computer and Communications Societies
(INFOCOM 1999), Vol. 1, March 1999, pp. 71–78.
17. H. Zhang and A. Arora. GS 3 : Scalable self-configuration and self-healing in wireless
networks. In 21st ACM Symposium on Principles of Distributed Computing (PODC 2002),
July 2002, pp. 57–68.
192
CLUSTERING IN WIRELESS SENSOR NETWORKS: A GRAPH THEORY PERSPECTIVE
18. D. G. Thaler and C. V. Ravishankar. Distributed top-down hierarchy construction. In
Proceedings of the Conference on Computer Communications, Seventeenth Annual Joint
Conference of the IEEE Computer and Communications Societies (INFOCOM 1998),
March 1998, pp. 693–701.
19. S. Banerjee and S. Khuller. A clustering scheme for hierarchical control in wireless
networks. In Proceedings of the Conference on Computer Communications, 20th Annual
Joint Conference of the IEEE Computer and Communications Societies (INFOCOM 2001),
Vol. 2, April 2001, pp. 1028–1037.
20. H. Chan and A. Perrig. ACE: An emergent algorithm for highly uniform cluster formation.
In Proceedings of the First European Workshop on Sensor Networks (EWSN 2004), January
2004, pp. 154–171.
21. B. Chen, K. Jamieson, H. Balakrishnan, and R. Morris. Span: An energy-efficient coordination algorithm for topology maintenance in ad hoc wireless networks. ACM Wireless
Networks Journal, 8(5):481–494, 2002.
22. G. Chen and I. Stojmenovic. Clustering and routing in mobile wireless networks. Technical
Report TR-99-05, SITE, June 1999.
23. S. Basagni. Distributed clustering for ad hoc networks. In Proceedings of the IEEE International Symposium on Parallel Architectures, Algorithms, and Networks (I-SPAN), June
1999, pp. 310–315.
24. L. Jia, R. Rajaraman, and T. Suel. An efficient distributed algorithm for constructing small
dominating sets. In Proceedings of the 20th Annual ACM Symposium on Principles of
Distributed Computing (PODC 2001), August 2001, pp. 33–42.
25. Z. J. Haas and B. Liang. Virtual backbone generation and maintenance in ad hoc network
mobility management. In Proceedings of the IEEE Conference on Computer Communications (INFOCOM 2000), Vol. 3, March 2000, pp. 1293–1302.
26. Y. P. Chen and A. L. Liestman. A zonal algorithm for clustering ad hoc networks. International Journal of Foundations of Computer Science, 14(2):305–322, 2003.
27. R. G. Gallager, P. A. Humblet, and P. M. Spira. A distributed algorithm for minimumweight spanning tree. ACM Transactions on Programming Languages and Systems, 5(1):
pp. 66–77, 1983.
28. Y. Bejerano. Efficient integration of multihop wireless and wired networks with QoS constraints. In Proceeedings of the eighth Annual International Conference on Mobile Computing and Networking (MOBICOM 2002), September 2002, pp. 215-226.
29. V. Chvatal. A greedy heuristic for the set-covering problem. Mathematics of Operations
Research, 4(3):233–235, 1979.
30. A. D. Amis, R. Prakash, T. H. P. Vuong, and D. T. Huynh. Max–min D-cluster formation in wireless ad hoc networks. In Proceedings of the IEEE Conference on Computer
Communications (INFOCOM 2000), Vol. 1, March 2000, pp. 32–41.
31. R. Virranoski and A. Savvides. TASC: Topology adaptive spatial clustering for sensor
networks. In Proceedings of the Conference on Computer Communications, 24th Annual
Joint Conference of the IEEE Computer and Communications Societies (INFOCOM 2005),
March 2005.
32. S. Ghiasi, A. Srivastava, X. Yang, and M. Sarrafzadeh. Optimal energy aware clustering
in sensor networks. Sensors 2002, 2(7):258–269, 2002.
33. Y. P. Chen and A. L. Liestman Maintaining. weakly-connected dominating sets for clustering ad hoc networks. Ad Hoc Networks, 3(5):629–642, 2005.
BIBLIOGRAPHY
193
34. L. Ramaswamy, B. Gedik, and L. Liu. A distributed approach to node clustering in decentralized peer-to-peer networks. IEEE Transactions on Parallel and Distributed Systems,
16(9):814–829, 2005.
35. J. Uchida, I. A. K. M. Muzahidul, Y. Katayama, W. Chen, and K. Wada. Construction and
maintenance of a cluster-based architecture for sensor networks. In Proceedings of the
39th Hawaii International Conference on System Sciences, January 2006.
CHAPTER 8
Position-Based Routing for Sensor
Networks: Approaches and Obstacles
MARWAN M. FAYED and HUSSEIN T. MOUFTAH
School of Information Technology and Information, University of Ottawa, Ottawa, Ontario K1N
6N5, Canada
8.1 INTRODUCTION
A growing number of sensor network applications require point-to-point services.
Intuitively, pure sensing applications, where environment is monitored or events are
tracked, require some geographical or geometric context for successful operation. In
such a context, data interpretation and management is often tied to node positions.
In addition to traditional sensing applications are a growing number of proposed
applications that require no knowledge of geography yet do require advanced pointto-point services [1–4]. Traditional Internet routing techniques are unsuitable in either
setting [5]. In this chapter we focus on the merits and challenges of algorithms and
protocols that provide point-to-point services through position-based routing, where
forwarding decisions are made by maximizing or minimizing some function of node
locations within a coordinate system. The focus of discussion is on those protocols
suited to static sensor networks where node positions within the network are known
and unknown.
The remainder of this chapter is organized as follows: In Section 8.2 we give some
context to our presentation as well as some background information relevant to its content. Our discussion of position-based routing protocols suitable to sensor networks
begins in Section 8.3 with those protocols where knowledge of sensor positions is
assumed to exist prior to protocol execution; while the presentation is cumulative, each
section may be read independently of the others. In Section 8.4 we present protocols
where no a priori knowledge of position exists; amongst such protocols the routing
element is coupled to the localization mechanism. We summarize our discussion in
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
195
196
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
Section 8.5; then we follow with open problems in Section 8.6 and, finally, exercises
in Section 8.7.
8.2 BACKGROUND AND OVERVIEW
In sensor networks, IP-like routing techniques face scalability issues since node identifiers, if available, share no topological or geographical similarities and, hence, no
addressing information. In addition, IP-like routing requires global cooperation and
dissemination of information which places undue demand on the energy constraints
that are inherent in such devices.
Current point-to-point routing solutions under investigation in the research community fall into one of two classes, on-demand routing [6–8], and position-based
(often referred to as geometric or geographic) routing [9–12]. Among the former
class, routes are found as needed, often without any prior communications. These
methods consume network resources and are known not to scale. In the latter category, geographic routing decisions are based on relative locations of sensors and are
generally greedy in nature (e.g., nodes forward to neighbors that are geographically
closer to the destination).
In position-based routing the next hop is decided by evaluating the coordinates
assigned to neighboring nodes relative to the coordinates of the destination and the
current node. For example, a node may choose to forward a message to its neighbor
that further minimizes the remaining Euclidean distance to the destination or that
maximizes the savings in energy. Position-based routing is particularly attractive due
to its scalability. In the best case a node need communicate only its position and only
store information regarding its immediate one hop neighborhood. Thus the storage,
length of communication, and data over which decisions are made is O(1).
However, position-based routing is not without its drawbacks and challenges.
When positions are known, for example, messages may get trapped in local minima
where no neighboring nodes further optimize the decision criteria. When positions are
not known, there exist the additional challenges of establishing and maintaining some
coordinate system in order to determine node positions in the network. The localized
nature of position-based routing may also render algorithms blind to obstacles that
might cause routing decisions to fail.
Until recently, the focus of position-based routing has been on the design of protocols that assume the existence of a globally known coordinate system—for example, through preprogramming or the Global Positioning System (GPS). More recent
advances demonstrate that it is possible to establish a network-centric coordinate system, the design of which is often coupled to the routing mechanism. Our attentions
concentrate on families of algorithms that are thought to be feasible and that, in the
case of references 10 and 13, have implementations in development.
Position-based routing algorithms for static sensor networks may be classified
according to the chart in Figure 8.1. This chart matches the categorization of positionbased routing protocols used to organize this chapter. Our discussion continues
in the next section by introducing the topic that follows the left branch of our
ROUTING WITH KNOWN POSITIONS
197
Position-based Routing Protocols
Unknown (Virtual) Coordinates
Known (Physical) Coordinates
Best Effort Delivery
Guaranteed Delivery
Routing Hole Avoidance
Distance-based
Hop-based
Routing Hole Recovery
Figure 8.1. A classification of position-based routing protocols.
categorization, those routing protocols that assume that sensors are aware of their
physical coordinates.
8.3 ROUTING WITH KNOWN POSITIONS
When investigating position-based routing algorithms for sensor networks, it is natural
to assume the a priori knowledge of location information via preprogramming or
GPS-like services. Thus, no efforts are made by these protocols to establish location
within a network. Current methods make some necessary assumptions. First, links
are bidirectional: If node x can receive messages from node y, then y can receive
messages from x. In addition, communication models often adhere to the simple unit
disk graph (UDG) model where all communication ranges are normalized to some
range r. In the UDG neighbors are pairs of nodes separated by a Euclidean distance
≤ r. The UDG model is relaxed later in our discussion. Feasible schemes must, at a
minimum, prove to be loop-free and scalable. We demonstrate some of these subtleties
by starting with a discussion of the naive approaches to position-based routing.
8.3.1 Naive Forwarding Mechanisms
Research into scalable forwarding methods for sensor networks has explored a variety
of forwarding schemes. Scalability is maintained by keeping knowledge only of the
nodes in communication range and choosing the next hop based on this knowledge.
All position-based schemes share a common theme: The next hop is determined by
maximizing or minimizing some criteria associated with local nodes’ positions. We
call this the progress criteria.
Notions of “Progress”. We begin with a formal definition of progress. Say a
node s holds a message to be forwarded to a destination location d (see Figure 8.2)
and has knowledge of all node locations within its communication radius. Then we
198
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
n
g
s
d
α
a
m
Figure 8.2. Progress.
have the following definition used to better understand the behavior of position-based
forwarding schemes.
Definition 8.3.1. The progress of a node x en route from sender s to destination d is
defined as the orthogonal projection of the location of x onto the line sd.
We use Figure 8.2 to better demonstrate various notions of progress as discussed
below. In this figure, sensor node s holds a message destined for sensor node d.
The large circle centered at s represents the communication range of s; thus all nodes
inside the circle are neighbors of s. The dotted horizontal line sd is the line onto which
orthogonal projections are made in order to evaluate and compare progress criteria.
The first forwarding mechanism based on the idea of progress was proposed by
Takagi and Kleinrock [14]. In their Most Forward within Radius scheme, or MFR, the
node with the greatest progress is chosen to be the next hop. Referring to Figure 8.2 we
can see that node m projects furthest onto the line that joins s and d. It is important to
note that m provides the greatest amount of progress though it is not the node closest
to the destination. The motivation behind the use of this myopic routing strategy was
to allow for tractable analysis. Recall that progress is defined as a projection onto a
line. Clearly, because we are working with projections on a line, the dot product of
dm · ds will be minimal over all other neighbors of s when using MFR. This simple
notion allowed the authors of reference 14 to determine an optimal neighborhood size
for their specific problem.
The converse scheme Nearest Forward Progress, or NFP was later proposed in
reference 15. In this work the node with the minimum progress, depicted as node n
in Figure 8.2, is selected as the next hop. This approach may seem counter-intuitive;
but consider that if the broadcast range is variable, then this method has the least
probability of collision as well as improved energy savings.
More recently, the direction of nodes was proposed as a criteria for progress in
reference 16. The node chosen to be the next hop is the node that is closest to the
ROUTING WITH KNOWN POSITIONS
199
direction of the destination. Referring to Figure 8.2, we can see that node c, with
angle ∠csd, is smallest among all of the neighbors of s. Thus, the goal is to minimize
the change in direction from the source to the destination.
The DREAM [17] and LAR [18] projects, simultaneously proposed, use an idea
similar to compass routing. However, these two approaches are best suited in networks where nodes are mobile. (Despite their application in mobile environments,
we provide an overview for completeness.) The node that holds a message m with
destination d calculates an angular range where the message must be forwarded. The
angular range is calculated using (i) the circle centered at d with radius equal to the
maximum movement of d since the last update and (ii) the tangents from the current
location to the aforementioned circle. All nodes within this angular range are sent a
copy of the message. Clearly, the success of both methods relies on some knowledge
of the global network, as well as duplication of messages. Such methods fall outside
of the domain of position-based routing.
Each of MFR, NFP, and Compass Routing is myopic. They are localized algorithms
that require knowledge only of the immediate neighborhood. Unfortunately, their
global behavior is such that none of these approaches can claim to be loop-free. Still,
consider that their forwarding decisions attempt to optimize some local criteria, the
effect of which is to approximate the shortest Euclidean path between the source
and destination. The shortest path may be approximated using another approach that
referred to as greedy forwarding. Greedy forwarding is a localized forwarding scheme
whose express goal is to traverse the Euclidean shortest path. It is this approach
on which most research and development, where there is a known and underlying
Euclidean coordinate system, has focused.
Greedy Forwarding. Greedy forwarding was first proposed in reference 19 as
a routing protocol for wired networks. Referred to as Cartesian routing, the next
hop was chosen to be the neighbor that is closest to the destination. Because this
work predated GPS and other localization services, knowledge of the global topology
was required. The same idea has been reapplied in wireless network settings as the
foundation of innumerable routing schemes and algorithms. It is known to behave
especially well in dense wireless networks such as those envisioned in many sensor
networks.
Greedy forwarding works as follows. Say node s has the neighborhood N =
n1 , ..., nk of size k where each ni is a potential next hop in a traversal that passes
through s. Any message that arrives at s has embedded within it the destination d. The
greedy approach says that the successor to s will be the neighbor ni that minimizes
the Euclidean distance to d. In Figure 8.2 we can see that node s will select g, the
node that most further reduces the distance to the destination among its neighbors.
It can be shown that greedy forwarding is loop-free (using the simple fact that every
hop reduces the distance to the destination). The delivery rate of greedy forwarding
is known to be quite high in dense networks, but diminishes quickly as the density
falls [20]. The problem is that a greedy path may terminate in a routing hole, or void,
as shown in the next section.
200
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
D
Z
D
C
Y
Y
B
S
B
X
A
S
A
X
(a)
(b)
Figure 8.3. Neighbors of stuck nodes may or may not make progress. (a) Convex case.
(b) Concave case.
8.3.2 Routing Holes
Greedy forwarding may be loop-free, but delivery is not guaranteed. A consequence
of greedy forwarding is that a route may terminate at local minima, where nodes have
no neighbors that further reduce the distance to the destination.
We refer our reader to Figure 8.3, which depicts the types of local minima that may
occur: The smaller dotted circle represents the communication range of node S, and
the larger circle centered at D is used to show that all neighbors of S are further from
D than from S itself. Consider that a message destined for node D reaches a minima
at node S. There are two cases to consider. The first, as shown in Figure 8.3a, occurs
where neighbors of S make no progress toward D according to Definition 8.3.1. This
is the obvious case. Less obvious is the case where neighbors of S may actually make
progress toward the destination, yet increase the distance from the current location.
This example is demonstrated in Figure 8.3b, where S clearly lies within the circle
centered at D, while the neighbors of S, x, and A lie outside of the same circle.
These local minima are commonly referred to as voids, holes, or stuck regions.
Their occurrence largely determines the performance of greedy forwarding, whose
performance varies with network density and distribution. A robust sensor network
routing protocol must perform well despite the occurrence of local minima. We proceed in our discussion with proposed solutions to the routing hole problem.
8.3.3 Algorithms for Recovery from Routing Holes
Greedy forwarding is known to perform well in sufficiently dense networks [10], yet
there are no delivery guarantees. Many sensor network applications are loss-sensitive
and have little to no tolerance for undelivered information. Thus greedy forwarding
schemes aiming to guarantee delivery demand that routing voids be circumvented or
ROUTING WITH KNOWN POSITIONS
201
avoided. Work in reference 20 presents evidence that the frequency and impact of
routing holes is manageable. Current solutions to this problem are generally categorized as either broadcasting, planar subgraph methods, or hole mappings, as discussed
below.
Broadcasting, BFS, and DFS Approaches. The naive solution to the routing
hole problem is simply to broadcast some number of hops from the stuck node until a
node is found that is closer to the destination. Two of the first solutions which aimed
to be more efficient were the Geographic Routing Algorithm (GRA) in reference 21
and the Optimal Transmission Ranges (OTR) approach in reference 22.
The authors of GRA implement greedy forwarding as the primary means of transport. If, however, a message reaches a stuck region, GRA launches a route discovery
packet from the stuck node. The route discovery itself takes the form of a breadthfirst search amounting to a limited broadcast, or a depth-first search that eventually
produces a single acyclic path. Each node visited during the recovery phase appends
its own location to the recovery packet, and all discovered paths are stored in routing tables. In networks of size n, routing tables are found to have a mean size of
O(L log n), where L is the average path length between two nodes. In addition, the
authors provide some mechanisms for dealing with location inaccuracy, node failure,
and node mobility.
The goal of OTR differs in that it creates Quality of Service (QoS) paths through
the network. The necessary characteristics are collected and disseminated during path
traversals, which occur from s to d via the depth-first search method. Routing tables
are not used, but instead each node stores the next and previous hop for each packet
until the status of that packet can be confirmed. Such methods face scalability issues as
traffic volume grows, and they are additionally challenged where acknowledgments
are not returned.
Hole Recognition and Mappings. A second approach to the recovery phase is
to recognize and map the hole boundaries in advance. This is the subject of work in
references 23–25. In reference 23, for example, a geometric relationship is described
and is labeled as the tent rule. The tent-rule is used by each sensor node in the network
to determine whether it lies on the boundary of a stuck region. We exemplify the tent
rule using the pictorial representation that appears in Figure 8.4. At each node, x
→ is left
neighbors are sorted angularly. For every pair of neighbors u and v where −
xu
→
of xv , now consider the perpendicular bisectors of each link (labeled b1 and b2 in
Figure 8.4). If the intersection of the bisectors falls outside of the range of x, then x
has no neighbors closer to the region bounded by the communication range and the
bisectors (i.e., the shaded region in Figure 8.4). Note that this method locates regions,
rather than nodes, where no neighbors are closer. This is particularly suitable in sensor
networks where sensing regions are of particular interest.
As stated previously, every node must evaluate its neighborhood using the tent
rule. Only at nodes where the tent rule fails may packets get stuck. Upon detecting a
failure, stuck nodes are responsible for initiating the following hole-mapping process.
202
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
O
u
v
x
b1
b2
Figure 8.4. Tent rule [23]: Node x has no neighbors closer to shaded region as determined by
the intersection of bisectors of adjacent links.
Assume that a packet destined for a node d gets stuck at x. Upon reaching x, the
packet is marked for discovery and forwarded to the neighbor whose connecting
→
edge is immediately counterclockwise, or left, from xd . Each node that receives the
discovery packet appends the newly traversed edge, and then it forwards the packet
along the next counterclockwise edge.
During the mapping process, a discovery packet may traverse an edge that intersects
with another edge that was previously traversed. Where edge intersections occur,
care must be taken. A traversal according to left-hand rule will always return to
the source node, yet intersections may cause the traversal to be led away from the
region that requires a map as can be seen in Figures 8.5a and 8.6a. If an intersection
tj
tj
ti
ti +1
t j +1
t0 = x
ti
t j +1
(a)
(b)
Figure 8.5. “Inside” intersections: (a) Before pruning and (b) after pruning.
t0 = x
ROUTING WITH KNOWN POSITIONS
t0 = x
t0 = x
ti
ti
t i +2
t j +1
t i +1
(a)
tj
t i +2
t i +1
203
tj
(b)
Figure 8.6. “Outside” intersections: (a) Before pruning and (b) After pruning.
appears during a traversal of nodes t0 · · · tk−1 , then we can say that tj tj+1 intersects
ti ti+1 , with j > i. There are only two such cases (the proof of which is available in
reference 23).
r Inside intersections where node tj is not visible to nodes ti and ti+1 . Upon
detection of this intersection, node tj replaces segments ti+1 ti+2 · · · tj tj+1 with
ti tj+1 tj, before forwarding to the node that next appears in counterclockwise
order from tj+1 . The effect of this pruning may be seen in Figure 8.5.
r Outside intersections where node ti is not visible to nodes tj and tj+1 . Upon
detection of this intersection, node tj ignores tj+1 and forwards instead to ti+1 ,
thereby ignoring the intersection all together. This example may be seen in
Figure 8.6.
The traversal terminates upon its return to x, having recorded the path P that maps
the complete stuck region. Once known, a recovery path is shared with all nodes along
the path so that resources required in processing and storing discovery packets may
be avoided in the future.
As an alternative, the boundhole algorithm may terminate upon reaching some node
u such that |ud| < |xd|. If the boundhole is to terminate early, then every recovery
packet must have encoded within it the subpath p ∈ P traversed from x. We call p
the escape path. Clearly, the number of hops in |p| is less than that in |P|; otherwise
x could not be a stuck node. Experiments in reference 20 reveal that the escape path
p is significantly smaller than P, perhaps enough to warrant early termination when
a stuck region is encountered instead of a complete mapping.
Hole recognition, in general, has thus far received little attention in the research
community. Additional algorithms are proposed in references 24 and 25. The former
uses a statistical approach to recognize holes in very dense, uniformly distributed
networks. The latter study is the first where the hole recognition algorithm does not
rely on location information. Neither provides methods for circumventing the stuck
regions, yet it is important to recognize that hole recognition is an important problem
in and of itself.
204
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
Planar Subgraph Methods. The recovery mechanisms so far described all rely
on resources that may be taxing on sensing devices given energy and scheduling
constraints. DFS, BFS, and mapping methods, for example, require storage either
in-device, in-packet, or both. Also, in the worst case, their communication amounts
to a limited broadcast. Such communication requires little to no additional memory,
but demands additional energy consuming transmissions. Drawbacks such as these
often render broadcast-like protocols unsuitable for point-to-point services in sensor
networks. To solve this challenge, many projects have investigated the restriction
of routing to subgraphs of the original network graph. This class of algorithm is
exemplified by one very desirable feature: Such algorithms are stateless; that is, a
node requires no knowledge of the network outside of its own neighborhood, yet is
able to guarantee delivery. All members of this class all share a simple characteristic:
They rely on the construction of the network’s planar subgraph. For our purposes a
planar subgraph is one that contains all the vertices of G, but where edge intersections
occur only at a vertex. (Rather, no edges overlap.)
The most prominent and best-known recovery algorithms route around the routing
hole face (or perimeter) in the planar subgraph. This method is equivalently known
as face routing [9, 26] and perimeter routing [10]. Face routing was first proposed by
Bose et al. [26] with some theoretical bounds. Karp et al. [10] independently proposed
an identical mechanism, but with work on a MAC-compatible implementation.
The inspiration behind these methods comes from the application of a simple rule
known to guarantee escape from a maze. In a maze, by keeping one hand against the
wall at all times, one is guaranteed to find an exit, or return to the point of origin if no
exit exists. This is referred to as the left- or right-hand rule. This works because of a
salient feature of maze construction: If we represent the walls of the maze as edges
in a graph and represent the intersection of those walls as vertices, then the resulting
graph is planar, where no edges intersect.
Wireless network graphs may consist of intersecting edges, so it is necessary for the
planar subgraph method to prune edges from the network graph so that it is planar and
so that it remains connected. Gabriel graphs (GGs) and relative neighborhood graphs
(RNGs) are planar graphs whose constructions are localized, a characteristic particularly suitable to sensor environments. Intersecting edges are eliminated by connecting
pairs of nodes through witness nodes, if such a node exists in a common region.
We refer the reader to Figure 8.7 for examples of each construction where large
circles centered at nodes u and v represent the communication ranges of nodes u
and v, respectively. The GG construction is depicted in Figure 8.7a. For any pair of
neighbors u, v, if a witness node w exists within the circle whose diameter is |uv|
(shown as the shaded region in Figure 8.7a), then uv is removed from the graph.
Similarly for the RNG in Figure 8.7b, if w appears within the intersection of the two
circles centered at u and v with radii equal to |uv| (again shown as the shaded region),
then uv is removed from the graph. In each, pruning is valid since communication
may continue through w. In references 9 and 26 and later in reference 10, it was
shown that if the unit disk graph is connected, then the intersections with Gabriel
graph UDG ∩ GG and with relative neighborhood graph UDG ∩ RNG remain
connected.
ROUTING WITH KNOWN POSITIONS
(a)
205
(b)
Figure 8.7. Localized constructions of planar subgraph. (a) Gabriel graph construction.
(b) Relative neighborhood graph construction.
Face- and perimeter-routing techniques choose to route greedily whenever possible. The recovery phase is initiated only when a message gets stuck at some node.
Upon receipt of message m destined for node t, node s inspects the message to reveal
it either in greedy or recovery mode. The corresponding algorithms, executed at each
node, are listed in Algorithms 1 and 2.
ALGORITHM 1. Greedy Mode
1:
2:
3:
4:
5:
6:
7:
8:
let s be the current node
let w be the neighbor closest to t
if (w, t) < (s, t) then
forward m to w
else {no such w exists; m is stuck at s}
mark packet as recovery with location of s
→t
forward to neighbor that is left of s,
end if
ALGORITHM 2. Face/Recovery Mode
1:
2:
3:
4:
5:
6:
7:
8:
9:
let u be the node from which m was received
let v be the current node
let w be the neighbor closest to t
if (w, t) < (s, t) then
mark packet as greedy
forward to w
else
−→
forward to neighbor that is left of v,
u
end if
206
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
Algorithms 1 and 2 assume that a stuck node has first constructed the portion of the
planar subgraph that occurs within its view, as above. If in greedy mode a packet is
forwarded according to Algorithm 1, where a sensor forwards to the neighbor closest
to the destination. If no such neighbor exists, then the sensor node forwards according
to Algorithm 2. Once stuck, the message m is marked recovery and is forwarded to
the neighbor that appears first in a counterclockwise direction. While in recovery
mode, each sensor that receives the message first checks for a neighbor closer to the
destination than the point at which the message was marked recovery. Returning to
Figure 8.3a, we can see s is a stuck node. In the case of the left-hand rule, sx is left
of sd. The recovery path sxy terminates upon finding z since z is closer to d than s,
where recovery began.
One special case occurs where an edge uv intersects sd during recovery. The
solution is left as an exercise.
These algorithms are especially suited to sensor networking environments: They
guarantee delivery, are localized in their operation, and are stateless in the sense
that they require no information outside of the location of their neighbors. However, we see in the next section that subplanar methods are not without their
drawbacks.
Drawbacks and Challenges of Planar Methods. Planar subgraph methods,
while promising, face three major challenges before their deployment may be considered feasible. First, planarization assumes locations are accurate, an assumption that
may be untrue. Second, the localized nature of the planarization process means that
one sensor node may be blind to environmental obstacles that are visible to neighboring sensors. Finally, the unit disk model on which these algorithms are constructed is
a poor representation of the real world. We use this section to touch on each of these
challenges.
Routing protocols that operate over pre-established coordinate systems are generally designed under the assumption that location information is accurate. This assumption is challenged by real-world limitations. GPS, for example, offers high-resolution
localization, but is subject to line-of-sight constraints rendering it ill-suited to underground, underwater, and undercoverage applications, to name a few. Furthermore, the
added expense incurred by supporting GPS in all nodes is restrictive. Many proposals exist to resolve this issue by supporting some small location-aware infrastructure
[27–30], from which all other nodes may learn their locations; yet even these fall prey
to poor resolution and estimation errors.
There are cases where localization and estimation errors have little adverse effect. For example, in reference 31 greedy routing was evaluated in simulated networks with localization errors. The results show that the performance of greedy
routing is largely unaffected by inaccuracies up to 40% of the radio range. Conversely, there are contexts in which localization errors can be destructive to correct
protocol operation. One such example is demonstrated in reference 32, which models and evaluates location error on face-routing techniques; here we find that errors of as little as 20% of the communication range caused high packet drop rates,
ROUTING WITH KNOWN POSITIONS
207
d
v
w
u
Figure 8.8. Planar methods may fail without inter-neighbor link information.
nonoptimal path selection, and looping. These experiments were reinforced in reference 33, which investigates the types of location errors that lead to performance
degradation in face-routing. In doing so, the authors were able to suggest a simple
check, consisting of mutual agreement between nodes, to resolve many of the problem
cases.
In addition to position errors, a second challenge faced by planarization processes
lies in its best feature, namely, that all operations are local and occur without need for
negotiation with neighboring nodes. The level of localization inherent in the planarization process leaves open the possibility for incorrect output. Specifically, because a
node planarizes its neighborhood using only the locations of its 1-hop neighbors, a
node may assume links exist where they do not. Consider the example in Figure 8.8
where a packet destined for node d gets stuck at node u. The planarization of the
neighborhood of u removes the link uv believing v is reachable via w. Since u is
unaware of any obstacles or interference between w and v, u’s planarization of its
neighborhood is incorrect. This phenomenon was first demonstrated in references 34
and 35.
Finally, we must address the likelihood that the unit disk graph correctly represents wireless network graphs. Experimental evidence in [36, 37] suggests that radio
ranges are inconsistent and irregular. Experiments in reference 38 conducted on two
sensor networks further demonstrate and enumerate the difficulties that occur because
the unit disk assumption is violated. The experimental evidence is used to design a
protocol that corrects the failings of the Gabriel and relative neighborhood graph
constructions. Protocols that avoid the unit disk assumption are discussed in the next
section.
208
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
8.3.4 Guaranteed Delivery in the Real World
Face-routing algorithms are attractive because they are localized and efficient. Yet
as previously discussed, they are known to be ill-suited to physical environments.
There are two reasons for this. First, the Gabriel and relative Neighborhood graph
constructions rely on the assumption that all communication ranges in the network
are identical and uniform (the attributes associated with the UDG model). Moreover,
these distributed constructions are unable to resolve links broken by obstacles or
interference. Recent breakthroughs have begun to surmount the impracticalities of
face-routing while maintaining delivery guarantees [13, 39].
The first-known protocol to guarantee delivery over a global coordinate system
without planarization or the use of left-hand rule is the Greedy Distributed Spanning
Tree Protocol (GDSTR) algorithm in reference 39. GDSTR builds on the fact that
any message can be successfully delivered via depth-first search if the network is
connected via a spanning tree. (This fact alone does not solve the problem: Delivery
would be inefficient, needing up to 2n−3 hops.) Leong et al. [39] describe a new
type of spanning tree, the hull tree, to route more efficiently. A hull tree is a spanning
tree with one added piece of information: Each node records the convex hull that
contains all of its descendants in the tree. (The convex hull of a set of points is the
smallest polygon that contains all the points.) In GDSTR, forwarding occurs greedily,
as with most position-based protocols. If a message reaches a void, a recovery mode
is initiated where convex hulls are used to determine the regions of the network that
contain unreachable destination. This information is used by GDSTR to route along
the spanning tree to forward to the appropriate convex hull. If a node is found en
route that is closer to the destination than the node where the message was stuck,
then GDSTR returns to greedy forwarding. GDSTR is known to scale well as the
neighborhood size grows. Furthermore, the use of multiple hull trees adds faulttolerance to the network; and if multiple trees are rooted at opposite ends of the
network, routing efficiency improves.
The Cross-Link Detection Protocol (CLDP) proposed in reference 13 circumvents
voids by face-routing, using left-hand rule over a planar subgraph of the network;
its design, however, is motivated by the observation that routing difficulties in planar
subgraph methods arise, in part, due to the constructions themselves. (Recall from
previous sections that successful local planar subgraph constructions rely on the unit
disk graph.) For this reason, CLDP proposes an alternate construction of planar subgraphs that assumes only that links are bidirectional. CLDP operates in a distributed
fashion, exchanging some localized operation for accurate information. The idea
behind CLDP is that each node is able to probe the vicinity for intersecting links. A
probe packet is initialized with the endpoints of the first link to be probed. Figure 8.9
shows the simplest example of a probe traversing a graph using the left-hand rule.
Say a probe starts at node d for link (d, a). Sensor node a then forwards to b as determined by the left-hand rule. When the probe reaches node b, the intersection (b, c)
is recorded, before the probe packet continues its traversal. (Recall that a traversal
according to the left-hand rule will eventually return to its starting point.) Prompted
by the return of the probe packet, d proceeds to prune links. Figure 8.9 depicts only
ROUTING WITH UNKNOWN POSITIONS
a
b
c
d
209
Figure 8.9. CLDP probes links using left-hand rule.
the naive case; the graph remains connected following a removal of either of the two
intersecting edges. We leave as an exercise the identification of remaining cases and
the care that must taken when pruning links to keep the graph connected. Furthermore,
to avoid the slow process of scheduling serial probing by neighboring nodes, a system
for concurrent probing is proposed. Concurrent probing is achieved by implementing a mechanism to “lock” links so that no more than one link is removed at a time
from any vicinity. CLDP is one of very few protocols to have been implemented on
testbeds [13]. The associated communication complexities and storage costs revealed
in this process (see reference 38) are motivation to develop alternative approaches to
guarantee delivery.
Protocols such as CLDP and GDSTR, in order to be feasible for physical networks,
sacrifice efficiency for accuracy. CLDP requires negotiation within each neighborhood in order to prune appropriate links, and GDSTR must broadcast information
to construct and maintain its hull trees. It remains an open question whether such
tradeoffs are a necessity.
8.4 ROUTING WITH UNKNOWN POSITIONS
By nature of its name, position-based routing protocols may be thought to assume that
position location is available for use by the routing protocol. Often this information is
not and cannot easily be made available from the outset: For example, the inclusion of
GPS in the manufacturing process is cost prohibitive; and despite this, sensor nodes
may be deployed in adverse environments where GPS information cannot penetrate.
This begs the question, Can position-based routing be applied to contexts where
location information is available to only a few nodes? This question is furthered by the
notion that many proposed applications do not require any knowledge of geography,
but do require advanced point-to-point routing services.
210
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
Applications such as data querying [1, 2, 40], reactive-tasking where events are
triggered by local events, and data-centric storage [3, 4] all demand robust routing
yet ignore physical location. Network applications such as these will identify a node
by its identifier or by the data it stores, each of which provides no means by which
to route using position. In such contexts a network may implement position-based
routing by use of a distributed hash table (DHT), which partitions keys/identifiers
from the owner of the key [41]. In a DHT sensor network, position-based routing is
an appealing means to bridge this partition.
In this section we explore the potential for position-based routing where location
information is generally unavailable. The solution that has received the widest attention is to develop protocols that construct their own maps and create their own
coordinate systems. Such protocols are often said to construct and rely upon virtual
coordinate systems. The routing in virtual coordinate systems is often coupled with the
construction of the coordinate system, a coupling that necessitates an understanding
of the coordinate construction itself.
One of the first feasible algorithms to construct a virtual coordinate system for the
purpose of routing in wireless systems appears in reference 42. This method requires
no infrastructure. However, its centralized nature and expensive cost (O(n3 )) render
it poorly suited for sensor networks.
The establishment of a coordinate system in a distributed fashion, in order to
route messages between nodes, is nontrivial. We introduce two methods under
investigation. The first is the distance-based approach where the aim is to create
some semblance of a Cartesian space (without which the direction between nodes
is unclear). The alternative is a hop-based protocol that addresses using the distance
between nodes, in hops. Each of these share drawbacks, a discussion we reserve until
later.
8.4.1 Distance-Based Coordinates and Routing
One of the earliest protocols promoting a virtual coordinate system construct with
a corresponding routing mechanism is the GEM infrastructure [43]. In GEM a labeled graph is constructed and embedded into the network topology in a distributed
fashion, where the label encodes a node’s position within the graph. The idea is demonstrated by coupled coordinate setup and routing protocols called VPCS (Virtual Polar
Coordinate Space) and VPCR (Virtual Polar Coordinate Routing), respectively. Using VCPS, the GEM system is able to embed a ringed tree into the network in
order to create a polar coordinate system. The VCPR algorithm designed to route
over the polar coordinates is the first to guarantee delivery in a point-to-point fashion with no a priori geographic information. While the initialization scales well,
the recovery process initiated on node failure or link degradation is expensive.
Recovery may force a large number of nodes to participate in a recomputation
process.
To better demonstrate the ideas behind distance-based constructions, we focus our
discussion on the NoGeo method proposed in reference 12. In it the authors provide a
mechanism to construct a virtual Cartesian space and implement the simplest form of
ROUTING WITH UNKNOWN POSITIONS
211
greedy routing. Their work is a derivation from previous work intended to test graph
connectivity [44], and it assumes that nodes may accurately measure the distances to
their immediate neighbors.
Virtual Coordinate Construction. For clarity, explanation of the NoGeo algorithm is presented in a cumulative fashion where at each additional step we remove
some knowledge from the system. Hence we begin by describing the coordinate construction under the assumption that nodes on the network boundary, or perimeter,
are aware of their position relative to the network as well as their exact location.
The algorithm in reference 12 is based on an iterative relaxation procedure from
reference 44 which is used to determine the locations of all remaining nodes in
the network. The procedure is such that each link is represented by a force that
pulls its adjoining neighbors together. The force in each of the x, y directions is
proportional to the difference in the x, y coordinates. At any iteration a node’s
neighbors are held with fixed position; its equilibrium position, where the sum of
the forces acting on it is zero, is the average of all its neighbors’ x coordinates
in the x direction, and similarly for the y direction. This relationship motivates
the iterative procedure repeated periodically at each node i using the relaxation
equations
xi =
yi =
k∈neighbor set xk
(8.1)
size of (neighbor set(i))
k∈neighbor set yk
(8.2)
size of (neighbor set(i))
in each of the x and y directions, respectively.
As an example, consider the wireless sensor network in Figure 8.10a. This network consists of 3200 nodes within a 200 × 200 unit space where each node has a
200
all nodes
perimeter nodes
200
150
150
100
all nodes
perimeter nodes
50
100
Y
Y
250
0
50
-50
0
-100
-50
-50
0
50
100
X
(a)
150
200
250
-150
-150 -100
-50
0
50
100
150
200
X
(b)
Figure 8.10. The (a) real and (b) virtual coordinates of a 3200-node network. (a) Physical
network in 200 × 200 space. (b) Virtual network where perimeter nodes are initially unknown.
212
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
250
250
all nodes
perimeter nodes
200
150
100
100
50
0
0
-50
-50
0
50
100
150
200
250
-50
-50
Y
150
100
50
all nodes
perimeter nodes
200
150
Y
Y
250
all nodes
perimeter nodes
200
50
0
0
50
100
X
X
(a)
(b)
150
200
250
-50
-50
0
50
100
150
200
250
X
(c)
Figure 8.11. Intermittent virtual coordinates of 3200 nodes in a 200 × 200 space where nodes
are known in advance to lie on the network’s perimeter. (a) After 10 iterations. (b) After 100
iterations. (c) After 1000 iterations.
communication range of 8 units. As the iterative procedure described above is executed over this network, sensor coordinates will gradually shift to match their true
coordinates within the network. This “shaping” is depicted following 10, 100, and
1000 iterations in Figures 8.11a – 8.11c, respectively. In this example the initial coordinates of each node have been set to the center of the network at (100, 100). (We
later return to compare the virtual and actual topologies.)
This coordinate construction may be prefaced with additional steps if there is no
advance knowledge of perimeter nodes, provided there are two beacons somewhere
in the network. [Beacons are distinguished from the remaining network because they
either (a) hold and disseminate information or (b) play some specific coordination
role required for successful setup and communications.] Either beacon may be used
to identify nodes on the perimeter, after which point the perimeter nodes exchange
messages to determine their locations.
The first step in accomplishing this task is to identify nodes on the perimeter.
Recall the assumption that two beacon nodes exist. Either beacon is designated as
the primary beacon, which broadcasts a HELLO message. Each node uses receipt
of the HELLO message to determine its location from the beacon, according to
the perimeter node criterion. The Perimeter Node Criterion says that if a node is
farthest away from the bootstrap beacon among all nodes in its two-hop neighborhood, then this node decides it lays on the perimeter of the network. This is by no
means an exact determination, but simulations show that it identifies a sufficient number of perimeter nodes. Once the nodes have identified their perimeter status, they
must coordinate to exchange information and calculate their coordinates according to
Algorithm 3.
Each perimeter node, following Step 3 in Algorithm 3, has established its own
virtual coordinate. One challenge remains. Each perimeter node has established its
own virtual coordinate, yet the network lacks a global orientation. The reason is that
any set of coordinates satisfying Eq. (8.3) may be rotated, translated, or transposed.
The result is that perimeter nodes cannot be guaranteed to share the same sense of
direction.
ROUTING WITH UNKNOWN POSITIONS
213
ALGORITHM 3. Coordinate Determination among Perimeter Nodes in NoGeo
1. Each perimeter node broadcasts a HELLO message so that each may calculate its distance
to every other perimeter node. The result is stored in a perimeter vector.
2. Each perimeter node broadcasts its perimeter vector to the network so that each perimeter
node knows the distance between every pair of perimeter nodes in the network.
3. Each perimeter node triangulates to find the coordinates of every perimeter node in the
network. Coordinates are chosen so as to minimize
i,j∈perimeter
(measured dist(i, j) − euclid dist(i, j))2
(8.3)
where measured dist(i, j) represents the distance between nodes i, j as measured in 1,
and dist(i, j) is the Euclidean distance between the virtual coordinates of i and j.
The solution is the reason why two beacons must exist. The two beacons share their
perimeter vectors and participate in the triangulation process. During this process,
each of the perimeter nodes calculates the center of gravity of the network. The
combination of the two beacons with the center of gravity provides a set of axes on
which all nodes can agree: The center of gravity becomes the origin, while each of
the two beacons define the positive x, y axes.
An example topology at equilibrium, following the execution of perimeter identification, coordinate determination, and dissemination, appears in Figure 8.10b. This
is the virtual network space that corresponds to the actual network in Figure 8.10a.
Using this scheme, the virtual coordinates maintain a structure similar to the actual
coordinates, but with some caveats. For example, the network appears to have been
rotated about the virtual center of gravity. However, the rotation is uniform and consistent, meaning that the rotation is network-wide. Note that holes in the virtual networks
constructed in Figures 8.10b and 8.11c are much larger in size than in the actual network of Figure 8.10a. As discussed below, this salient feature does not reflect itself
in the routing mechanism.
Routing in the virtual Cartesian coordinate system takes the form of pure greedy
routing as described in Section 8.3.1: Messages are forwarded to the neighbor
that most reduces the distance to destination until no such neighbor exists or the
destination is found. Despite the inaccuracies of location determination and the
increase in hole sizes (as reflected in physical space), greedy routing over virtual
coordinate systems performs better than it does over physical coordinate systems.
The rate of successful delivery rate is higher when routing over virtual coordinate
systems.
The underlying reason for an increase in the rate of delivery is that locations are
assigned coordinates relative to the locations of other nodes in the network, not relative
to the space that the network occupies. Hence, a virtual coordinate system accurately
reflects not network geography, but rather network connectivity. This relationship,
214
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
when reflected in actual space, is responsible for the apparent growth in hole sizes.
(Despite this fact, the physical coverage area is unaffected, a necessary property of
sensor networks.) It is also responsible for the improvement in the performance of
greedy routing because it reflects network paths and connections more than node
location in the physical space.
One drawback to approaches in this class of position-based routing is the expense
incurred by the computation and transmission necessary to establish a reasonable
coordinate system. Consider √
in NoGeo, for example, that it is somewhat impractical
that initialization requires O( n) nodes to flood the network, and it is also inpractical
for√
at least√this number of nodes to store O(n) state (since the distance vectors occupy
O( n × n) space). We proceed with an approach that reduces this expense in
exchange for a loss in topological accuracy.
8.4.2 Hop-Based Coordinates and Routing
One alternative approach to constructing virtual coordinates foregoes Euclidean
approximations, altogether. Instead, designated coordinates are comprised of a vector
that contains a set of hop-distances to beacons located around the sensor network
[45–48]. As is the case with NoGeo, the goal of these methods is to deliver packets
in an environment with no a priori knowledge of node locations, in a point-to-point
manner. We demonstrate this approach to position-based routing using Beacon Vector
Routing (BVR) in reference 47 and reserve discussion for the differences with other
projects until later.
BVR is a protocol that assigns routing coordinates and defines a distance function used in forwarding decisions. A node’s coordinate is a tuple recording the hop
distance to each of a subset of available beacons, information that is disseminated
using reverse path tree constructions. (A reverse path tree construction occurs when
a beacon broadcasts its existence to the network and all remaining nodes record the
shortest hop distance to that beacon.) A distance function is used to route greedily.
In the event that greedy routing fails, a correction mechanism exists to guarantee
delivery.
Let r denote the total number of beacons, and let qi denote the distance in hops
from a node to beacon i. The position of node q is the tuple (q1 , q2 , . . . , qr ). By this
definition, it is possible for multiple nodes to share the same coordinate so a node
identifier is necessary to disambiguate between nodes with identical coordinates. The
distance function must favor greedy forwarding to maintain a high level of efficiency.
When evaluating the distance function, the BVR metric aims to minimize the difference in coordinates componentwise. This metric is based on the idea that it is better
to move toward a beacon close to the destination than it is to move away. Hence, the
distance function δ is designed to move a message toward a beacon if the destination is
closer to the beacon than the current node; that is, it also moves a message away from
a beacon if the destination is further away. (Note that using this intuition, movement
toward a beacon always reduces the distance to the destination but moving away is
not: The destination may sit on the other side of the beacon from the current node.)
ROUTING WITH UNKNOWN POSITIONS
215
Let the distance function δ(p, d) measure the goodness of node p as a next hop to
d. The above intuition is encapsulated into the distance function using the sums
δ+
k (p, d) =
δ−
k (p, d) =
max(pi − di , 0)
(8.4)
max(di − pi , 0)
(8.5)
i∈Ck (d)
i∈Ck (d)
where Ck (d) is the set of k beacons closest to d. The metric is a sum of differences
derived from Eqs. (8.4) and (8.5). δ+
k is the sum of the differences of beacons closer
to the destination than the node p, while δ−
k is the sum of the differences of beacons
further away.
BVR routes greedily as follows. The next hop is chosen to be the node that mini−
mizes δ+
k ; any tie that may occur is broken by minimizing δk . Note that the k beacons
may number fewer than the total number of beacons in the system and that the smallest
difference δmin encountered during a traversal must be stored in the message header
for reference.
ALGORITHM 4. Overview of Beacon Vector Routing (BVR)
1. (Greedy) Where possible, forward to the neighbor that minimizes Eq. (8.4), breaking
ties using Eq. (8.5).
2. (Recovery) If no such neighbor exists, record the current distance in the packet as δmin
and forward to the beacon closest to the destination.
3. (Recovery) If message has reached a beacon without reverting back to greedy mode,
broadcast with a time-to-live equal to hop distance from the destination node.
A global view of the main BVR algorithm is summarized in Algorithm 4. As is the
case with other greedy schemes, there are occasions where forwarding may terminate
prematurely, failing to find a neighbor that improves on δmin . BVR is able to guarantee
delivery using a two-tier recovery mode. First, if a node has no neighbor closer to
the destination than itself, it will forward the message to the beacon closest to the
destination. The idea is that if a sensor node is unable to find the destination, then
it should send the message in the direction of a node that can. Interim nodes that
receives the message will forward to the destination as if recovery never occurred,
if possible. Second, if a beacon is unable to further minimize δmin , then it initiates
a scope flooding to find the destination. (The scope of the flood is known since
the destination coordinates record the hop-distance from each beacon.) While this
recovery mechanism is an expensive means of guaranteeing delivery, it is found to
occur infrequently in simulations.
216
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
Similar ideas have been proposed in references 45 and 46. FT-BVR in reference 46,
for example, extends BVR using fault-tolerant techniques to improve on a variety of
performance metrics. Logical Coordinate Routing (LCR), in reference 45, is similar
to BVR in its coordinate construction and routing mechanisms. Where it differs is
in recovery. Where BVR ultimately resorts to a scoped broadcast, LCR backtracks
along the path-thus-far until an alternate path is found or the message returns to the
originating node where delivery is deemed to have failed. Clearly, LCR avoids the
expense of a broadcast in exchange for additional state (either in the message or stored
at interim nodes).
8.4.3 Drawbacks and Challenges with Unknown Coordinates
Generally speaking, mechanisms that construct and route over virtual coordinate
systems manage to overcome many of the challenges that face routing in physical
coordinate systems. To wit, virtual coordinate routing schemes make no assumptions
pertaining to the unit disk graph, an assumption that is shown to be violated in practice [36, 37, 49], nor do they require GPS-like services that may be unavailable. This
benefit does come at some expense, however.
The first limitation is related to the infrastructure necessary for correct operation.
Constructs that create coordinate systems that reflect connectivity require the use of
beacons. (Beacons are defined as nodes that have knowledge of their location via
global positioning systems that have some preprogramming or that are placed strategically.) Each of these incurs a cost increase on the manufacturing and deployment
process that is intended to otherwise be cheap. In addition, many environments wellserved by sensor-network applications can be volatile. In such environments a small
number of beacons may be lost, destroyed, or otherwise disabled. The use of beacons
may be restrictively expensive and complicated when deployed broadly, yet present
fault-tolerance and failure issues when deployed narrowly. If beacons are required,
does it suffice to select them randomly? If not, what determines the goodness of a
beacon? Despite these challenges, beaconing may be an effective solution to a difficult
problem.
In addition, virtual coordinate and routing methods also incur additional complexity in communication, and sometimes computation. Coordinates in the network cannot
be learned unless information is broadcast so that all nodes share similar knowledge.
Energy consumption is a greater concern since transmission is known to be a highenergy operation, and broadcasts increase the chance of collision (though there are
proposals to intelligently broadcast such as in reference 50. Often the determination
of coordinates requires additional computation.
Finally, we address the accuracy of virtual coordinates. Coordinates that record
hop counts are likely to be duplicated throughout the network, and so additional care
must be taken to deliver messages to the intended recipient. Coordinate systems that
measure physical distances reflect network connectivity well, but are subject to limits in resolution. In reference 51, work on Nagpal’s algorithm, a set of algorithms
to construct and improve coordinates based on distances from three distinct beaπ
cons, reveals the smallest resolution to be 4n
r, where n is the average neighborhood
TABLE 8.1. Summary of Various Protocol Attributes
MFR,
GRA,
NFP,
Greedy
DFS-QoS
Compass
UDG Assumption
Loop-free
Critical nodes
Beacon requirements
Multiple routes
Startup coordination
Void recovery cost
Guaranteed delivery
Obstacle-resilient
Coords reflect connectivity
aA
No
No
No
No
No
Local
NA
No
Yes
No
Yes
Yes
No
No
No
Local
NA
No
Yes
No
No
NA
No
No
No
Local
Medium
Yes
Yes
No
Mapping
Planar
Subgraph
Methods
CLDP
GDSTR
NoGeo
BVR
Yes
NA
No
No
Yesb
Extended
Medium
Yes
Yes
No
Yes
Yesa
No
No
No
Local
Low
Yes
No
No
No
NA
No
No
NA
Extended
High
NA
Yes
No
No
Yesa
Yes
No
Yesb
Global
Medium
Yes
No
No
No
Yes
Yes
Yes
No
Global
NA
No
Yes
Yes
No
Yes
No
Yes
No
Global
Medium
Yes
Yes
Yes
loop only occurs when a packet returns to its origin, indicating that the destination is unreachable.
choice of route occurs only when recovering from a void region.
NA, not available.
b The
217
218
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
size and r is the radius. If achievable, this limit may or may not be acceptable in
practice, see Table 8.1 for a summary of routing algorithms.
8.5 SUMMARY
In this chapter we have explored algorithms and mechanisms for position-based routing in sensor networks, whose communication graphs are large and dense relative to
traditional wireless ad hoc networks. Position-based routing occurs where forwarding
decisions are based on a metric that reflects the position of nodes in reference to the
physical space or to reference nodes.
Our discussion began under the assumption that location information is available
and accurate. In such environments it is necessary to contend with the appearance
of local minima if any delivery guarantees are to be made. This problem has simple
solutions that are impractical in physical deployment. It has thus far been necessary to
sacrifice simplicity for accuracy in order to resolve the challenges of implementation
and deployment.
Furthermore, we presented algorithms that require no a priori location information.
Nodes that use such algorithms must cooperate to establish their own coordinates
within the network. The construction of the coordinate system is often coupled with
routing. Local minima are less of a problem since coordinates emphasize the network
topology and reflect connectivity. However, their success relies on the use of beacons.
This raises issues related to fault-tolerance, selection, deployment, and additional
communication and computational cost in network initialization and maintenance.
For this reason we present some open problems and future directions.
8.6 FUTURE WORK AND OPEN PROBLEMS
There are few position-based routing algorithms implemented as protocols, making
difficult the task of measurement and evaluation. Many additional open problems
remain. Beacon selection remains a challenging task. Current research assumes that
beacons may be chosen efficiently at random or that a selection infrastructure exists.
A random selection may yield less-than-optimal resolution, while any specialized
infrastructure removes from the uniformity of the network and the ease of deployment.
Another current challenge that faces the community is to efficiently guarantee pointto-point services without planarity. As described in this chapter, current solutions
require broadcasting or locking mechanisms, each of which is costly. Lastly, there
is little work which allows us to understand the sacrifices necessary in efficiency
and resource conservation that are necessary to provide reasonable accuracy without
resorting to broadcasts. Open avenues of exploration exist where recognition and
circumvention of local minima are concerned. Also, it is entirely unclear how positionbased routing in general will affect (or can be made to interact with) other physical
and application layers.
EXERCISES
219
Despite these challenges, the growing research into position-based routing and its
associated challenges is reflective of its potential for providing scalable and efficient
data transport in sensor networks.
8.7 EXERCISES
1. In mechanisms that map void regions in advance, a path results. It is possible for
multiple nodes along this path to identify themselves as stuck nodes, in which
case each stuck node initiates the mapping protocol. This is a redundant exercise.
How might mapping techniques be augmented to deal with this occurrence?
2. Assume the unit-disk graph (UDG) is connected. Show that the Gabriel (GG)
and relative neighborhood (RNG) graphs do not disconnect the unit-disk graph.
In other words, show the intersections GG ∩ UDG and RNG ∩ UDG are connected.
3. In face-routing methods, a packet in recovery mode is routed greedily as soon
as possible (i.e. some node is found that is closer to the destination than where
recovery began). However, it is conceivable that only an edge sits closer to the
destination (as shown in Figure 8.12). Modify face-routing so that it resolves
this problem and maintains its delivery guarantees.
4. Nodes in CLDP probe incident links in order to find and remove intersections.
However, there are cases where concurrent probing may disconnect the network graph. Enumerate these cases (there are 3) and establish a mechanism to
guarantee a disconnection does not occur.
5. In our discussion of BVR and other mechanisms relying on beacons, we left
open any mechanism to replace a failed beacon. Assume that a minimum number
r of beacons is required for correct operation. There are two operations at issue.
e
d
t
c
a
b
s
Figure 8.12. A message stuck at s will traverse edge (ba) according to left-hand rule before
returning to s.
220
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
(a) How might a sensor node in the network detect that less than r beacons are
functional?
(b) Once it is detected that too many beacons have failed, how does a node
decide to become a beacon itself? What if multiple nodes simultaneously
decide to become beacons?
BIBLIOGRAPHY
1. D. Ganesan, D. Estrin, and J. Heidemann. Dimensions: Why do we need a new data handling
architecture for sensor networks? SIGCOMM Computer Communication Review (CCR),
33(1):143–148, 2003.
2. B. Greenstein, D. Estrin, R. Govindan, S. Ratnasamy, and S. Shenker. Difs: A distributed
index for features in sensor networks. Elsevier Journal of Ad Hoc Networks, 2:39–46, 2003.
3. S. Ratnasamy, B. Karp, S. Shenker, D. Estrin, R. Govindan, L. Yin, and F. Yu. Data-centric
storage in sensornets with ght, a geographic hash table. Mobile Network Applications,
8(4):427–442, 2003.
4. S. Shenker, S. Ratnasamy, B. Karp, R. Govindan, and D. Estrin. Data-centric storage in
sensornets. SIGCOMM Computer Communication Review, 33(1):137–142, 2003.
5. D. Estrin, R. Govindan, J. Heidemann, and S. Kumar. Next century challenges: scalable coordination in sensor networks. In Proceedings of the fifth annual ACM/IEEE International Conference on Mobile Computing and Networking, Seattle, WA, 1999, 263–
270.
6. D. B. Johnson and D. A. Maltz. Dynamic source routing in ad hoc wireless networks.
In T. Imielinski and H. Korth, editors, Mobile Computing. Kluwer Academic Publishers,
Norwell, MA, 1996, pp. 000–000.
7. C. Perkins and P. Bhagwat. Highly dynamic destination-sequenced distance-vector routing
(DSDV) for mobile computers. In ACM SIGCOMM’94 Conference on Communications
Architectures, Protocols and Applications, 1994, pp. 234–244.
8. C. E. Perkins and E. M. Royer. Ad hoc on-demand distance vector routing (aodv). In
Proceedings of the 2nd IEEE Workshop on Mobile Computing Systems and Applications,
February 1999, pp. 90–100.
9. P. Bose, P. Morin, I. Stojmenović, and J. Urrutia. Routing with guaranteed delivery in
ad hoc wireless networks. Wireless Networks, 7(6):609–616, 2001.
10. B. Karp and H.T. Kung. Gpsr: Greedy perimeter stateless routing for wireless networks. In
Proceedings of ACM MobiCom, Boston, MA, September 2000.
11. F. Kuhn, R. Wattenhofer, Y. Zhang, and A. Zollinger. Geometric ad-hoc routing: Of theory
and practice. In Proceedings of the Symposium on Principles of Distributed Computing
(PODC), 2003, pp. 63–72.
12. A. Rao, S. Ratnasamy, C. Papadimitriou, S. Shenker, and I. Stoica. Geographic routing
without location information. In Proceedings of ACM MobiCom, San Diego, CA, September
2003.
13. Y.-J. Kim, R. Govindan, B. Karp, and S. Shenker. Geographic routing made practical. In
Proceedings of the 2nd USENIX Symposium on Networked Systems Design and Implementation (NSDI), Boston, MA, May 2005.
BIBLIOGRAPHY
221
14. H. Takagi and L. Kleinrock. Optimal transmission ranges for randomly distributed packet
radio terminals. IEEE Transactions on Communications, 32(3):246–257, 1984.
15. T. C. Hou and V. O. K. Li. Transmission range control in multihop packet radio networks.
IEEE Transactions on Communications, 34(1):38–44, 1986.
16. E. Kranakis, H. Singh, and J. Urrutia. Compass Routing on Geometric Networks. In
Proceedings of the 11th Canadian Conference on Computational Geometry, Vancouver,
August 1999, pp. 51–54.
17. S. Basagni, I. Chlamtac, V.R. Syrotiuk, and B.A. Woodward. A distance routing effect
algorithm for mobility (dream). In Proceedings of the 4th Annual ACM/IEEE International
Conference on Mobile Computing and Networking (MobiCom, pages 76–84, 1998.
18. Y. B. Ko and N. H. Vaidya. Location-aided routing (lar) in mobile ad hoc networks. In
Proceedings of the 4th Annual ACM/IEEE International Conference on Mobile Computing
and Networking (MobiCom), 1998, pp. 66–75.
19. G. G. Finn. Routing and addressing problems in large metropolitan-scale internetworks.
Technical Report ISI/RR-87-180, Information Sciences Institute (ISI), March 1987.
20. M. Fayed and H. T. Mouftah. Characterizing the Impact of Routing Holes on Geographic
Routing. In Proceedings of Systems Communications 2005 (ICW / ICHSN / ICMCS / SENET
2005), August 2005, pp. 401–406.
21. A. Puri, R. Jain, and R. Sengupta. Geographical routing using partial information for wireless ad hoc networks. IEEE Personal Communications, 8(1):48–57, 2001.
22. I. Stojmenovic, M. Russell, and B. Vukojevic. Depth first search and location based localized
routing and qos routing in wireless networks. In Proceedings International Conference on
Parallel Processing (ICPP), Washington, DC, 2000, IEEE Computer Society, New York,
p. 173.
23. Q. Fang, J. Gao, and L. Guibas. Locating and Bypassing Routing Holes in Sensor Networks.
In Proceedings of IEEE/ACM Infocom, Hong Kong, China, March 2004.
24. S. P. Fekete, A. Kröller, D. Pfisterer, S. Fischer, and C. Buschmann. Neighborhood-based
topology recognition in sensor networks. In Algorithmic Aspects of Wireless Sensor Networks (ALGOSENSORS), Vol. 3121 of Lecture Notes in Computer Science, Springer, Berlin,
2004, pp. 123–136.
25. S. Funke. Topological hole detection in wireless sensor networks and its applications. In
Proceedings of the 2005 Joint Workshop on Foundations of Mobile Computing (DIALMPOMC ’05), New York, 2005, pp. 44–53.
26. P. Bose, P. Morin, I. Stojmenović, and J. Urrutia. Routing with guaranteed delivery in
ad hoc wireless networks. In Workshop on Discrete Algorithms and Methods for Mobile
Computing and Communications (DialM), 1999.
27. J. Hightower and G. Borriello. Location systems for ubiquitous computing. IEEE Computer
Magazine, August: 57–66, 2001.
28. L. Hu and D. Evans. Localization for mobile sensor networks. In Proceedings of the 10th
Annual International Conference on Mobile Computing and Networking (MobiCom, 2004,
pp. 45–57.
29. T. Li, A. Ekpenyong, and Y. Huang. A location system using asynchronous distributed
sensors. In Proceedings of IEEE/ACM Infocom, Hong Kong, China, March 2004.
30. L. M. Ni, Y. Liu, Y. Lau, and A. P. Patil. LANDMARK: Indoor location sensing using
active RFID. In Proceedings of IEEE International Conference on Pervasive Computing
(PerCom), Fort Worth, TX, March 2003.
222
POSITION-BASED ROUTING FOR SENSOR NETWORKS: APPROACHES AND OBSTACLES
31. T. He, C. Huang, B. M. Blum, J. A. Stankovic, and T. Abdelzaher. Range-free localization
schemes for large scale sensor networks. In Proceedings of the 9th Annual International
Conference on Mobile Computing and Networking (Mobicom), 2003, pp. 81–95.
32. Y. Kim, J.-J. Lee, and A. Helmy. Modeling and analyzing the impact of location inconsistencies on geographic routing in wireless networks. SIGMOBILE Mobile Computer and
Communication Review (MC2R), 8(1):48–60, 2004.
33. K. Seada, A. Helmy, and R. Govindan. On the effect of localization errors on geographic
face routing in sensor networks. In Proceedings of the 3rd International Symposium on
Information Processing in Sensor Networks (IPSN), 2004, pp. 71–80.
34. B. Karp. Challenges in geographic routing: Sparse networks, obstacles, and traffic provisioning. Presented at DIMACS Workshop on Pervasive Networking, May 2001.
35. C. Lochert, M. Mauve, H. Fler, and H. Hartenstein. Geographic routing in city scenarios.
ACM SIGMOBILE Mobile Computing and Communications Review (MC2R), 9(1):69–72,
2005.
36. A. Woo, T. Tong, and D. Culler. Taming the underlying challenges of reliable multihop
routing in sensor networks. In Proceedings of the 1st International Conference on Embedded
Networked Sensor Systems (SenSys), 2003, pp. 14–27.
37. J. Zhao and R. Govindan. Understanding packet delivery performance in dense wireless sensor networks. In Proceedings of the 1st International Conference on Embedded Networked
Sensor Systems (SenSys), 2003, pp. 1–13.
38. Y.-J. Kim, R. Govindan, B. Karp, and S. Shenker. On the pitfalls of geographic face routing.
In Proceedings of the 2005 Joint Workshop on Foundations of Mobile Computing (DIALMPOMC, 2005, pp. 34–43.
39. B. Leong, B. Liskov, and R. Morris. Geographic routing without planarization. In Proceedings of the 3rd USENIX Symposium on Networked Systems Design and Implementation
(NSDI), San Jose, CA, May 2006.
40. X. Li, Y. J. Kim, R. Govindan, and W. Hong. Multidimensional range queries in sensor
networks. In Proceedings of the 1st International Conference on Embedded Networked
Sensor Systems (SenSys), 2003, pp. 63–75.
41. H. Balakrishnan, M. F. Kaashoek, D. Karger, R. Morris, and I. Stoica. Looking up data in
p2p systems. Communications of the ACM, 46(2):43–48, 2003.
42. Y. Shang, W. Ruml, Y. Zhang, and M. P. J. Fromherz. Localization from mere connectivity.
In Proceedings of the 4th ACM International Symposium on Mobile Ad Hoc Networking &
Computing (MobiHoc), New York, 2003, pp. 201–212.
43. J. Newsome and D. Song. Gem: Graph embedding for routing and data-centric storage in
sensor networks without geographic information. In Proceedings of the 1st International
Conference on Embedded Networked Sensor Systems (SenSys), 2003, pp. 76–88.
44. N. Linial, L. Lovasz, and A. Widgerson. Rubber bands, convex embeddings, and graph
connectivity. Combinatorica, 8(1):91–102, 1988.
45. Q. Cao and T. Abdelzaher. A scalable logical coordinates framework for routing in wireless
sensor networks. In Proceedings of the 25th IEEE Real-Time Systems Symposium (RTSS),
2004, pp. 349–358.
46. L. Demoracski and D. R. Avresky. Performance analysis of fault-tolerant beacon vector
routing for wireless sensor networks. In MSWiM ’05: Proceedings of the 8th ACM International Symposium on Modeling, Analysis and Simulation of Wireless and Mobile Systems,
New York, 2005, pp. 40–44.
BIBLIOGRAPHY
223
47. R. Fonseca, S. Ratnasamy, J. Zhao, C. T. Ee, D. Culler, S. Shenker, and I. Stoica. Beacon
vector routing: Scalable point-to-point routing in wireless sensornets. In Proceedings of the
2nd USENIX Symposium on Networked Systems Design and Implementation (NSDI ’05),
Boston, MA, May 2005.
48. Y. Zhao, B. Li, Q. Zhang, Y. Chen, and W. Zhu. Efficient hop id based routing for sparse
ad hoc networks. In Proceedings of the 13TH IEEE International Conference on Network
Protocols (ICNP), 2005, pp. 179–190.
49. D. S. J. De Couto, D. Aguayo, B. A. Chambers, and R. Morris. Performance of multihop
wireless networks: Shortest path is not enough. SIGCOMM Compututer Communication
Review (CCR), 33(1):83–88, 2003.
50. W. Heinzelman, J. Kulik, and H. Balakrishnan. Adaptive protocols for information dissemination in wireless sensor networks. In Proceedings of the 5th Annual ACM/IEEE International Conference on Mobile Computing and Networking (Mobicom), 1999 pp. 174–185.
51. R. Nagpal, H. Shrobe, and J. Bachrach. Organizing a global coordinate system from local
information on an ad hoc sensor network. In the 2nd International Workshop on Information
Processing in Sensor Networks (IPSN). Springer Verlag Lecture Notes in Computer Science
LNCS 2634, April 2003.
CHAPTER 9
Node Positioning for Increased
Dependability of Wireless Sensor
Networks
MOHAMED YOUNIS
Department of Computer Science and Electrical Engineering, University of Maryland Baltimore
County, Baltimore, MD 21250
KEMAL AKKAYA
Department of Computer Science, Southern Illinois University, Carbondale, IL 62901
9.1 INTRODUCTION
Advances in microelectronics have enabled the development of very tiny sensor
nodes that have the ability of measuring ambient conditions such as temperature,
pressure, humidity, light intensity, vibration, and so on. The sensed data can then
be transmitted through an onboard radio transmitter to a single or multiple base
stations where it can be further processed. The cost and size advantage of such
emerging sensor nodes has encouraged practitioners to explore using them collaboratively in a network formed in ad hoc manner. Such networked sensor systems are
not only cost effective but can also provide fast and accurate information gathering
in remote and risky areas. Figure 9.1 depicts a typical sensor network architecture.
The base station acts as a gateway for linking the sensors to multiple command
nodes.
The past few years have witnessed increased interest in the potential use of wireless
sensor networks (WSNs) in applications such as disaster management, combat field
reconnaissance, border protection, and security surveillance [1, 2]. Sensors in these
applications are expected to be remotely deployed and to operate autonomously in
unattended environments. While the initial view of the community was that WSNs
will play a complementary role that enhances the quality of these applications, recent research results have encouraged practitioners to envision an increased reliance
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
225
226
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
Figure 9.1. A sensor network for a disaster management application.
on WSNs. In order to best realize their potential, dependable design and operation of
WSNs have to be ensured.
Dependability is a property that indicates the ability of a system to deliver services
to the user subject to a required level of quality. Dependability can be specified in
terms of attributes, such as responsiveness, availability, safety, security, and so on.
Dependability in WSNs is complicated by many factors including the following: (1)
Sensors are significantly constrained in the amount of available resources such as
energy, storage, and computation; (2) sensors are expected to be deployed in very
large numbers in normal as well as forbidding environments; and (3) WSNs suffer
from structural weakness and limited physical protection. Moreover, dependability
requirements may vary according to a network’s mission, such as field of deployment
(e.g., hostile versus friendly), type of application (e.g., monitoring, tracking, data
collection), mode of operation (e.g., normal, exception, post-event recovery), and
time.
The bulk of the research on WSNs has focused on the effective support of the
functional (e.g., data latency) and the nonfunctional (e.g., data integrity) requirements
while coping with the resource constraints and on the conservation of available energy
in order to prolong the life of the network. Contemporary design schemes for WSNs
pursue optimization at the various layers of the communication protocol stack. Popular
optimization techniques at the network layer include multihop route setup, in-network
data aggregation, and hierarchical network topology [3]. For medium access control,
collision avoidance, minimizing idle listening of radio receivers, and output power
control are a sample of proposed schemes [1, 4]. At the application layer, examples
include adaptive activation of nodes, lightweight data authentication and encryption,
load balancing, and query optimization [5, 6].
INTRODUCTION
227
One of the design optimization strategies is to deterministically place the sensor
nodes in order to meet the desired performance goals. In such cases, the coverage
can be ensured through careful planning of node densities and field of view, and
thus the network topology can be established at setup time. However, in most WSN
applications, sensors deployment is random and little control can be exerted in order to
ensure coverage and yield uniform node density while achieving strongly connected
network topology. Also, the location of the base station can have great influence on
the network performance. For example, routing data to a base station that is distant
from the source sensor usually involves numerous relaying nodes and thus increases
the aggregate delay and energy consumption and risks a packet loss due to link
errors. Therefore careful selection of the base-station location may affect various
performance metrics such as energy consumption, delay, and throughput. Unlike
sensors deployment, positioning of the base stations can be somewhat controlled and
is feasible in many application setups.
Optimal node placement is a very challenging problem that has been proven to be
NP-complete for base stations [7] and for most of the formulations of sensors’ deployment [8–10]. To tackle such complexity, several heuristics have been proposed to find
suboptimal solutions [7, 11–14]. However, the context of these optimization strategies
is mainly static in the sense that assessing the quality of candidate positions is based
on structural quality metric such as distance, network connectivity, and/or basing the
analysis on a fixed topology. Therefore, we classify them as static approaches. We
note, however, that dynamically adjusting nodes’ location can further increase the
dependability of WSNs since the optimality of the initial positions may become void
during the operation of the network depending on the network state and various external factors. For example, traffic patterns can change based on the monitored events,
load may not be balanced among the nodes causing bottlenecks, application-level
interest can vary over time, and the available network resources may change due to
the depletion of energy of some nodes and/or the addition of more nodes.
In this chapter we opt to categorize the various strategies for statically positioning
base stations and sensor nodes in WSNs. We contrast a number of published approaches highlighting their strengths and limitations. Our aim is to help applications
designers identify alternative solutions and select appropriate strategies. In addition
to surveying static positioning approaches, we show that dynamically repositioning
nodes while the network is operational can be a very effective means for boosting
the network’s dependability attributes. We describe scenarios for which node relocation can be pursued to counter holes in coverage, achieve/maintain strong network
connectivity, preserve sensor’s energy, increase data timeliness, and boost the node’s
physical security. We highlight the issues, report on the state of the art, and outline
open esearch problem.
The chapter is organized as follows. The next section is dedicated to static strategies
for node positioning. We separately cover approaches for sensor placement as well as
single and multiple base stations positioning. In Section 9.3 we turn our attention to
dynamic positioning schemes. We highlight the technical issues and describe sample
techniques that exploit sensors and base-station repositioning to enhance the network
dependability. The bulk of the discussion applies to a single base-station setup or to
228
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
clustered network architectures where multiple base stations operate independently.
Section 9.4 highlights the challenges of coordinated repositioning of multiple nodes
and describe few attempts to tackle these challenges. Finally, Section 9.5 concludes
the chapter.
9.2 STATIC POSITIONING OF NODES
Node placement schemes prior to network startup usually base their choice of the
particular nodes positions on metrics that are independent of the network state or
assume a fixed network operation pattern that stays unchanged throughout the lifetime
of the network. Examples of such static metrics are area coverage, internode distance,
and so on. Static network operation models often assume periodic data collection over
preset routes. In this section we give an overview of the different categories of static
node positioning and discuss sample work from the literature. We cover both sensor
and base-station placement in distinct subsections.
9.2.1 Sensor Node Placement
As mentioned before, the position of nodes have dramatic impact on the effectiveness
of the WSN and the efficiency of its operation. In this section we discuss contemporary sensor placement strategies in the literature. We classify them according to
the deployment schemes, the optimization objective of the placement, and how they
relate to the network topology.
Deployment Schemes. Sensors can generally be placed in an area of interest either deterministically or randomly. The choice of the deployment scheme depends
highly on the type of sensors, application, and the environment in which the sensors
will operate. Controlled node deployment is viable and often necessary when sensors
are expensive or when their operation is significantly affected by their position. Examples of such scenarios include when populating an area with highly precise seismic
nodes and when placing imaging and video sensors. On the other hand, in some applications, random deployment schemes are the only feasible option. This is particularly
true for harsh environments such as a battlefield or a disaster region. Depending on the
node distribution and the level of redundancy, random node deployment can achieve
the required performance goals.
Deterministic deployment is usually pursued for indoor applications of WSNs.
Examples of indoor networks include the Active Sensor Network (ASN) project at
the University of Sydney in Australia [15], the Multiple Sensor Indoor Surveillance
(MSIS) project at Accenture Technology Labs in Chicago [16], and the Sensor Network Technology projects at Intel [17]. The ASN and MSIS projects gear for serving
surveillance applications such as secure installations and enterprise asset management. At Intel, the main focus is on applications in manufacturing plants and engineering facilities—for example, preventative equipment maintenance (Figure 9.2).
STATIC POSITIONING OF NODES
229
Figure 9.2. Sensors are mounted to analyze the vibration and assess the health of equipment
at a semiconductor fabrication plant. (Photographs are from 17.)
Hand-placed sensors are also used to monitor the health of large buildings in order
to detect corrosions and overstressed beams that can endanger the structure integrity
[18, 19]. Another notable effort is the Sandia Water Initiative at Sandia National Lab
which addresses the problem of placing sensors in order to detect and identify the
source of contamination in the air or water supplies [20, 21].
Deterministic placement is also very popular in applications of range finders, imaging, and video sensors. In general, these sensors are involved in three-dimensional
(3-D) application scenarios, which is much more difficult to analyze compared to
two-dimensional deployment regions. Poduri et al. [10] investigated the applicability
of contemporary coverage analysis and placement strategies pursued for 2-D space to
3-D setups. They concluded that many of the popular formulation such as art-gallery
and sphere-packing problems, which are optimally solvable in 2-D, become NP-hard
in 3-D. Most placement approaches for these types of sensor strive to enhance the
quality of visual images and/or accuracy of the assessment of the detected objects.
For example, Gonzalez-Banos and Latombe [22] studied the problem of finding
the minimum number of range finders, which estimate the proximity of a target, or
video sensors and their location in order to cover an area. Unlike other sensors, such
230
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
as acoustic or temperature, the authors had to account for the restricted capabilities of
the range finders, which provide lower and upper bounds only, and for the reliability
of detecting objects at grazing angles. The problem is formulated as an Art-Gallery
model, for which the fewest guards are to be placed to monitor a gallery. Similarly,
the work of Navarro et al. [23] addresses the orientation of video cameras so that high
quality images of target objects are captured.
Randomized sensor placement often becomes the only option. For example, in
application of WSNs in reconnaissance missions during combat, disaster recovery,
and forest fire detection, deterministic deployment of sensors is very risky and/or
infeasible. It is widely expected that sensors will be dropped by helicopter, grenade
launchers, or clustered bombs. Such means of deployment lead to random spreading
of sensors, although the node density can be controlled to some extend. Although it
is somewhat unrealistic, many research projects (e.g., reference 24), have assumed
uniform node distribution when evaluating the network performance. The rationale
is that with the continual decrease in cost and size of microsensors, a large population of nodes is expected and thus a uniform distribution becomes a reasonable
approximation.
Ishizuka and Aida [25] have investigated random node distribution functions by
trying to capture the fault-tolerance properties of stochastic placement. They compared three deployment patterns (Figure 9.3, taken from reference 25); namely simple
diffusion (two-dimensional normal distribution), uniform, and R-random where the
nodes are uniformly scattered with respect to the radial and angular directions from
the base-station. The R-random node distribution pattern resembles the effect of an
exploded shell. The experiments tracked coverage and node reachability as well as
data loss in a target tracking application. The simulation results indicated that the initial placement of sensors has a significant effect on network dependability measured
in terms of tolerance of node failure that may be caused by damage and battery exhaustion. The results also showed that the R-random deployment is a better placement
strategy in terms of fault-tolerance.
While a flat architecture is assumed in reference 25, Xu et al. [26] have considered
a two-tier network architecture in which sensors are grouped around relaying nodes
that directly communicate with the base station. The goal of the investigation is to
identify appropriate node deployment strategies in order to maximize the network
lifetime. They first showed that uniform node distribution often does not extend the
Figure 9.3. (a) Simple diffusion. (b) Constant (uniform) placement. (c) R-random placement.
STATIC POSITIONING OF NODES
231
Figure 9.4. An illustration of weighted random deployment (from reference 26). The density
of relay nodes inside the circle (close to the base station) is lower and the connectivity is weaker
than outside.
network lifetime since relay nodes will consume energy at different rates depending on their proximity to the base station. Basically, the further away the relays are
from the base station, the higher the energy they deplete in transmission. To counter
such shortcomings, a weighted random node deployment strategy is then proposed
to account for the variation in energy consumption rate in the different regions. Such
strategy, as illustrated in Figure 9.4, increases the density of relays away from the base
station to split the load among more relays and thus extend their average lifetime.
Although it has a positive impact on network lifetime, the weighted random distribution may leave some relay nodes disjoint from the base station since some relays
may be placed so far that the base station becomes out of their transmission range.
Finally, a hybrid deployment strategy is introduced to balance the network lifetime
and connectivity goals. The analysis is further extended in reference 27 for the case
where relay nodes reach the base station through a multihop communication path.
The conclusion regarding the three strategies was found to hold in the multihop case
as well.
Primary Objectives for Deployment. Application developers surely like the
sensors to be deployed in a way that leverages the overall design goals. Therefore,
most of the proposed deployment schemes in the literature focused on increasing the
coverage, achieving strong network connectivity, prolonging the network lifetime,
and/or boosting the data fidelity. A number of secondary objectives such as tolerance
of node failure and load balancing were also considered. Most of the work strives
to maximize the deployment objectives using the least amount of resources (e.g.,
number of nodes). Obviously, meeting the design objectives using random deployment schemes is an utmost challenge. Meanwhile, although intuitively deterministic
232
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
placement can theoretically meet all primary and secondary objectives, the quest for
minimizing the required network resources keeps the problem very hard.
Area Coverage is the deployment objective that has received the most attention in
the literature. Assessing the coverage varies based on the underlying model of sensor’s
field of view and the metric used to measure the collective coverage of deployed
sensors. The bulk of the published work (e.g., reference 28) assumed a disk coverage
zone centered at the sensor with a radius that equals its sensing range. However, some
recent work has started to employ more practical sensor’s field of view in the form of
irregular polygons [29]. Some of the published papers, especially early ones, use the
ratio of the covered area to the overall deployment region as a metric for the quality
of coverage [28]. Recent work, however, has focused on the worst-case coverage,
usually referred to least exposure, measuring the probability that a target would travel
across an area or an event would happen without being detected [30]. The advantage
of exposure-based coverage assessment is the inclusion of a practical object detection
probability that is based on signal processing formulations (e.g., signal distortion) as
applicable to specific sensor types.
As mentioned earlier, optimized sensor placement is not an easy problem, even
for deterministic deployment scenarios. Complexity is often introduced by the quest
for employing the least number of sensors for meeting the application requirements
and by the uncertainty in a sensor’s ability to detect an object due to distortion that
may be caused by terrain or the sensor’s presence in a harsh environment. Dhillon
and Chakrabarty [13] considered the placement of sensors on a grid approximation
of the deployment region. They formulated a sensing model that factors in the effect
of terrain in the sensor’s surroundings and inaccuracy in the sensed data (Figure 9.5).
The model is then used to identify the grid points on which sensors are to be placed,
so that an application-specific minimum confidence level on object detection is met.
They proposed a greedy heuristic that strives to achieve the coverage goal through
the least number of sensors. The algorithm is iterative. In each iteration, one sensor
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
Figure 9.5. Knowing the coordinate of obstacles, the sensor’s field of view is adjusted. In the
shown example, redrawn from reference 13, the sensor on grid point 14 is not sufficient to
cover the area in direction to point 11. The same applies for (2,7), (6,3), and (10,15).
STATIC POSITIONING OF NODES
233
Figure 9.6. The dark circle marks the location of sensors. The weights on the individual line
segments are based on the probability that a target is detected by all sensors (combined exposure). The least exposure path, marked in dark line, is found by applying Dijkstra’s algorithm.
is placed at the grid point with the least coverage. The algorithm terminates when the
coverage goal is met or a bound on the sensor count is reached.
Clouqueur et al. [14] also studied the problem of populating an area of interest
with the least number of sensors so that targets can be detected with the highest
probability. Unlike reference 13, random deployment is assumed in this work. The
authors propose a metric called path exposure to assess the quality of sensor coverage.
The idea is to model the sensing range of deployed nodes and establish a collective
coverage map of all sensors based on a preset probability of false alarm (detection
error). The map is then checked in order to identify the least exposure path on which a
target may slip by, with the highest probability of being undetected. Figure 9.6, taken
from reference 14, illustrates the idea on a grid structure. Employing such a metric,
the authors further introduced a heuristic for incremental node deployment so that
every target can be detected with a desired confidence level using the fewest sensors
count. The idea is to randomly deploy a subset of the available sensors. Assuming
that the sensors can determine and report their positions, the least exposure path is
identified and the probability of detection is calculated. If the probability is below a
threshold, additional nodes are deployed in order to fill holes in the coverage along
the least exposure path. This procedure would be repeated until the required coverage
is reached. The paper also tried to answer the question of how many additional nodes
are deployed per iteration. On the one hand, it is desirable to use the least number
of sensors. On the other hand, the means for sensor deployment may be expensive
or risky (e.g., sending a helicopter). The authors derive a formulation that accounts
for the cost of deploying nodes and the expected coverage as a function of sensors
count. The formulation can be used to guide the designer for the most effective way
to populate the area.
234
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
The sensor placement problem considered by Biagioni and G. Sasaki [31] is more
difficult. They opt to find a placement of nodes that achieves the coverage goals using
the least number of sensors and also maintains a strongly connected network topology
even if one node fails. The authors review a variety of regular deployment topologies
(e.g., hexagonal, ring, star, etc.) and study their coverage and connectivity properties
under normal and partial failure conditions. They argue that regular node placement
simplifies the analysis due to their symmetry despite the fact that they often do not
lead to optimal configuration. The provided analytical formulation can be helpful in
crafting a placement as a mix of these regular topologies and in estimating aggregate
coverage of nodes.
Network connectivity is also a major concern when deploying sensor nodes. Unlike
coverage, which has constantly been an objective or constraint for node placement,
connectivity was deemed a non-issue in some of the early work based on the assumption that the transmission range Tr of a node is much longer than its sensing range
Sr . The premise is that good sufficient coverage will yield a connected network when
Tr is multiple of Sr . However, if the communication range is limited (e.g., Tr = Sr ),
connectivity becomes an issue unless there is a substantial redundancy in coverage. It
is worth noting that some work tackled the connectivity concern through deploying
relay nodes that have long haul communication capabilities. Such approaches will be
discussed in detail later in this section.
Kar and Banerjee [32] considered sensor’s placement for complete coverage and
connectivity. Assuming that the sensing and radio ranges are equal, the authors first
define an r-strip as shown in Figure 9.7a. In an r-stripe, nodes are placed so that
neighbors of a sensor along the x-axis are located on the circumstance of the circle that defines the boundary of its sensing and communication range. Obviously,
nodes on an r-strip are connected.
The authors then tile the entire plane with r√
strips on lines y = k(0.5 3 + 1)r such that the r-strips are aligned for even values of the integer k and shifted horizontally r/2 for odd values of k, as illustrated
in Figure 9.7b. The goal is to fill gaps in coverage with the least overlap among
the r-disks that define the boundary of the sensing range. To establish connectivity among nodes in different r-strips, additional sensors are placed along the yaxis, which marked by the shaded disks in Figure
√ 10.7b. For every
√ odd value of
the integer k, two sensors are placed at [0, k(0.5 3 + 1)r ± 0.5 3r] to establish
connectivity between every pair of r-strips. For a general convex-shaped finite-size
region, connectivity among nodes in horizontal r-strips is established by another
r-strip placed diagonally within the boundary of the region (Figure 9.7c). The authors
generalize their scheme for the case where points of interest are to be covered rather
than the whole area. However, unless the base station is mobile and can interface with
the WSN through any node, having a strongly connected network is not essential in
WSNs since data are gathered at the base station. Therefore, ensuring the presence of
a data route from a node to the base station would be sufficient, and thus fewer nodes
can be employed to achieve network connectivity than the presented approach would
use. In addition, vertically placed nodes or diagonal r-strips can become a communication bottleneck since they act as gateways among horizontal r-strips, which may
require the deployment of more sensors to split the traffic.
STATIC POSITIONING OF NODES
(a)
235
region
boundary
r-strip #0
Even r -strips
are vertically
a ligned
r-strip #1
r-strip #2
r-strip #3
Odd r-strips
are shifted
r-strip #4
r-strip #5
r-strip #6
(b)
(c)
Figure 9.7. Illustration of the placement algorithm in a plane and a finite size region. Figure
is redrawn from reference 32.
The focus of reference 33 is on forming K-connected WSNs. K-connectivity implies that there are K independent paths among every pair of nodes. For K > 1, the
network can tolerate some node and link failures and guarantee certain communication capacity among nodes. The authors study the problem of placing nodes to
achieve K-connectivity at network setup time or to repair a disconnected network.
They formulate the problem as an optimization model that strives to minimize the
number of nodes required to maintain K-connectivity. They show that the problem is
NP-hard and propose two approximation algorithms with varying degree of complexity and closeness to optimality. The algorithms are graph-theory-based. The idea is to
compute a weighted complete graph on the same set of vertices (nodes) and then find
an approximate minimum-weight K-vertex-connected subgraph g. Finally, missing
links (edges) in g are established by deploying the least number of nodes. Again in
most WSNs, it is not necessary to achieve K-connectivity among sensors unless the
base station changes its location frequently.
Network lifetime has been the optimization objective for most of the published
communication protocols for WSNs. The positions of nodes significantly impact the
network lifetime. For example, variations in node density throughout the area can
eventually lead to unbalanced traffic load that causes the rapid drain of the energy
reserve of some sensors [26]. In addition, uniform node distribution may lead to
the depletion of energy of nodes that are close to the base station at a higher rate
than other nodes and thus shorten the network lifetime [34]. Some of the published
work—such as reference 27 which we discussed earlier—has focused on prolonging
the network lifetime rather than area coverage. The implicit assumption is that there
is a sufficient number of nodes or that the sensing range is large enough such that no
holes in coverage may result.
Chen et al. [35] studied the effect of node density on network lifetime. Considering
the one-dimensional placement scenario, the authors derived an analytical formulation
236
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
for the network lifetime per unit cost (deployed sensor). They also argued that the
network lifetime is not growing proportionally to the increased node population and
thus a careful selection of the number of sensors is necessary to balance the cost and
lifetime goals. Considering the network to be functional until the first node dies, an
optimization problem was defined with the objective of identifying the least number
of sensors and their positions so that the network stays operational for the longest
time. An approximate two-step solution is proposed. In the first step, the number
of sensors is fixed and their placement is optimized for maximum network lifetime.
They formulate such optimization as a multivariant nonlinear problem and solve it
numerically. In the second step, the number of sensors is minimized in order to achieve
the highest network lifetime per unit cost. A closed-form solution is analytically
derived for the second step.
Hou et al. [36] considered a two-tier sensor network architecture where sensors
are split into groups; each is led by an aggregation-and-forwarding (AFN) node
(Figure 9.8). A sensor sends its report directly to the assigned AFN, which aggregates
the data from all sensors in its group. The AFNs and the base station form a second
tier network in which an AFN sends the aggregated data report to the base station
over a multihop path. The authors argue that AFNs can be very critical to the network
operation and that their lifetime should be maximized. Two approaches were suggested to prolong the AFNs lifetime. The first is to provision more energy to AFNs.
The second is to deploy relay nodes (RNs) in order to reduce the communication
energy consumed by an AFN in sending the data to the base station. The RN placement and energy provisioning problem was formulated as a mixed-integer nonlinear
programming optimization. For a pool of an E energy budget and M relay nodes, the
objective of the optimization is to find the best allocation of the additional energy to
Base
Station
Upper tier
AFN
Lower tier
MSN
Figure 9.8. Logical view of the assumed two-tier network architecture (redrawn from reference
36). MSN stands for micro-sensor nodes.
STATIC POSITIONING OF NODES
237
existing AFNs and the best positions for placing the M relays. To efficiently solve the
optimization problem, the formulation was further simplified through a two-phase
procedure. In the first phase a heuristic is proposed for optimized placement of the M
relay nodes. Given the known positions of the RNs, in the second phase the energy
budget is allocated to the combined AFNs and RNs population, which is a linear
programming optimization.
Data fidelity is obviously an important design goal of WSNs. A sensor network
basically provides a collective assessment of detected objects by fusing the readings of multiple independent, and sometimes heterogeneous, sensors. Data fusion
boosts the creditability of the reported incidents by lowering the probability of false
alarms and of missing a detectable object. From a signal processing point of view,
data fusion tries to minimize the effect of the distortion by considering the readings
from multiple sensors so that a high-fidelity assessment can be made regarding the
detected phenomena. Although increasing the number of sensors reporting in a particular region will sure boost the accuracy of the fused data, redundancy in coverage
would require an increased node density, which can be undesirable due to cost and
other constraints such as the potential of detecting the sensors in a combat field.
Zhang and Wicker [37] looked at the sensor placement problem from a data fusion
point of view. The observation that they made is that there is always an estimation
distortion associated with a sensor reading which is usually countered by getting many
samples. They thus mapped the problem of finding the appropriate sampling points
in an area to that of determining the optimal sampling rate for achieving a minimal
distortion, which is extensively studied in the signal processing literature. In other
words, the problem is transformed from the space to the time domain. The optimal
sampling points are actually the best spots where sensors can be placed. The approach
is to partition the deployment area into small cells, and then optimal sampling rate
per cell is determined for minimal distortion. Assuming that all sensors have the same
sampling rate, the number of sensors per cell can be determined.
Similar to Zhang and Wicker [37], Ganesan et al. [38] studied sensors placement
to meet some application quality goals. The problem considered is to find nodes
positions so that the fused data at the base station meets desired level of fidelity.
Unlike Zhang and Wicker [37], a tolerable distortion bound is imposed as a constraint,
and minimizing energy consumption in communication is set as an objective of the
optimization formulation. Also, in this work the number of sensors is fixed and their
position is to be determined. Given the consideration of energy consumption, data
paths are modeled in the formulation, making the problem significantly harder. The
authors first provided a closed-form solution for the one-dimensional node placement
case and used it to propose an approximation algorithm for node placement in a
circular region. Extending the approach to handle other regular and irregular structures
is noted as future work.
Wang et al. [39] also exploited similar ideas for a WSN that monitors a number
of points of interest. Practical sensing models indicate that the ability of detecting
target/events diminishes with increased distance. One way to increase the creditability
of the fused data is to place the sensors so that a point of interest would be in the
238
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
high-fidelity sensing range of multiple nodes. Given a fixed number of sensors, there is
a tradeoff between deploying a sensor in the vicinity of one point of interest to enhance
the probability of event detection and the need to cover other points of interest. The
probability of event detection by a sensor is called the utility. The utility per point of
interest is thus the collective utility of all sensors that cover that point. The authors
formulate a nonlinear optimization model to identify the locations of the sensors so
that the average utility per point of interest is maximized. To limit the search space,
the area is represented as a grid with only intersection points considered as candidate
positions.
Finally we would like to note that the work of Clouqueur et al. [14], which we
discussed earlier, can also be classified under data-fidelity-based sensor placement.
They estimate the creditability of fused data from multiple sensors and use it to identify
the position of sensors for maximizing the probability of target detection. Table 9.1
categorizes the sensor nodes placement mechanisms discussed in this section.
9.2.2 Base-Station Placement
A considerable research has been done on optimal initial (i.e., at network setup time)
positioning of single or multiple base stations in WSNs. Published work generally
differs based on the assumptions made, the considered network model, the available network state information, and the metrics to be optimized. Popular objectives
of base-station positioning include maximizing the overall network lifetime, minimizing the longest data path, and achieving maximal data rate. Given the location,
onboard energy supply, and number of sensors, such objectives are optimized using
techniques like integer linear programming, network flow, and computational geometry. In this section, we categorize prior work on static single and multiple base-station
positioning.
Optimized Positioning of a Single Base Station. The limited energy supply
onboard a sensor node has made network longevity a key performance metric. Some
published work exploited the flexibility in base station positioning in order to extend
the network lifetime. Nonetheless, multiple variants of the base-station positioning
were pursued. The difference is due to either the definition of a network lifespan,
the network operation model, and the network state parameters that are included
in the formulation. While some considered the network to be functional until the
first sensor node dies [7], many used the failure of a percentage of the deployed
sensors [12, 40, 41] as indicative of the network lifespan. Other work strived to extend
the network lifetime through minimizing the total power consumed in collecting
the readings of all sensors [11]. All these static base-station positioning approaches
have assumed a periodic data collection model for the network. That is, each sensor
node transmits a certain amount of data at a fixed rate. Some schemes also factored in
the necessary transmission scheme for the sensor nodes [7]. However, no in-network
data aggregation has been considered; that is, each node should transmit the packets it
received without any concatenation, suppression, or compression. While most of these
TABLE 9.1. A Comparison Between the Various Approaches for Sensors Placement
Space
Deployment
3-D
3-D
3-D
3-D
Deterministic
Deterministic
Deterministic
Deterministic
3-D
Deterministic Data fidelity
3-D
Deterministic Coverage
3-D
Deterministic Coverage
22
23
24
25
26
27
13
14
31
Surveillance
Surveillance
Manufacturing
Structural health
monitoring
Structural health
monitoring
Contamination
detection
Contamination
detection
Generic
Generic
Outdoor
Outdoor
Outdoor
Outdoor
Surveillance
Outdoor
Outdoor
3-D
3-D
2-D
2-D
2-D
2-D
2-D
2-D
2-D
Deterministic
Deterministic
Deterministic
Random
Random
Random
Deterministic
Random
Deterministic
Data fidelity
Data fidelity
Min. relay count
Coverage and connectivity
Network lifetime
Network lifetime
Coverage
Data fidelity and coverage
Coverage and connectivity
35
36
32
33
37
38
Generic
Generic
Outdoor
Outdoor
Outdoor
Generic
Deterministic
Deterministic
Deterministic
Deterministic
Random
Deterministic
Network lifetime
Network lifetime
Coverage and Connectivity
Connectivity
Data fidelity
Data fidelity
39
Surveillance
1-D
2-D
2-D
2-D
2-D
1-D
2-D
2-D
Reference
15
16
17
18
19
20
21
Application
Primary Objective
Coverage
Coverage
Data fidelity
Data fidelity
Deterministic Data fidelity
Secondary Objective
Constraint
Connectivity
Data fidelity
Connectivity
Connectivity
—
—
Connectivity and Fault-tolerance
—
—
—
Fixed sensors count
Delay
Fixed sensors count
Min. sensor count
—
Fault-tolerance
Fault-tolerance
—
—
Minimum sensor count
Minimum sensor count
Minimum sensor count and
fault-tolerance
—
—
Minimum sensor count
Fault-tolerance
—
Minimal energy consumption in
communication
—
—
—
Connectivity
—
—
—
—
—
—
Coverage
Fixed relays count
—
—
—
Lower bound on tolerable
distortion; fixed sensors count
—
239
240
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
approaches (e.g., see reference 41) rely on the availability of exact nodes locations
obtained through GPS or simply by deterministic sensor deployment, a few such as
those in references 12 and 42 use internode proximity estimates provided through
ranging technology.
The considered network topology and system model are also differentiating factors among published static base-station positioning approaches. In a flat network,
topology sensors are homogeneous having the same amount of initial energy, and
usually form multihop routes to relay their data to the base station. The base-station
positioning problem gets more challenging in case of multihop routing since each
sensor node not only transmits its own data but also forwards data from its neighbors
when requested [7]. The problem becomes simpler when sensors directly transmit to
the base station [42]. However, given the limited transmission range for the sensor
nodes, this assumption cannot be applied to WSNs that cover large areas. To support
scalability in such setups, hierarchical network topologies are often pursued. The
most common hierarchical topology is based on grouping sensors into clusters with
a designated cluster head. In such cases, the scope of the base-station positioning
problem is reduced to the inter-cluster-head network and thus becomes simpler [12].
It is worth noting that if base stations act as cluster heads, the optimization becomes
a multi-base-station positioning problem.
Depending on the assumptions and network models, most of the optimal basestation positioning problem formulations are NP-complete. A common way to counter
such complexity is to employ approximation. For instance, in reference 7 the search
space is restricted to the sensor locations and the best position “s” among them
in terms of network’s lifetime is picked. This solution is shown to be a constant
approximation of the optimal solution; for example, it achieves a fraction of the
optimal network lifespan. The approximation ratio was further improved to (1 - ε),
where ε > 0 is any desired error bound, by factoring in the routes and transmission
schedule. However, the improvement came at the cost of increased computation for
solving multiple linear programs. To limit such high computation, a technique that
explores the potential overlap among the elements of the search space is proposed [43].
The idea is to replace an infinite search space for each variable by a finite-element
search space with guaranteed bound on possible loss in performance. Specifically,
the search space increases exponentially with the increasing number of variables,
and such an increase can be reduced by exploring the potential overlap among the
elements of the search space. In order to determine the potential overlap, the variables
are expressed in the form of a geometric progression and a common factor among
theses geometric progression is identified.
The approaches in references 12 and 42 do not consider data relaying, and thus
the problem becomes solvable in a polynomial time. In order to minimize the total
communication power, the base station is to be located such that the maximum distance
to a sensor node is minimized [42]. A computational geometry-based algorithm,
whose complexity is linear in the number of nodes n, is proposed. This algorithm tries
to determine the circle with the least diameter that encloses the nodes. Such circle
can be formed with at most three points picked among the locations of the sensors.
The base station will then be positioned at the center of the circle [42]. The same
241
STATIC POSITIONING OF NODES
2
2
1
1
1
0
3
0
0
Figure 9.9. Finding the minimum enclosing circle for six application nodes [12]. The base
station marked as a triangle is placed at the center of the smallest disk that contains all application
nodes (the small circles).
algorithm is further extended for a two-tiered WSN where special “application” nodes
are designated as cluster heads [12]. The application node interfaces the cluster with
the base station. The minimum enclosing circle is thus found for the application nodes
as shown in Figure 9.9, redrawn from reference 12. Although the time complexity for
these approaches are quite promising, they imply a centralized network management
strategy, including routing and/or MAC, which may not be desirable particularly in
large-scale settings.
Positioning Multiple Base-Stations. Multi-base-station positioning is even
more challenging given the higher scale and the fact that individual sensors can select
among multiple destinations to which they send the data. The positioning problem
is typically defined as the optimal layout for a known number of base stations in
order to maximize some performance metric such as total communication energy and
throughput [11] or area coverage [44]. In some cases, the number of base stations may
not be known in advance, and thus the optimal number and location of base stations
are to be found [40]. In the context of wireless sensor and actor networks, the base
stations (actors) need to be positioned for maximized area coverage and reduced data
delivery latency [44].
In general, the complexity of the multi-base-station positioning problem varies
based on the planned network architecture. When a flat network topology is pursued,
the problem stays NP-complete [7]. The same applies to clustered networks in which
some sensors are designated as cluster heads forming a two-tier topology. However,
when the nodes are grouped into clusters that are led by the individual base stations,
the complexity depends on the order of the network clustering and base-station positioning procedures. If the sensors are assigned to base stations prior to placing or
finalizing the positions of base stations, the scope of the problem becomes local to
the individual clusters and concerns only each base station independently from the
others [40]. In other words, the problem becomes similar to a single base-station positioning. However, if the base-station positioning precedes the network clustering,
the complexity remains NP-complete [11].
242
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
To counter the high complexity, most published approaches try to limit the search
space (i.e., the cardinality of the set of candidate locations) in the hope of converging to
some locations that achieve near-optimal performance. For example, in reference 41,
candidate locations are determined a priori by the application designers. Meanwhile,
in reference 11 the feasible set is restricted to where sensors are. The latter case also
applies when the base-station positioning is combined with the network clustering
scheme. For example, it may be desired for the base stations to be placed so that each
sensor would reach a base station in at most K hops. By deciding to place base stations
next to some sensors, the positioning/clustering problem in that case becomes finding
the K-dominating set that has some polynomial time solution [45]. It is worth noting
that if the number of base stations is fixed, the problem stays NP-complete.
Possible approaches for the multi-base-station positioning include approximation
algorithms [11] and integer programming [41]. To maximize the achievable rate of
collecting sensor’s data, the base-station positioning problem is solved by using two
approximation strategies, namely, greedy and local search [11]. In the greedy algorithm the base-station position is restricted to the location of sensors and then the
base stations are individually placed in an arbitrary order for increased data rate. The
local search starts with a random configuration of base stations and is then followed
by a search for a better layout that boosts the data rate. An example for local search
for four base stations on a 10 × 10 grid with a sensor transmission radius of 2.2
units is shown in Figure 9.10, which is redrawn from [11]. On the other hand, the
Integer Linear Programming (ILP) formulation proposed in reference 41 is geared for
splitting the data routing load among the sensors as evenly as possible. The objective
function is to minimize the maximum energy consumption at the individual sensors
while minimizing the total communication energy. The constraints of the ILP formulation include a bound on the total energy consumed by a node in a data collection
round and a restriction on the candidate base-station location to be picked from a set
of predetermined positions. Other constraints are also specified to ensure a balanced
Figure 9.10. Optimal solution with local search.
STATIC POSITIONING OF NODES
3 actors
243
4 actors
Figure 9.11. Initially, base stations (actors) are uniformly placed in the area of interest. Circles
define the acting range of an actor.
flow through the individual nodes and to allow transmission of messages to a feasible
site only if a base station exists at that site. The base-station positions are recomputed
periodically to cope with changes in the network state.
In addition to energy metric considered in references 11 and 41, area coverage
can also be optimized through careful positioning of multiple base stations (actors)
[44]. In some applications, scenarios like disaster management and combat field reconnaissance base stations not only collect and process the data but also can do some
reactive actions such as extinguishing a fire and de-mining a travel path. In this case,
minimizing the delivery latency of sensors’ data and the time for a base station to
reach the spot that needs an action would be design objectives. Initially the actors
are positioned uniformly in order to maximize the coverage (i.e., minimize the overlap among the action ranges) of the area as shown in Figure 9.11. Sensors are then
grouped into clusters; each is led by an actor. After clustering, each actor considers
the positions of its assigned sensors as vertices and computes the vertex 1-center [7].
Relocating the actor at the vertex 1-center location ensures minimum delay from the
farthest sensor node; that is, it minimizes the maximum latency for data delivery.
However, when relocating an actor to its 1-center location, it may lose connection
with the other actors in the network. In order to also ensure network connectivity,
the approach in reference 44 is further extended in reference 46. Connectivity is
maintained by moving an actor close to the vertex 1-center of its cluster as much as
possible without breaking the links with its neighbors. The relocations are strived to
be performed in a global order based on the IDs of actors so as not to disconnect the
network with simultaneous relocations of neighboring actors.
Table 9.2 compares the characteristics of the approaches discussed in this section. It is worth noting that the approach of reference 41 reevaluates the optimality of the picked base-station positions unlike the previously discussed static approaches. However, it still falls short from being a dynamic approach since it simply
pauses the network operation between successive base-stations placements. In the
next section, we argue that if this repositioning can be done dynamically without
disrupting the network operation, it not only can improve certain performance metrics but also adjust the network topology based on the changes in the event area.
244
TABLE 9.2. Comparison of Approaches for Static Base-Station Positioning
Reference 7
Reference 12
Reference 11
Data rate of sensors
Network lifetime
definition(s)
Fixed
First sensor
node to die
Fixed
K-of-N sensor
nodes to die
Fixed
Total power
Number of BS
1
1
Fixed and known
Initial candidate BS
Locations
NA
NA
The set of sensor
locations
Consider routing
and MAC
schemes
Clusters
Network topology
Yes
No
Yes
No
No restriction
Sensor location
Available
No
Regular grid,
random graph,
preferentia1
attachment graph
Available
Multihop routes
Solution
Yes
Integer programming
Yes
Two-tier (sensors,
cluster head)
with random
graphs
Available Not
available
No
Computational
Geometry
Yes
Local search and
greedy heuristics
NA, not available.
Reference 40
NA
Percentage of
unreachable
nodes
Fixed (known or
unknown)
Depends on the
clustering
algorithm
No
Reference 41
Reference 44
Fixed
Percentage fo dead
sensors
NA
NA
Fixed and known
Predefined random
set of locations
Fixed and
known
Uniformly
distributed
Yes
No
Yes
Uniform random
graph
No
Random graph
Yes
Two-tier
uniform
Available
Available
Available
Yes
Computational
geometry
Yes
Integer
programming
Yes
Vertex 1-center
heuristics
DYNAMIC REPOSITIONING OF NODES
245
9.3 DYNAMIC REPOSITIONING OF NODES
Most of the protocols described above initially compute the optimal location for the
nodes and do not consider moving them once they have been deployed. Moreover,
the context of the pursued optimization strategies is mainly static in the sense that
assessing the quality of candidate positions are based on performance metrics like
the data rate, sensing range, path length in terms of the number of hops from a sensor
node to the base station, and so on. In addition, the placement decision is made at the
time of network setup and does not consider dynamic changes during the network
operation. For example, traffic patterns can change based on the monitored events,
load may not be balanced among the nodes causing bottlenecks, application-level
interest can vary over time, and the available network resources may change due to
the depletion of energy of some nodes and/or the addition of more nodes.
Therefore, dynamically repositioning the nodes while the network is operational
is necessary to further improve the performance of the network. For instance, when
many of the sensors in the vicinity of the base-station become dysfunctional due to the
exhaustion of their batteries, it is better for the base station to reposition itself in order
to become easily and reliably reachable to data sources. Such repositioning can boost
the network longevity and reduce the effect of packet drops caused by link and node
failures. Similarly, instead of repositioning the base station, some redundant sensors
from other parts of the monitored region can be identified and relocated to replace
the dead sensors in the vicinity of the base-station to improve the network lifetime.
Such dynamic relocation can also be very beneficial in a target tracking application
where the target is mobile. For instance, some of the sensors can be relocated close to
the target to increase the fidelity of the sensor’s data. Moreover, in some applications
it may be wise for the base station to keep a distance from harmful targets (e.g., an
enemy tank) by relocating to safer areas in order to ensure its availability.
Relocating the nodes during regular network operation is very challenging. Unlike
initial placement, such relocation is pursued in response to a network- or environmentbased stimulus and thus requires continual monitoring of the network state and performance as well as analysis of events happening in the vicinity of the node. In addition,
the relocation process would need careful handling since it could potentially cause
disruption in the data delivery. The basic issues can be enumerated as follows: when
it would make sense for a node to relocate, where it should go, and how the data
will be routed while the node is moving. In this section we discuss these issues in
detail and discuss sample published approaches on dynamic sensor and base-station
repositioning.
9.3.1 Relocation Issues
Relocation. The decision for a node movement has to be motivated by either (a)
an unacceptable dependability measure despite setting up the most efficient network topology or (b) a desire to boost such measures beyond what is achievable
at the present node position. Motives vary based on the targeted dependability attributes. Examples include the observation of bottlenecks in data relaying, a decrease
246
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
in node coverage in an area, an increase in packet latency, excessive energy consumption per packet delivery, an increased risk to important nodes such as the
base station from an approaching target, and so on. A weighted average may also
be pursued to combine multiple dependability metrics based on the application in
hand.
Once a node has the motive, it will consider moving to a new position. Such
consideration does not necessarily lead to an actual relocation. The node would need
to qualify the impact of repositioning at the new location on the network dependability.
Therefore the “when” and “where” issues of the node movement are very closely
interrelated. In addition, the node would have to assess the relocation overhead. Such
overhead can be incurred by the node and the network. For example, if the base station
or the sensor node is a robot, the energy consumed by the mechanical parts in the
movement is a significant overhead to the lifetime of the robot’s battery and thus
should be minimized. Moreover, when energy and timeliness metrics are of utmost
concern, the impact on the lifetime of individual sensors and on route maintenance
has to be considered, respectively.
Where to Relocate. When having a motive to relocate, the node needs to identify
a new position that would boost the network dependability. Again, the qualification
of the new position and possibly the search criteria may vary based on the dependability attribute. Finding an optimal location for the node in a multihop network is
a very complex problem. The complexity is mainly resulting from two factors. The
first is the potentially infinite number of possible positions to which a node can be
moved. The second factor is the overhead of keeping track of the network and the
node state information for determining the new location. In addition, if the base station is to be relocated, for every interim solution considered during the search for
an optimal position, a new multihop network topology may need to be established
in order to qualify that interim solution in comparison to the current or previously
picked positions.
A mathematical formulation of the base-station relocation problem may involve a
huge number of parameters including the positions of all deployed nodes, their state
information such as energy reserve, transmission range, and so on, and the data sources
in the network. Similar observations can be made for sensor repositioning problems.
In that case, a sensor may need to know the boundaries of the monitored region, the
current coverage ratio of the network, the location of dead sensor nodes, and so on, in
order to determine its new location. Given the large number of nodes typically involved
in applications of WSNs, the pursuance of exhaustive search will be impractical. In
addition, the dynamic nature of the network makes the sensor state and sources of
data variant, and thus the optimization process may have to be repeated frequently.
Moreover, it may be undesirable to involve the nodes in complex computation in order
to ensure sufficient computation capacity for application level processing (e.g., data
fusion) or save enough energy for movement of the node. Therefore, approximate
and local solutions or search heuristics are more attractive in the context of WSNs [7,
47, 48].
DYNAMIC REPOSITIONING OF NODES
247
Managing and Justifying the Move. Once the new location of the node has
been picked and confirmed to enhance some desired dependability attributes, the
node should identify a travel path to this new location. The main contributing factors
to the path selection are the total distance to be traveled, suitability of the terrain,
the path safety, and the risk of disrupting the network operation. Minimizing the
travel distance for both the base station and the sensor nodes is very crucial since the
energy consumed by the mechanical parts in such a movement is much more than
the communication and computation energy. Therefore, the shortest possible path
should be identified to reach to the new location. However, the node also has to pick a
path that is physically feasible to travel over. The node may need to consult a terrain
map or rely on the skills of a carrier (e.g., a robot with cameras or a human being) to
avoid obstacles and dead ends. The other concern is protecting the node during the
move. Since a WSN is usually deployed in harsh environments to detect and track
dangerous targets/events, the node should avoid exposure to harm or getting trapped.
For example, the node should not go through a fire to reach the new location.
The node should also minimize any negative impact on the network operation.
While the node is on the move, it must ensure that data continue to flow. For example,
if the base station is relocating, it has to arrange for sensor transmission to cover the
planned travel path in order to make sure that packets will continue to reach it. Continual data delivery prevents the relocation from causing an application level failure
by missing important reports. Such application level robustness is a dependability
attribute in itself. Therefore, it is desirable to restrict changes to the network topology. Avoiding radical changes to the data routes limits the disruption of ongoing data
traffic and also curtails the overhead that the relocation introduces. Again, the node
performs a tradeoff analysis between the gain achieved by going to a new location
and the overhead in terms of additional energy consumption that the motion imposes
on sensors and the base station. If the motion is justified, the node can physically
relocate.
The last issue is whether there are constraints on the time duration that the node
budgets for the move. These constraints may arise in very dynamic environments in
which the traffic pattern changes frequently. In such a case the gains achieved by
going to a location may be lost or degraded very quickly and the node would find out
that it has to move yet to a third location or even get back to the old position. In the
worst case, the node keeps switching back and forth. Therefore, a gradual approach
to the new location may be advisable in order to prevent this scenario.
9.3.2 Sensor Repositioning Schemes
While the bulk of published work envisioned sensors to be stationary, some investigated the possibility of attaching sensors to moveable entities such as robots [49, 50].
Sensor’s mobility has been exploited to boost the dependability of WSNs. For example, sensors can re-spread in the area to ensure uniform coverage, move closer
to loaded nodes in order to prevent bottlenecks, increase bandwidth by carrying
data to the base station, and so on [51–55]. Proposed schemes for dynamic sensors positioning in the literature can be categorized into two groups based on when
248
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
relocation is exploited: (1) post-deployment and (2) on-demand relocation. We discuss
these two categories of relocation in details in the following subsections.
Post-Deployment Sensor Relocation. This type of relocation is pursued at
the conclusion of the sensor deployment phase when the sensor nodes are being
positioned in the area. As we discussed earlier, in most of the WSN applications, sensor
deployment is performed randomly due to the inaccessibility of the monitored areas.
However, this random configuration usually does not provide an adequate coverage of
the area without deploying an excessive number of nodes. Alternatively, the coverage
quality can be improved by moving the sensor nodes if they are able to do so. In that
case, the sensor nodes can be relocated to the regions that do not have the desired
level of coverage or even are not covered at all. Given the energy cost of mechanical
movement and the communication messages involved in directing the motion, the
relocation process should be lightweight and should conclude in a reasonable time.
Wang et al. [48] utilizes sensor’s ability to move to distribute the sensor nodes as
evenly as possible in the region. The goal is to maximize the area coverage within
the least time duration and with minimal overhead in terms of travel distances and
inter-sensor message traffic. The main idea is that each sensor assesses the coverage
in its vicinity after deployment and decides on whether it should move to boost the
coverage. To assess the coverage, a sensor node creates a Voronoi polygon with respect
to neighboring sensors, as illustrated in Figure 9.12. Every point in a Voronoi polygon
is closer to the sensor of that polygon (i.e., Si in Figure 9.12) than any other sensor.
The intersection of the disk that defines the sensing range and the Voronoi polygon
would identify any uncovered area, which would motivate a sensor to move.
In order to decide where to reposition a sensor, three methods were proposed:
vector-based (VEC), Voronoi-based (VOR), and minimax. The main idea of the VEC
method is borrowed from electromagnetics where close particles are subject to an
Voronoi Polygon
for Si
B
Si
A
sing
Sen for S i
ge
Ran
Figure 9.12. Every sensor Si forms Voronoi polygon with respect to the position of its
neighboring sensor. The part of the polygon that lies outside the sensing range is not covered by Si .
DYNAMIC REPOSITIONING OF NODES
249
expelling force to keep them apart. In the context of WSNs, virtual forces are applied
to a sensor node by its neighbors and by the boundaries of its Voronoi polygon in order
to change its location. While in VEC the nodes are pushed away from the densely
populated areas, VOR pulls the sensors to the sparsely populated areas. In VOR, the
sensor node is pulled toward the farthest Voronoi vertex to fix the coverage hole in
the polygon, point “A” in Figure 9.12. However, the sensor will be allowed to travel
only a distance that equals half of its communication range, point “B” in Figure 9.12,
in order to avoid stepping into the area handled by another sensor that was out of
reach prior to the move (i.e., is not a current neighbor of Si ,), which can lead to an
unnecessary move backward later on. In the minimax method, a sensor also gets closer
to its farthest Voronoi vertex. However, unlike VOR, the minimax approach strives
to keep most of the other vertices of the Voronoi polygon within the sensing range. It
thus relocates the sensor to a point inside the Voronoi polygon whose distance to the
farthest Voronoi vertex is minimized. The minimax scheme is more conservative in
the sense that it avoids creating coverage holes by going far from the closest vertices,
leading to more regularly shaped Voronoi polygon.
The conserved departure from current sensor location leads to a gradual relocation,
round by round as shown in Figure 9.13. This usually causes zigzag movement of each
sensor rather than directly going to the final destination. In order to shorten the total
travel distance, a proxy-based approach is proposed in reference 54. In this approach,
the sensor nodes do not move physically unless their final destination is computed.
The authors consider a network with stationary and mobile sensors. Mobile sensors
are used to fill coverage holes identified in a distributed way by stationary nodes. Thus,
mobile sensors only move logically and designate the stationary sensor nodes as their
proxies. With this approach, significant improvements can be made to the total and
average distance traveled by mobile nodes while at the same time achieving exactly
the same level of coverage reported in reference 48. The approach only increases
the message complexity. However, given that movement is more costly in terms
50
50
40
40
30
30
20
20
10
10
0
0
10
20
30
Round 0
40
50
0
0
10
20
30
40
50
Round 1
Figure 9.13. Sensors pursue relocation iteratively until no improvement in coverage can be
achieved. (Taken from reference 51.)
250
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
1
2
3
2
1
3
2
2
2
2
2
2
10
10
10
10
10
10
2
22
3
7
56
18
18
18
18
18
18
18
10
10
10
10
10
10
4
12
2
30
1
5
9
9
9
9
9
9
10
10
10
10
10
10
3
7
2
1
2
3
3
3
3
3
3
3
10
10
10
10
10
10
1
5
19
3
21
77
21
21
21
21
21
21
10
10
10
10
10
10
6
7
10
10
4
5
7
7
7
7
7
7
10
10
10
10
10
10
(a)
(b )
(c)
Figure 9.14. Steps for SMART: (a) Initial 2D mesh. (b) Row scan. (c) Column scan.
of energy, such increase can be justified. Nonetheless, this process still can be very
slow and hence prolong the deployment time.
With the objective of reducing the overall deployment time, Wu and Yang [53]
proposed another solution to the same problem based on two-dimensional scan of
a clustered network, called SMART. The approach adopts a popular scheme for
balancing load among nodes in parallel processing architectures by assigning an
equal number of tasks to each processor. This idea is applied to a multicluster WSN
where each cluster is represented with a square cell forming a 2D mesh, as seen in
Figure 9.14, which is redrawn from reference 53. The number of sensors annotated
on every cell represents the load of that cluster. Each cluster head knows only its location within the mesh (i.e., row and column indices) and the number of sensors in its
cluster. It is assumed that a cluster head can only communicate with its counterparts in
neighboring cells. Achieving uniform coverage is then mapped to the load-balancing
problem with a goal of evening the distribution of sensors among the clusters. To
achieve this goal, each cluster head performs both a row-based and a column-based
scan to exchange the load information. In a row-based scan, the leftmost cluster head
forwards its load (i.e., number of sensors) to its right neighbor. Each neighbor on the
row adds the received load to its own load and forwards it until the rightmost cluster
head is reached. This cluster head computes the average load for its row and sends a
message back until the leftmost cluster head gets such average (Figure 9.14b). After
the scan process, the sensors are relocated to match the desired node count per cluster.
That is, the overloaded clusters give sensors while the underloaded clusters take sensors. The same procedure is applied for each column (Figure 9.14c). The approach
also handles possible holes in the network when there are clusters with no sensors.
The simulation results compared VOR (discussed above) and SMART with respect to
the number of moves made by the sensors and the number of rounds until termination.
SMART was shown to provide the minimum number of moves. Although it was also
shown that SMART converges in a fewer number of rounds for densely populated
WSNs, VOR was found to be superior for sparsely populated networks.
Another similar post-deployment relocation work for improving the initial coverage and providing uniform distribution of sensors is presented in reference 56.
Although the idea is similar to the VEC mechanism reported in reference 48, this
time it is inspired by the equilibrium of particles in Physics. The particles follow
DYNAMIC REPOSITIONING OF NODES
251
the law of Coulomb and push themselves to reach equilibrium in an environment.
Therefore, the authors define forces for each sensor node in the network based on the
internode distances and the local density of nodes. The partial force on a node i from
node j at time t is expressed as follows:
j
i,j
ft
j
= d(R − |pit − pt |)
pt − pit
j
|pt − pit |
where d is the density, R is the transmission range, and pt is the location at time t.
Each sensor’s movement is decided by the combined force applied to that sensor by
all neighboring nodes. A sensor keeps on moving until it travels a distance below a
certain threshold at a given time. In some cases, the node can move back and forth
between two locations leading to an oscillation. If such oscillation is noticed to occur
more than a preset limit, the node stays at the center of gravity of the oscillation
points.
To validate the performance, the approach is implemented and compared to a simulated annealing-based solution, which provides an optimal coverage. The validation
results indicated that the proposed self-spreading approach performs very close to
optimal in terms of coverage and delivers a superior performance in terms of the total
distance traveled and the time to converge.
On-Demand Repositioning of Sensors. Instead of relocating the nodes at the
deployment phase, sensor relocation can be used on demand to improve certain performance metrics such as coverage, network lifetime, and so on. This can be decided
during the network operation based on the changes in either application-level needs
or the network state. For instance, the application can be tracking a fast-moving
target that may require repositioning of some sensor nodes based on the new location of the target. Furthermore, in some applications, there can be an increasing
number of dysfunctional nodes in a particular part of the area necessitating the redistribution of available sensors. In addition to improving coverage, the energy consumption can be reduced through on-demand relocation of sensors in order to reach
the best efficient topology.
The approach presented in reference 47 performs sensor relocation to counter
holes in coverage caused by sensors failure. The idea is simply to identify some spare
sensors from different parts of the network that can be repositioned in the vicinity
of the faulty nodes. The selection of the most appropriate choice among multiple
candidate spare nodes is based on the recovery time and overhead imposed. Both
criteria would favor close-by spares over distant ones. Minimizing the recovery time
can be particularly crucial for delay sensitive applications. The overhead can be in the
form of energy consumption due to the node’s travel and due to the message exchange,
especially if spares are picked in a distributed manner. In order to detect the closest
redundant sensor with low message complexity, a grid-based approach is proposed.
The region is divided into cells with a designated head for each cell. Each cell head
advertises/requests redundant nodes for its cell. A quorum-based solution is proposed
252
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
S2
Choice 1:
S3
S0
S1
S2
Choice 2:
S3
S0
S1
Figure 9.15. Cascaded Movement of Sensors; S3 replaces S2, S2 settles in S1’s position and
S1 move to where S0 is.
to detect the intersection of advertisements and requests within the grid. Once the
redundant sensor is located, it is relocated to the desired cell without disrupting the
data traffic and affecting the network topology.
Since directly moving the node can drain significant amount of energy, a cascaded
movement is proposed. The idea is to determine intermediate sensor nodes on the
path and replace those nodes gradually. That is, the redundant sensor will replace
the first sensor node on the path. That node will also move and replace the second
sensor node, and so on. For the example shown in Figure 9.15 (which is redrawn from
reference 47), rather than directly moving S3 to the location of S0, in choice 2 all sensors S3 , S2 , and S1 move at the same time and replace S2 , S1, and S0, respectively, in
order to minimize the relocation time. The path is selected such that it will minimize
the total mechanical movement energy and at the same time maximize the remaining energy of sensor nodes. In order to determine such a path, Dijkstra’s least-cost
path algorithm is used. The overall solution is also revisited to provide a distributed
approach for determining the best cascading schedule.
When validated, the approach outperformed VOR of reference 48 with respect
to the number of sensors involved in the relocation, the total consumed energy, and
the total remaining energy. In addition, cascaded movement delivered much better
performance than direct movement in terms of relocation time, energy cost, and
remaining energy. However, obviously the cost of maintaining a grid, selection of the
cell head and redundant nodes will grow dramatically with the increasing number of
nodes. For scalability, a hierarchical solution might be needed to restrict the size of
the region and the cost of movements.
Coverage improvement was the objective of relocating imaging sensors in reference 57. Stationary cameras may not provide the desired coverage when there
DYNAMIC REPOSITIONING OF NODES
253
Obstacle
Figure 9.16. An obstacle reducing the coverage of two cameras directed at different areas.
are environmental peculiarities (e.g., moving obstacles) in the event area as seen in
Figure 9.16. Thus, moving the cameras in order to avoid obstacles that block their
vision would increase the coverage. The mobility for the camera nodes is made possible through providing a traction mechanism under each camera that will enable motion
in one dimension, as is also implemented in Robomote [49]. This mobility feature
on cameras is actuated when the coverage of the monitored areas falls below a certain
ratio. The experiments performed with real-life target tracking applications verified
that the mobile cameras can increase the coverage of the monitored area and thus
would decrease the target miss ratio significantly when compared to stationary-node
based setups.
Dasgupta et al. [58] exploited sensor relocation to improve the lifetime of the
network rather than coverage. The maximum lifetime sensor deployment problem
with coverage constraints has been investigated. The authors assumed a network
operation model in which every sensor periodically sends its data report to the base
station. The network is required to cover a number of points of interest for the longest
time. The average energy consumption per data collection round is used as a metric
for measuring the sensor’s lifetime. The problem is then transformed to minimizing
the average energy consumption by a sensor per round by balancing the load among
sensors. The idea is to spread the responsibility of probing the points of interest
among the most number of sensors and to carefully assign relays so that the data are
disseminated using the least amount of energy. A heuristic was proposed that tries to
relocate sensors in order to form the most efficient topology. First, sensors are sorted
decently according to the point of interest that they cover. Starting from the top of
the sorted list, the algorithm iterates on all sensors. In each iteration, the sensor is
254
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
checked for whether it can move to another location to serve as a relay. The new
location is picked based on the traffic flow and the data path that this node is part of or
will be joining. Basically the relocating node should reduce its energy consumption
by getting close to its downstream neighbor. A sensor repositioning is allowed only
if it is not risking a loss in coverage.
9.3.3 Base-Station Relocation for Increased Dependability
In this section, we report on some of our investigation of the base-station relocation problem. Three dependability attributes are considered: network longevity, data
delivery timeliness, and base-station physical security (safety). For each of these attributes, we explain how we address the three questions; when to relocate, where to
place the base station, and how the move is managed.
Relocation for Increased Network Longevity. Although energy-aware multihop routing does dynamically adapt to changes in sensor’s energy and traffic pattern,
sensors nearby the base station die quickly because they are the most utilized nodes
in the network. Consequently, nodes that are further away from the base station are
picked as substitute relays as depicted in Fig. 9.17, and the node’s energy utilized for
communication with the base station would be considerably higher. Such effects can
spread in a spiral manner, thus draining the sensors energy and hence shortening the
lifetime of the network. To stop such pattern of energy depletion, the base station is
repositioned [59, 60].
The main idea is to move the base station toward the sources of largest traffic.
The traffic density (PT) times the transmission power (ETR ) is used as a metric for
monitoring the network operation and searching for the best base-station location.
The idea is to track changes in (a) the nodes that act as the closest hop to the base
station and (b) the traffic density going through these hops. If the distance between the
base station and some of the nodes that are in direct communication is smaller than a
Base-station
Inactive Sensor
Active Sensor
One hop Sensor
Dead Sensor
Figure 9.17. Nodes close to the base station die rather quickly due to overload forcing the
more distant nodes to relay the data to the base station.
DYNAMIC REPOSITIONING OF NODES
255
threshold value, the base station will qualify the impact of these nodes on the overall
network lifetime by considering the number of packets routed through them. If the
PT * ETR is greater than a certain threshold, the base station will consider relocating
to a new position.
Such an approach has multiple advantages. First, the base station will be close to
the area where more nodes are collecting data and thus communication-related energy
consumption will be reduced. It is thus expected that the average energy per packet
gets reduced. In addition, nodes collecting the most data will be closer to the base
station and most probably fewer hops will be involved lowering the overall latency
time for data collection. Moreover, the packet throughput will be higher since it is
expected that most messages pass through fewer hops and travel shorter distances,
making them less likely to be dropped. In summary, such an approach for relocating
the base station not only can increase network longevity but also can enhance other
performance metrics like latency and throughput.
While such positioning will be ideal for high-traffic paths, it can worsen the performance on paths with lower traffic density or which are topologically opposite
to the direction of the base-station motion. Therefore, before confirming the move,
the base station validates the overall impact on transmission energy by factoring
in the possible extension of some data paths and the cost of signaling overhead to
those sensor nodes affected by the move. In summary, the cost of base-station relocation should be justified before being pursued. When the base station starts to move,
the data gathering process still continues and thus the routes should be adjusted before
the base station gets out of range of some sensors (if any). Unlike static approaches
discussed in Section 9.3, routes adjustment is an issue in dynamic base-station positioning. This issue can be handled by either increasing the transmission power or
designating additional forwarder sensors. The change in base-station position may
also introduce shadowing or multipath fading to some links. Slow or gradual advance
toward the new position can be effective in avoiding unexpected link failures that may
cause negative performance impacts, there by allowing the base station to rethink the
suitability of the newly selected position and/or the decision to move further.
Enhancing Timeliness of Delay-Constrained Traffic. In addition to boosting
network longevity, the repositioning of the base station also becomes influential when
real-time traffic with certain end-to-end delay requirements is involved. For instance,
when routes to the base station get congested, most requests for establishing paths
for real-time data may be denied or the deadline miss rate of real-time packets may
increase significantly. Traffic congestion can be caused by an increase in the number
of real-time data packets coming from nodes close to a recent event. In such circumstances, it may be infeasible to meet the requirements for real-time data delivery.
Repositioning the base station can then be beneficial in order to spread the traffic on
additional hops and increase the feasibility for meeting the timeliness requirements.
A trigger for such relocation can be the unacceptable increase in the miss rate of
real-time packets or just a desire to increase timeliness even if the miss rate is at a
level that is tolerable by the application. To boost timeliness, the base station can
move to the location of, or close to, the most heavily loaded node and try to split the
256
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
Figure 9.18. (a) The base station (denoted as B) is relocated to the location of A if the delay is
not extended on the path from C to B. (b) If real-time traffic through C is affected, B is relocated
close to C while still splitting traffic at A.
incoming traffic passing through that node without extending the delay experienced
by real-time packets over other routes as shown in Figure 9.18 [60, 61]. Such loaded
nodes are picked based on the real-time traffic service rate, often determined during
the route setup in order to allocate bandwidth to both real-time and non-real-time
traffic. Again, the pros of the relocation have to be qualified to make sure that the
overhead is justified. It is also worth noting that in this approach the impact on link
quality is not a major concern when the base station moves close to a heavily loaded
node since it is unlikely that the node is experiencing a disruptive level of interference
while being able to relay a high volume of real-time packets.
Handling the base-station motion is similar to the approach of the network
longevity relocation, described above. As long as the base station remains within the
transmission range of all the last-hop nodes, the current routes can be maintained by
only adjusting the transmission power. If it is expected that the new location will put the
base station out of the transmission range of some of the last hop nodes in the current
routes, new forwarder nodes that are not involved in any routing activity are selected.
Such unused nodes will introduce very little queuing delay, which is desirable for
on-time delivery of all real-time packets that use these nodes as relays. Note that designating new forwarder nodes is still more costly than just adjusting the transmission
power. Therefore, if the new location imposes excessive topology changes, alternative
positions that will cause no or minimal topology changes are to be considered.
Protecting the Base Station. As discussed earlier, moving the base station toward
highly loaded nodes would have the potential for enhancing network performance in
terms of energy consumption, throughput, and delay. However, not all locations would
be safe for the base station even if a substantial gain in performance is perceived. In
many situations, sensors are placed in hostile environments where there is always a
lurking danger to the base station as it gets closer to the data sources. For example, in
DYNAMIC REPOSITIONING OF NODES
257
a disaster management application, sensors may be reporting about fires, collapsing
building, gas leaks, and so on. Similarly, in combat field surveillance, the sensors
may be reporting tanks or enemy troops. Moving too close to the reported events in
these scenarios would be very risky. Thus, handling such scenarios would be subject
to performance and safety tradeoff. In such cases, the base station would have to
consider the potential performance degradation while getting away from danger.
To assess the safety implication of repositioning the base station, a stochastic or a
cognitive formulation can be pursued [62]. The idea is to track the base-station safety
levels at different locations and use them to define the parameters of the base-station
safety model. Then the threat implication is estimated as a function of the proximity
to reported events and the severity of these events. If the location of the event is
not accurately known, the data volume related to that event and the location of the
reporting sensors are factored in. An objective function is then formed to balance
safety and performance goals and used to guide the search for the new location of the
base station. Figure 9.19 shows some performance improvements through the basestation repositioning and its safety aware version named SAFER. In these experiments,
sensor nodes are randomly placed in a 500 × 500-m2 area while the network is tasked
with a target tracking mission. The base station is dynamically repositioned during
the network operation based on the location of high volume traffic in the event area.
Figure 9.19. Performance improvements in terms of throughput, energy and delay by repositioning of the base-station and its safety aware version.
258
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
The same experiments are performed by taking into consideration the existing risks
in the event region while the base station is on the move.
Finally we would like to note that all computation related to the presented heuristics
is performed at the base station, which usually is not resource-constrained when
compared to sensors. Therefore, the base station should be able to track/collect the
necessary state information of the network in order to trigger relocation and determine
the new position. The base station would identify the necessary network information
based on the goal of the relocation (e.g., energy, timeliness, safety, etc.). We argue
that this information can be obtained even in distributed route setups by keeping track
of packets arriving from the sensor nodes over time. For example, the base station can
monitor the number of received packets from a particular node, estimate the remaining
energy, determine the length of the path, and so on.
9.4 COORDINATED MULTINODE RELOCATION
In many application setups, nodes coordinate among themselves in order to efficiently
and effectively handle application-level requirements. Examples of such applications
include robotic-based land-mine detection and deactivation, multi-rover exploration
of distance planets, employing a sensor fleet for oceanic studies, and so on. However,
unlike the case discussed in the previous section, such application-level coordination
requires the relocation to care for internode networking issues. For instance, consider
the scenario depicted in Figure 9.20a. A set of sensors are deployed in an area of interest. Four base stations are employed to serve an application based on data collected
from these sensors. The sensors have been partitioned into nonoverlapping clusters;
each is managed by a distinct base station.
Given the commonly uneven distribution of sensor nodes, some events would be
hard to monitor or would overburden the scarce network resources in the proximity
of the event. In Figure 9.20a, two sample events articulate such an issue. The depicted
events, may be targets or fires, are reported by very few sensors for which the events
happen to be within their detection range. Clearly, the intracluster network topology
for both clusters 1 and 4 are not efficient since the data are routed over many hops
and may not be arriving at B1 and B4 reliably and/or on time. In addition, some of the
nodes that are involved in relaying data may not have abundant energy reserve to keep
the path stable for an extended duration. The following are some of the challenges
that base stations may face when trying to boost the efficiency of the data collection
and dependability of the intracluster network:
1. B1 may decide to relocate to better serve the event tracked by sensors in its
cluster—for example, monitor the event in a timely manner or to minimize
the energy consumption at relaying sensor nodes. However, such repositioning
may get B1 out of the communication range of B4 . Such a move can only be
acceptable if B1 does not need to interact with B4 , which is unlikely, or when an
alternative path can be established with an acceptable delay bound. Referring to
Figure 9.20a, relocating B1 would be feasible if the communication path (B1 ,
B3 , B4 ) meets the timeliness requirements.
COORDINATED MULTINODE RELOCATION
259
Figure 9.20. (a) A sample multi-base-station clustered sensor network architecture, where
each cluster is handled/managed by a distinct base station. (b) B3 moves to maintain the communication link with B4 , probably losing connectivity to B2 . However, base station still form
a connected graph.
260
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
2. The possibility of handling the event by a different cluster is another issue of
consideration. For example, while B1 can relocate to better serve the event, the
data can also be routed from Si to B2 over a shorter and more energy-efficient
path. That would require changing the association of Si from cluster 1 to cluster
2, at least temporarily. Such modification of cluster membership would sure
impose an overhead. Such a tradeoff will be unavoidable.
3. When relocation of a base station is the only option for boosting the dependability of the operation within a cluster, a ripple effect may be caused throughout the network. Consider, for example, the relocation of B4 close to Sj . Although such a close proximity to the event would have positive impact on the
operation in cluster 4, moving B4 that far makes it unreachable to other base
stations. Such a scenario is not acceptable in a collaborative computing environment. A more complex alternative is to relocate multiple base stations in
order to maintain connectivity among base stations. Figure 9.20b illustrates
a possible multi-base-station relocation that better serves the event reported
by Sj . Basically, B3 gets closer to the new location of B4 in order to prevent
their communication link from getting broken, possibly at the expense of losing connectivity to B2 . This solution could have been more involved if B1
does not have a link to B2 forcing it to move as well to be in the range of B3
and B2 .
Obviously, a multinode relocation can be significantly more complex and can
introduce lots of overhead. We envision coordinated multinode relocation to be a
promising research direction. Only little attention has been paid to tackle the multinode
relocation challenges. One of the few attempts was reported in reference 63, where a
dynamic multi-base-station positioning is proposed in order to improve the network
longevity. The motion of base stations is further restricted to maintain the connectivity
of the inter-base-station network as seen in Figure 9.21. When a BS is to be relocated,
its links with the neighbors are checked first before the relocation is performed. If
changing the position of a base station is to cause partitioning in the network, its
BS1
Cluster 3
Cluster 1
BS3
BS5
Cluster 5
BS2
BS4
Cluste r 4
Cluster 2
Figure 9.21. Connected inter-base-station network.
EXERCISES
261
neighbors are to move to restore broken links. For example, in Figure 9.21, if BS3
is to move toward BS2 , BS5 would follow through to avoid being disconnected from
the network. In order to prevent simultaneous relocations, a mutual-exclusion-based
mechanism is used. The idea is to access to a global token exclusively in order to
perform the relocation. A base station will cease any motion until a token is granted.
9.5 CONCLUSION
Wireless sensor networks (WSNs) have attracted lots of attention in recent years due
to their potential in many applications such as border protection and combat field
surveillance. Given the criticality of such applications, maintaining a dependable
operation of the network is a fundamental objective. However, the resourceconstrained nature of sensor nodes and the ad hoc formation of the network, often
coupled with an unattended deployment, pose nonconventional challenges and motivate the need for special techniques for dependable design and management of WSNs.
In this chapter, we have discussed the effect node placement strategies on the dependability of WSNs. We categorized the various approaches in the literature on sensor and
base-station positioning for enhanced dependability. We have further highlighted the
potential of dynamic repositioning of the base-station and sensor nodes, as a viable
means for increasing the dependability of WSNs. Unlike the initial careful placement,
node repositioning can assist in dealing with dynamic variations in the network
resources and surrounding environment. We have identified the technical issues pertaining to relocating the nodes—namely, when to reposition a node, where to move
it, and how to manage the network while the node is in motion. We have discussed
sample techniques that employ sensor and base-station relocation for boosting the
coverage, tolerating node failure, extending network lifetime, increasing responsiveness in data delivery, and protecting the network assets. We have further identified
the coordinated multinode repositioning problem as an open area for research.
9.6 EXERCISES
1. List some examples of sensor network applications and enumerate the dependability requirements that these applications are subject to. Discuss how node
placement affects the network dependability attributes in the context of these
applications.
2. What are the most important metrics to be considered in performing sensor
relocation? Discuss possible tradeoffs among these metrics.
3. When would dynamic node placement be most appropriate?
4. How would the communication range of relay and sensor nodes affect the complexity and placement strategy of relay nodes?
5. Describe some scenarios for which dynamic placement of nodes can hurt the
network performance rather than enhancing it.
262
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
6. Based on the analysis and deployment techniques discussed in this chapter, how would you design a sensor network for the following applications?
(i) Underwater surveillance, (ii) Border protection, and (iii) Lunar exploration.
7. Pick a randomly deployed sensor network with five nodes. Show the final configuration of the sensor network after applying the Voronoi-based sensor relocation
(VOR) technique of reference 48.
8. Discuss the relation between network clustering and the multiple base-station
placement problems. How would the order of applying the clustering and placement techniques affect the network design/performance?
BIBLIOGRAPHY
1. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks:
A survey. Computer Networks, 38:393–422, 2002.
2. C.-Y. Chong and S. P. Kumar. Sensor networks: Evolution, opportunities, and challenges.
Proceedings of the IEEE, 91(8):1247–1256, 2003.
3. K. Akkaya and M. Younis. A survey on routing protocols for wireless sensor networks.
Elsevier Journal of Ad Hoc Networks, 3(3):325–349, 2005.
4. P. Naik and K. M. Sivalingam. A survey of MAC protocols for sensor networks. In Wireless
Sensor Networks, C. S. Raghavendra, K. M. Sivalingam, and T. Znati, editors. Kluwer
Academic Publishers, Norwell, MA, 93–107.
5. N. Sadagopan, B. Krishnamachari, and A. Helmy. Active query forwarding in sensor networks (ACQUIRE). Elsevier Journal of Ad Hoc Networks, 3(1):91–113, 2005.
6. A. Cerpa and D. Estrin, ASCENT: Adaptive self-configuring sensor networks topologies.
IEEE Transactions on Mobile Computing, 3(3):272–285, July-August 2004.
7. A. Efrat, S. Har-Peled, and J. S. B. Mitchell. Approximation algorithms for two optimal
location problems in sensor networks. In Proceedings of the 3rd International Conference
on Broadband Communications, Networks and Systems (Broadnets 2005), pp. 714–723,
Boston, MA, October 2005.
8. X. Cheng, DZ Du, L. Wang, and B. Xu. Relay sensor placement in wireless sensor networks.
ACM/Springer Journal of Wireless Networks, 14(3):347–355, June 2008.
9. E. L. Lloyd and G. Xue. Relay node placement in wireless sensor networks. IEEE Transactions on Computers, 56(1):134–138, January 2007.
10. S. Poduri, S. Pattem, B. Krishnamachari, and G. S. Sukhatme. Sensor network configuration
and the curse of dimensionality. In Proceedings of the 3rd IEEE Workshop on Embedded
Networked Sensors, Cambridge, MA, May 2006.
11. A. Bogdanov, E. Maneva, and S. Riesenfeld, Power-aware base station positioning for
sensor networks. In Proceedings of the 23rd International Annual Joint Conference of
the IEEE Computer and Communications Societies (INFOCOM 2004), pp. 575–585,
Hong Kong, March 2004.
12. J. Pan, L. Cai, Y. T. Hou, Y. Shi, and S. X. Shen. Optimal base-station locations in twotiered wireless sensor networks. IEEE Transactions on Mobile Computing, 4(5):458–473,
September/October 2005.
BIBLIOGRAPHY
263
13. S. S. Dhillon and K. Chakrabarty. Sensor placement for effective coverage and surveillance
in distributed sensor networks. In Proceedings of the IEEE Wireless Communications and
Networking Conference (WCNC ’03), New Orleans, LA, March 2003.
14. T. Clouqueur, V. Phipatanasuphorn, P. Ramanathan, and K. K. Saluja. Sensor deployment
strategy for target detection. In Proceedings of the 1st ACM international Workshop on
Wireless Sensor Networks and Applications (WSNA ’02), pp. 42–48, Atlanta, GA, September 2002.
15. A. Brooks, A. Makarenko, T. Kaupp, S. Williams, and H. Durrant-Whyte. Implementation
of an indoor active sensor network. In Proceedings of the 9th International Symposium on
Experimental Robotics, Singapore, June 2004.
16. V. A. Petrushin, G. Wei, O. Shakil, D. Roqueiro, and V. Gershman. Multiple-sensor indoor
surveillance system. In Proceedings of the 3rd Canadian Conference on Computer and
Robot Vision (CRV’06), pp. 40–47, Québec City, June 2006.
17. L. Krishnamurthy et al. Design and deployment of industrial sensor networks: Experiences
from a semiconductor plant and the North Sea. In Proceedings of the 3rd ACM Conference on Embedded Networked Sensor Systems (SenSys ’05), pp. 64–75, San Diego, CA,
November 2005.
18. J. Paek, K. Chintalapudi, J. Cafferey, R. Govindan, and S. Masri. A wireless sensor network
for structural health monitoring: Performance and experience. In Proceedings of the Second IEEE Workshop on Embedded Networked Sensors (EmNetS-II), pp. 63–74, Syndney,
Australia, May 2005.
19. K. Mechitov, W. Kim, G. Agha, and T. Nagayama. High-frequency distributed sensing for
structure monitoring. In Proceedings of the First International Conference on Networked
Sensing Systems, Tokyo, Japan, June 2004.
20. J. Berry, L. Fleischer, W. E. Hart, and C. A. Phillips. Sensor placement in municipal water
networks. In Proceedings of the World Water and Environmental Resources Conference,
Philadelphia, PA, June 2003.
21. J.-P. Watson, H. Greenberg, and W. E. Hart. A multiple-objective analysis of sensor placement optimization in water networks. In Proceedings of the World Water and Environment
Resources Congress, Salt Lake City, UT, June 2004.
22. H. H. Gonzalez-Banos and J. C. Latombe. A randomized art-gallery algorithm for sensor placement. In Proceedings of the 17th ACM Symposium on Computational Geometry
(SoCG’01), pp. 232–240, Boston, MA 2001.
23. L. Navarro, J. Dolan, and P. Khosla. Optimal sensor placement for cooperative distributed
vision. In Proceedings of the International Conference on Robotics and Automation (ICRA),
pp. 939–944, New Orleans, LA, April 2004.
24. J. Tang, B. Hao, and A. Sen. Relay node placement in large scale wireless sensor networks.
Computer Communications (special issue on wireless sensor networks), 29:490–501,
2006.
25. M. Ishizuka, and M. Aida. Performance study of node placement in sensor networks.
In Proceedings of the 24th International Conference on Distributed Computing Systems
Workshops, W7:EC (Icdcsw’04), Vol. 7, pp. 598–603, Hachioji, Tokyo, Japan, March 2004.
26. K. Xu, H. Hassanein, G. Takahara, and W. Wang. Relay node deployment strategies in
heterogeneous wireless sensor networks: Single-hop communication case. In Proceedings
of the IEEE Global Telecommunication Conference (Globecom ’05), pp. 21–25, St. Louis,
MO, November 2005.
264
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
27. K. Xu, H. Hassanein, G. Takahara, and Q. Wang. Relay node deployment strategies in heterogeneous wireless sensor networks: Multiple-hop communication case. In Proceedings
of 2nd IEEE Conference on Sensor and Ad Hoc Communications and Networks (SECON
’05), pp. 238–241, Santa Clara, CA, September 2005.
28. C.-F. Huang and Y.-C. Tseng. The coverage problem in a wireless sensor network. In
Proceedings of the ACM 9th Annual International Conference on Mobile Computing and
Networking (MobiCom 2003), pp. 115–121, San Diego, CA, September 2003.
29. A. Boukerche, X. Fei, and R. B. Araujo. A coverage preserving and fault tolerant based
scheme for irregular sensing range in wireless sensor networks. In Proceedings of the
49th Annual IEEE Global Communication Conference (Globecom’06), San Francisco,
CA, November 2006.
30. S. Meguerdichian, F. Koushanfar, M. Potkonjak, and M. B. Srivastava. Coverage problems
in wireless ad-hoc sensor networks. In Proceedings of the IEEE Annual Joint Conference
of the IEEE Computer and Communications Societies (INFOCOM’01), pp. 1380–1387,
Anchorage, AL, April 2001.
31. E. S. Biagioni and G. Sasaki. Wireless sensor placement for reliable and efficient data
collection. In Proceedings of the 36th Annual Hawaii international Conference on System
Sciences (HICSS’03), Track 5, Vol. 5, Big Island, HI, January 2003.
32. K. Kar and S. Banerjee. Node placement for connected coverage in sensor networks.
In Proceedings of the Workshop on Modeling and Optimization in Mobile, Ad Hoc and
Wireless Networks (WiOpt’03), Sophia Antipolis, France, 2003.
33. J. Bredin, E. Demaine, M. Taghi Hajiaghayi, and D. Rus. Deploying sensor networks with
guaranteed capacity and fault tolerance. In Proceedings of the 6th ACM International Symposium on Mobile Ad Hoc Networking and Computing, pp. 309–319, Urbana-Champaign,
IL, May 2005.
34. M. Younis, M. Youssef, and K. Arisha. Energy-aware management in cluster-based sensor
networks. Computer Networks, 43(5):649–668, 2003.
35. Y. Chen, C. Chuan, and Q. Zhao. Sensor placement for maximizing lifetime per unit cost in
wireless sensor networks. In Proceedings of the IEEE Military Communication Conference
(MILCOM ’05), Vol. 2, pp. 1097–1102, Atlantic City, NJ, October 2005.
36. Y. T. Hou, Y. Shi, and H. D. Sherali. On energy provisioning and relay node placement for
wireless sensor networks. IEEE Transactions on Wireless Communications, 4(5):2579–
2590, 2005.
37. X. Zhang and S. B. Wicker. How to distribute sensors in a random field? In Proceedings
of the 3rd International Symposium on Information Processing in Sensor Networks (IPSN
’04), pp. 243–250, Berkeley, CA, April 2004.
38. D. Ganesan, R. Cristescu, and B. Beferull-Lozano, Power-efficient sensor placement and
transmission structure for data gathering under distortion constraints. In Proceedings of the
3rd International Symposium on Information Processing in Sensor Networks (IPSN ’04),
pp. 142–150, Berkeley, CA, April 2004.
39. Q. Wang, K. Xu, G. Takahara, and H. Hassanein. Deployment for information oriented
sensing coverage in wireless sensor networks. In Proceedings of the 49th IEEE Global
Telecommunication Conference (Globecom’06), San Francisco, CA, November 2006.
40. E. I. Oyman and C. Ersoy. Multiple sink network design problem in large scale wireless
sensor networks. In Proceedings of the IEEE International Conference on Communications
(ICC 2004), pp. 3663–3667, Paris, June 2004.
BIBLIOGRAPHY
265
41. S. R. Gandham, M. Dawande, R. Prakash, and S. Venkatesan. Energy efficient schemes for
wireless sensor networks with multiple mobile base stations. In Proceedings of the 46th
IEEE Global Telecommunication Conference (Globecom’03), pp. 377–381, San Francisco,
CA, December 2003.
42. J. Pan et al. Locating base-stations for video sensor networks. In Proceedings of the IEEE
Vehicular Technology Conference (VTC-Fall’03), Vol. 5, pp. 3000–3004, Orlando, FL,
October 2003.
43. Y. Shi, Y. T. Hou, and A. Efrat. Algorithm design for base station placement problems in
sensor networks. In Proceedings of IEEE International Conferenfsfce on Quality of Service
in Heterogeneous Wired/Wireless Networks, Waterloo, Ontario, Canada, August 7–9, 2006.
44. K. Akkaya and M. Younis. COLA: A coverage and latency aware actor placement for wireless sensor and actor networks. In Proceedings of IEEE Vehicular Technology Conference
(VTC-Spring ’06), Montreal, Canada, September 2006.
45. T. W. Haynes, S. T. Hedetniemi, and P. J. Slater. Domination in Graphs: Advanced Topics.
Marcel Dekker, New York, 1998.
46. K. Akkaya and M. Younis. Coverage and latency aware actor placement for wireless sensor
and actor networks. Technical Report, Department of Computer Science and Electrical
Engineering, University of Maryland Baltimore County, April 2006.
47. K. Akkaya, M. Younis and M. Bangad. Sink repositioning for enhanced performance in
wireless sensor networks. Computer Networks, 49(1):512–534, 2005.
48. G. Wang, G. Cao, T. La Porta, and W. Zhang. Sensor relocation in mobile sensor networks. In Proceedings of the IEEE Annual Joint Conference of the IEEE Computer and
Communications Societies (INFOCOM’05), pp. 2302–2312, Miami, FL, March 2005.
49. http://www-robotics.usc.edu/∼robomote/
50. http://www.symbio.jst.go.jp/symbio/morph.html
51. G. Wang, G. Cao, and T. La Porta. Movement-assisted sensor deployment. In Proceedings
of IEEE INFOCOM, Hong Kong, March 2004.
52. A. Kansal, E. Yuen, W. J. Kaiser, G. J. Pottie, and M. B. Srivastava. Sensing uncertainty
reduction using low complexity actuation. In Proceedings of Information Processing in
Sensor Networks (IPSN 2004), Berkeley, CA, April 2004.
53. J. Wu and S. Yang. SMART: A scan-based movement assisted sensor deployment method
in wireless sensor networks. In Proceedings of the IEEE Annual Joint Conference of the
IEEE Computer and Communications Societies (INFOCOM’05), Miami, FL, March 2005.
54. G. Wang, G. Cao, and T. La Porta. Proxy-based sensor deployment for mobile sensor
networks. In Proceedings of the 1st IEEE International Conference on Mobile Ad-hoc and
Sensor Systems (MASS), Fort Lauderdale, FL, October 2004.
55. A. Kansal et al. Controlled mobility for sustainable wireless sensor networks. In Proceedings of IEEE Sensor and Ad Hoc Communications and Networks (SECON), Santa Clara,
CA, October 2004.
56. N. Heo and P. K. Varshney. A distributed self spreading algorithm for mobile wireless sensor
networks. In Proceedings of IEEE Wireless Communications and Networking Conference,
WCNC 2003, New Orleans, LA, March 2003.
57. A. Kansal, E. Yuen, W. J. Kaiser, G. J. Pottie, and M. B. Srivastava. Sensing Uncertainty
Reduction Using Low Complexity Actuation. In Proceedings of Information Processing
in Sensor Networks (IPSN 2004), Berkeley, CA, April 2004.
266
NODE POSITIONING FOR INCREASED DEPENDABILITY OF WIRELESS SENSOR NETWORKS
58. K. Dasgupta, M. Kukreja, and K. Kalpakis. Topology-aware placement and role assignment
for energy-efficient information gathering in sensor networks. In Proceedings of 8th IEEE
Symposium. on Computers and Communication (ISCC ’03), Kemer-Antalya, Turkey, July
2003.
59. M. Younis, M. Bangad, and K. Akkaya, Base-station repositioning for optimized performance of sensor networks. In Proceedings of the IEEE VTC 2003—Wireless Ad hoc, Sensor,
and Wearable Networks, Orlando, FL, October 2003.
60. K. Akkaya, M. Younis, and M. Bangad. Sink repositioning for enhanced performance in
wireless sensor networks. Computer Networks, 49(1):512–534, 2005.
61. K. Akkaya and M. Younis. Relocation of gateway for enhanced timeliness in wireless
sensor networks. In Proceedings of the IEEE Workshop on Energy-Efficient Wireless Communications and Networks (EWCN 2004), in conjunction with the 23rd IEEE International Performance Computing and Communications Conference (IPCCC’04), Phoenix,
AZ, April 2004.
62. W. Youssef, M. Younis, and K. Akkaya. An intelligent safety-aware gateway relocation
scheme for wireless sensor networks. In Proceedings of the IEEE International Conference
on Communications (ICC 2006), Istanbul, Turkey, June 2006.
63. J. English, M. Wiacek, and M. Younis. CORE: Coordinated relocation of sink nodes in
wireless sensor networks. In Proceedings of the 23rd IEEE Biennial Symposium on Communications (QBSC 2006), Kingston, Ontario, Canada, May 2006.
CHAPTER 10
Mobility in Wireless Sensor Networks
STEFANO BASAGNI
ECE Department, Northeastern University, Boston, MA 02115
ALESSIO CAROSI and CHIARA PETRIOLI
Dipartimento di Informatica, Università di Roma “La Sapienza,” Roma 00198, Italy
10.1 WIRELESS SENSOR NETWORKS: GENERALITIES
AND APPLICATIONS
Technological advances as well as the advent of 4G communications and of pervasive
and ubiquitous computing have fostered a renewed interest in multihop (ad hoc)
communications [1]. In particular, the interest is in self-organizing wireless multihop
networks composed of a possibly very large number of nodes. These nodes can be
either static or mobile and are usually constrained as for the most critical resources,
such as power and computation capabilities.
Wireless sensor networks (WSNs) are a typical example of this kind of
networks [2, 3]. In this case, the well-known paradigm of ad hoc networking specializes to consider the following characteristics.
Mobility. Whereas mobility is a fundamental aspect of all nodes in an ad hoc
networks, mobility in WSNs pertains mostly to a subset of the network elements,
and it is more specifically application-dependent.
Volume of Nodes. WSNs are usually comprised of a higher number of nodes (in
the thousands) rather than the few hundreds that are typical of ad hoc networks.
Resource Availability. Even if an ad hoc node is portable, power and computational resources are not usually crucial elements. Laptops and PDAs have
rechargeable batteries, have plenty of volatile and durable memory, and run the
same software available for bigger, static computers. WSN nodes are heavily
resource-constrained. Sensor nodes are usually unreplaceable, and they become
unusable after failure or energy depletion. Available memory is in the order of
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
267
268
MOBILITY IN WIRELESS SENSOR NETWORKS
KBs, and computational capabilities are limited. This imposes custom operative
systems, reduced protocol complexity, and highly specialized software.
Data Communications. Ad hoc data communication is typically peer-to-peer. One
node that wishes to communicate with one or more nodes does so using one
of the many available protocols. In WSNs the emphasis is on data transport
from the sensor to specific data collection nodes (sinks). Nodes are informed
about events of interest to the users. When such an event is detected by the
sensor, a corresponding data packet is created and sent to the sinks possibly via
a multihop route (data routing).
Given these very specific characteristics of WSN, it is crucial to devise WSN
protocols for topology organization, interest dissemination, and data routing that
are energy-conserving, scalable, and able to prolong the overall network longevity,
especially in networks with a large number of devices. Differently from what is
commonly understood and found in the WSNs literature, such protocols often have
to deal with the mobility of some of the network components. This is the case of
applications recently been proposed, which include:
r Underwater monitoring, such as submerging a network of sensors in an ocean
bed to detect debris from plane crashes for recovery and identification purposes.
In this case, beyond (partially) mobile sensors, the network comprises unmanned
or autonomous underwater vehicles (AUV) that are sent roaming through the
network for data collection [4].
r Networks in support of mobile small-scale robot squads that coordinate for performing a common task. While performing their operations, the robots exchange
information among each other and/or transmit collected measures to remote centers by using the network [5, 6].
r Sensor networks for collecting information about the location of a user/piece
of equipment, and so on, like tracking objects as well as humans in a hospital,
items in a warehouse, and so on [7].
r Enabling autonomous albeit controlled living of elderly people (independent
assisted living) through e-health systems made up of wearable sensors as well as
telemedicine equipment, home networking (domotics) techniques, and sensorbased smart spaces [8].
All these applications involve mobility: Sensor nodes can move (e.g., assisted
living through wearable sensors, tracking, etc.) as well as the data collectors (robots
or AUVs).
In this chapter we motivate and illustrate the use of mobility in WSNs. Furthermore,
we demonstrate how mobility is actually advantageous for improving the performance
of these networks.
WIRELESS SENSOR NETWORKS: GENERALITIES AND APPLICATIONS
269
10.1.1 Motivating the Use of Mobility in WSNs
Most of the research in WSNs concerns networks whose nodes do not move and
cannot be replaced. Nodes sense events of interest, and some energy-efficient routing
protocol is used for delivering the sensed data to static sinks. In this scenario, it
has been observed that the nodes that more than all the others have their energy
drained from data communication are those closer to the sinks. These nodes relay
data for all the other nodes in the network as well as packets from the sinks to the
sensors. As a consequence, nodes that are closer to the sinks soon “die” from energy
depletion resulting in the disconnection of the sinks from the rest of the network.
The problem of energy drainage at the sink neighbors is referred to as the “sink
neighbors problem.” The consequences of this problem are illustrated in Figure 10.1,
which shows the average nodal residual energy when the nodes closest to the sink
die. The figure refers to networks with 400 homogeneous nodes, each with a short
transmission range (25 m), placed on a 20 × 20 grid. The static sink is (optimally)
located at the center of the deployment area. Packets are periodically sent to the sink
by using a shortest path-like multihop routing protocol [9, 10]. The picture shows the
quite uneven node average residual energy (percentage of the nodal initial energy) at
the time the four sensor nodes that relay packets to the sink die, leaving the sink unable
to receive any more data from the network. The remarkably high variance among the
residual energies is due to the different distance of each node from the sink and, in
general, to the different number of sensor-to-sink routes to which a node belongs,
which implies different number of packets to relay. Nodes along the “cross” centered
at the sink tend to be the preferred data relays. The closer these nodes are to the sink,
the higher the number of packets they receive and transmit, and consequently the
higher their energy consumption. These are the nodes with the lowest residual energy
in the figure. In particular, when the nodes around the sink die, the energy at the nodes
along the cross arms averages at 71.07% of their initial energy, while 42.75% of the
100
90
80
70
60
50
40
30
20
10
0
Figure 10.1. Node residual energy (%).
270
MOBILITY IN WIRELESS SENSOR NETWORKS
network nodes have more than 95% of their initial energy available! This incapacity
of balancing node energy consumption results in short network lifetime (the time
until packets are no longer deliverable to the sink) and inefficient use of available
resources. The sink is soon disconnected from the network, while a large number of
the deployed nodes are still fully operational!
This chapter explores solutions for mitigating the sink neighbors problem illustrated above, by exploiting the mobility of some of the network components, thus
improving network performance.
Given the typical mobile nature of general ad hoc networks, the impact of mobility
on their performance has been extensively explored [11–19]. The primary objective
of these works is that of exploiting the mobility of some of the nodes for message
delivery in disconnected ad hoc networks, for improving network throughput, and for
studying mobility-assisted routing in general.
This chapter concerns approaches designed specifically for WSNs, and in particular
those solutions where the mobility of some network elements brings considerable
improvement in key performance metrics that are typical of sensor networking. In
the rest of the chapter we describe research on mobility in WSNs proposed recently.
In particular, we will survey works where (a) the sensor nodes can move, so that
batches of “fresher” nodes are kept close to the sink for providing uninterrupted
data forwarding; and (b) relays are sent throughout the network to collect data and
bringing them to the sink. Finally, (c) we will describe works where the sink itself
moves to collect sensed data. The last two approaches appear to be the more promising
for energy efficiency and longer network lifetime, since sink and relays are usually
considered resource-rich. Therefore, energy consumption and network lifetime are
not impacted by the energy needed to move them. In the case where the sensor nodes
move, a great deal of the nodes’ energy is spent on the movement itself, thus having a
detrimental impact on the lifetime of the nodes. The final part of the chapter concerns
a quite thorough discussion on how to compare different approaches to mobility in
WSNs, focusing on sink and on relay mobility. In particular, we show how, by using
simulations, one can effectively assess pros and cons of different solutions and gain
useful insights for defining and testing new ones.
10.2 SENSOR NODES MOBILITY
Mobility of the sensor nodes has been exploited for improving, or enabling altogether,
sensing and communication coverage [20–22]. The idea presented in reference 20 is
to have the sensors move into positions that minimize the energy cost of reporting
streams of data to the sink, which is statically placed. The protocols proposed by
Wang et al. [21] aim at moving mobile sensors from densely deployed areas to areas
with coverage holes, where for some reasons a limited number of sensors have been
deployed. The three protocols are proven by simulations to be effective in terms
of coverage, deployment time, and moving distance. Minimization of the energy
consumption of moving nodes has been subsequently addressed by the authors in
reference 22 by letting the node “move logically”—that is, only after they can decide
MOBILE RELAYS
271
whether their moving maximizes the coverage or not. The idea of using mobile sensors
has also been explored within the robotic community, where mobile robots are (also)
equipped with sensors. Typical examples are the works by Howard et al. [23, 24].
In reference 23 an algorithm for the deployment of the members of a robotic team
into an unknown environment is given. The aim of this algorithm is the maximization
of the coverage area, while maintaining line-of-sight contact among the robots. In
reference 24 the same authors draw from the theory of potential fields to distribute
the mobile sensors throughout a given area. The fields are constructed in such a way
that each node is repelled by obstacles and other nodes, thereby forcing the node to
spread throughout the area.
Finally, distributed algorithms for the mobility of sensor nodes have been investigated in reference 25. In this work, mobility algorithms are proposed that move
the nodes to positions that reduce the transmission power needed to send the data to
the (static) sink. The positions for the moving sensors are determined via distributed
simulated annealing, as opposed to a greedy strategy that could lead to a suboptimal
placement. By using distributed simulated annealing, a node based only on information on its current neighbors accepts a “bad move” with a positive probability. This
move could be locally nonoptimal, but could benefit the network globally in the longer
run.
Works that consider mobile sensors and robots are mostly concerned with sensor
deployment time and with sensing coverage. The costs associated with sensor movements as well as the cost of transmitting sensed data are often not considered, and
network lifetime is rarely a metric of interest.
10.3 MOBILE RELAYS
Chatzigiannakis and Nikoletseas [26] explore the possibility of using the coordinated
motion of a small number of users in the network for efficient communication between
any pair of other mobile nodes. Some of the nodes act as mobile relays in that they
carry packets for other nodes. Packets are exchanged when the source node and the
relay are neighbors (namely, they are in the radio vicinity of each other), and they are
then delivered to the intended destination when the relay passes by it.
This is basically what has been introduced to WSNs by Shah et al. in their works
on data MULEs [15, 27]. The MULEs are mobile nodes roaming among the sensors
to act as forwarding agents. Energy conservation is possible because of single-hop
communications (from a sensor to the MULE that is passing by) rather than in a
multihop way (from the sensor to the sink). The packet now reaches the sink when
the MULE eventually passes by it and transfers all collected sensed data to it. MULEs
are effective for energy conservation in the so-called delay tolerant networks [28].
Energy is traded off for latency; that is, the energy needed to communicate a packet
to the sink is decreased at the cost of waiting for a MULE to pass nearby and at the
cost of waiting for the MULE to move to the vicinity of a sink. Other problems, such
as scheduling sensor-to-sink transmissions within this model, have been studied in
reference 29.
272
MOBILITY IN WIRELESS SENSOR NETWORKS
Unmanned vehicles for data collection have been further investigated in
reference 30. Nodes send their packets to nearby cluster heads via multihop routing. The moving relays then pass by the cluster heads to collect the data according to
some predefined schedule. Therefore, the collector has to visit only the cluster heads
and not all the nodes. Furthermore, multihop routing is reduced to a smaller number
of hops since the data are sent from a sensor node to a clusterhead who is nearby. Data
delivery to the static sink occurs periodically, when the collectors return to the sink
to drop their packets and recharge. This concept and architecture has been further
explored in reference 31. In this case the authors consider different classes of nodes,
where the collectors roam (controllably) among the nodes, grouped into clusters, and
can be (uncontrollably) mobile. The goal here is that of determining the schedules
of the collectors visits to the nodes that minimize transmission energy consumption,
data latency, and nodal buffer requirements.
A first discussion on how to include controllable mobile relays into the network
has been presented in reference 32. In the paper, an implementation of a sensor network with an autonomous mobile relay (a robot) is presented. The robot visits the
(static) sensors, collects their data (single-hop exchange), and delivers them to the
sink, similarly to what happens for the MULEs. However, in this case the movements
of the robot adapt to the network application priorities, which dictate data collection
performance parameters. In other words, the robot is part of the system, and it is
the system that controls its mobility. The testbed-based experimental results concern
the evaluation of methods for controlling the speed of the robot for optimizing data
collection. The robot traverses networks with different densities following a straight
trail and collects the data. The data are then delivered to the sink. Methods are defined
for routing the sensed data to nodes that are one hop from the robot route when the robot
itself does not pass sufficiently close to some sensors. Further development of this
work with multiple controlled mobile elements (here called explicitly data MULEs)
has been presented in reference 33. This work considers the two cases where nodes
are deployed uniformly and randomly in a given geographic area and when, more
realistically, they are distributed differently. In the first case, criteria are given for the
choice of the number of MULEs and for dealing with nodes that can be served by
multiple mules. In the case of nonuniform nodal distribution, a load balancing algorithm is introduced for distributing the number of sensor nodes to the various MULEs
so that each MULE serves approximately the same number of sensors. MULEs roam
through the network in straight lines and gather information about the nodes they
can reach. Then the MULEs, which can wirelessly talk to each other directly, elect
a leader and send the information they gather to it. Based on this information, the
leader executes the load balancing algorithm and associate nodes to MULEs. Data
collection is finally performed by the MULEs that travel through the network (in
straight lines) and explicitly poll the assigned sensor nodes for collecting their data.
The problem of scheduling the visit to the sensor nodes of a single relay (called ME,
for Mobile Element) so that there is no data loss (due to buffer overflow) is tackled
in reference 34. The corresponding Mobile Element Scheduling (MES) problem is
proven to be NP-complete, and centralized analytical model (ILP) and algorithms are
given for solving the problem. In particular, given as input the data generation patterns
MOBILE SINKS
273
(sensing rates) at the sensors, the corresponding buffer overflow times, and a matrix
cost for the ME movements, the model and algorithms determine the sequence of
visit to the nodes so that none of the buffers overflow. General discussions on the
advantages and challenges of controlled mobility are conducted in references 35 and
36. In the first paper a new approach to exploiting mobility, termed Morph, is defined, and the benefits to sensor networks sustainability are shown and discussed. In
reference 36, considerations about the use of controllable mobility are more general
and are followed by a detailed description of how mobility induces improvements in
WSNs performance.
The problem of controlled data MULE-like mobility has been also recently addressed in reference 37. The authors first propose an algorithm for avoiding sensor
nodes buffer overflow while minimizing the speed of the mobile relay. They also
extend this algorithm to the case where some of the packets have delay constraints
(i.e., they are “urgent messages” that have to be delivered to the sink within a given
time since their generation). Finally, an investigation of the controlled use of relay
nodes for data collection and subsequent report to the sink has also been proposed
in reference 38. Although the authors recognize that moving the sink directly yields
better resource utilization and hence longer lifetime, they argue that for certain applications, moving the sink is infeasible. Therefore, having one or more resource-rich
mobile relay nodes is remarkably helpful. Given that the sensor nodes know about
the current location of the relay node, routing protocols are presented for delivering
the data from the sensors to the relay and from the relay to the sink and, finally, for
determining the route of the relay. Improvements on network lifetimes are fourfold
with respect to the case of a static sink.
10.4 MOBILE SINKS
A data dissemination protocol, termed SEAD (Scalable Energy-Efficient Asynchronous Dissemination), is defined by Kim et al. [39] where a tree-like communication
structure is built and maintained. Differently from previous data centric solutions à la
directed diffusion [40], which also use tree structure for data and interest gathering and
dissemination, according to SEAD the sink moves (randomly) to certain sensor nodes
in the tree (access points) from which it collects the data. Communication between
the sink and the access points can be multihop. This happens when the sink moves
away from the access points. The tradeoffs that SEAD attempts to obtain concern data
latency and the energy needed for tree reconfiguration so that the access points are
closer to the current position of the sinks. In this way, SEAD is shown to outperform
other solutions for data dissemination in WSNs where the sink does not move, such
as directed diffusion [40], TTDD [41], and ADMR [42].
Building and maintaining routes to a mobile sink is the topic of references [43–
45]. The aim is that of minimizing these operations overhead. In the first paper,
local update techniques are described for detecting disconnections and perform route
repair in “sink-oriented trees.” The ERUP protocol is proposed in reference 44 for
conducting route rediscovery only in the vicinity of the damaged route. In reference
274
MOBILITY IN WIRELESS SENSOR NETWORKS
45, initial routes are constructed from the nodes to the sink according to any viable
WSN routing. Given that the sink is moving, if the routes are no longer valid, forwarder
nodes are designated to extend the current routes.
All these works have in common that the mobility of the sink is unpredictable
and references uncontrollable. For example, in references 39 and 46, sinks move
according to the random waypoint model.
The use of mobile sinks with predictable mobility has been more recently presented in 47–49 and 50. In these works the sinks (airplanes) fly over the sensor
field and gather the sensed data periodically. While the movement of the sink is
fully controllable, it is external to the network infrastructure; that is, the trajectories
are not determined by network components and activity. The main contribution of
these papers is the energy-efficient transmission to the passing sink [47, 48, 51]. In
reference 49 the authors consider heterogeneous sensor networks made up of two
types of nodes. Type 0 nodes do the basic sensing, perform short-range communications, and are partitioned into clusters whose cluster head is a type 1 node. The
cluster heads take care of receiving data from type 0 nodes (possibly through multihop routes), do some sensing, aggregate data, and perform long-range transmissions
to the aircraft. Each passing of the aircraft triggers a new data sensing and data collection cycle. The aim of the paper is to determine the optimum node deployment (i.e.,
the densities of each type of nodes) and the nodal energy needed to achieve a given
network lifetime (numbers of data gathering cycles) while ensuring sensing coverage
and radio connectivity with high probability.
Inherent patterns of the sink movement are exploited in reference 52 for the
design of robust and energy-efficient routing. This work assumes that there is a certain degree of predictability in the sink movement, such as the routine route of a
ranger patrolling a forest. Sensor nodes learn about the sink whereabouts at given
times via statistics techniques as well as methods from distributed reinforcement
learning.
In reference 50 a model for sink movements is presented where the sinks, called
“observers,” move along the same route repeatedly and collect data from the sensors. The sensed data are collected while the observer traverses the network. In
particular, when passing by sensor nodes, the observer wakes them up and receives their data (if any). The authors describe a prototype system developed at
Rice University where the observers are carried by shuttles, and the sensors are
spread out throughout the campus. In particular, the authors determine the transmission range needed to collect data from a predefined percentage of the sensor
nodes, given the observer speed, the time required to transmit a packet, and different
traffic patterns. The correlation among the various system parameters is investigated
analytically.
Network-controlled sink mobility for reducing energy consumption and for maximizing the lifetime of a sensor network has been considered initially in reference 53
and then in references 54–56, 9 and 10. In these works the sink moves among the
(static) sensor nodes and, while sojourning at given locations, collects data that are
sent to it via multihop routes. The first work concerns minimizing energy consumption. An ILP model is presented that determines the locations of multiple sinks as well
MOBILE SINKS
275
as the routes from the sensors to the sinks. Time is divided into rounds. At the beginning of each round, information on the nodes’ residual energy is centrally gathered
and the ILP problem is solved for finding the new locations for the sink that minimize
the maximum energy consumption spent at the nodes during that round. Minimizing
the energy consumption results in increased network longevity. No constraints are
enforced on the sink movements, and there is no relation between the number of the
sinks and their position in subsequent rounds: The model is solved from scratch every
time.
The problem of network lifetime maximization through controlled sink mobility
is explicitly addressed in reference 54 for networks with a single sink. Sink locations
and sink sojourn times at those locations are determined that maximize the network
lifetime via a new LP formulation of the problem. Maximizing the network lifetime
is maximizing the sum of sojourn times of the sink at the visited locations (in this
case they coincide with the sensor locations). The experiments in reference 54 refer
to scenarios where n = L2 nodes are arranged in a L × L grid. The sink has no
limitation on the time tk ≥ 0 it can spend at sensor k and can move from any location
to any other location in the network. The network lifetime is improved by a factor
of 5. This happens when the sink sojourns at the nodes located at the four corner areas
and in the central area of the grid.
Papadimitriou and Georgiadis [55] present another centralized solution for the
problem of maximizing network lifetime by combining the model presented in reference 54 and the LP formulation for maximum lifetime routing described in reference
57. A constant of the model in reference 54 is here turned into a variable so that the
resulting model presented solves both (a) the problem of determining the sink sojourn
times at the given sites and (b) the routing of the packets to the current position of
the sink. This routing-dependent solution achieves improvements with respect to the
lifetime values of reference 54 that are twofold.
Lifetime maximization as a min–max problem is the idea explored by Luo and
Hubaux [56]. The authors consider together sink mobility and data routing. A load
balancing solution is presented that, while keeping the sink moving along the external
perimeter of the network, achieves lifetimes 500% higher than when the sink stays
still in the center of the network.
These centralized solutions concern how to drive the sink to places in the
network where data collection has the least impact on the sensor nodes’ residual energy. However, several aspects of sensor networking (such as the possibility of collision and corresponding energy cost) are not considered. Similarly, the
cost of building, maintaining, and releasing routes to the current position of the
sink and its impact on network lifetime and other metrics is never taken into account. A Mixed Integer Linear Programming (MILP) model has been presented
in references 58 and 59 and in more detail in references 9 and 10, where more
realistic data communication costs and constraints of sensor networking and on
sink mobility are explicitly considered. Moreover, those papers introduce and evaluate the first distributed and localized heuristic for controlled sink mobility. The
MILP model and the distributed solution are described thoroughly in the next
section.
276
MOBILITY IN WIRELESS SENSOR NETWORKS
10.5 COMPARING DIFFERENT APPROACHES TO MOBILITY
The most promising use of mobility for enhancing the performance of WSNs has
been shown to be obtained when resource-rich network components move—that is,
when either the sink itself or mobile relays are left free to roam through the network
for collecting data.
This part of the chapter is dedicated to describe and discuss how to compare two
very different approaches to mobility in WSNs. Our aim is twofold. First of all,
we want to gain an understanding of which among the many proposed solutions are
the most promising ones in terms of usage and performance enhancement. More
than specific mobility protocols, we are interested in determining which network
architecture is the best for supporting efficiently the application listed above. Then, we
want to establish clear metrics and tools for comparing mobility solutions and obtain
from this performance comparison new insights for protocol design and effective use
of mobility.
By the study of what proposed in the literature, we could observe that when
considering mobility of sinks and relays, the major difference in performance is
made by whether routing between the sensor nodes and a sink/relay is single-hop or
multihop. This choice for the routing greatly affects essential metrics such as nodal
energy consumption, packet latency, and probability of successfully delivering data to
the sink/relay. It has also immediate consequences on the design and cost of the wireless sensor node: When a multihop data routing protocol needs to be implemented, a
node has to provide resources (storage, computation, etc.) for enabling the forwarding
of packets to the data collector; that is, a protocol stack has to be available. When
routing is single-hop, instead, a simpler nodal architecture suffices for storing the
packets and for transmitting them to a relay when it passes by.
In the following we describe and investigate two network architectures that clearly
show the impact of choosing multihop versus single-hop routing when sinks or relays roam through the network. We consider single-hop routing according to the
data MULE approach. We will consider the original model [27], which exploits the
uncontrolled mobility of multiple relays, and more recent works where it is shown
that the relay mobility can be controlled by the network current conditions [33]. We
then show the pros and cons of multihop routing by describing an analytical model
and distributed heuristics for sink mobility. As in the previous case, we consider
and compare solutions where the mobility of the sink is controlled by the network
itself with those where sink mobility does not depend on what is going on in the
network [9, 10].
10.5.1 The Data MULE Approach to Mobility in WSNs
The idea of a Mobile Ubiquitous LAN Extension (MULE) [15, 27] stems from that
of deploying mobile agents for carrying information [60], especially among nodes in
possibly disconnected networks. The concept is quite simple: MULEs are resourcerich nodes that roam freely throughout the network. When as a result of its motion
a MULE gets to be in the radio proximity of a wireless sensor node, it receives all
COMPARING DIFFERENT APPROACHES TO MOBILITY
277
the packets generated by that sensor so far (if any). Upon getting close to one of
the network access points (i.e., the sinks), the MULE transfers to it all the collected
packets. In this way, whether there is a route from a sensor to one of the sinks or not,
(i.e., independently of network connectivity), a packet is eventually delivered to the
sink. The key concept here is that sensor-relay routing as well as relay-sink routing
is single-hop.
The data MULE architecture is made up of three layers. The highest layer comprises
one or more access points—that is, the sinks. These are the resource-rich, static
components to whom the MULEs deliver the data collected throughout the network.
The middle layer is the set of the MULEs. These are nodes characterized by mobility,
large storage, and renewable power and by the ability to communicate with the sensor
nodes as well as with the sinks wirelessly. Most importantly, MULEs movements
cannot be predicted in advance and are considered completely random (uncontrolled
mobility). Examples of MULE nodes are (a) animals roaming in a given area (as in
an environmental monitoring application [11]) and (b) vehicles traveling in a city or
university campus. Finally, the bottom tier is comprised of a possibly large number
or resource-constrained static sensor nodes that can communicate with a MULE that
is passing by.
Communication in the MULE architecture is single-hop: A node stores the sensed
data until a MULE passes by. Once the MULE arrives, the node transmits all the
stored packets to the MULE. When the MULE passes by the sink, it transmits to it
all the packets it has collected.
Clear advantages of the MULEs approach to data gathering and delivery in wireless
sensor networking include the following.
r Robustness. No node depends on any single MULE. Any of the roaming MULEs
would do. If one of the MULEs fails, a node can still count on another MULE
to pick up its packets.
r Lower Complexity. The one-hop nature of the sensor-to-MULE communication
allows very simple nodal protocol stack. Unsophisticated MAC is often enough
to guarantee safe transfer of data packets from their source to the passing MULE.
r Decreased Energy Consumption. The typical overhead associated with multihop
ad hoc routing is not present in the MULE system. This overhead is oftentimes
the major culprit of energy consumption in WSNs.
However, the flip side of this quite simple and effective system presents nonnegligible drawbacks. The first that comes to mind is the average latency incurred
by the packets from the time they are created to when they are delivered to the
sink. Given the serendipitous nature of the MULEs movements, there is no guarantee
on deterministic delay bounds. It is not possible to surely state when and where a
MULE will show up close to a sensor or to the sink. Therefore, a packet can sit in
a sensor queue, or in the queue of a MULE for quite a long time, which imposes
very high overall sensor-to-sink latency. The uncontrolled MULE mobility has also
a detrimental effect on the packet delivery ratio; that is, the ratio between all packets
278
MOBILITY IN WIRELESS SENSOR NETWORKS
that are created in the network over those that are successfully delivered to the sink.
Sensor nodes, having limited storage, can buffer only a limited amount of packets. If
the inter-arrival time of a MULE in the radio vicinity of a sensor is too large, chances
are that the sensor is forced to discard all the packets in excess of its buffer size. This
forces down the amount and type of applications that could use the MULE architecture
for data gathering and dissemination. Only applications that are delay-tolerant can
endure the possibly very high latencies imposed by the MULE system.
Recent studies have shown that it is possible to mitigate the drawbacks of the MULE
architecture by having the relay movement guided by the network itself—that is,
by controlling the mobility of the MULEs. One of the first attempts is described by
Kansal et al. [32]. As mentioned earlier, this work has quite an experimental flavor
and is concerned with a network where one MULE (a mobile robot) travels along a
straight trail up and down the sensor node deployment area to collect data. Variation
in the speed of the MULE is determined by the network load: The robot will go slower
if more data are to be transmitted by the nearby sensors, but will go faster otherwise.
Since the MULE’s route is always the same straight line, it is highly likely that
some sensors are not able to transmit their data to the MULE directly. For this reason
the solution proposed in reference 32 organizes the network operation into two parts.
The first, named the network algorithm, takes care of how the sensor nodes interact
with each other so that packets from those nodes that are not sufficiently close to the
passing MULE are delivered to it by nodes that are close. The second part specifies how
the MULE moves, and it is called the motion control algorithm. Here is a description
of the two parts.
The Network algorithm defines the way in which those nodes that cannot directly
transmit to the sink find data routes to send their packets via intermediate nodes. The
network algorithm is performed in three steps: initialization, local multihop, and data
collection.
1. Initialization. During the first run through the network deployment area, along
its only route, the MULE beacons a simple “hello”-like packet. All nodes that
receive this packet are neighbors to the passing MULE, set a hop-count variable
to 1, and broadcast the hello packet further, updating the hop-count field to its
own. A node that receives multiple hello messages from neighboring nodes
updates its own hop-count variable to the least values received plus 1, and it
keeps forwarding the message (with the updated hop-count). At the same time,
it records the node closer to the MULE path that sent the hello packet with
the least values (possible ties are easily broken). In this way each node become
aware of (a) its hop distance from a node that can directly transmit to the passing
MULE and (b) a route to that node. Overall, a forest (set of disjoint trees) is built
as a first step of the network algorithm, with the nodes closer to the MULE’s
route as the roots.
2. Local Multihops. Data packets are sent at the root nodes from any sensor
node in the corresponding tree. Any suitable routing algorithm (e.g., directed
diffusion [40] or a simpler “Convergecasting,” such as the one defined in
reference 61) can be used for implementing this step.
COMPARING DIFFERENT APPROACHES TO MOBILITY
279
3. Data Collection by the MULE. After its first passage through the network, the
MULE keeps moving up and down its route collecting data from the nodes
along its path. Nodes polled by the passing MULE transmit their packets along
with those of the nodes in their trees.
Motion Control Algorithm. Since the route of the MULE is fixed, motion control
reduces to speed regulation. This is done by fixing the round trip time (RTT) of the
MULE. Then, three possible strategies are described in reference 32 that correspond
to different strategies of MULE movements.
1. The path is traversed at a certain fixed speed. Let us assume that at this speed
the MULE takes T < RTT to travel its route (and back). The remaining
“spare time” RTT-T is divided equally among the nodes along the MULE’s
route, and the MULE stops at each node for this time. The MULE might not
know about the node location. Therefore, the MULE stops as soon as it hears
from the node.
2. The MULE travels the route at the constant speed of (length of route)/RTT and
does not stop at the nodes.
3. The speed of the MULE adapts to the current situation of (the queue of) the
nodes. This can be realized by having the MULE traveling at the speed 2∗(length
of route)/RTT. The extra amount of time is divided among those nodes from
which the MULE has collected a limited amount of data (determined by a given
threshold) in the previous round. (The set of nodes the MULE stops at changes
at each round.)
Further details about the described algorithms as well as its implementation on a
real-hardware testbed can be found in reference 32.
Here we describe the natural evolution of this single-MULE solution to controlled
mobility in the “mostly single-hop” routing realm: multiple controlled mobile elements (data MULEs) for data collection [33]. The motivations are clear: The single
MULE approach does not scale well. If the number of nodes in the network increases,
there are more nodes from which data must be collected in the same amount of time.
Therefore, packets might be dropped (buffer overflow) because the MULE cannot
get to visit certain nodes on time. Deploying multiple MULEs solves this scalability
problem. If the nodes are scattered randomly and uniformly throughout the deployment area, the obvious solution consists in dividing the area into same-size parts
and having a MULE in each of these parts. Given the deployment distribution, each
MULE would have roughly the same amount of nodes to serve. In this case, there are
two main problems to deal with: deciding the number of data MULEs and deciding
what to do with nodes that are shared by two MULEs.
r Number of Data MULEs. The number of data MULEs needed to collect data
while avoiding packet loss is computed based on the RTT of a MULE (defined
as the time it takes to the MULE for traversing the assigned area along a straight
line) and on the buffer fill time, which is the time needed to a node to fill its
280
MOBILITY IN WIRELESS SENSOR NETWORKS
buffer. The RTT is defined as follows:
(l/s) + n × st + (l/s)
where l is the side of the deployment area (a square), s is the MULE speed, n is
the number of nodes in the network, and st is the service time—that is, the time
needed to transfer packets from a node’s buffer to the servicing data MULE. It
is assumed that the MULE stops at each node while traversing the networks for
st time. This happens only one way. Once the MULE has reached the end of
the line, it goes back without collecting data. In the case where there are nodes
that are not in the transmission range of any of the MULEs, multihop routing
is used to send the packet to a node (root of the data distribution tree, as in the
single-MULE case described above) close to the MULE route. In order not to
lose packets because of buffer overflow, the number of data MULEs required
will be
RTT
buffer fill time
(One MULE suffices when RTT ≤ buffer fill time.)
r Sharing Nodes. Whenever a node can be serviced by two MULEs, most likely
because it is equidistant (in hops) from the MULEs routes, ties are broken randomly or by simple techniques. For instance, the node might decide to send one
packet to one MULE and the next one to the other MULE, or it may just flip a
coin every time it has a packet to send, and decide which MULE is going to get
the packet accordingly.
Deployment of multiple MULEs is also described in case nodes are not scattered
randomly and uniformly. This can happen, for instance, when nodes are manually
placed in specific places, or when the terrain induces a nonuniform nodal distribution. The MULEs move along straight lines that are not necessarily equally spaced
throughout the area. Assuming that each node is at most one hop from a MULE’s
route (i.e., that every sensor is directly served by at least a MULE), nodes are divided
into two classes: nonshareable (NS) nodes and shareable (SH) nodes. The first set
comprises those nodes that can be served only by one of the MULEs. The set SH
instead contains nodes that are in the transmission range of multiple MULEs (at least
two). For NS nodes there is no option other than to be served by the only MULE that
passes close to them. Shareable nodes should instead be assigned to only one MULE.
The challenge here is to define a way to assign nodes in SH to the MULEs so that the
number of nodes served by each MULE is roughly the same. The algorithm for data
gathering in this case is made up of five phases: initialization, leader election, load
balancing, node assignment, and data collection. The following is a brief description
of the algorithm (details can be found in reference 33).
COMPARING DIFFERENT APPROACHES TO MOBILITY
281
r Initialization. The MULEs go on their first trip beaconing around their presence.
Nodes that receive this beacon reply back with their ID. By the end of this phase,
the MULEs are back at their starting points and know of all the nodes they can
serve.
r Leader Election. All the MULEs are assumed to be able to communicate with
each other (i.e., at least when at their initial position, the topology of the “MULE
network,” is a clique). One of the MULEs is elected as the leader of the bunch
(for instance, the MULE with the smaller ID). At this point, every MULE sends
to the leader the information gathered in the first phase.
r Load Balancing. The leader divides the network nodes into shareable and
nonshareable. The shareable nodes are further classified according to the MULEs
that can serve them. This is the most delicate part of the algorithm, all performed
locally by the leader MULE. It is divided into two main parts: The first part takes
care of calculating the average number of nodes that could be assigned to each
MULE. The second takes care of the actual load balancing [33], determining the
nodes that could be assigned to each MULE.
r Node Assignment. As a result of the previous phase, some of the shareable nodes
can be served by two MULEs. In the node assignment, phase nodes that are shared
by two MULEs are ordered according to their unique ID. At this point the first
part of the set is assigned to the MULE of the two with the smallest ID, and the
remaining nodes are assigned to the other MULE. After the assignment has been
calculated, the leader MULE communicates to all the other MULEs the nodes
they have been given to serve.
r Data Collection. In the data collection phase the MULEs start their journey along
their routes, looking for data. At this time the nodes do not know the MULE
to which they are assigned; and hence when they hear the MULE beacon, they
respond with their packet. They will receive an acknowledgment only from the
MULE that is supposed to collect their packets. A node records this information,
and in the future it will respond only to its assigned MULE.
Simulation results are shown in reference 33 that demonstrate how effective this
algorithm is in distributing nodes to MULEs.
10.5.2 Multihop Routing and Controlled Sink Mobility
We consider here the case of a mobile sink traveling through the network and receiving
data at its current location via multihop routing. We start by describing in detail a
solution where the sink moves according to the current network condition (controlled
mobility), with the aim of balancing energy consumption among the network nodes
and thus prolonging the network lifetime.
This solution, presented in references 9 and 10, illustrates two fundamental research steps: mathematical modeling (for the centralized determination of an optimal
solution) and the definition of distributed heuristics. A mathematical model is defined as follows: Taking as input realistic parameters such as the cost of route setup,
282
MOBILITY IN WIRELESS SENSOR NETWORKS
maintenance, and release, the cost of transmitting and relaying packets to the sink,
and limitations in sink movement will determine where the sink should go and how
much it should stay there so that network lifetime is maximized. The solution, clearly
centralized, is based on a Mixed Integer Linear Programming (MILP) modeling.
Although MILP models are computationally hard to solve, the model defined in reference 10 is solvable for realistically large instance of the problem. The optimum sink
route determined by solving the MILP model is used for comparison with the route
determined by a distributed heuristic which is more suited for use in WSNs. The distributed heuristic, termed Greedy Maximum Residual Energy (GMRE), determines
sink movements and sojourn times by greedily sending the sink from one point to the
one surrounded by the nodes with the highest residual energy.
Here is the typical scenario. A large number |N| of resource constrained, static
nodes with sensing and wireless communication capabilities are scattered in a given
geographic area. Data packets are generated periodically at the sensor nodes: At a
given data rate ri , node i ∈ N transmits packets that are delivered to the sink for
processing.
While the nodes are static, the sink can be mobile. More specifically, a set S =
{1, . . . , q} of q sink sites is considered which are the points within the geographic
area the sink can visit. For instance, Figure 10.2 shows a typical scenario where
36 nodes (represented by circles) are placed on a 6 × 6 grid and 25 sink sites (squares)
Figure 10.2. Sensor nodes, sink sites, and sink movements.
COMPARING DIFFERENT APPROACHES TO MOBILITY
283
are organized according to a 5 × 5 grid. A link between two nodes indicates that those
two nodes are neighbors (i.e., they can hear each other’s transmissions). A lightercolored link between two sites indicates that the sink can move from one site to
the other and vice versa. Because of the sink neighbors problem, the sink moves
throughout the network in an attempt to balance the energy consumption among the
nodes. A problem that arises is that of how to let the nodes know what is the new
position of the sink. To this aim, when the sink arrives at a new site, it broadcasts
a packet f to all the network nodes, making them aware of its current site. When
a node has a packet to send, it now sends it toward the new site of the sink. Every
routing scheme that works with the topological information provided by f , such as
geographic or shortest paths-based routing, is a viable routing for data delivery to the
sink. The main advantages of routing independence are the following. It guarantees
the longest possible network lifetime given the specific routing. It allows the network
users to design or choose the routing algorithm that best meets the WSN application
requirements, which go beyond improving network lifetime and may consider other
metrics. Every time the sink leaves a site, it again broadcasts a packet to all nodes to
communicate that it has moved. Upon receiving this packet, a node stops forwarding
data (remaining packets are buffered) and waits to receive a new packet f from the
sink, carrying its new location.
While the sink is traveling the senser do not transmit and they buffer the packets.
The farther the sink travels, the more it takes to get to the new site and the higher
the delay suffered by the data. For containing this delay, a new parameter called
dMAX is introduced that represents an upper bound on the distance that the sink can
travel from a site to the following one. The pair (S, dMAX ) uniquely defines a graph
of sink sites where there is a link between two sites if and only if their (Euclidean)
distance is ≤ dMAX . Figure 10.2 shows the four sites (darker squares) the sink (the
triangle) can reach from its current position. The lighter lines between the sites of
Figure 10.2 indicate that the sink can only move horizontally or vertically in the 5 × 5
grid.
In the case of high rate sink mobility and low data traffic, the energy cost for route
construction and release can be significant. Therefore, this cost should be explicitly
taken into account. Finally, in order to evaluate the impact of different sink mobility
rates, the parameter tmin is introduced to represent a mandatory minimum time the
sink has to sojourn at a site. (High tmin values slow the sink down, while low tmin
values allow it to move faster.) The problem we set out to solve here is the following:
Determine the starting site and the route for the mobile sink over the graph (S, dmax ),
together with the sojourn times tk ≥ tmin of the sink at each visited site k ∈ S so that
network lifetime is maximized.
MILP formulation. Here are the sets, the parameters, and the variables used for
formalizing the problem.
r Let S be the set of sink sites—that is, the locations at which the sink may sojourn:
S = {1, . . . , q}.
284
MOBILITY IN WIRELESS SENSOR NETWORKS
r N is the set of the network nodes: N = {1, . . . , n}.
r e0 : Initial energy (Joules) of each node.
r bik : Energy consumption (Joules) at node i ∈ N for setting up/releasing routes
when the sink moves to site k ∈ S.
r cik : Power consumption (Watts) for receiving and transmitting packets at node
i ∈ N when the sink sojourns at site k ∈ S.
r tmin : Mandatory minimum time (seconds) for which the sink is required to stay
at site k ∈ S.
r djk : Euclidean or shortest path distance (meters) between any two sink sites
j, k ∈ S.
r dmax : Maximum distance (meters) the sink is allowed to travel each time it
moves.
r A: The set of directed edges joining sink sites whose distance is less than or
equal to dmax , that is, A = {(j, k) ∈ S × S : j =
/ k, djk ≤ dMAX }.
r O: The set of directed edges (0, k), k ∈ S, joining a fictitious site 0 (origin) with
the sites in S.
r D: The set of directed edges (k, q + 1), k ∈ S, joining the sites in S with a
fictitious site q + 1 (final destination).
r X: The union of A, O, and D.
The following variables are also considered:
r tk : Sojourning time (seconds) of the sink at site k ∈ S.
r zk : Binary variable taking the value 1 if the sink sojourns at site k ∈ S (tk > 0);
0 otherwise (tk = 0).
r xjk : Binary variable indicating the status of (j, k) ∈ X. xjk = 1 if and only if arc
(j, k) is on the sink movement route; xjk = 0 otherwise.
r vk : Auxiliary variable used to enforce a unique sink route.
The MILP formulation is defined by the following objective function and constraints.
Max
(10.1)
tk
k∈S
subject to:
k∈S
cik tk +
bik zk ≤ e0
(10.2)
k∈S
tmin zk ≤ tk ≤ Mzk
x0k = 1
k∈S
(i ∈ N)
(k ∈ S)
(10.3)
(10.4)
COMPARING DIFFERENT APPROACHES TO MOBILITY
xk,q+1 = 1
285
(10.5)
k∈S
xjk =
j∈S∪{q+1}
(j,k)∈O∪A
(k,j)∈A∪D
xjk = zk
(k ∈ S)
xkj
j∈S∪{0}
(k ∈ S)
(10.6)
(10.7)
j∈S∪{0}
(j,k)∈O∪A
vj − vk + qxjk ≤ q − 1
vk ≥ 0
zk ∈ {0, 1}
xjk ∈ {0, 1}
(k ∈ S)
(k ∈ S)
((j, k) ∈ X)
((j, k) ∈ A)
(10.8)
(10.9)
(10.10)
(10.11)
The objective function (10.1) maximizes the sink total time at sojourning sites,
k tk , which is the effective network lifetime.
Constraint (10.2) states that the combined energy
spent at node i for data
delivery
(
k∈S cik tk ) and data route construction
and release ( k∈S bik zk ) during k tk (the time before the death of the first node)
should not exceed the node’s initial energy e0 . The right part of double inequality (10.3) forces zk to take the value 1 if the sink sojourns at site k (tk > 0), thus
linking the binary variable zk [constraint (10.10)] with the continuous variable tk . M
is a significantly large number. The left part of double inequality (10.3) restricts the
sojourn time tk to be at least equal to the mandatory minimum sojourn time tmin if
the sink sojourns at site k (zk = 1) and at the same time forces zk to take the value
0 if the sink does not sojourn at site k (tk = 0). The first sojourning site in the sink
movement route is allowed to be any site in S. To implement this, a fictitious fixed
initial site 0 (origin) is introduced. At the beginning of the sensor network’s lifetime,
the sink moves in zero time (and cost) from the origin to some site α ∈ S, determined
by the model. This is that particular site such that x0α = 1 [Eq. (10.4)]; that is, it is
the optimum starting point of the sink journey. Then, the sink sojourns at that first site
and at subsequent other sites in S to be determined by the model. Finally, from the
last sojourning site ω the sink moves to a second fictitious site “q + 1” (destination),
again in zero time [Eq. (10.5)]. The site ω completes the sink route started at site α.
This is the last site at which the sink sojourns, and it marks the end of the sensor
network lifetime. The arcs (j, k) ∈ X on the sink route are associated with binary
variables xjk equal to 1. The variable xjk is equal to 0 for all the (j, k) ∈ X that do
not belong to the route. Equivalently, one can think of a unit of flow moving from
the origin to the destination. Constraint (10.4) induces a unit of flow from the origin
to some node α ∈ S, while constraints (10.5) cause the destination to absorb a unit
of flow coming from some node k ∈ S. Constraint (10.6) forces flow conservation at
all sites k ∈ S, thus ensuring the generation of a route. Constraint (10.7) ensures that
the sites k ∈ S on the generated route are sites at which the sink sojourns (k|zk = 1).
To elaborate, if zk in constraint (10.7) equals 1, then the sink sojourns at site k, and
286
MOBILITY IN WIRELESS SENSOR NETWORKS
therefore there must be one and only one arc on the sink movement route reaching
site k. On the other hand, if zk equals 0, then there will not be any incoming arc to
that site. Finally, constraint (10.5) induces a unit of flow from the last node in the sink
route (ω) to the fictitious final node q + 1.
A possible optimum sink route is depicted in Figure 10.3a. This route goes from
the initial site α = a to the final site ω = g. Constraints (10.4) and (10.5) ensure that
there is only one initial site and one final site for the sink route (x0α = xωq+1 = 1),
independently of any α and ω. The two sets of constraints (10.6) and (10.7) generate
the route from α to ω that passes through all the sites where the sink has to sojourn
in order to maximize the network lifetime. More precisely, constraints (10.6) enforce
that for every site k ∈ S the global number of outgoing visited arcs xkl equal the
incoming visited ones xjk (with the exception of the two fictitious sites 0 and q + 1).
For instance, this is the case of site c in Figure 10.3a: In the route from α to ω, there is
only one incoming arc and one outgoing visited arc (xbc = 1 and xcd = 1). According
to constraints (10.7), for every site k on the sink route, there must be a way to get
there; that is, there must be exactly one site j (which includes the fictitious site 0)
from which k is reachable directly (xjk = 1 and zk = 1). At the same time, the other
sites not belonging to the sink route are not visited by the sink. For instance, sites
k ∈ {a, b, c, d, e, f, g} in Figure 10.3a are all and only those for which zk = 1. These
are all and only the sites in the sink route. All other sites h are such that zh = 0.
Constraints (10.6) and (10.7) ensure flow conservation. However, they do not prevent the formation of cycles disjoint from the sink route from α to ω. The (disjoint,
non-simple) route depicted in Figure 10.3b comprising nodes in {a, b, c, d, e, f, g}⊂S
and nodes {h, i, j, k, l, m}⊂S (cycle) is possible according to this model up to constraints (10.7), since none of the constraints from (10.2) to (10.7) is violated by having
zk = 1 for k ∈ {h, i, j, k, l, m} as well as xh,i = xi,j = xj,k = xk,l = xl,m = 1. This
situation is not realistic and is undesirable. It is practically impossible, for instance,
to have the sink moving from site d (a site in the connected route from site a to
site g) to site i (a site in the cycle), since the distance between these two sites is
greater than dmax . Constraints (10.8) ensure that no such cycles are formed. (Similar
constraints have been used in the integer programming formulation of the Traveling
Salesman Problem (TSP) to avoid subtours [62].) According to constraints (10.9), a
site k is associated with a “weight” vk ≥ 0. Constraints (10.8) introduce a site ordering in the sink sojourns. The sites visited by the sink are traversed in increasing
order; that is, if xjk = 1, then vj < vk . In this way it is clearly impossible to return to
the same node and, hence, to generate sink routes with disjoint cycles like the one in
Figure 10.3b.
Some comments are in order. Parameter tmin has been introduced for investigating
the effect of higher or lower sink mobility rates on the network performance. The
model solution produces sink route and sojourn times tk ≥ tmin at site k that maximizes network lifetime. Varying tmin enables us to explore a number of tradeoffs.
For example, higher tmin values should induce lower overhead for route construction
and release. Lower tmin values instead should enable a finer tuning of sojourn times
at different sites (which may be useful in achieving a longer network lifetime), at the
price of increasing overhead.
COMPARING DIFFERENT APPROACHES TO MOBILITY
287
Figure 10.3. Sink optimum routes produced by constraints (10.4) to (10.7). (a) A sink route.
(b) A disjoint cycle in the sink route.
288
MOBILITY IN WIRELESS SENSOR NETWORKS
Figure 10.4. Two adjacent physical sites, n + m logical sites and their interconnections.
The power consumption rate cik of each node i ∈ N when the sink sojourns at
site k ∈ S can be computed in a different way (e.g., analytically [54]), or provided as
input to the model from simulations or from real-data traffic traces [9, 10]. In other
words, the model can be customized to find the optimum lifetime for different routing
protocols (by computing the corresponding values of bik and cik ).
The model does not allow the sink to pass twice for the same site. However, this
can be made possible by having a single “physical” represented by h “logical” sites,
where h is the number of times we want the sink to be able to pass through that site.
The logical sites have no arcs between them and are connected to all the (logical) sites
of adjacent (physical) sites. Figure 10.4 concerns the case of two adjacent physical
sites and their logical sites. With this simple modification, optimal lifetime is obtained
where the sink is allowed to visit each site at most h times.
The MILP formulation is an improvement over previously proposed models. The
model is independent of a number of factors such as sensor node deployment and
sensor density, the sink site topology, the size and shape of the geographic area of
deployment, and the sensor node technical features such as transmission radius, energy
model, and so on. The formulation also includes a number of realistic constraints,
such as the noninstantaneous movement of sink between sites potentially far apart
from each other. Most importantly, and differently from all previously proposed LP
solutions, this formulation explicitly considers costs for changing location.
The Greedy Maximum Residual Energy (GMRE) Protocol. The solution presented above determines the movements and the sojourn times of the mobile sink at
different sites so that network lifetime is maximized. Movements and times are determined by solving the described analytical model by providing as input a host of
information concerning the whole networks. In other words, this solution is centralized: Information is collected at a “solver” and the resulting output is the best route
for the sink.
Collecting information about network condition can be overwhelmingly expensive in terms of energy and time, and most of the times it is unfeasible in resource constrained networks like WSNs. Therefore, protocols for controlled mobility
have to be designed and deployed that can realistically be deployed in WSNs. The
optimality is traded off for feasibility, as often happens. This motivates the definition
of the following heuristic.
COMPARING DIFFERENT APPROACHES TO MOBILITY
289
In the Greedy Maximum Residual Energy (GMRE) protocol [9, 10] the sink
periodically moves to a new site. More specifically, every tmin , it decides whether to
move or to stay at the current site. If it moves, the sink selects the site within dmax
from its current position surrounded by nodes that have the most energy left. The
choice is performed greedily—that is, moving toward the site that at the current time
appears the best to move to. In time, this should highly likely result in balanced
energy consumption at the network nodes and therefore result in longer network
lifetime. For deciding whether to move or not, the sink collects information about
the residual energy at the nodes around each of the potential future sites (we call this
energy value the residual energy at the site) and compares it with the residual energy
at the current site. If there are adjacent sites with a residual energy higher than that at
the current site, the sink moves to the site with the highest residual energy (selecting
randomly among sites with the same residual energy in case of ties). Otherwise the
sink stays at the current location. The communication to the sink of the energy level
at other network locations proceeds in two phases.
First Phase. For each of the adjacent sites the sink identifies one sentinel sensor
node that measures and reports the residual energy at the site when requested by the
sink. The second phase concerns the sink inquiring the sentinels. This happens every
time the sink has to decide to move. To implement the first phase, we take advantage
of the flooding performed by the sink when it makes the nodes aware of its new
location. For this heuristic protocol we assume that a node that is in the “transmission
vicinity” of a site (i.e., whose Euclidean distance from a site is less than or equal to
the nodes transmission range) is aware of that. This can be obtained by endowing
the nodes with a suitable localization mechanism (such as one of those described in
references 63 and 64). The flooding message contains the coordinates of the current
location of the sink. Upon receiving this packet, a node knows if it is in the vicinity
of a possible future sink site. If this is the case, it sends to the sink a (small) packet to
candidate itself as sentinel. When the sink receives these packets, it decides which is
the sentinel for a given site. An example of the sentinel identification mechanism is
depicted in Figure 10.5. The sink current site E is indicated by a triangle. The squares
A, B, C, D, F , G, H, and I identify the adjacent sink sites. The figure shows what
happens upon flooding the route construction message. Potential sentinels for sink
site G are marked as black circles. Their distance from G is less than or equal to the
node transmission range. When such potential sentinels receive the flooded message,
they answer back to the sink, identifying themselves as candidates. It is the sink that
selects the (single) sentinel for a given possible future site.
Second Phase. At the time the sink has to decide whether to move or not, it
interrogates the selected sentinels about the residual energy at their sites by sending
them a (small) packet. At this time the sentinels query their neighboring sensor nodes
about their residual energy and communicate back to the sink the minimum of the
obtained values or any suitable function that can express how critical (for the network
lifetime) it is to place the sink in that area. This information enables the sink to select
the next site, depending on the residual energy of its area.
290
MOBILITY IN WIRELESS SENSOR NETWORKS
Figure 10.5. Sentinels at adjacent sites.
10.5.3 Multihop Routing and Uncontrolled Mobility
The mathematical model and the distributed heuristic (GMRE) are clear examples of
controlled mobility: The sink moves, or attempts to move, toward network locations
where nodal energy is available.
For the sake of benchmarking, it is interesting also to compare controlled
mobility—and especially the more realistic, distributed GMRE—with a simple protocol for sink mobility that captures the case of uncontrolled, random mobility (RM)
of the sink à la data MULEs. The difference here is that routing from the sensors to
the sink is multihop. The heuristic is extremely simple: Every tmin the sink selects
randomly and uniformly the new location among all the sites within distance dmax
from the current. In case a site different from the current is selected, the sink moves to
that site. We observe that even in this case, where the sink does not necessarily move
to areas with high residual energy, improvements in network lifetime are obtained
with respect to the case with an optimally placed static sink.
There are cases where the mobility, although neither random nor controlled by
the network, can be exploited to improve network lifetime. Luo and Hubaux [56]
successfully demonstrate this by simply moving the sink around the deployment
area. Specifically, by searching the space of all possible sink mobility strategies that
are periodic (recurrent sink movements within a constant period), the authors claim
that a good mobility strategy is obtained by having the sink traveling at the periphery
of the network. While controlled, the sink movements are not guided by the network
situation/activities.
COMPARATIVE EVALUATION
291
Inspired by the solution by Luo and Hubaux, we devise a mobility strategy, termed
Perimeter Mobility (PM). According to PM, the sink visits cyclically the sites along
the perimeter of the (square) deployment area. It stays at each site for tmin and then
moves onto the next site in a counterclockwise fashion.
It makes sense to compare this different cases of controlled and uncontrolled
mobility, evaluating which solution performs more vigorously with respect to selected
metrics. This is the natural next step: After having discussed different types of mobility
and the corresponding algorithmic solutions (models and protocols), we proceed at
selecting scenarios and metrics for comparing them.
10.6 COMPARATIVE EVALUATION
A fundamental part of the creation and investigation of solutions for networking
consists in testing the new solutions (protocols, methods, techniques, algorithms) and
comparing them. We describe here the way this is usually done, giving some hints
and insights on the process of evaluating the performance of a WSN.
Since experiments on real testbeds are costly and usually scale to a few tens of
nodes, it is common to (start to) evaluate a protocol for WSNs via simulations. In the
attempt to consider the different layers of a WSN architecture, we use the VINT project
network simulator ns2 [65]. This widely used software tool allows us to consider some
physical layer properties and a MAC protocol CSMA/CS-like which are typical of
wireless sensor networking.
Concerning the protocols for mobility presented here, the first question one is
lead to investigate is about the properties of GMRE and of data MULEs mobility in a
realistic, typical WSN scenario (static sensor nodes). The comparison should be about
properties corresponding to relevant metrics. These include node energy consumption,
network lifetime, end-to-end packet latency, and probability of successfully delivering
the monitored information to the sink.
The next natural step is that of investigating in deeper details the data MULE
architecture and the case of mobility with multihop routing (GMRE, RM, and PM)
with the aim of identifying the impact of varying scenarios and protocol parameters
on their performance.
Overall, the purpose of the following sections is that of introducing the reader to
a deeper understanding of the performance trends when considering different application scenarios.
10.6.1 The Choice of a Suitable Scenario
It is meaningful to choose a scenario that resembles a realistic deployment of a WSN
with static nodes. Therefore, parameters like the number or network nodes, the size
of the deployment area, the nodal transmission radius and initial energy, and so on,
should be chosen so that what is depicted is substantially similar to a scenario one
could find in real life. Here is a suitable example: 400 sensor nodes are deployed on a
20 × 20 regular grid covering a squared area of side L = 400 m. Sensor nodes have
292
MOBILITY IN WIRELESS SENSOR NETWORKS
a transmission radius equal to 30 m, which imposes a maximum of 8 neighbors per
node. Nodes periodically generate packets at a rate of 0.5 bps (low traffic scenario,
typical of sensor networking). Packet size is equal to 512 B. The buffer size of the
sensor nodes varies among 10, 50, and 100 packets. Energy consumption follows
the TR1000 specifications [66]. An ideal awake/asleep schedule is assumed in which
nodes consume energy only when transmitting or receiving packets. The channel is
ideal (no errors occurs when transmitting) and it has a data rate equal to 250 Kbps.
In the case of the data MULE architecture, the number of available MULEs is an
important parameter: 1, 2, and 4 data MULEs can be a good starting point. In the case
of data MULEs the deployment area is divided into 81 cells. The MULEs move from
the center of one cell (source) to the center of another one (destination) according
to the random waypoint model [27]. The MULE travels at a speed of 1 m/s. While
traveling from the source to the destination cell, a MULE stops at intermediate cells
for data collection, gathering packets from all (and only) the sensors in a cell. The
sojourn time spent in each cell is 1 s. MULE queues are considered unbounded (in
fact, they can contain 1000 packets, which ensure no overflow, i.e., data loss). The
sink to which the MULEs deliver the collected packets is (statically) placed at the
center of the deployment area. Once at the sink, a MULE stays in its proximity for a
time necessary to empty its queue.
Meaningful parameters for GMRE and the other heuristics are as follows. The sink
moves among 8 × 8 sink sites (grid deployment) at the speed of 1 m/s. A shortest
path-like routing is used to deliver data in a multihop fashion from the sensor to the
current site of the sink. The parameter tmin is set to 50 K (best-performing value
among the many tested). The parameter dmax is equal to 190 m.
In all the experiments, every node generates over 3000 packets during the simulation time. Finally, in order to obtain statistically meaningful results, the same experiment (simulation run) should be performed on different networks for an enough
number of times. That is why results are obtained by averaging over experimental
outcomes from 100 different network topologies.
Table 10.1 gives a bird’s-eye view of the results of the comparison between GMRE
and the data MULE solutions.
When the network traffic is low and the network deployment area size is limited
(as considered here), we observed that both GMRE and the data MULEs deliver all
TABLE 10.1. GMRE and Data MULEs, General Comparison
Data latency
Energy consumption
Packet delivery ratio
Computational needs
Low •
High • • • • •
GMRE
Data MULEs
•
•••
•••••
•••
•••••
•
•••
•
COMPARATIVE EVALUATION
293
generated packets to the sink. As a rule of thumb, successful packet delivery for
low/medium network traffic is more challenging for the MULE solution than for the
multihop approach (GMRE). The packet delivery ratio degrades as the deployment
area grows in size, since the inter-arrival time of a MULE at a cell grows and overflow
of sensor node queues can. On the other hand, GMRE is always able to deliver packets
to the sink.
The advantage of using solution data MULEs-like is clear when energy consumption and sensor node computational capability are considered. Given the single-hop
nature of data exchange between the sensors and the MULEs, nodes are not required
to implement a full protocol stack. Basic physical and MAC layer functions are sufficient for all data communications. This implies lower nodal energy consumption and
lower node complexity and cost. In the scenario described above, sensor nodes in a
data MULE setting consume one order of magnitude less energy than in the multihop
scenario. However, the gain in energy conservation is heavily paid in terms of endto-end data latency. The difference in this case is as high as four orders of magnitude!
This is due to (a) the extremely long time it takes for a MULE to visit the same cell
twice and (b) the time needed to go back to the sink to deliver the packets. These two
factors are of course dependent on the number of MULEs, their speed, and the size
of the deployment area. We observed that varying MULEs speed from pedestrian (as
shown here) to slow vehicular speed would not lead to improvements of more than one
order of magnitude. Only introducing a high number of data MULEs would satisfy
the latency requirements of many WSN applications. This is sometimes impossible
and is certainly costly.
10.6.2 Experiments for Data MULEs
The kind of experiments that are meaningful to perform in the data MULE architecture
mainly concern the average number of packet discarded, and hence lost, because of
buffer overflow and the sensor-to-sink packet data latency. It is interesting to notice
that in order to interpret correctly the experimental results on these fundamental
metrics, it is sometimes necessary to investigate other metrics and patterns, such as
(a) the inter-arrival times of the MULEs at the different cells and (b) the average
sojourn time of a MULE in a cell.
Here is an example of experimental investigation in the basic scenario depicted
above. We observed that the average number of packets discarded by the sensor nodes
in the case of one MULE roaming among sensors with buffer size set to 10 (packets)
is not high at all, especially for nodes that are toward the center of the deployment
area. There are a few nodes, however, that, given the presence of a single MULE and
small buffers, have to drop some packets. In this case the problem is the movement of
the MULE itself, given that the random waypoint mobility leads the MULE toward
the center of the deployment area with high probability. The longer at the center, the
longer the MULE inter-arrival times at the cells along the perimeter. This is the reason
for the observed 4.5% packet loss at the nodes in the corners.
This behavior is confirmed by an investigation of the inter-arrival times of 1, 2, and
4 MULEs at the same cell. In all the considered scenarios the inter-arrival times values
294
MOBILITY IN WIRELESS SENSOR NETWORKS
are extremely high (in the thousands of seconds or more), explaining why the data
MULEs approach results in high packet latency. As the number of MULEs increases,
it takes less and less time before a MULE arrives at a given cell, as expected. At the
same time, given that the MULEs tend more to visit central cells than peripheral ones,
the inter-arrival times are higher at the area borders.
MULEs visit each cell for the enforced 1 s. The only cell that experiences a longer
sojourn time is the one where the sink is located. The higher the traffic and the area
size and the lower the number of MULEs, the larger the amount of data stored in a
MULE queue, which imposes longer times to unload the data to the sink. Therefore,
the time needed for data transfer may exceed 1 s.
The study of these last two metrics makes us understand the results about what
it is perhaps the most important metric for assessing whether or not the data MULE
architecture is suitable for certain applications: data latency. Measures on sensor-tosink latency should take into account the latency incurred by data packets from the
sensor to the sink, in a MULE queue and in a (source) node queue. While it has
an impact on the number of packets that are dropped, the size of a node queue has
little influence on the packet latency. Packets are transferred from the source queue
(node or MULE) to the destination queue (MULE or sink) all at once, and the transfer
time is quite small. As a consequence, the end-to-end delay is given almost entirely
by the time from when the packet is generated and a MULE arrives at the cell, plus the
time it takes to the MULE to deliver it to the sink. The time a packet sits on its node
queue depends on the number of MULEs, as expected. When the number of MULEs
doubles, this time halves. We observe that this component is the most significant one.
Given that the sink is placed at the center, the inter-arrival time of a MULE at the sink
is on average shorter than that at a (noncentral) cell. The time spent by a packet on a
MULE queue is basically independent of the number of MULEs. As a consequence,
the end-to-end packet latency almost halves with doubling the MULEs, since only
one of the delay component halves. In any case, and independently of the number of
MULEs, the average latency of a packet is in the thousands of seconds. As expected,
this is quite high.
The trends observed in the basic scenario are also confirmed in a setting where
while the nodes are as dense as in the case with L = 400 m, the deployment area is
larger (L = 572 m).
10.6.3 Experiments on GMRE
As outlined in Section 10.6.1, controlled sink mobility and multihop data communication lead to higher energy consumption than in the data MULE case. However, this
is the solution for satisfying more stringent latency constraints. What we show here is
one possible, meaningful choice for evaluating a distributed heuristic like GMRE. In
particular, we explore GMRE to explain the impact of different protocol parameters
on its performance and to assess the effectiveness of an energy-aware sink mobility
heuristic over other sink mobility schemes. It makes sense to compare GMRE to the
following schemes.
COMPARATIVE EVALUATION
295
r Static sink placement (termed STATIC in the following). In this case the sink is
optimally placed at the center of the deployment area.
r The sink mobility as generated by solving the MILP model presented in Section 10.5.2 (OPT).
r RM, the mobility heuristic outlined in Section 10.5.3, which is similar to the
scenario with a randomly moving MULE, but with multihop routing.
r The scheme that we term PM (for Perimeter Mobility), according to which
the sink moves cyclically along the perimeter of the deployment area (Section 10.5.3).
Experiments should be performed in different settings. For instance, one should
vary the transmission radius (from 25 m to 30 m), the data routing protocol (shortest
path-like versus geographic), and the way nodes are deployed (grid positioning versus
random scattering). The results obtained are consistent with what we describe here
(which refers to the basic scenario with transmission ranges of 30 m and 25 m).
The conclusion is that GMRE is a simple but effective solution, achieving a network
lifetime close to OPT’s—that is, the best that can be obtained by exploiting sink
mobility in a multihop WSN. RM and PM have similar performance in terms of
network lifetime with PM outperforming RM at higher tmin values. Both schemes
improve considerably over STATIC, which advocates for the use of mobility in WSNs.
However, the lack of energy awareness does not allow us to match the results seen
with GMRE (see references 10 and 59 for further details).
Table 10.2 shows the average network lifetimes achieved by STATIC, RM, PM,
GMRE, and OPT in networks whose nodes have transmission ranges of 25 m and
30 m. (The network lifetime has been defined as the time until the first node dies
because of energy depletion.)
Improvements with respect to the static case can be as high as 350% when the
sink moves according to GMRE in scenarios with 64 sink sites. In this case the
GMRE lifetime is never more than 22% (25%) shorter than the OPT lifetime when
the transmission range is 25 m (30 m) and tmin is kept below 250 K. Improvements on
network lifetime are obtained even when the sink moves randomly (RM heuristic) or
TABLE 10.2. Average Network Lifetime for Different Transmission Ranges
(seconds ×106 )
Radius 25 m
Radius 30 m
tmin
50 K
250 K
500 K
1M
50 K
250 K
500 K
1M
OPT
GMRE
RM
PM
STATIC
36.2
30.9
23.18
21.9
7
36
28.79
20.44
20.9
7
35.9
24.4
17.6
19.9
7
33.43
21.9
13.08
16.17
7
41.2
31.9
28.9
28.7
6.7
41
31
25.7
26.1
6.7
40.8
29
21.9
24
6.7
39.3
27.9
16.4
19.67
6.7
296
MOBILITY IN WIRELESS SENSOR NETWORKS
when the sink moves along the area perimeter (PM heuristic). The latter experiences
performances only slightly better than RM when the transmission range is 25 m. RM
leads to a network lifetime halfway between the STATIC lifetime and that obtained
by OPT mobility.
In general, for GMRE, RM, and PM we notice that the lower the tmin value, the
higher the network lifetime. At higher tmin values, the sink cannot finely decide the
time to spend at each site, being stuck at a site for a longer time. This implies a
less balanced energy consumption at the nodes. When tmin is very high, it is not even
possible for the sink to sojourn at all network sites, since the lifetime is reached before
the sink can visit them all.
Even in the case of OPT mobility, lower tmin values correspond to longer lifetimes.
In this case, however, the decrease in the network lifetime corresponding to higher
tmin values is not as evident as for GMRE and RM. At this point, one becomes curious
about why this happens. First of all, in the OPT case, tmin is simply a lower bound on
the sojourn time: The sink has to stay at a site for that time, but does not have to move
after it. It can stay an (optimum) amount of time after tmin and then move. In case
of GMRE, RM, and PM sink mobility, the decision about whether to move or not is
due every tmin (after which the sink often moves). This has a twofold consequence.
On one side, increasing tmin is more restrictive for GMRE than for OPT in terms
of fine-tuning the sojourn time at a site. Furthermore, being able to decide optimum
sojourn times implies much lower sink mobility (OPT case), which corresponds to
lower overhead for route management and hence to lower energy consumption with
respect to that incurred by GMRE, RM, and PM. Secondly, GMRE, RM, and PM do
not have a global view of the network topology and do not know the network traffic—
that is, how the nodes energy consumption evolves over time. This induces decreased
performance with respect to OPT. The RM and PM heuristics do not enforce any
energy-based criterion for sink movement, resulting in the worst performance among
all the schemes with mobile sink we investigated. Even if GMRE takes into account
nodes residual energy, the decision about whether to move and where is based on the
current status of the network and on a local view of the residual energy. According
to the best “greedy” tradition, this could lead to a bad move with respect to global
network lifetime maximization. The impact of such a bad move is clearly higher for
high tmin values: The wrong toll is paid for a longer time.
The OPT performance also degrades for higher values of tmin . In this case it
converges to values that are typical of when the sink is kept static. For instance, for
values of tmin approaching 7.013.801s the MILP model positions the sink at one of
the sites in the center of the deployment area and leaves it there (static). However,
for what explained above, increasing values of tmin are less critical in the case of
OPT mobility than in the case of GMRE, RM, and PM. OPT network lifetime values
start to decrease steeply at very high tmin values. Considering that OPT needs global
information for deriving optimum sink mobility and sojourn times, and considering
also the more “philosophical” algorithmic differences between OPT and GMRE, the
fact that the heuristic never obtains network lifetimes more than 25% lower than
OPTs, for relatively low sink mobility rate (low tmin values), can really be considered
an excellent result.
COMPARATIVE EVALUATION
297
Now, onto the flip side of this mobile story. Increases in network lifetime due to
the sink mobility are paid in term of increased data packet latency. The reasons are
quite clear. First of all, packets that are generated while the sink is moving, as well
as those in transit toward the sink, have to wait until routes to the new position of the
sink are established. Furthermore, in order to balance energy depletion, the sink will
spend time not only at the center of the deployment area but also along the borders.
This imposes longer average route length and hence a higher packet latency than
the latency experienced when the sink is statically placed at the center. (The latter is
actually the dominant reason for increased latency in low sink mobility scenarios.) In
any case, the average packet latency is always (i.e., for all the solution investigated)
below half a second! This is far beyond the measures of latency observed in the data
MULEs case, which were in the thousands of seconds (independently of the number
of MULEs). The improvement is four/five orders of magnitude!
Table 10.3 reports the average data latency incurred by packets in OPT, GMRE,
RM, PM, and STATIC for tmin =50,000 s and 1,000,000 s. (Intermediate values are
consistent with these ones.) Although small and always below .5 s, there are differences among the data latencies of the four schemes that we have investigated
and compared. It is interesting to see why. When the sink stays on the perimeter or at the corner sites (as in GMRE, OPT, and PM), we know that lifetime increases. However, these are also the cases when the average length of the routes
to the sink increases, which implies, in turn, a (slightly) higher packet latency. It
is thus reasonable to expect that a packet experiences lower latency when the sink
is statically placed at the center of the sensor deployment area. The RM heuristic, which tends to move the sink to sites located centrally, is the second best mobility scheme in terms of latency. As tmin increases, the sink tends to stay less at
central sites, leading to higher average latencies experienced by RM packets. The
opposite trend is observed for GMRE. For low tmin values, the sink stays at sites
on the corners and on the perimeter, which leads to and average packet latency
up to 30% higher than those experienced in the RM case. When tmin increases,
the sink sojourns less at corner sites, which implies a decrease of the average
latency.
Although the above discussion allows us to understand the subtle differences between the performance of the investigated schemes, there is very little variance among
TABLE 10.3. Average Packet Latency (seconds)
8 × 8 Sink Sites
tmin
50 K
1M
OPT
GMRE
RM
PM
STATIC
.315
.32
.25
.30
.19
.315
.29
.26
.30
.19
298
MOBILITY IN WIRELESS SENSOR NETWORKS
the measured latencies. All values range between .19 s (STATIC) and .32 s (GMRE
and OPT).
As a result of this simple investigation, which the reader is encouraged to reproduce
and vary, we obtain that GMRE shows a good tradeoff between network lifetime
gains and latency increases, promising to be a suitable solution for those application
scenarios that cannot be covered by the data MULEs mechanism.
10.7 SOME OPEN PROBLEMS AND RESEARCH DIRECTIONS
Research on mobile WSNs is quite interesting and recent. Mobility in these networks
can be tackled from several points of view. Many research directions could be taken.
For instance, as demonstrated in the case of OPT and GMRE, having mathematical
models is quite an important benchmark. Although centralized, and also often computationally tough, an optimal solution is quite useful for giving an idea of what can
be obtained by designing and deploying a more realistic distributed heuristic. While
progresses are being done in the multihop routing/mobile sink realm, this is quite
uncharted territory in the case of the data MULEs approach.
In the initial approach to controlling the mobility of the MULEs, as described
here (Section 10.5.1), it is pointed out that when not enough MULEs are available,
some nodes might not fall close enough to a MULE route. A proposed solution is
that of introducing multihop routes (organized as trees) to nodes that can directly
communicate with a MULE when it passes. This approach could be studied in more
detail and improvement can be obtained by determining, for instance, the depth of
the tree that best keeps energy consumption low while allowing packets to get to the
MULE—that is, without discarding them. One can also think of dividing the data
packets into priority classes, so that the MULE trajectory is controlled by the network
in such a way that the MULE goes first where high-priority data are to be collected.
10.8 CONCLUSIONS
This chapter explores ways for using the mobility of network components in large
networks of resource constrained devices, like wireless sensor networks (WSNs). The
aim is that of exploiting mobility for improving the performance of these networks
in terms of network lifetime, throughput, and connectivity without significantly affecting data routing and end-to-end latency. Whereas energy-efficient solutions have
been proposed for WSNs with statically placed sensors and one or more static data
collection points (the sinks), our investigation delves into the various ways the sensor
nodes, some mobile relays, or the sinks can be made mobile for improving network
performance. We review the most recent proposals for mobility use and management
in WSNs, considering the pros and cons, as well as the costs and tradeoffs, of using
mobile sensors, mobile relays, and moving sinks. In the final part of the chapter we
give examples of how, once we have selected typical solutions for dealing with the
mobility of sinks and relays, these solutions are compared. The aim is to (a) introduce
EXERCISES
299
the reader to the selection of suitable scenarios and performance metric and (b) illustrate a general approach to “reading” experimental results. In particular, we illustrate
controlled and uncontrolled mobility both in the case of networks with single-hop
routing of the packets to a passing mobile relay (data MULEs architecture) and in the
case of controlled and uncontrolled mobility of a sink roaming through the network
and having the data delivered to it in a multihop fashion. The two different network
architectures, with single-hop versus multihop data routing, are described in details
and compared. Our comparison concerns key metrics of interest for WSNs, which
include network lifetime and sensor-to-sink data latency. Besides providing evidence
of the advantages of using mobility, the chapter also provides insights on different
techniques for protocol design and performance evaluation and comparison.
10.9 EXERCISES
1. Review questions. (a) Why is it less convenient to use resource-constrained
mobile sensor nodes rather than resource-rich mobile elements like relays and
sinks? (b) Give an example that clearly motivates the use of mobility in WSNs.
(c) What metric is most affected by the use of mobile relays rather than mobile
sinks? (d) For what kind of application it is clearly more advisable to use the data
MULE approach instead of a multihop solution with mobile sink? (e) List five
metrics of interest that should be evaluated when dealing with mobile WSNs. (f)
Describe pros and cons of developing a mathematical formulation of a problem
for WSNs like the one described in Section 10.5.2. (g) What are the major
differences, in terms of mobility, when comparing GMRE, RM, and PM? (h)
If the MULEs move according to the random waypoint mobility model, where
are they more likely to be found? What does this imply in terms of network
performance? (That is, which metric is the most affected in this case?)
2. Consider the MILP model presented in Section 10.5.2. What are the problems
that would occur if constraints (10.8) are removed? Why is this removal a
problem?
3. Consider the MILP model presented in Section 10.5.2. Imagine that instead
of knowing the sink sites, several different routes for the sink are available.
Rewrite the model in such a way that the output is the sequence of routes that
maximize network lifetime.
4. Consider again the MILP model presented in Section 10.5.2. Rewrite the model
so that the sink can pass for the same site at most five times.
5. Consider the pattern of energy consumption depicted in Figure 10.6 (which
refers to the basic scenario described in Section 10.6.1 with 8 × 8 sink sites).
The darker the area, the higher the energy consumption. Given this pattern,
describe where an heuristic will more likely send the sink (i.e., to which sites).
Justify your answer.
6. What is the main difference between the data MULE approach and the PM
scheme described in Section 10.5.3?
300
MOBILITY IN WIRELESS SENSOR NETWORKS
Figure 10.6. Average nodal energy consumption when the sink sojourns at different sites.
(a) Corner site. (b) Perimeter site. (c) Central site.
ACKNOWLEDGMENTS
This work was supported in part by the European FP6 027227 IP Project “E-Sense
(Capturing Ambient Intelligence for Mobile Communications through Wireless Sensor Networks) and by NSF grant CNS-0738720.”
BIBLIOGRAPHY
1. S. Basagni, M. Conti, S. Giordano, and I. Stojmenovic, editors. Mobile Ad Hoc Networking.
IEEE Press and John Wiley & Sons, Piscataway, NJ, and Hoboken, NJ, 2004.
2. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks:
A survey. Computer Networks, vol. 38(4):393–422, 2002.
3. E. H. Callaway, Jr. Wireless Sensor Networks: Architectures and Protocols. Auerbach
Publications, Boca Raton, FL, 2003.
4. I. F. Akyildiz, D. Pompili, and T. Melodia. Underwater acoustic sensor networks: Research
challenges. Elsevier’s Journal of Ad Hoc Networks, 3(3):257–279, 2005.
5. G. Kantor, S. Singh, R. Peterson, D. Rus, A. Das, V. Kumar, G. Pereira,
and J. Spletzer. Distributed search and rescue with robot and sensor team.
In Proceedings of the 4th International Conference on Field and Service Robotics, FSR
2003, Lake Yamanaka, Japan, July 2003, pp. 327–332.
6. K. Kotay, R. Peterson, and D. Rus. Experiments with robots and sensor networks for
mapping and navigation. In Proceedings of the International Conference on Field and
Service Robotics, FSR 2005, Port Douglas, Australia, July 29–31, 2005.
7. R. Brooks, P. Ramanathan, and A. Sayeed. Distributed target classification and tracking in
sensor networks. Proceedings of the IEEE, 91(8):1163–1171, 2003.
8. G. Virone, A. Wood, L. Selavo, Q. Cao, L. Fang, T. Doan, Z. He, R. Stoleru, S. Lin, and J. A.
Stankovic. An assisted living oriented information system based on a residential wireless
sensor network. In Proceedings of the 1st Distributed Diagnosis and Home Healthcare
(D2H2) Conference, Arlington, VA, April 2–4, 2006, pp. 95–100.
BIBLIOGRAPHY
301
9. S. Basagni, A. Carosi, E. Melachrinoudis, C. Petrioli, and M. Z. Wang. A new MILP
formulation and distributed protocols for wireless sensor networks lifetime maximization.
In Proceedings of the IEEE International Conference on Communications, ICC 2006,
Vol. 8, Istanbul, Turkey, June 11–15 2006, pp. 3517–3524.
10. S. Basagni, A. Carosi, E. Melachrinoudis, C. Petrioli, and M. Z. Wang. Controlled sink mobility for prolonging wireless sensor networks lifetime. ACM/Elsevier Wireless Networks,
2008.
11. P. Juang, H. Oki, Y. Wang, M. Martonosi, L.-S. Peh, and D. Rubenstein. Energy-efficient
computing for wildlife tracking: Design tradeoffs and early experiences with zebranet. In
Proceedings of the 10th International Conference on Architectural Support for Programming Languages and Operating Systems, ASPLOS-X, San Jose, CA, October 5–9, 2002,
pp. 96–107.
12. Q. Li and D. Rus. Sending messages to mobile users in disconnected ad-hoc wireless networks. In Proceedings of 6th ACM Annual International Conference on Mobile Computing
and Networking, MobiCom 2000, Boston, MA, August 6–11 2000, pp. 44–55.
13. M. Grossglauser and D. N. C. Tse. Mobility increases the capacity of ad-hoc wireless
networks. IEEE/ACM Transactions on Networking, 10(4):477–486, 2002.
14. W. Zhao, M. H. Ammar, and E. W. Zegura. A message ferrying approach for data delivery
in sparse mobile ad hoc networks. In Proceedings of the 5th ACM International Symposium
on Mobile Ad Hoc Networking and Computing, MobiHoc 2004, Roppongi Hills, Tokyo,
Japan, May 24–26, 2004, pp. 187–198.
15. S. Jain, R. C. Shah, W. Brunette, G. Borriello, and S. Roy. Exploiting mobility for energyefficient data collection in sensor networks. In Proceedings of the IEEE Workshop on
Modeling and Optimization in Mobile, Ad Hoc and Wireless Networks, WiOpt 2004, Cambridge, UK, March 24–26, 2004.
16. W. Zhao, M. H. Ammar, and E. W. Zegura. Controlling the mobility of multiple data
transport ferries in a delay-tolerant network. In Proceedings of IEEE Infocom 2005, Vol. 2,
Miami, FL, March 13–17, 2005, pp. 1407–1418.
17. H. Jun, W. Zhao, M. H. Ammar, E. W. Zegura, and C. Lee. Trading latency for energy
in wireless ad hoc networks using message ferrying. In Proceedings of the 3rd IEEE
International Conference on Pervasive Computing and Communications, PerCom 2005,
Kawai, HI, March 8–12, 2005, pp. 220–225.
18. M. Mukarram Bin Tariq, M. H. Ammar, and E. W. Zegura. Message ferry route design for
sparse ad hoc networks with mobile nodes. In Proceedings of the 7th ACM International
Symposium on Mobile Ad Hoc Networking and Computing, MobiHoc 2006, Firenze, Italy,
May 22–25, 2006, pp. 37–48.
19. T. Spyropoulos, K. Psounis, and C. S. Raghavendra. Performance analysis of mobilityassisted routing. In Proceedings of the 7th ACM International Symposium on Mobile Ad
Hoc Networking and Computing, MobiHoc 2006, Firenze, Italy, May 22–25, 2006, pp.
49–60.
20. D. K. Goldenberg, J. Lin, A. S. Morse, B. E. Rosen, and Y. R. Yang. Towards mobility
as a network control primitive. In Proceedings of the 5th ACM International Symposium
on Mobile Ad Hoc Networking and Computing, MobiHoc 2004, Roppongi Hills, Tokyo,
Japan, May 24–26, 2004, pp. 163–174.
21. G. Wang, G. Cao, and T. F. La Porta. Movement-assisted sensor deployment. In Proceedings of IEEE INFOCOM 2004, Vol. 4, Hong Kong, March 7–11, 2004, pp. 2469–
2479.
302
MOBILITY IN WIRELESS SENSOR NETWORKS
22. G. Wang, G. Cao, and T. F. La Porta. Proxy-based sensor deployment for mobile sensor
networks. In Proceedings of the First IEEE International Conference on Mobile Ad Hoc
and Sensor Systems, MASS 2004, Fort Lauderdale, FL, October 25–27, 2004, pp. 493–502.
23. A. Howard, M. J. Matarić, and G. S. Sukhatme. An incremental deployment algorithm for
mobile sensor networks. Kluwer Autonomous Robots, Special Issue on Intelligent Embedded Systems, G. S. Sukhatme, editor, Vol. 13, no. 2, pp. 113–126, September 2002.
24. A. Howard, M. J. Matarić, and G. S. Sukhatme. Mobile sensor networks deployment using
potential fields: A distributed, scalable solution to the area coverage problem. In Proceedings of the 6th International Symposium on Distributed Autonomous Robotic Systems,
DARS 2002, Fukuoka, Japan, June 25–27, 2002.
25. R. Rao and G. Kesidis. Purposeful mobility for relaying and surveillance in mobile ad hoc
sensor networks. IEEE Transactions on Mobile Computing, 3(3):225–232, 2004.
26. I. Chatzigiannakis and S. Nikoletseas. An adaptive compulsory protocol for basic communications in highly changing ad-hoc mobile networks. In Proceedings of the International
Parallel and Distributed Processing Symposium, IPDPS 2002, Fort Lauderdale, FL, April
15–19 2002, pp. 193–202.
27. R. C. Shah, S. Roy, S. Jain, and W. Brunette. Data MULEs: Modeling a three-tier architecture for sparse sensor networks. In Proceedings of the First IEEE International Workshop
on Sensor Network Protocols and Applications, SNPA 2003, Anchorage, AK, May 11,
2003, pp. 30–41.
28. T. Small and Z. J. Haas. Resource and performance tradeoffs in delay-tolerant wireless
networks. In Proceedings of the ACM SIGCOMM 2005 Workshop on Delay-Tolerant Networking and Related Networks, WDTN 2005, Philadelphia, PA, August 22–26, 2005.
29. L. Song and D. Hatzinakos. Dense wireless sensor networks with mobile sinks. In
Proceedings of the IEEE International Conference on Acoustic Speech, and Signal
Processing, ICASSP 2005, Vol. 3, Philadelphia, PA, March 18–23, 2005, pp. 677–
680.
30. Y. Tirta, Z. Li, Y.-H. Lu, and S. Bagchi. Efficient collection of sensor data in remote fields
using mobile collectors. In Proceedings of the 13th International Conference on Computer
Communications and Networks, ICCCN 2004, Chicago, IL, October 11–13, 2004, pp.
515–519.
31. Y. Tirta, B. Lau, N. Malhotra, S. Bagchi, Z. Li, and Y.-H. Lu. Controlled mobility for efficient data gathering in sensor networks with passively mobile nodes. In IEEE Monograph
on Sensor Network Operations, S. Phoha, T. F. La Porta, and C. Griffin, editors, IEEE Press
& John Wiley & Sons, Piscataway, NJ, and New York, 2005, 49–58.
32. A. Kansal, A. A. Somasundara, D. D. Jea, M. B. Srivastava, and D. Estrin. Intelligent
fluid infrastructure for embedded networks. In Proceedings of the 2nd ACM/SIGMOBILE
International Conference on Mobile Systems, Applications, and Services, MobySys 2004,
Boston, MA, June 6–9, 2004, pp. 111–124.
33. D. D. Jea, A. A. Somasundara, and M. B. Srivastava. Multiple controlled mobile elements
(data mules) for data collection in sensor networks. In Proceedings of the 1st IEEE International Conference on Distributed Computing in Sensor Systems, DCOSS 2005, LNCS
series, V. K. Prasanna, S. Iyengar, P. G. Spirakis, and M. Welsh, editors, Vol. 3560, Marina
del Rey, CA, June 30–July 1 2005, pp. 244–257.
34. A. A. Somasundara, A. Ramamoorthy, and M. B. Srivastava. Mobile element scheduling for efficient data collection in wireless sensor networks with dynamic deadlines. In
BIBLIOGRAPHY
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
45.
46.
47.
303
Proceedings of the 25th IEEE International Real-Time Systems Symposium, RTSS 2004,
Lisbon, Portugal, December 5–8 2004, pp. 296–305.
A. Kansal, M. Rahimi, D. Estrin, W. J. Kaiser, G. J. Pottie, and M. B. Srivastava. Controlled
mobility for sustainable wireless sensor networks. In Proceedings of the First IEEE Annual
Conference on Sensor and Ad Hoc Communications and Networks, SECON 2004, Santa
Clara, CA, October 4–7, 2004, pp. 1–6.
A. A. Somasundara, A. Kansal, D. D. Jea, D. Estrin, and M. B. Srivastava. Controllably
mobile infrastructure for low energy embedded networks. IEEE Transactions on Mobile
Computing, 5(8):958–973, 2006.
E. Ekici, Y. Gu, and D. Bozdag. Mobility-based communications in wireless sensor networks. IEEE Communications Magazine, vol. 44(7):56–62, 2006.
W. Wang, V. Srinivasan, and K.-C. Chua. Using mobile relays to prolong the lifetime of
wireless sensor networks. In Proceedings of the 11th Annual ACM/SIGMOBILE International Conference on Mobile Computing and Networking, MobiCom 2005, Cologne,
Germany, August 28–September 2, 2005, pp. 270–283.
H. S. Kim, T. F. Abdelzaher, and W. H. Kwon. Minimum energy asynchronous dissemination to mobile sinks in wireless sensor networks. In Proceedings of the First International
Conference on Embedded Networked Sensor Systems, SenSys 2003, Los Angeles, CA,
November 5–7, 2003, pp. 193–204.
C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva. Directed diffusion
for wireless sensor networking. IEEE/ACM Transactions on Networking, 11(1):2–16, 2003.
F. Ye, H. Luo, J. Cheng, S. Lu, and L. Zhang. A two-tier data dissemination model for large
scale wireless sensor networks. In Proceedings of the 8th ACM Annual International Conference on Mobile Computing and Networking, MobiCom 2002, Atlanta, GA, September
23–28, 2002, pp. 148–159.
J. G. Jetcheva and D. B. Johnson. Adaptive demand-driven multicast routing in multi-hop
wireless ad hoc networks. In Proceedings of the 2nd ACM International Symposium on
Mobile Ad Hoc Networking & Computing, MobiHoc 2001, Long Beach, CA, October 4–5,
2001, pp. 33–44.
K. Hwang, J. In, Y. Yun, and D. Eom. Dynamic sink oriented tree algorithm for efficient
target traking of multiple mobile sink users in wide sensor fields. In Proceedings of the
60th IEEE Vehicular Technology Conference, VTC 2004-Fall, Vol. 7, Los Angeles, CA,
September 26–29, 2004, pp. 4607–4610.
X. Hu, Y. Liu, M. J. Lee, and T. N. Saadawi. A novel route update design for wireless sensor networks. ACM Mobile Computing and Communications Review, 8(1):18–26,
2004.
K. Akkaya and M. Younis. Energy-aware to mobile gateway in wireless sensor networks. In
Proceedings of the IEEE Globecom 2004 Workshops, Dallas, TX, November 29–December
3, 2004, pp. 16–21.
H. L. Xuan and S. Lee. A coordination-based data dissemination protocol for wireless sensor networks. In Proceedings of the 2004 Intelligent Sensors, Sensor Networks
and Information Processing Conference, ISS-NIP 2004, Melbourne, Australia, December
14–17, 2004, pp. 13–18.
L. Tong, Q. Zhao, and S. Adireddy. Sensor networks with mobile agents. In Proceedings
of the IEEE Military Communication Conference, MILCOM 2003, Vol. 1, Boston, MA,
October 13–16, 2003, pp. 705–710.
304
MOBILITY IN WIRELESS SENSOR NETWORKS
48. P. Venkitasubramaniam, S. Adireddy, and L. Tong. Sensor networks with mobile agents:
Optimal random access and coding. IEEE Journal on Selected Areas in Communications,
22(6): 1058–1068, 2004.
49. V. P. Mhatre, C. Rosenberg, D. Kofman, R. Mazumdar, and N. Shroff. A minimum cost
heterogeneous sensor network with a lifetime contraint. IEEE Transactions on Mobile
Computing, 4(1), 4–15, 2005.
50. A. Chakrabarti, A. Sabharwal, and B. Aazhang. Using predictable observer mobility for
power efficient design of sensor networks. In Proceedings of the Second International
Workshop on Information Processing in Sensor Networks, IPSN 2003, LNCS series,
Vol. 2634, F. Zhao and L. Guibas, editors, Palo Alto, CA, April 22–23, 2003, pp. 129–
145.
51. P. Venkitasubramaniam, Q. Zhao, and L. Tong, Sensor networks with multiple mobile
access points. In Proceedings of the 38th Annual Conference on Information Sciences and
Systems, CISS 2004, Princeton, NJ, March 17–19, 2002.
52. P. Baruah, R. Urgaonkar, and B. Krishnamachari, Learning-enforced time domain routing to
mobile sinks in wireless sensor fields. In Proceedings of the 29th Annual IEEE International
Conference on Local Computer Networks, LCN 2004, Tampa, FL, November 16–18, 2004,
pp. 525–532.
53. S. R. Gandham, M. Dawande, R. Prakash, and S. Venkatesan. Energy efficient schemes
for wireless sensor networks with multiple mobile base stations. In Proceedings of IEEE
Globecom 2003, Vol. 1, San Francisco, CA, December 1–5, 2003, pp. 377–381.
54. Z. M. Wang, S. Basagni, E. Melachrinoudis, and C. Petrioli. Exploiting sink mobility for
maximizing sensor networks lifetime. In Proceedings of the 38th Hawaii International
Conference on System Sciences, Big Island, HI, January 3–6, 2005.
55. I. Papadimitriou and L. Georgiadis. Maximum lifetime routing to mobile sink in wireless sensor networks. In Proceedings of the 2005 International Conference on Software,
Telecommunications and Computer Networks, SoftCOM 2005, Split, Croatia, September
15–17, 2005.
56. J. Luo and J.-P. Hubaux. Joint mobility and routing for lifetime elongation in wireless
sensor networks. In Proceedings of IEEE Infocom 2005, Vol. 3, Miami, FL, March 13–17,
2005, pp. 1735–1746.
57. J.-H. Chang and L. Tassiulas. Maximum lifetime routing in wireless sensor networks. IEEE
Transactions on Networking, 12(4): 609–619, 2004.
58. S. Basagni, A. Carosi, E. Melachrinoudis, C. Petrioli, and Z. M. Wang. Controlling sink mobility in wireless sensor networks: A new model and protocols. Poster at ACM/SIGMOBILE
MobiCom 2005, Cologne, Germany, August 28–September 2, 2005.
59. S. Basagni, A. Carosi, E. Melachrinoudis, C. Petrioli, and M. Z. Wang. Protocols and model
for sink mobility in wireless sensor networks. ACM Mobile Computing and Communication
Review, (MC2 R), 10(4): 28–30, 2006.
60. A. Vahdat and D. Becker. Epidemic routing for partially connected ad hoc networks. Duke
University, Durham, NC, Technical Report CS-200006, April 2000.
61. S. Basagni, M. Elia, and R. Ghosh. ViBES: Virtual backbone for energy saving in wireless sensor networks. In Proceedings of the IEEE Military Communication Conference,
MILCOM 2004, Monterey, CA, October 31–November 3, 2004.
62. C. E. Miller, A. W. Tucker, and R. A. Zemlin. Integer programming formulation of traveling
salesman problems. Journal of the ACM, 7(4): 311–325, 1960.
BIBLIOGRAPHY
305
63. A. Savvides and M. B. Srivastava. Location discovery. In Mobile Ad Hoc Networking,
S. Basagni, M. Conti, S. Giordano, and I. Stojmenovic, editors, IEEE Press and John
Wiley & Sons, Piscataway, NJ, and Hoboken, NJ, 2004, chapter 8, pp. 231–254.
64. M. Battelli and S. Basagni. Localization for wireless sensor networks: Protocols and perspectives. In Proceedings of IEEE CCECE 2007, Vancouver, Canada, April 22–26, 2007,
pp. 1074–1077.
65. The VINT Project, The ns Manual. http://www.isi.edu/nsnam/ns/, 2002.
66. ASH transceiver designer’s guide. www.rfm.com, May 19, 2004.
CHAPTER 11
Localization Systems for Wireless
Sensor Networks
AZZEDINE BOUKERCHE
School of Information Technology and Engineering, University of Ottawa, Ottawa, Ontario K1N
6N5, Canada
HORACIO A. B. F. OLIVEIRA
School of Information Technology and Engineering, University of Ottawa, Ottawa, Ontario K1N
6N5, Canada; Federal University of Minas Gerais, Brazil; and Federal University of Amazonas,
Brazil
EDUARDO F. NAKAMURA
Federal University of Minas Gerais, Brazil; and FUCAPI—Analysis, Research, and Technology
Innovation Center, Brazil
ANTONIO A. F. LOUREIRO
Department of Computer Sciences, Federal University of Minas Gerais, Belo Horizonte, Brazil
11.1 INTRODUCTION
Wireless sensor networks (WSNs) [1–4] are composed of a large number of sensor
nodes used to monitor an area of interest. This type of network has become popular
due to its applicability, which includes several areas such as environmental, health,
industrial, domestic, agricultural, meteorological, spatial, and military. Despite the
fact that the main goal of a WSN is to monitor an area of interest, several secondary
objectives, or prerequisites, have to be achieved in order to reach the main objective (Figure 11.1).
The definition of a localization system [5–16] among the sensor nodes is one of
these prerequisites needed in order to make viable many of the WSNs applications.
The localization problem consists in identifying the physical location (e.g., latitude,
longitude, and altitude) of a determined object. Such a problem is very ample and
extensive, relating areas such as robotics, ad hoc networks, wireless sensor networks,
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
307
308
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
Figure 11.1. Several areas of the WSNs that work together in order to achieve one common
goal: to monitor an area of interest.
cellular telephony, military, aviation, and astronomy. In this chapter the localization
problem will be addressed under the viewpoint of the WSNs.
The localization systems have been identified as a key technology to the development and operation of the WSN [2]. Although its use is not exclusive to these
networks, the localization systems, especially the ones with multihop, differentiates
from the other areas because its applicability had been greatly extended with the
advent of the WSN.
11.1.1 Problem Statement
In this work, we consider a WSN as being composed of n nodes, with a communication
range of r, and distributed in a two-dimensional squared sensor field Q = [0, s] ×
[0, s]. For the sake of simplification, we consider symmetric communication links;
that is, for any two nodes u and v, u reaches v if, and only if, v reaches u and with
the same signal strength w. Thus, we represent the network by the Euclidean graph1 :
r V = {v1 , v2 , . . . , vn } is the set of sensor nodes;
r i, j ∈ E iff vi reaches vj ; that is, the distance between vi and vj is less than r;
r w(e) ≤ r is the weight of edge e = i, j, that is, the distance between vi and vj .
In an Euclidean graph, each node has a coordinate (xi , yi ) ∈ R2 in two-dimensional
space, which represents the location of the node i in Q. For the sake of simplicity,
we will only consider two dimensions in this chapter, but the methods explained here
can be easily extended to provide position information in three dimensions.
Some terms can be used to designate the current state of a node:
1 A Euclidean graph is a weighted graph where the weights are equal to the Euclidean lengths of the edges.
INTRODUCTION
309
Definition 1 (Unknown Nodes, U). Also known as free or dumb nodes, U refers to
the nodes of the network that do not know their localization information. To allow
these nodes to estimate their position is the main goal of the localization systems.
Definition 2 (Settled Nodes, S). These nodes were initially unknown nodes, but
managed to estimate their positions using the localization system. The number of
settled nodes and the estimated position error of these nodes are the main parameters
of quality of a localization system.
Definition 3 (Beacon Nodes, B). Also known as landmarks or anchors, these are
the nodes that do not need the localization system in order to estimate their physical
positions. Their localization is obtained by manual placement or by external means
such as the GPS. These nodes form the base of the majority localization systems for
WSN.
Definition 4 (Reference Nodes, R). These are the nodes in which the localization
information will be used by an unknown node to estimate its location. A reference
node must also be a beacon or a settled node.
The localization problem can then be simplified and defined by the following
definition.
Definition 5 (Localization Problem). Given a multihop network G = (V, E) and a
set of beacon nodes B and their positions (xb , yb ), for all b ∈ B, we want to find the
position xu of almost all unknown node u ∈ U, transforming these unknown nodes
into settled nodes S.
11.1.2 Importance of the Localization Systems
The position of sensor nodes need not to be engineered or predetermined. This allows
random deployment in inaccessible terrains or disaster relief operations [1]. Thus,
a localization system is required in order to provide the position information to the
nodes. The importance of the localization information arise from several factors, many
of them related only to the WSNs. Some of these factors include:
r Gathered Data Identification: This consists in mapping the data/events to their
location of gather/occurrence. One of the main goals of a WSN is to monitor an
area of interest. However, once the data are collected, it becomes important to
identify the region to which these data belong.
r Gathered Data Correlation: This allows the intermediate nodes to correlate and
perform information fusion of the data gathered on the same region while these
data are forwarded through the network [17–19].
r Nodes Addressing: This refers to the possibility of using the physical location
of the nodes that have their unique identification in the network [20, 21].
310
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
r Network Management: This allows the management and query of nodes localized in a determined region [22], evaluation of the nodes coverage [23], and
energy maps generation [24].
r Geographic Algorithms: These are algorithms that use the localization information of the nodes to optimize the use of the network resources. Some of these
algorithms include routing [22, 25–27], density control [23], and object tracking [28] algorithms.
11.1.3 Requirements of a Localization System
Due to its limitations and applications, the WSNs impose a number of prerequisites
that must be taken into consideration by a localization system. Some of these prerequisites include:
r Auto-organization, Which is independent of any infrastructure.
r Scalability, Where the algorithms can be applicable to large-scale and/or dense
sensor networks.
r Robustness, which consists of tolerance to communication problems and also to
inaccurate distance and position information.
r Efficiency on using the network resources, because even being indispensable to
the most WSNs, the localization system is not the main goal of these networks.
11.1.4 The Components of the Localization Systems
The localization systems can be divided into three distinct components:
1. Distance/Angle Estimation: This component is responsible for estimating information of distances and/or angles between two nodes. Such information will
be used by the other components of the localization system.
2. Position Computation: This component is responsible for computing a node’s
position based on the available information of distances/angles and positions of
the reference nodes.
3. Localization Algorithm: This is the main component of a localization system. It
determines how the available information will be manipulated in order to allow
most or all the nodes of the WSN to estimate their positions.
Figure 11.2 depicts this component division. Besides being a didactic viewpoint,
the importance of such a division into components comes, as we will see, from the
need to recognize that the final performance of the localization systems depends
directly on each one of these components. Also, each component has its own goal
and methods of solution. They can be seen as subareas of the localization problem
that need to be separately analyzed and studied.
DISTANCE/ANGLE ESTIMATION
311
Figure 11.2. The division of the localization systems into three distinct components:
distance/angle estimation, position computation, and localization algorithm. The arrows indicate the dependency relations—that is, the information flow from one component to another.
The rest of this chapter defines and discusses each one of these components by
showing, analyzing, and comparing the main methods used by each one of them.
This way, in Section 11.2 the first component, the Distance/Angle Estimation, and
its techniques are studied. Section 11.3 presents the second component, the Position
Computation, also with some of its techniques. The third component, the Localization
Algorithm, is discussed in Section 11.4 along with some known localization algorithms proposed in the literature. Finally, Section 11.5 presents a summary along with
some concluding remarks.
11.2 DISTANCE/ANGLE ESTIMATION
The distance/angle estimation is obtained by identifying the distance or angle between
two nodes. Such estimates constitute an important component of the localization
systems, because they are used by the position computation and also by the localization
algorithm.
Different methods can be used to estimate such information. Some of them are
very accurate, but with higher costs (in terms of hardware, energy, and processor
resources), while others are not accurate, but already available on most sensor nodes.
In the following sections, some of the main methods used by the localization
systems to estimate distances/angles will be studied. These methods include RSSI,
ToA/TDoA, AoA, and the communication range. Finally, some general comments
will be made regarding this component and its methods of estimation.
11.2.1 Received Signal Strength Indicator (RSSI)
RSSI derives the distance between two nodes based on the strength of the signal
received by one of the nodes. As depicted in Figure 11.3, a sender node sends a signal
with a determined strength that gets reduced as this signal is propagated. The greater
312
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
Figure 11.3. Decrease in the signal strength. The signal is sent with a determined strength
that decreases, theoretically, proportionally with the square of the distance. In the real world,
a number of factors collaborate in the faster and varied decrease in this signal strength.
the distance of the receiver node, the lower the signal strength when it arrives this
node.
The signal strength is commonly measured in dBm (decibel in reference to one
milliwatt) or in watts. Theoretically, this signal strength decreases as the inverse of
the squared distance, and a known radio propagation model (Table 11.1) can be used
to convert the signal strength into distance. However, in real-world environments, this
indication is highly influenced by noises, by obstacles, and by the type of the antenna,
which makes it hard to be modeled by a mathematical formula. In these cases, it is
normal to make a system calibration [29], where values of RSSI and distances are
previously evaluated in a controlled environment.
This method, like the others, has some advantages and some disadvantages. The
main advantage is its low cost, because most receivers are capable of estimating the
received signal strength. The disadvantage of this method is that it is very subject to
noises and interferences, which results in higher inaccuracies on the distance estimations. Some experiments [6] show errors from 2 to 3 m in scenarios where all nodes
are placed in a plane field, 1.5 m from the ground, and with a communication range
of 10 m.
Although the RSSI show some plausible results in simulations and controlled experiments, its use in real-world applications is still questionable [32]. But, considering
TABLE 11.1. Two Known Radio Propagation Models
Model
Description
Formula
Free Space [30]
Consider the ideal propagation condition
without interferences or obstacles.
Pr (d) =
Two-Ray Ground [31]
Like the Free Space, but consider the
possibility of signal reflection in the ground.
Pr (d) =
Pt Gt Gr λ2
(4π)2 d 2 L
Pt Gt Gr h2t h2r
d4 L
Pt and Pr are the transmitted and received signal power (strength), Gt and Gr are the antenna gains of the
transmitter and receiver, λ is the wavelength, d is the distance between the nodes, L is the system loss, and
ht and hr are the heights of the transmit and receive antennas.
DISTANCE/ANGLE ESTIMATION
313
its low cost, it is possible that a more sophisticated and precise use of the RSSI (e.g.,
with better transmitters) could become the most used technology of distance estimation by the cost/precision viewpoint [33]. But this technology is not yet available.
11.2.2 Time (Difference) of Arrival (ToA/TDoA)
Different methods try to estimate the distance between two nodes using time-based
measures. The most simple and intuitive is the ToA (Time of Arrival) [34]. In this
case, the distance between two nodes is directly proportional to the time that the signal
takes to propagate from one point to another. This way, if a signal was sent in time
t1 and reached the receiver node in time t2 , the distance between the sender and the
receiver is
d = sr (t2 − t1 )
(11.1)
where sr is the propagation speed of the radio signal (speed of light), and t1 and
t2 are the times when the signal was sent and received (Figure 11.4a). This type of
estimation requires precisely synchronized nodes [35], and the time when the signal
leaves the node must be in the packet that is sent.
The TDoA (Time Difference of Arrival) is based on (a) the difference of the times
that a single signal from a single node arrives in three or more nodes or (b) the
difference of the times that multiple signals from a single node arrive in another node.
The first case, more common in cellular networks, requires precisely synchronized
receiver nodes (in this case, base stations). In the second case, more common and
suitable for WSNs, the nodes must be equipped with an extra hardware capable
of sending two types of signals simultaneously. These signals must have different
propagation speeds, like radio/ultrasound [36] or radio/acoustic [29]. Usually, the
first signal is the packet itself, which propagates in the speed of light (∼ 300, 000
Figure 11.4. Methods to derive the distance from the signal arrival time: (a) ToA (Time of
Arrival), where the time that the signal takes to leave the transmitter and arrive in the receiver
is computed. (b) TDoA (Time Difference of Arrival), where the difference of the arrival times
of two different signals sent simultaneously is used.
314
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
km/s), and the second signal is some kind of sound, because of its slower propagation
(≈340 m/s), which is six orders of magnitude slower then the first signal.
An example of TDoA that is suitable for WSN is used by reference 6 and depicted
in Figure 11.4b, where an ultrasound pulse is sent simultaneously with a radio signal.
In this case, nodes compute the difference in the time of arrivals of the two signals.
The distance can now be computed by the following formula:
d = (sr − ss ) ∗ (t2 − t1 )
(11.2)
where sr and ss are the propagation speed of the radio and ultrasound signals, and t1
and t2 are the arrival times of the radio and ultrasound signals, respectively.
The errors in the distance estimations obtained by TDoA are measured in centimeters. Experiments with ultrasound made in reference 6 indicate errors of about 2
or 3 cm (smaller than the sensor node) in scenarios where the nodes were in a distance
of 3 m. In reference 37 acoustic sound was studied and the results showed errors of
abouts 23 cm, with nodes in a distance of 2 m.
Despite the lower errors, these systems have some disadvantages. The first is the
need of extra hardware to send the second signal, which increases the node cost. The
second disadvantage is the range of the second signal, which is normally low between
3 m and 10 m with more powerful transmitters.
11.2.3 Angle/Direction of Arrival (AoA/DoA)
The angle of arrival of the signal [36, 38] can also be used by the localization systems.
This angle can be in relation to the node itself, in relation to a electronic compass, or
in relation to a second signal received by the node.
The estimation of the angle of arrival is done by using directive antennas or using
an array of receivers—usually three or more—uniformly separated. In the last case,
based on the times of arrival of the signal to each one of the receivers, it becomes
possible to estimate the angle of arrival of this signal (Figure 11.5).
Figure 11.5. Angle of arrival of the signal. The difference in the times of arrival of the signal
to each one of the receivers, as well as the difference in the position of these receivers, allows
the node to estimate the angle of arrival of the signal.
POSITION COMPUTATION
315
Experiments show that this method has an inaccuracy of a few degrees (about 5◦ in
reference 36. The need of extra hardware and the need of a minimal distance among
the receivers result in some disadvantages in terms of cost and size of the sensor
nodes.
11.2.4 Communication Range
In some cases, the only information available to estimate a distance is the communication range of the sensor nodes. If a node receives a data packet from another node, then
the distance between these nodes is between zero and the maximum communication
range.
Usually, techniques that use this method of distance estimation do not need an
accurate distance, but only an interval. To get only one distance (and not an interval),
we can choose one point from the interval, like the middle point, for example. In this
last case, the maximum error of this estimation will be one-half the communication
range.
This method of distance estimation has the advantage of being the most simple and
with the lowest cost. No extra hardware is required, nor is extra computation needed
to estimate a distance. On the other hand, an error of half the communication range
for each distance estimation is not viable for the most localization systems. Consider,
for example, a communication range of 100 m. In this case, the error of this method
can be about 50 m for each distance estimation.
11.2.5 Comments About the Distance/Angle Estimation
The choice of what method to use to estimate distance between nodes in a localization
system is an important factor that influences the final performance of the system.
Usually, as will be shown in the next section, to estimate a position, a node uses at
least three distance estimations, each of them with an associated error. On the other
hand, if only the accuracy of such methods were important, we could just use a TDoA
that has lower errors. But factors such as the size and cost (in terms of hardware,
processor, and energy) of the nodes must also be taken into consideration. Thus, the
chosen method used to estimate distances will depend on the application requirements
and also on the available resources. Table 11.2 compares each one of the methods
described in this section.
11.3 POSITION COMPUTATION
When a node has enough information of distances and/or angles and positions, it can
compute its own position using one of the methods that will be studied in this section.
Several methods can be used to compute the position of a node. Such methods
includes trilateration, multilateration, triangulation, probabilistic approaches, bounding box, and the central position. The choice of which method to use also impacts the
final performance of the localization system. Such a choice depends on the available
316
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
TABLE 11.2. Comparison of the Methods Used to Estimate Distances/Angles
Between Two Nodesa
Method
RSSI
ToA
Precision
Meters (2–4 m)
Centimeters
(2–3 cm)
TDoA
Centimeters
(2–3 cm)
AoA
A few degrees
(5◦ )
Communication Half the
Range
communication
range
a The
Maximum
Distance
Extra
Hardware
Communication
range
None
Communication
range
few meters
(2–10 m)
Communication
range
Communication
range
None
Ultrasound
transmitter
Set of
receivers
None
Challenges
Variation
of the RSSI,
interferences
Nodes
synchronization
Maximum
distance of work
Work on small
sensor nodes
—
chosen method depends on the application, scenario, and available resources.
information and on the processor limitations. In the subsequent sections, the cited
methods will be studied. After that, some general comments will be made regarding
this component and its methods.
11.3.1 Trilateration and Multilateration
Trilateration is the most basic and intuitive method. This method computes a node
position by the intersection of three circles, as depicted in Figure 11.6. To estimate
Figure 11.6. Theoretical model of the trilateration: The position of the unknown node corresponds to the intersection of the three circles formed by the positions and distances to the
reference nodes.
POSITION COMPUTATION
317
its position using trilateration, a node needs to know the positions of three reference
nodes and the distance to each of these nodes. The distances can be estimated using
one of the methods explained in the previous section.
The circles formed by the position and distance to each one of the references can
be represented by the formulas:
(x̂ − x1 )2 + (ŷ − y1 )2 = d12
(x̂ − x2 )2 + (ŷ − y2 )2 = d22
(x̂ − x3 )2 + (ŷ − y3 )2 = d32
where (x̂,ŷ) is the position we want to compute, (xi ,yi ) is the position of the ith
reference, and di is the distance of the ith reference node to the unknown node. In
this case, we have three quadratic equation with two unknowns, which can be solved,
theoretically, into one solution.
In real-world applications, the distance estimation inaccuracies as well as the inaccurate position information of the reference nodes make it difficult to compute the
position. As depicted in Figure 11.7a, the circles do not intersect into only one point,
resulting in an infinite set of possible solutions.
Furthermore, when a larger number of reference points are available, we can use
the multilateration to compute the nodes position. In this case, an overdetermined
system of equations, where we have more equations than unknowns, must be solved.
Figure 11.8 depicts this case. Usually, overdetermined systems do not have a unique
Figure 11.7. A more realistic model of the trilateration: (a) The inconsistencies of the positions
and distances generate a system with infinite solutions. (b) The residual value, as the sum of
the squared differences between the estimated and computed distances.
318
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
Figure 11.8. Multilateration: similar to the trilateration, but more than three references can be
used and an overdetermined system is solved. A method to solve this system includes the least
squares method.
solution. When considering n reference points and also the error of the distance
estimations, which makes di = d̂i − ǫ, the system of equations becomes
(x̂ − x1 )2 + (ŷ − y1 )2 = d̂12 − ǫ
..
.
(x̂ − xn )2 + (ŷ − yn )2 = d̂n2 − ǫ
where ǫ is normally considered to be an independent normal random variable with zero
mean. This system can be linearized, by subtracting the last equation, into Ax ≈ b,
or
2(x1 − xn ) 2(y1 − yn )
x12 − xn2 + y12 − yn2 + dn2 − d12
x̂
..
..
..
.
.
.
ŷ ≈
2
2
2
2(xn−1 − xn ) 2(yn−1 − yn )
xn−1
+ dn2
− yn2 − dn−1
− xn2 + yn−1
(11.3)
This linear system can be easily solved using standard methods like the least
squares approach [39, 40]. This can be done by the following parameter estimation:
x = (AT A)−1 (AT b)
(11.4)
The main idea of this method is to minimize the sum of the squares of the differences
between the estimated (e.g., using the RSSI) and the computed distances (using the
estimated position). This sum of the differences is known as the residuals, as depicted
in Figure 11.7b. In mathematical terms, the computed position is
n
(x̂, ŷ) = min
i=1
2
(x̂ − xi )2 + (ŷ − yi )2 − di
(11.5)
POSITION COMPUTATION
319
where (xi , yi ) is the position of the ith reference node, and di is the estimated distance. (x̂ − xi )2 + (ŷ − yi )2 is the distance between the computed position and the
position of the ith reference node, which is the computed distance.
The number of floating point operations needed to compute a position depends on
the method used to solve the system of equations. In the case of the method of least
squares, (m + n/3)n2 floating point operations (where m is the number of unknowns
and n is the number of equations) are required to estimate a position [39].
11.3.2 Triangulation
In triangulation [36, 38], information of angles are used instead of distances. The
position computation can be made remotely or by the node itself. In both cases, the
position is computed using the trigonometry laws of sines and cosines.
In the first case—remote positioning, depicted in Figure 11.9a—at least two reference nodes estimate the angle of arrival and remotely compute the position of the
unknown node as the point where the lines along the angles from each reference node
intersect. This type of triangulation is mostly used in cellular networks.
But for sensor networks, the most important is that the node itself computes its
own position. In this case, depicted in Figure 11.9b, at least three reference nodes are
required. The unknown node estimates its angle to each one of the three reference
nodes and, based on these angles and on the positions of the reference nodes (which
form a triangle), computes its own position using simple trigonometrical relationships.
This technique is similar to trilateration. In fact, based on the angles of arrival, it is
possible to derive the distances to the reference nodes [38].
11.3.3 Probabilistic Approaches
The uncertainty in the distance estimations motivated the appearance of probabilistic
approaches to compute a node position. In the probabilistic approach, the position
computation does not result in one single point, like in the other cases, but in a set of
points and their probabilities of being the real position of the unknown node.
Figure 11.9. Triangulation. (a) Remote positioning: The reference nodes exchange information between them to compute the position of the unknown node using the angle of arrival.
(b) self-positioning: Where the unknown node estimates its angle to at least three reference
nodes, the unknown node itself is able to compute its own position.
320
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
A example of a probabilistic approach is proposed in reference 41. In this work,
the error of the distance estimation is modeled as a normal random variable. When
an unknown node receives a packet from a reference node, it can be in any place
around the reference node with probabilities as depicted in Figure 11.10a. When
Figure 11.10. Probabilistic approach. (a) Probability of a node’s position using one reference node: The unknown node has equal probability of being on around the reference node.
(b) Probability when using two reference nodes: A pair of points with greater probabilities
results from the system. (c) probabilities when using more than two references: A unique point
with greater probability results from the system.
POSITION COMPUTATION
321
Figure 11.10. (Continued)
another packet is received from another reference node, the unknown node computes
its position again as depicted in Figure 11.10b. When new position information of
other nodes is received, it becomes possible to identify the probable location of the
unknown node, as depicted in Figure 11.10c .
When an application requires a single position, the point with greater probability
can be computed. The main problem of this approach is the high computational cost
and the space required to store the information. In reference 42 it is shown that if
we consider the sample size as a grid of d × d, the complexity of this method would
be O(3d 2 ). One possible application of this method consists in sending the gathered
information to a central and more powerful node in order to compute the positions.
11.3.4 Bounding Box
Bounding box, proposed in reference 43, uses squares—instead of circles as in
trilateration—to bound the possible positions of a node. An example of this method
is depicted in Figure 11.11.
For each reference node i, a bounding box is defined as a square with center in the
position of this node (xi ,yi ), with sides of size 2di (where d is the estimated distance)
and with coordinates:
(xi − di , yi − di )
and
(xi + di , yi + di )
(11.6)
The intersection of all bounding boxes can be easily computed without the need
for floating point operations by taking the maximum of the low coordinates and the
322
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
Figure 11.11. Bounding box illustration. The intersection of all bounding boxes (shaded) is
computed without the need for floating point operations. The position of the node is computed
as the center of this rectangle.
minimum of the high coordinates of all bounding boxes:
(max(xi − di ), max(yi − di ))
and
(min(xi + di ), min(yi + di ))
(11.7)
This is the shaded rectangle in Figure 11.11. The final position of the unknown node
is then computed as the center of the intersection of all bounding boxes:
(x̂, ŷ) =
max(xi − di ) + min(xi + di ) max(yi − di ) + min(yi + di )
,
2
2
(11.8)
Despite the final error of this method, which is greater than that of trilateration,
a smaller amount of processor resources is used compute the intersection of squares
than to compute the intersection of circles.
11.3.5 Central Position in Relation to the Reference Nodes
By the assumption that the most probable position of a node is the central point among
all the reference nodes, we can compute the position of an unknown node without the
need of estimating distances or angles, but only by using the communication range,
as explained in Section 11.2.4.
POSITION COMPUTATION
323
In this case, the position of a node is computed by using the following equation [32]:
(x̂, ŷ) =
n
i=1 xi
n
,
n
i=1 yi
n
(11.9)
where n is the number of reference nodes.
This method is the most simple in terms of computational resources and required
information. Only 2n + 2 float point operations (where n is the number of reference
nodes) are required to compute a position. On the other hand, the obtained solutions
are not accurate, mainly when the number of reference nodes is low.
11.3.6 Comments About the Position Computation
In previous subsections, we showed a number of methods that can be used to compute
the position of an unknown node based on the gathered information of positions and
distances to the reference nodes.
A number of other methods exist that aim to compute the position of a node.
Location Fingerprinting is a method where the signal characteristics obtained from a
set of locations are cataloged, and then the position computation of a node now consists
in comparing its signal characteristics with the ones that were previously cataloged.
This technique is used by Bahl and Padmanabhan [44] and others indoors localization
systems, but the need for generating a signal signature database makes this technique
unfeasible for the most scenarios of the WSNs. Some works use a variant of the
trilateration that uses only two reference nodes. In this case, two possible solutions,
which is the intersection of two circles, result from the system. The choice between
these two possible positions is made by using extra multihop information or even the
direction of the localization recursion, as done by Oliveira and co-workers [11, 12].
In the APIT algorithm, He et al. [32] use triangles formed by three beacon nodes, and
a node decides if it is inside or outside these triangles by comparing its signal strength
measurements with the measurements of its neighbors. The position of the node is
computed by finding the centroid of the intersection of the beacon triangles that the
node is inside. Other works concentrate all information of distances among the nodes
into a central node and uses mathematical optimization techniques to compute the
positions of the nodes. As examples, we have (a) the work of Doherty et al. [5], which
formulated the localization problem as a convex optimization problem based only in
connectivity-induced constraints and used semidefinite program (SDP) to solve the
problem, and (b) the work of Shang et al. [7, 45] that used multidimensional scale
(MDS).
The choice of what method to use can also impact the final performance of the
localization system. In the next section, some localization algorithms will be studied.
Depending on the used localization algorithm, this error in the position computation
can harm in greater or minor degree the localization system as a whole. In some
algorithms, for example, the newly computed positions are used to assist the other
unknown nodes to compute their positions. In this case, a small error in the position
computation can result in a localization system with high errors.
324
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
TABLE 11.3. Comparison of the Methods Used to Compute Positionsa
Method
# Refs
Dist? Angle?
Comp.
Challenges
Trilateration
3
Yes
No
O(1)
Multilateration
n≥3
Yes
No
O(n3 )
Susceptible to inaccurate
distances
Computational complexity
Triangulation
Probabilistic
3
n≥3
No
Yes
Yes
No
Bounding Box
Central Position
n≥2
n≥1
Yes
No
No
No
O(1)
O(3d 2 )
(d=grid)
O(n)
O(n)
Require extra hardware
Computational and space
complexity
Final Position Error
Final position error
a Again, there is no ideal solution that works in all scenarios. The choice of what method to use will depend
on the gathered information and on the available processor resources.
The information of positions and distances gathered by a node and the available
processor resources also restrict the choice of what method to used. Table 11.3 summarizes and compares the main characteristics of the position computation methods
explained in this section.
11.4 LOCALIZATION ALGORITHM
The localization algorithm is the main component of a localization system. This component determines how the information of distances and positions will be manipulated
in order to allow most or all the nodes of the WSN to estimate their positions.
The localization algorithms can be classified into some categories:
r Distributed or Centralized Position Computation. The positions of the nodes
can be computed in a distributed way by the network nodes (selfpositioning) [8, 9, 34] or by a single central node (e.g., a more powerful node,
the sink node—remote positioning) [5, 46].
r With or Without an Infrastructure. If there is no need for infrastructure [8, 9] or
if there is the need to redesign the previous infrastructure in order to allow the
functioning of the localization algorithm (e.g., manual placement of the beacon
nodes) [34, 36, 47],
r Relative or Absolute Positioning. The computed positions can be related to
global coordinates (e.g., latitude, longitude) [9, 34] or related to a node or point
of the network [48, 49].
r Indoor or Outdoor Scenarios. If the system is more appropriated for indoors [36, 47] or outdoors [34] scenarios.
r One Hop or Multi Hop. If all unknown nodes have direct communication with
the beacon nodes [34, 47] or if a multihop communication is needed [8, 9].
LOCALIZATION ALGORITHM
325
Many times in this chapter, the word “performance” was used to make reference
to aspects of quality of the localization systems. The following aspects can be used
to evaluate this performance:
r Mean Error and Consistence. This identifies is the mean error of the position
estimates and also shows if this mean error is repeated in similar but different scenarios (consistence of the mean). This mean error limits the usage of
the localization system to the applications where this level of inaccuracies is
acceptable.
r Communication Cost. This identifies the algorithm complexity in terms of packets exchanged. It also identifies the cost of the localization system to the sensor
network.
r Number of Settled Nodes. This defines the percentage of network nodes that
were able to compute their positions at the end of the localization algorithm.
The ideal is that all nodes should be able to compute their position, but in many
cases it is not possible.
r Number of Beacon Nodes. This identifies the number of beacon nodes required
to make the algorithm work. Beacon nodes are usually more expensive than the
normal nodes, and their usage should be minimized.
Some network characteristics can affect the performance of the localization systems. It is important to make experiments for each proposed localization system to
evaluate their behaviors when varying these characteristics, which includes:
r Network Density. In high dense networks, we have lower distances among the
nodes, which results in lower errors in distance estimations and also in the error
of the localization system. Besides, the higher number of neighbors results in
more information that can be used by an unknown node to better compute its
position.
r Network Scale. Increasing the number of nodes (and keeping the network density, which increases the area of the sensor field) results in a higher number
of hops. Usually, a higher number of hops results in more inaccurate positions
computation, increasing the mean error of the localization system.
r Number of Beacon Nodes. When deploying a higher number of beacon nodes
in the network, the mean error of the localization system tends to decrease and
the number of settled nodes tends to increase.
r GPS Accuracy. Although considered by many works, the GPS does not provide
perfect localization, especially in sensor networks. Because most of the beacon
nodes use the GPS to get their position, the GPS accuracy will impact the final
position error of the localization systems that depend on this service.
In the next sections, some proposed localization algorithms will be studied and
analyzed. These algorithms are: Ad Hoc Positioning System, Recursive Position
Estimation, Localization with a Mobile Beacon, Global Positioning System, and the
326
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
Cricket Location Support System. After that, some general comments will be made
regarding this component and its methods.
11.4.1 Ad Hoc Positioning System (APS)
In the APS [8], a reduced number of beacon nodes (e.g., three or more) are deployed
with the unknown nodes. The author proposes, then, a localization algorithm where
each node estimates its distance to the beacon nodes in a multihop way. Once these distances are estimated, the nodes can compute their positions using trilateration. Three
methods of hop-by-hop distance propagation are proposed: DV-Hop, DV-Distance,
and Euclidean.
In the DV-Hop, the beacon nodes start the propagation of their position information (Figure 11.12a). Working as an extension of the distance vector algorithm, all
nodes receive the position information of all beacon nodes as well as the number of
hops to these beacons. When a beacon node receives a position information from the
other beacon nodes, it has enough information to compute the average size of one
hop based on its own position, on the position of the other beacon nodes, and also
on the number of hops among them (Figure 11.12b). This last value is then flooded
in a controlled way into the network as a correction factor. When an unknown node
receives the correction, it is able to convert its distance to the beacon nodes from
number of hops to meters (Figure 11.12c). The complete DV-Hop algorithm is shown
in Algorithm 1. The complexity of message exchanging of this algorithm is driven
by the total number of beacon and normal nodes, which is O(n ∗ (m + 1)), where n
is the number of nodes and m is the number of beacon nodes.
DV-Distance works like DV-Hop. But, instead of propagating the number of
hops, it propagates the estimated distances (e.g., using RSSI); and each node, before
forwarding the position information of the beacon nodes, adds its estimated distance
to the one contained in the packet and then forwards this packet. In this case, there is
Figure 11.12. Example and phases of the APS: DV-Hop. (a) Initially, the beacon nodes broadcast their positions and (b) Compute the average size of hop. (c) This last value is sent to the
network, and the nodes receiving it compute their positions.
LOCALIZATION ALGORITHM
327
ALGORITHM 1. APS DV-HoP Localization Algorithm
䉯 Variables:
1: positionsi = ∅; {Set of beacon positions.}
2: correctioni = −1; {Correction computed/received by this node.}
䉯 Input:
3: msgi = nil.
Action:
4: if ni ∈ B then {If this node is a beacon node.}
5:
(xi , yi ) := getGpsPosition();
6:
Send beaconPos(i, xi , yi , 0) to all nj ∈ Neigi .
7: end if
䉯 Input:
8: msgi = beaconPos(k, xk , yk , hk ).
Action:
9: if k ∈
/ Ri then {If this node did not already receive this packet.}
10:
Ri := Ri ∪ {k};
11:
positionsi := positionsi ∪ {(xk , yk , hk )};
12:
if ni ∈ B then {If this is a beacon node, wait for more position packets}
Re
Start waitTimer.
14:
end if
15:
Send beaconPos(k, xk , yk , hk + 1) to all nj ∈ Neigi .
16: end if
䉯 Input:
17: waitTimer timeout. {Executed only by the beacon nodes.}
Action:
18: correctioni =
√
(xi −xj )2 +(yi −yj )2
hj
19: Send correctioni to all nj ∈ Neigi .
for all (xj , yj , hj ) ∈ positionsi ;
䉯 Input:
20: msgi = correctionk .
Action:
21: if ni ∈ U then {If this node is an unknown node.}
22:
correctioni := correctionk ;
23:
hj := hj ∗ correctioni for all (xj , yj , hj ) ∈ positionsi ;
24:
(xi , yi ) := positionComputation(positionsi ); {Becomes a settled node.}
25:
Send correctioni to all nj ∈ Neigi .
26: end if
no need for a correction factor, because the distances to the beacon nodes are already
in meters.
The Euclidean method works by propagating the true Euclidean distance of the
unknown nodes to the beacon nodes. To allow this, a node needs at least another
328
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
two nodes that already estimated their distances to a beacon node. It also needs its
own distance to these last two nodes, and also the distance between these two nodes.
The distance estimation of the unknown node to the beacon node is made using the
Pythagorean theorem on the triangles generated by the lines that is formed by the
distances.
An advantage of the APS is that its localization algorithm requires a low number
of beacon nodes in order to work. However, the way the distances are propagated,
(especially in the DV-Hop and DV-Distance), as well as the way the distances are converted from hops to meters in DV-Hop, results in an erroneous position computation,
which increases the final localization error of the system.
11.4.2 Recursive Position Estimation (RPE)
In the algorithm of the RPE [9], the nodes estimate their positions based on a set of initial beacon nodes (e.g., 5% of the nodes) using only local information. The localization
information iteratively increases as the newly settled nodes become reference nodes.
The RPE algorithm can be divided into four phases as depicted in Figure 11.13.
In the first phase, a node determines its reference nodes. In the second phase, the
node estimates its distance to the reference nodes using, for example, the RSSI. In
the third phase, the node computes its position using trilateration (becoming a settled
node). In the final phase, the node becomes a reference node by broadcasting its newly
estimated position to its neighbors.
When a node becomes a reference, it can assist the other nodes to also compute their
positions. Figure 11.13 depicts this behavior of the RPE. In Figure 11.13a, node 14
has only two reference nodes (nodes 9 and 13), which does not allow the node to
compute its position. On the other hand, node 8 also does not know its position, but
can compute it (Figure 11.13b), becoming the third reference (Figure 11.13c) that
node 14 needs to compute its own position. The complete RPE algorithm is shown in
Algorithm 2, and its communication complexity is the same as that of flooding.
Figure 11.13. Example and phases of the RPE. (a) Initially, the node chooses its reference
nodes; then (b) this node estimates its distance to each one of the reference nodes, (c) Computes
its position using trilateration, and (d) broadcasts its newly estimated position to assist the other
nodes.
LOCALIZATION ALGORITHM
329
ALGORITHM 2. RPE Localization Algorithm
䉯 Variables:
1: positionsi = ∅ {Set of received positions.}
2: referencesi = ∅ {Set of reference nodes.}
䉯 Input:
3: msgi = nil.
Action:
4: if ni ∈ B then {If this node is a beacon node.}
5:
(xi , yi ) := getGpsPosition();
6:
Send position(xi , yi , 0) to all nj ∈ Neigi .
7: end if
䉯 Input:
8: msgi = position(xk , yk , rk ) such that distk = distanceEstimation(msgi ).
Action:
9: if ni ∈ U then {If this node is an unknown node.}
10:
positionsi := positionsi ∪ {(xk , yk , rk , distk )};
Re
Start waitTimer.
12: end if
䉯 Input:
13: waitTimer timeout.
Action:
14: if size(positionsi ) >= 3 then {If there is enough positions.}
15:
referencesi := chooseThreeBestPositions(positionsi )
16:
(xi , yi , ri ) := positionComputation(referencesi ); {Becomes a settled node.}
17:
Send position(xi , yi , ri ) to all nj ∈ Neigi . {Becomes a reference node.}
18: end if
An advantage of this algorithm is that the number of reference nodes quickly
increases, in such a way that the majority of the nodes can compute their position.
But this technique has the disadvantage of propagating the localization error. This
means that the low inaccurate position estimation of a node can be used by other
nodes to estimate their positions, increasing this inaccuracy. One example of this
problem is that node 14 in Figure 11.13 will get a localization error greater than
that of node 8. Furthermore, a node must have at least three reference neighbors to
compute its position.
11.4.3 Directed Position Estimation (DPE)
By adding some restriction to a recursive localization system like that in the RPE,
we can make the localization recursion go from a single point (recursion origin)
and follow a determined and known direction (Figure 11.14a). Once this behavior
330
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
is guaranteed, it is possible to estimate a node position using only two reference
neighbors.
When a node has a position information of only two reference neighbors, a pair
of possible points results from the system: One is the right position of the unknown
node and the other is a wrong estimate (Figure 11.14b). Once the direction of the
Figure 11.14. (a) The DPE doing a directed localization recursion. and (b) A position estimate
using only two reference neighbors. A pair of possible solutions result from the system. The
right position of the node is the most distant point from the recursion origin.
LOCALIZATION ALGORITHM
331
Figure 11.15. Example and phases of the DPE. (a) First, the beacon nodes starts the recursion.
(b) Then a node determines its (two) reference nodes, (c) estimates its position, and (d) then
becomes a reference by broadcasting this information.
localization recursion is kept, it is easy to choose between the two possible solutions:
The most distant point from the recursion origin is the right position of the unknown
node. This is the base of the DPE [11, 12].
The DPE algorithm is divided into four phases (Figure 11.15). In the first phase,
the recursion of such a system is started from a single point by the beacon structure (Figure 11.15a). In the second phase, a node determines its (two) reference nodes
and estimate its distances to these nodes (Figure 11.15b). In the third phase, the node
computes its position (Figure 11.15c), and then becomes a reference by sending this
information to its neighbors (fourth phase, Figure 11.15d). This way, the recursion
of such a system goes from the center to the edge of the WSN. The complete DPE
algorithm is shown in Algorithm 3, and its communication complexity is the same as
that of flooding.
This approach leads to a localization system that can work in a low dense sensor
network. Besides, the controlled way in which the recursion is made will also lead to
a system with lower and predictable errors. But like the RPE, this technique has the
disadvantage of propagating the localization error.
11.4.4 Localization with a Mobile Beacon (LMB)
Some recent works [42, 46] have proposed the use of mobile beacons to assist the
nodes of the WSN in estimating their positions. A mobile beacon is a node that is
aware of its position (e.g., equipped with a Gps receiver) and that has the ability to
move among the sensor field. This beacon can be a human operator, an unmanned
vehicle, an aircraft, or a robot. The localization algorithm proposed in reference 42
uses mobile beacons to allow the nodes to compute their positions, while in reference
46 the mobile beacon itself computes the position of the nodes.
The system operation in reference 42 is quite simple. Once the nodes are deployed,
the mobile beacon travels through the sensor field broadcasting messages that contain
its current coordinates. When a free node receives more than three messages from the
mobile beacon, it computes its position, using a probabilistic approach, based on the
received coordinates and on the RSSI distance estimations. Figure 11.16 illustrates
332
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
ALGORITHM 3. DPE Localization Algorithm
䉯 Variables:
1: positionsi = ∅ {Set of received positions.}
2: referencesi = ∅ {Set of reference nodes.}
䉯 Input:
3: msgi = nil.
Action:
4: if ni ∈ B then {If this node is a beacon node.}
5:
(xi , yi ) := getGpsPosition();
6:
Send position(xi , yi ) to all nj ∈ Neigi .
7: end if
䉯 Input:
8: msgi = position(xk , yk ) such that distk = distanceEstimation(msgi ).
Action:
9: if ni ∈ U then {If this node is an unknown node.}
10:
positionsi := positionsi ∪ {(xk , yk , distk )};
Re
Start waitTimer.
12: end if
䉯 Input:
13: waitTimer timeout.
Action:
14: if size(positionsi ) >= 2 then {If there is enough positions.}
15:
referencesi := chooseTwoBestPositions(positionsi )
16:
(xi , yi ) := mostDistantFromOrigin(intersectCircles(referencesi )); {Becomes a
settled node.}
Send position(xi , yi ) to all nj ∈ Neigi . {Becomes a reference node.}
18: end if
17:
Figure 11.16. Operation and possible trajectories for the localization with a mobile beacon.
(a) The mobile beacon moving along the sensor field in a straight line. (b) A less rectilinear
trajectory. (c) A trajectory in spiral form.
LOCALIZATION ALGORITHM
333
ALGORITHM 4. MBL Localization Algorithm
䉯 Variables:
1: positionsi = ∅ {Set of position information.}
䉯 Input:
2: msgi = nil.
Action:
3: if ni ∈ U then {If this node is a beacon node.}
4:
StartWalking();
5:
Start posTimer.
6: end if
䉯 Input:
7: posTimer timeout.
Action:
8: (xi , yi ) := getGpsPosition();
9: Send position(xi , yi ) to all nj ∈ Neigi .
10: Restart posTimer.
䉯 Input:
11: msgi = position(xk , yk ) such that distk = distanceEstimation(msgi ).
Action:
12: positionsi := positionsi ∪ {(xk , yk , distk )};
13: if size(positionsi ) >= 3 then {If there is enough references.}
14:
(xi , yi ) := positionComputation(positionsi );
15: end if
this scenario and three possible trajectories for the mobile beacon. Algorithm 4 details
the functioning of the system. The communication cost for the WSN is null, since the
nodes (except the mobile beacon) do not need to send any packets.
An advantage of LMB is that the position estimates are computed based on the
same node (mobile beacon) keeping the mean localization error low and preventing
the propagation of this error. In addition, LMB avoids the use of nodes equipped with
GPS, except for the mobile beacon. On the other hand, in this technique a sensor node
can estimate its position only when the mobile beacon passes near this node, which
may take a long time depending on factors as the size of the sensor field, the beacon
mobility capacity, and the node trajectory. Yet, the mobile beacon may never pass
nearby some nodes, either because of the trajectory or because of a problem with the
mobile beacon.
An important aspect that directly influences the position estimates is the trajectory
of the mobile beacon. The less rectilinear the trajectory, the better the estimates.
The reason is that the lower the collinearity among the reference points, the lower the
estimate error [3]. Thus, rectilinear trajectories such as Figure 11.16a must be avoided.
334
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
11.4.5 The Global Positioning System (GPS)
The GPS [34, 50] is a system composed of 24 satellites in operation that orbit around
the earth. Each satellite circles the earth at a height of 20.200 km and makes two
complete rotations every day. The orbits were defined in a way that in each region of
the earth we could “see” at least four satellites in the sky.
A GPS receiver is able to receive the information constantly sent by the satellites,
estimate its distance to at least four known satellites using ToA, and, finally, compute
its position using trilateration. Once these procedures are executed, the receiver is
able to inform its latitude, longitude, and altitude.
One of the solutions for the localization problem in WSN is to equip each one
of the sensor nodes with a GPS receiver. One of the main advantages would be
the relatively small (2–15 m, depending on the receiver) and precise localization
error, because all nodes would have a similar error. However, this solution has many
disadvantages [5, 6, 8, 48]: (a) The cost and size of the sensor nodes, are increased, (b)
it cannot be used when there is no satellite visible (e.g., indoor scenarios, under water,
under climatic conditions, mars exploration), and (c) the extra hardware consumes
energy. Due to these disadvantages, the usage of the GPS is usually limited to a small
fraction of the nodes (e.g., beacon nodes).
11.4.6 The Cricket Location Support System
Cricket [47] combines active beacons and passive ultrasonic receivers to provide a
localization system. Cricket is designed for mobile nodes in an indoor environment,
but it may be used with static nodes. The active beacons broadcast their location
information over an RF (radio frequency) channel together with an ultrasonic pulse.
The other nodes use the TDoA method to estimate their distances to the beacons.
When the node has enough information of positions and distances, it can compute its
position through multilateration.
To make it possible for the algorithm to work, a grid of beacon nodes must be
previously created so that all nodes have at least three beacon nodes in their communication range. This infrastructure can be considered as a disadvantage of the system,
but Cricket is very accurate and able to work in indoor scenarios.
11.4.7 Comments About the Localization Algorithm
The localization algorithm is the main component of a localization system. This
component defines how the available information that is provided by the beacon nodes,
by the distances estimations, and by the position computations will be manipulated
in order to allow the localization information to expand from the beacon nodes to the
nodes of the sensor network.
As previously mentioned, the localization systems, especially the ones with multihop, had been extensively studied with the advent of the WSN. This way, a number
of other localization algorithms have recently been proposed that focus on different aspects like errors, number of beacons, number of settled nodes, or GPS usage,
FINAL REMARKS
335
TABLE 11.4. Localization Algorithms Comparisona
Algorithm
Number of
Beacons
Position
InfrasComputation tructured?
Positioning Scenarios
Multihop?
APS
RPE
DPE
LMB
GPS
Cricket
≥3
5% nodes
4
1 mobile
—
Grid
Distributive
Distributive
Distributive
Distributive
Distributive
Distributive
Absolute
Absolute
Relative
Absolute
Absolute
Relative
Yes
Yes
Yes
No
No
No
No
No
No
No
Yes
Yes
Outdoors
Outdoors
Outdoors
Outdoors
Outdoors
Indoors
a Some
characteristics of the localization algorithms identify the possible scenarios that can be applicable.
The choice of what algorithm to use depends on the application requirements and on the available resources.
among other things. A number of works try to reduce or completely remove the
need of GPS receivers on beacon nodes. Bulusu et al. [48] manually placed multiple beacon nodes with overlapping coverage regions so the nodes could directly
use the beacons position to compute their positions, but the manual placement of
the beacon nodes is not feasible for the most WSNs scenarios. Capkun et al. [49]
proposes a distributed, infrastructure-free localization algorithm that does not rely
on GPS. The key point of the algorithm is to show that it is possible to build a relative coordinate system without centralized knowledge of the network topology. In
APIT proposed by He et al. [32], beacon nodes equipped with high-powered transmitters in a heterogeneous network broadcast their positions generating a series of
triangles formed by the positions of three beacon nodes. A node can estimate the
triangles in which it is inside and compute its position using the intersection of these
triangles.
The choice of which algorithm to use depends on the available resources, on the
scenario, on the requirements of the application, and also on the mean localization
error acceptable by the nodes. Table 11.4 compares the main characteristics of each
one of these algorithms.
11.5 FINAL REMARKS
In this chapter we have discussed the localization problem for wireless sensor networks. We have divided the design of the localization systems for WSNs into the
following three components:
1. Distance/Angle Estimation
r Received Signal Strength Indicator (RSSI)
r Time [Difference] of Arrival (ToA/TDoA)
r Angle/Direction of Arrival (AoA/DoA)
r Communication Range
336
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
2. Position Computation
r Trilateration
r Multilateration
r Triangulation
r Probabilistic Approaches
r Bounding Box
r Central Position
3. Localization Algorithm
r Ad Hoc Positioning System (Aps)
r Recursive Position Estimation (Rpe)
r Directed Position Estimation (Dpe)
r Localization with a Mobile Beacon (Lmb)
r The Global Positioning System (Gps)
r The Cricket Location Support System
The importance of dividing the localization systems into components, as mentioned
before, comes from the necessity of recognizing that the final performance of the
localization systems depend directly on each one of these components. For example,
a localization system should achieve better results if the TDoA method is used instead
of the RSSI to estimate distances. The same principle applies to the other components.
These components can be seen as subareas of the localization problem that need to
be separately studied.
A general rule in WSN is that there is no perfect solution to a problem that performs
best in any situation or application. The same rule applies to the localization problem. This chapter showed a number of proposed localization systems, each of them
with emphasis in a specific scenario and/or application. The necessity of different
solutions for different applications and also the high number of possible applications
of the WSN have greatly motivated the study and proposals of new solutions to the
localization problem shown in this chapter. However, there are a number of other
localization systems for WSNs which the reader can find in the provided bibliography and in the current literature. For instance, in military applications, WSNs can
be deployed in remote—possibly hostile—environments in order to perform tasks
such as battlefield surveillance, enemy tracking, and security monitoring of military
facilities. In these cases, security techniques must be implemented in order to provide
a secure localization system. These techniques, as shown in Chapter 19, are used to
propose a whole new set of localization systems for WSNs.
11.6 EXERCISES
1. What is the importance of localization systems in wireless sensor networks?
What are the main applications for WSNs that differ from the applications in
Ad Hoc Networks?
BIBLIOGRAPHY
337
2. List and discuss the requirements of a localization system to be implemented
in wireless sensor networks.
3. What is RSSI? What are that main advantages and disadvantages of using this
technique?
4. An unknown node received the positions and estimated the distances from three
different beacon nodes. The received positions were (6,3), (3,4), and (5,6) while
the estimated distances were 2.3, 1.5, and 2, respectively. Show the system of
equations obtained from the received data when computing the unknown node’s
position using trilateration. Solve the system using MATLAB.
5. Why is GPS not a good solution for the localization problem in WSNs?
6. Explain the key ideas of the APS algorithm and identify the main disadvantages
and disadvantages using this APS localization algorithm?
BIBLIOGRAPHY
1. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cyirci. Wireless sensor networks:
A survey. Computer Networks, 38(4): 393–422, 2002.
2. D. Estrin, L. Girod, G. Pottie, and M. Srivastava. Instrumenting the world with wireless
sensor networks. In International Conference on Acoustics, Speech, and Signal Processing
(ICASSP 2001), Salt Lake City, Utah, June 2001, pp. 2033–2036.
3. M. Ilyas and I. Mahgoub. Handbook of Sensor Networks: Compact Wireless and Wired
Sensing Systems, CRC Press LLC, Boca Raton, FL, 2004, Chapter 20.
4. G. J. Pottie and W. J. Kaiser. Wireless integrated network sensors. Communications of the
ACM, 43(5):51–58, 2000.
5. L. Doherty, K. S. J. Pister, and Laurent E. Ghaoui. Convex position estimation in wireless sensor networks. In IEEE Conference on Computer Communications 2001, Vol. 3,
Anchorage, AK, April 2001, pp. 1655–1663.
6. A. Savvides, C.-C. Han, and M. B. Strivastava. Dynamic fine-grained localization in ad-hoc
networks of sensors. In 7th ACM/IEEE International Conference on Mobile Computing
and Networking, Rome, Italy, 2001, pp. 166–179.
7. Y. Shang, W. Ruml, Y. Zhang, and M. P. J. Fromherz. Localization from mere connectivity.
In Proceedings of the 4th ACM International Symposium on Mobile Ad Hoc Networking
and Computing, Annapolis, MD, 2003, pp. 201–212.
8. D. Niculescu and B. Nath, Ad hoc positioning system (aps). In IEEE Global Communications Conference (GlobeCom ’01), San Antonio, TX, USA, November 2001, pp. 2926–
2931.
9. J. Albowicz, A. Chen, and L. Zhang, Recursive position estimation in sensor networks. In
The 9th International Conference on Network Protocols, November 2001, pp. 35–41.
10. N. Bulusu, J. Heidemann, D. Estrin, and T. Tran. Self-configuring localization systems: Design and experimental evaluation. Transactions on Embedded Computing Systems, 3(1):24–
60, 2004.
11. H. A. B. F. Oliveira, E. F. Nakamura, A. A. F. Loureiro, and A. Boukerche. Directed position
estimation: A recursive localization approach for wireless sensor networks. In Proceedings
of the 14th IEEE International Conference on Computer Communications and Networks
338
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
(IC3N ’05), S. R. Thuel, Y. Yang, and E. K. Park, editors, San Diego, CA, October 2005,
IEEE, New York, pp. 557–562.
H. A. B. F. Oliveira, E. F. Nakamura, A. A. F. Loureiro, and A. Boukerche. Error analysis
of localization systems in sensor networks. In Proceedings of the 13th ACM International
Symposium on Geographic Information Systems (GIS ’05), C. Shahabi and O. Boulcema,
editors, Bremen, Germany, November 2005, ACM Press, New York, pp. 71–78.
A. Boukerche, H. A. B. F. Oliveira, E. F. Nakamura, and A. A. F. Loureiro. A Voronoi
approach for scalable and robust dv-hop localization system for sensor networks. In Proceedings of 16th International Conference on, Computer Communications and Networks,
2007 (ICCCN 2007), 2007, pp. 497–502.
A. Boukerche, H. A. B. F. Oliveira, E. F. Nakamura, and A. A. F. Loureiro. Localization
systems for wireless sensor networks. IEEE Wireless Communications—Special Issue on
Wireless Sensor Networks, 14:6–12, 2007.
H. A. B. F. Oliveira, E. F. Nakamura, A. A. F. Loureiro, and A. Boukerche. Localization in
time and space for sensor networks. In AINA ’07: 21st IEEE International Conference on
Advanced Information Networking and Applications, Niagara Falls, Canada, May 2007,
pp. 539–546.
A. Boukerche, H. A. B. F. Oliveira, E. F. Nakamura, and A. A. F. Loureiro. Vehicular
ad hoc networks: A new challenge for localization-based systems. Elsevier Computer
Communications, 31:2838–2849, 2008.
R. R. Brooks and S. S. Iyengar. Multi-Sensor Fusion: Fundamentals and Applications,
Prentice Hall, Englewood Cliffs, NJ, 1998.
C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed diffusion: A scalable and robust
communication paradigm for sensor networks. In Proceedings of the 6th ACM International
Conference on Mobile Computing and Networking (MobiCom ’00), Boston, MA, August
2000, , ACM Press, New York, pp. 56–67.
E. F. Nakamura, C. M. S. Figueiredo, and A. A. F. Loureiro. Information fusion for data
dissemination in self-organizing wireless sensor networks. In Proceedings of the 4th International Conference on Networking (ICN 2005), Reunion Island, April 2005.
J. C. Navas and T. Imielinski, GeoCast—geographic addressing and routing. In Mobile
Computing and Networking, 1997, pp. 66–76.
J. Heidemann, F. Silva, C. Intanagonwiwat, R. Govindan, D. Estrin, and D. Ganesan. Building efficient wireless sensor networks with low-level naming. In SOSP ’01: Proceedings of
the 18th ACM Symposium on Operating Systems Principles, New York, 2001, ACM Press,
New York, pp. 146–159.
Y. Yu, R. Govindan, and D. Estrin. Geographical and energy aware routing: A recursive data
dissemination protocol for wireless sensor networks. Technical Report CSD-TR-01-0023,
UCLA Computer Science Department, 2001.
Y. Xu, J. S. Heidemann, and D. Estrin. Geography-informed energy conservation for ad
hoc routing. In Mobile Computing and Networking, 2001, pp. 70–84.
M. Do Val Machado, O. Goussevskaia, R. Aparecida de Freitas Mini, A. A. F. Loureiro,
G. R. Mateus, and J. M. S. Nogueira. Data dissemination using the energy map. In The
Second Annual Conference on Wireless on demand Network Systems and Services (WONS
2005), St. Moritz, Switzerland, 2005. CD-ROM.
A. Boukerche, H. A. B. F. Oliveira, E. F. Nakamura, and A. A. F. Loureiro. A novel
location-free greedy forward algorithm for wireless sensor networks. In Proceedings of
BIBLIOGRAPHY
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
339
the 2008 IEEE International Conference on Communications (ICC 2008), Beijing, China,
May 2008.
A. Boukerche, H. A. B. F. Oliveira, E. F. Nakamura, and A. A. F. Loureiro. Towards an
integrated solution for node localization and data routing in sensor networks. In ISCC ’07:
12th IEEE Symposium on Computers and Communications, Aveiro, Portugal, July 2007,
pp. 449–454.
Brad Karp and H. T. Kung. Gpsr: greedy perimeter stateless routing for wireless networks.
In Proceedings of the 6th Annual International Conference on Mobile Computing and
Networking, Boston, MA, 2000, pp. 243–254.
S. Kumar, C. Alaettinglu, and D. Estrin. Scalable object-tracking through unattended techniques (scout). In ICNP ’00: Proceedings of the 2000 International Conference on Network
Protocols, Washington, DC, 2000, IEEE Computer Society, New York, p. 253.
K. Whitehouse and D. Culler. Calibration as parameter estimation in sensor networks.
In WSNA ’02: Proceedings of the 1st ACM International Workshop on Wireless Sensor
Networks and Applications, New York, 2002, ACM Press, New York, pp. 59–67.
H. T. Friis. A note on a simple transmission formula. In Proceedings IRE, 1946, p. 34.
T. S. Rappaport. Wireless Communications, Principles and Practice, Prentice Hall, Englewood Cliffs, NJ, 1996.
T. He, C. Huang, B. M. Blum, J. A. Stankovic, and T. Abdelzaher. Range-free localization
schemes for large scale sensor networks. In MobiCom ’03: Proceedings of the 9th Annual
International Conference on Mobile Computing and Networking, New York, 2003, ACM
Press, New York, pp. 81–95.
J. Bachrach and C. Taylor. Localization in sensor networks. In Handbook of Sensor Networks: Algorithms and Architectures, I. Stojmenovic, editors, John Wiley & Sons, Hoboken, NJ, 2005.
B. Hofmann-Wellenho, H. Lichtenegger, and J. Collins. Global Positioning System: Theory
and Practice, 4th edition, Springer-Verlag, Berlin, 1997.
J. Elson. Time synchronization in wireless sensor networks. Department Computer Sciences, University of California, Ph.D. dissertation, Los Angeles, 2003.
N. B. Priyantha, A. K. Miu, H. Balakrishnan, and S. Teller. The cricket compass for contextaware mobile applications. In 7th ACM International Conference on Mobile Computing
and Networking. Rome, Italy, July 2001. CD-ROM.
K. Whitehouse. The design of calamari: An ad hoc localization system for sensor networks.
M.S. thesis, University of California at Berkeley, 2002.
D. Niculescu and B. Nath. Ad hoc positioning system (aps) using aoa. In Proceedings of
INFOCOM 2003, San Francisco, CA, 2003.
G. H. Golub and C. F. Van Loan. Matrix Computations, 3rd edition, Johns Hopkins University Press, Baltimore, MD, 1996.
R. Jain. The Art of Computer Systems Performance Analysis: Techniques for Experimental
Design, Measurement, Simulation, and Modeling, John Wiley & Sons, New York, 1991.
V. Ramadurai and M. L. Sichitiu. Localization in wireless sensor networks: A probabilistic
approach. In Proceedings of the 2003 International Conference on Wireless Networks
(ICWN 2003), Las Vegas, NV, June 2003, pp. 275–281.
M. L. Sichitiu and V. Ramadurai. Localization of wireless sensor networks with a mobile
beacon. In Proceedings of the 1st IEEE International Conference on Mobile Ad Hoc and
Sensor Systems (MASS 2004), FL, October 2004, pp. 174–183.
340
LOCALIZATION SYSTEMS FOR WIRELESS SENSOR NETWORKS
43. S. Simic and S. Sastry. Distributed localization in wireless ad hoc networks. Technical
Report UCB/ERL M02/26, UC Berkeley, 2002.
44. P. Bahl and V. N. Padmanabhan. Radar: an in-building rf-based user location and tracking
system. In Proceedings of the IEEE Infocom 2000, Vol. 2, Tel Aviv, Israel, March 2000,
IEEE, New York, pp. 775–784.
45. Y. Shang and W. Ruml. Improved mds-based localization. In IEEE Conference on Computer
Communications 2004, Vol. 4, March 2004, pp. 2640–2651.
46. P. Pathirana, N. Bulusu, S. Jha, and A. Savkin. Node localization using mobile robots in
delay-tolerant sensor networks. IEEE Transactions on Mobile Computing, 4(3):285–296,
2005.
47. N. B. Priyantha, A. Chakraborty, and H. Balakrishnan. The cricket location-support system.
In Mobile Computing and Networking, Boston, MA, August 2000, pp. 32–43.
48. N. Bulusu, J. Heidemann, and D. Estrin. Gps-less low cost outdoor localization for very
small devices. IEEE Personal Communications Magazine, 7(5):28–34, 2000.
49. S. Capkun, M. Hamdi, and J.-P. Hubaux. Gps-free positioning in mobile ad hoc networks.
Cluster Computing, 5(2):157–167, 2002.
50. E. D. Kaplan, editors. Understanding GPS: Principles and Applications, Artech House,
Norwood, MA, 1996.
CHAPTER 12
Location Discovery in Sensor
Networks
ASIS NASIPURI
Department of Electrical and Computer Engineering,
The University of North Carolina at Charlotte, Charlotte, NC 28223
12.1 INTRODUCTION
Location estimation is a fairly mature topic that has interested researchers for many
years for applications such as maritime and aircraft navigation, transportation, geological explorations, robotics, tactical missions, emergency systems, and more. It led
to the development of a number of different location discovery systems that use a variety of techniques. Location discovery techniques range from using the direction and
elevation of stars to sophisticated distance ranging techniques applied to RF, acoustic,
and ultrasonic signals. The most popular of these location estimation systems is perhaps the geographical positioning system (GPS), in which a device uses RF signals
received from a system of satellites to calculate its absolute geographical location.
Currently, GPS receivers cost less than $100 and can provide position estimates with
errors within a few tens of meters. Despite these advances, location discovery has
attracted renewed interest in recent times for a special area of application—that of
wireless sensor networks. This is driven by special requirements and constraints in
these networks that prohibit the use of existing technology such as GPS. Consequently,
a number of novel approaches for location discovery have emerged.
Wireless sensor networks is an exciting new concept that emerged from advancements in embedded systems, wireless, and sensor technologies. A wireless sensor is an
embedded system that may have an array of sensors (such as temperature, sound, light,
magnetic field, chemical, etc.) as well as on-board microprocessor and memory that
can be used for signal processing, storage, and on-the-fly networking. Such devices
may be manufactured at low cost and very small form factor, making it possible for
disposable use. A large number of such wireless sensors can form a self-configuring
wireless sensor network for sensing multiple types of signals over wide regions for
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
341
342
LOCATION DISCOVERY IN SENSOR NETWORKS
extended periods of time without direct human intervention. Such networks have
tremendous prospects for distributed monitoring applications at low installation and
operating costs. It is envisioned that in the future, dust-sized smart sensors will be
embedded in the physical world (buildings, roadways, automobiles, etc.) to perform a
variety of tasks that include sensing, computation, communication, and actuation [1].
Examples of applications that have already received attention include environmental
monitoring, agriculture, transportation systems, security systems, military applications, industrial monitoring and control, and many others.
The field of sensor networks opened an array of challenging research problems
that include design issues on sensor technology, RF communications, networking
protocols, information processing, database, algorithm development, power harvesting, and many more. Many of these issues are focused on developing techniques that
enable maximum utilization of the limited resources in individual sensors—namely,
processing, memory, battery, and so on—to optimize the sensing tasks of the network
as a whole. A review of design issues and research directions on sensor networks may
be found in reference 2.
In this chapter we focus on the issue of location discovery, which is a very important aspect in sensor networks. For instance, when a number of networked sensors
are used for environmental monitoring, it is important to know the locations of the
signals obtained (temperature, humidity, etc.). In addition, many of the networking
protocols for sensor networking also require location information. Examples include
geographic query forwarding [3], location aware routing [4], and multisensor information fusion [5]. Since sensor networks consist of a large number of nodes, it is not
practical to deploy the nodes in preestablished locations. Moreover, due to limitations
in hardware, cost, and size, many of the preexisting location estimation techniques
cannot be applied to sensor nodes. Hence, there has been tremendous interest for
developing new low-cost and low-complexity techniques that would enable wireless
sensor nodes to determine their locations after they have been deployed in an area of
interest. This problem is commonly referred to as localization [6] within the scientific
community involved with research on wireless sensor networks. The specific design
challenges, state of the art, and research issues on localization in sensor networks are
presented in this chapter.
12.2 LOCATION DISCOVERY CONCEPTS
Existing localization schemes rely mostly on the principle of triangulation. This refers
to the process of using distances or angles from three reference points to an unknown
point to compute its position on two-dimensional space. For three-dimensional space,
an additional reference point is needed.
Lateration: If the distances from an unknown location to three reference points
are known, the location of the unknown point can be obtained using geometrical calculations known as multilateration as shown in Figure 12.1a. Here, the
unknown location is obtained as the point of intersection of three circles
LOCATION DISCOVERY CONCEPTS
343
N
N
Reference point
B
Reference point
A
A
B
NAP
NBP
R B,U
R A,U
R C ,U
Unknown location
P
Reference point
C
(a)
Unknown location
P
(b)
Figure 12.1. Triangulation using (a) lateration, which shows that the unknown location is
obtained as the intersection of three circles, and (b) angulation, where the unknown location
is obtained from the intersection of lines with appropriate angles.
centered at the reference points having radii equal to the distances from these
reference points.
Angulation: Alternatively, the angles from two known reference points may be
used to determine the location of an unknown point, as shown in Figure 12.1b.
This mechanism is known as angulation. Here, the unknown location is obtained
by the point of intersection of the lines drawn from the reference point with
appropriate angles.
The key issue for implementing a system that employs triangulation is to measure
either the distances or angles from a set of fixed stations acting as reference points.
We next describe the methods that can be applied for measuring these.
12.2.1 Mechanisms for Estimating Distances
Some of the known methods for estimating the distance between two points are as
follows:
r Received Signal Strength Indicator (RSSI). This utilizes the predicted propagation characteristics of a wireless signal to determine the distance from a transmitter. If the power of the transmitted signal is known, a receiver can estimate
the distance from the transmitter by measuring the power of the signal that it
receives from the transmitter. Although simple and cost effective, this technique
suffers from significant errors as the path loss characteristics of wireless signals
is highly dependent on the physical environment. For instance, the path loss
344
LOCATION DISCOVERY IN SENSOR NETWORKS
characteristics in free space is modeled as 1/d 2 , whereas in terrestrial conditions it is determined as 1/d α , where 2.5 < α < 4.5 depending on the terrain
and other environmental factors [7]. The path loss factor α may be obtained
experimentally for the specific environment where it is to be used. This is often
a cost-effective method for distance estimation.
r Time of Flight: A useful technique for determining the distance between an
RF transmitter and a receiver is to measure the time taken for the RF signal
to cover the distance. Here, a receiver measures the time of arrival (ToA) of
a signal transmitted from reference point to compute the distance using the
velocity of propagation of the signal. This requires the receiver to have accurate
time synchronization with the transmitter, which is often hard to implement. An
alternative is use an additional transmitter to act as a frame of reference, where
the receiver estimates the time differences of arrivals (TDoA) between signals
from multiple transmitters to determine distances. These distances can be used
for localization using multilateration. Such a technique is fairly accurate when
the distances involved are long, such as between a ground-based receiver and a
system of satellites orbiting the earth, as used in GPS. However, distance ranging
from time of flight is hard to implement for short distances because RF signals
propagate in the speed of light and it requires extremely accurate hardware to
detect small time differences.
12.2.2 Mechanisms for Estimating Angles
Most of the earlier methods for angle estimation were based on optical methods, such
as determining the direction of stars for air and sea navigation. Some of the more
technologically advanced methods for angle estimation are as follows:
r Use of Directional Antennas. Large phased-array antennas have gain patterns
that can be sufficiently directional to allow determination of the angle of arrival
of an RF signal. This technique was proposed and evaluated in reference 8 for
designing a system for tracking the locations of transportation vehicles. It used
rotating directional antennas in each vehicle to determine the angles of arrival
of RF signals from fixed base stations for localization. However, even the most
expensive phased array antennas can produce errors in angle estimation, which
is the primary drawback of the system. It also involves complex and bulky
instrumentation (rotating directional antenna arrays) at the receiver.
r Use of Angular Signatures. Another example of localization using angulation
is the VHS omnidirectional ranging or VOR systems that are used for aircraft
navigation. Here, specific RF signals are transmitted from ground-based VOR
stations that allow receivers on flying aircrafts to determine the angles from
the stations [9]. A simple implementation of this idea involves a combination
of periodic transmissions of bursts of an omnidirectional signal along with a
continuously rotating beam emitted from the VOR station. A receiving aircraft
can determine the angular orientation with respect to the VOR station from the
ISSUES ON LOCALIZATION IN SENSOR NETWORKS
345
observed time difference between the flash and the rotating beam. The combination of angles estimated from two VOR stations and their locations can be used
to determine the position of the aircraft.
12.3 ISSUES ON LOCALIZATION IN SENSOR NETWORKS
With these ideas on location discovery methods, we now turn our attention to the
special requirements and issues for location discovery in sensor networks. In most
cases, sensor network applications require that the wireless sensor nodes are deployed
randomly in an area of interest. Applications may also require sensor nodes to be
deployed by dropping them from flying aircrafts and requiring the nodes to selfconfigure into a wireless network. Because of such deployment methods and the
large number of nodes involved, special techniques are required for the nodes to
determine their locations after deployment.
Why Not GPS?. Although GPS receivers have evolved to become smaller in size,
cheaper, and accurate, several reasons prohibit the use of GPS in most sensor network
applications. Firstly, GPS requires the receiver to receive data from four or more
satellites for functioning. This is not possible in indoor environments or in locations
where foliage and other cover obstructs satellite signals. Secondly, limitations of size,
battery, and hardware resources of sensor nodes is prohibitive for using GPS hardware
in every node. This is particularly true for wireless sensor nodes of the future, which
are envisioned to be of the size of a few cubic millimeters [10]. Thirdly, the addition of
GPS hardware would make sensor nodes costly, which clashes with one of the design
objectives of sensor nodes. Finally, in most applications of wireless sensor networks,
only relative locations of the nodes are needed and not their absolute geographical
positions. Hence, a GPS device is considered as an overkill in terms of cost and
hardware requirements for applications involving wireless sensor nodes.
12.3.1 Requirements for Localization in Sensor Networks
The special requirements for localization schemes for sensor networks generally depend on the nature of the applications and the constraints imposed by hardware and
network infrastructure. Based on these, some of the specific issues concerning the
design of the localization scheme are as follows:
1. Absolute versus Relative Locations. Systems such as the GPS determine the
absolute location in terms of the latitude, longitude, and altitude with respect to
the earth’s coordinates. Alternatively, locations may be obtained with respect
to a given frame of reference, such as the location of a base station that acts as
the coordinator of the sensor network. It could also be an arbitrary reference
point at the scene of application, such as a location inside a building. In almost
all applications of sensor networks, only relative locations are required for
346
LOCATION DISCOVERY IN SENSOR NETWORKS
performing the monitoring, processing, and communication tasks. Note that
a relative location can always be transformed to an absolute location if the
absolute location of the reference point is known.
2. Accuracy and Precision of Localization. Accuracy implies the closeness of
the location estimate to the ground truth. This may depend on a number of
factors. Each system has its own granularity of measurements, which describes
the smallest measurable distance. Depending on the technology and mechanism
used, localization schemes may have a granularity of measurements within a few
inches or within a room (tens of meters) or even larger. The required granularity
of localization in sensor networks depends on the application. In most cases, it
depends on the relative distances between nodes in the network. A second issue
is the precision, which describes the consistency of the estimates. As a frame
of reference, some expensive GPS receivers can estimate locations with errors
smaller than 10 m with 95% precision.
3. Scale of Measurements. A different, but perhaps related issue is the scale of
measurements, which describes the largest distance over which the localization
system can work. GPS can obtain location estimates over the whole world. The
required scale of localization in sensor network is the total area of the network,
which makes the implementation simpler.
4. Dynamics of the Nodes. If the nodes of a network are mobile, the localization
system has to face the challenge of taking repeated measurements to continuously track its location. The repetition rate will depend on the dynamics of the
network. Most sensor network applications require the nodes to be static after
deployment, although dynamic scenarios also exist (such as using wireless sensors on robotic platforms). Hence, it is generally required that the localization
system for a sensor network should be able to track displacements of the sensor
nodes when necessary.
5. Communication Requirements. In addition to using passive measurements of
some reference signals for localization, such as those used for schemes employing pure triangulation, localization schemes may benefit from information
exchanged between the nodes in the network and/or some reference nodes, such
as beacons. Existing systems vary in the communications required for position
estimation. For instance, a GPS receiver does not transmit but performs localization by only receiving the RF signals from satellites. On the other hand,
most cellular telephone localization schemes rely on two-way communication
between the cell phone and the base stations. Communication between a sensor
node and a reference node or other sensor nodes can provide important benefits such as time synchronization and improvements in accuracy and precision.
However, a central issue in wireless sensor networks is the minimization of
communication requirements in the nodes to conserve battery. This introduces
special considerations for designing the localization scheme as well.
6. Self-Localization versus Remote Localization. The most desirable scheme for
localization in this respect is one in which the sensor nodes can self-localize—
that is, determine their own positions from signals received from external
ISSUES ON LOCALIZATION IN SENSOR NETWORKS
347
sources. Alternatively, localization schemes can allow a remote station, such
as a base station, to determine the locations of all sensor nodes in the network
from signals received from them. Such remote localization schemes suffer from
scalability and communication cost problems. It should be noted that the resulting difference between self- and remote positioning is where the computations
take place and the resultant need for communication between the sensor nodes
and the remote station.
7. Cost. This is an important issue, since the requirements for designing largescale sensor networks are to (a) keep the cost (and complexity) of each node
low and (b) benefit from the collective sensing and computation capabilities of
a large number of nodes in the network. Hence, it is highly desirable that the
localization system does not require expensive hardware at the sensor nodes.
The cost of building external infrastructure for enabling localization also plays
a role. However, that is usually considered less critical because it does not
increase with the size of the network.
8. Form Factor. Wireless sensor nodes must have a small form factor, which eliminates localization schemes that require large components such as antenna arrays. The size of the sensor node thus plays a critical role in determining the
mechanism that can be effectively used for localization in sensor networks.
9. Passive versus Active Localization. Some location estimation schemes require
the unknown node to play an active role for position estimation, while others
can work even without any action from the target node requiring localization.
In sensor networks, the sensor nodes can allow active localization within the
limitations of its hardware and cost constraints. For instance, a sensor node
may act as a transponder for a transmitted RF signal to estimate round-trip
transmission delay between a remote transmitter and the node. However, it may
be difficult to implement a system that requires the sensor node to estimate
the direction of arrival of the received signal and send that information back
because that usually requires complex antenna arrays.
12.3.2 Challenges
From the above discussion, it can be concluded that the main challenges for designing
a localization scheme for wireless sensor nodes arise from the need to deal with
the low hardware complexity and cost of implementation, small form factor of the
nodes, and their arbitrary locations (indoor, outdoor, and in uncharacterized regions).
Typically, the natural choice for location estimation is to use triangulation, which
requires estimation of distances or angles from fixed reference points. In this section,
we discuss the challenges involved with obtaining these estimates from the perspective
of sensor networks.
Ranging Issues. Of the two primary techniques for ranging, measuring time of
flight using RF signals is difficult for applications in sensor networks. This is because
RF signals travel at the speed of light and measuring the extremely short travel times
348
LOCATION DISCOVERY IN SENSOR NETWORKS
within the spacial domain of a sensor network (that can be room sized or of equivalent
size) is a technical hurdle. It would require extremely accurate clock synchronization
between the sensor node and the transmitter that remains to be a technological challenge. Consequently, time-of-flight measurements have been explored using acoustic
and ultrasonic signals that have a much slower propagation speed than RF. Acoustic
ranging has been used in many different distance measuring devices due to their low
cost and accuracy for indoor use. These devices can provide accuracies within 10 cm.
The comparatively higher frequency utrasonic signals (frequencies typically within
24–40 kHz) can be used to provide accuracies within 2–5 cm. However, they have a
smaller range as compared to acoustic signals. Usage of acoustic or ultrasonic signals
for ranging in sensor networks also faces other challenges. Firstly, acoustic sources
and detectors are generally larger in size as compared to RF sources, due to their larger
wavelength. This poses a problem for small sensor nodes. Secondly, again due to their
larger wavelength, acoustic signals cannot propagate through physical obstructions.
They also suffer from severe multipath effects, making it hard to design a reliable
distance estimation system for arbitrarily placed sensor nodes in unknown environments. Systems such as Cricket [11] and BAT [12] use the time-difference-of-arrivals
of ultrasound and RF signals from the same source to perform indoor localization.
Assuming that the RF signal is received instantaneously, the corresponding delay of
the much slower ultrasound signal provides the needed distance estimate.
Usage of RSSI for ranging requires the knowledge of the corresponding signal
propagation model. However, even with extensive channel estimation and modeling,
ranging using RSSI faces inaccuracies caused by shadowing, multipath reflections,
refractions, and scattering effects. Nevetheless, because of the ease of implementation,
localization schemes using RSSI have been researched extensively [6]. It has been
applied in a number of localization schemes such as RADAR [13], AHLoS [14], and
APS [15]. The RADAR indoor location system applies a wall attenuation factor (WAF)
and signal strength maps that are obtained from extensive offline measurements of
indoor signal propagation characteristics. AHLoS assumes that a limited number of
nodes know their locations, either from using GPS or from manual configurations.
Other nodes use a combination of RSSI and ToA ranging techniques to determine their
positions with respect to the beacons. AHLoS utilizes collaborative multilateration, a
mechanism that is explained in the next section. A decentralized approach to RF-based
localization in indoor environments was presented in reference 16.
Angle of Arrival (AOA) Estimation. Estimating the direction of arrival of a wireless signal requires an antenna with extremely small beamwidth. This can be achieved
with an antenna array, which can be prohibitvely large for use in wireless sensor nodes.
Hence alternative techniques need to be used for AOA estimation for applications in
sensor networks. A possible approach to reduce the size of the antenna is to use
ultrasound signals, which is used in the Cricket Compass project [17] to determine
angles from phase difference and time difference of arrivals of an ultrasound pulse
on multiple detectors that are placed in a specific pattern within a space of few centimeters. Such a device can determine the angle of arrival with accuracy of 5◦ within
LOCALIZATION USING RANGING IN SENSOR NETWORKS
349
an angle of ±40◦ . A key requirement for the success of this mechanism is having
line of sight from the source. Since such AOA estimation heavily relies on phase
differences, multipath and scattering can also cause problems. AOA estimation has
been used in references 15, 18 and 19. In reference 18, rotating optical beacons are
used, and the angle is measured at the sensors by the times at which the optical signal
is detected. The concept used in reference 19 is similar, except that it used RF signals
from an 802.11 transmitter using a directional antenna. To reduce errors in AOA estimation caused by the nonzero beamwidth of the RF signal, the center of the beam
was detected from the angle where the signal strength is maximum.
12.3.3 Other Technical Challenges
The errors in distance or angle measurements need to be minimized either by multiple
measurements at the same node or by combining measurements from other nodes in
the vicinity, a method that is known is collaboration. In either case, a common requirement is the need for optimization operations that requires extensive computations and
data handling. This must be done within the limitations of processor capabilities,
memory, and other hardware of the wireless sensor nodes. In addition, collaborative
computation also taxes the battery of the nodes from communication. Additional complications arise when collaboration requires routing, which is dependent on location
information, and consequently, nodes need to have localization done before they can
collaborate.
12.4 LOCALIZATION USING RANGING IN SENSOR NETWORKS
A significant amount of work has been reported on the development new techniques
that are applicable for location discovery of small, low-powered wireless nodes in
a sensor network. Several of these ideas have also been demonstrated in proof-ofconcept implementations and commercially available products. In this section, we
describe a few of these developments that use ranging as the primary mechanism
for localization. Other localization schemes using AOA and alternative schemes are
presented in the following sections.
12.4.1 Localization Using Ranging
Localization schemes that use ranging generally address the multilateration problem
where the distance estimates have errors that are caused by problems with accurate
ranging as described in the previous sections. With ranging errors, the three circles
centered at the reference points with radii equal to the corresponding distance estimates (shown in Figure 12.1a) do not intersect at a common point. The solution to the
localization problem then shifts from a geometric solution to one of estimation, where
one has to determine the optimum location coordinates that minimizes an objective
function involving the estimated distances and the unknown node location. This is
explained below with reference to the scenario depicted in Figure 12.2.
350
LOCATION DISCOVERY IN SENSOR NETWORKS
3
4
2
R3,u
R2,u
u
R1,u
Rn ,u
1
n
Figure 12.2. Illustration of atomic multilateration for estimation of a location from erroneous
distance measurements.
Let (xu , yu ) be the unknown location of a node u that is to be determined, and
let (xi , yi ) be the location of the ith reference node, i = 1, 2, · · · , n. Most often
the reference nodes are called beacons as they generate RF, acoustic, or ultrasound
signals to allow range estimation from the sensor node [14]. If Ri,u be the estimated
distance between the unknown sensor node and the ith beacon, then the error in this
estimate is given by
fi (xu , yu ) = Ri,u −
(xi − xu )2 + (yi − yu )2
(12.1)
and consequently, the objective function to minimize for estimating the unknown
location (xu , yu ) can be expressed as
F (xu , yu ) =
n
fi2 (xu , yu )
(12.2)
i=1
The solution has been addressed in detail in reference 14, where it is termed as atomic
multilateration. A solution to linearizing the equations for obtaining the minimum
mean square estimate (MMSE) of the unknown location is also proposed under the
assumption that the ranging errors in Eq. (12.1) has a Gaussian distribution. The
following are some important issues in this regard:
r The solution requires at least three reference nodes or beacons from which distance estimates have been obtained at the sensor node. The locations of these
beacons must be known. Also, these beacon nodes should not be located on a
straight line.
r If the speed of propagation of the beacon signal that is used for ranging is
unknown, it can be estimated by using mesurements of time-of flight of the
beacon signal from at least four beacon nodes.
LOCALIZATION USING RANGING IN SENSOR NETWORKS
351
r In a typical scenario of a wireless sensor network, it is likely that a few of the
nodes may know their locations using GPS or some other mechanism (perhaps
even from manual deployment). These nodes may act as beacons for other nodes
to localize in the network. This introduces the issue of solving the problem in a
multihop network environment.
12.4.2 Localization over Multiple Hops
A more general scenario for a localization problem in sensor networks is depicted in
Figure 12.3. Here, the dark nodes are assumed to know their locations and the rest
(unfilled) nodes are not aware of their locations. The problem is to determine the
locations of as many of the unfilled nodes as possble from the information at the dark
nodes. Each node may transmit beacon signals for other nodes in its neighborhood
to estimate its distance from it. Two solutions to this muiltihop localization problem
have been explored in reference 14, referred to as the iterative multilateration and
collaborative multilateration algorithms.
Iterative Multilateration. This algorithm uses the method of determining the locations
of unknown nodes from those of the known nodes using atomic multilateration and
then propagating this knowledge to other unknown nodes. An unknown node acts as a
beacon after its location has been determined. The concept is depicted in Figure 12.4.
With sufficient number of beacon nodes to start the process, this algorithm can allow
all nodes to self-localize iteratively. Moreover, the problem can be solved in a totally
distributed manner. A concern with this algorithm is that errors tend to accumulate
from one iteration to the next as more unknown nodes start acting as beacons. An alternative where all distance measurements are passed to a central node for centralized
processing for localization of all nodes in the network can also be considered. This
has the advantage that the algorithm can start from the node that has the maximum
number of beacons in its neighborhood, thereby reducing the error at the starting
point.
Beacon nodes
Sensor nodes that
have localized
Sensor nodes with
unknown locations
Figure 12.3. Illustration of a multihop network with beacons, nodes with known locations and
nodes with unknown locations.
352
LOCATION DISCOVERY IN SENSOR NETWORKS
Stage i+1
Stage i
Node with known position/beacon node
Node with unknown position)
Figure 12.4. Illustration of the iterative multilateration principle.
Collaborative Multilateration. The collaborative multilateration algorithm addresses
the problem that one or more unknown nodes may not have a minimum of three beacon
nodes in its neighborhood to perform localization using atomic multilateration. An
example is shown in Figure 12.5. The proposed solution uses all distance estimates
over multiple hops and solves for the unknown locations simultaneously. For the
scenario depicted in Figure 12.5, where Ri,j represents the estimated distance between
nodes i and j, and (xi , yi ) is the location of node i, the collaborative multilateration
algorithm attempts to solve the following problem:
min
x3 ,y3 ,x4 ,y4
2
2
2
2
2
+ f4,3
+ f4,5
+ f4,6
f2,3
+ f1,3
(12.3)
where fi,j represents the expression for error in the estimated distance between nodes
i and j, that is,
fi,j = Ri,j −
(xi − xj )2 + (yi − yj )2
1
5
3
4
2
6
Node with known position/beacon node
Node with unknown position)
Figure 12.5. Illustration of the collaborative multilateration principle.
(12.4)
353
LOCALIZATION USING ANGLE ESTIMATION IN SENSOR NETWORKS
Obviously, this problem is more complex to solve. A solution using initial estimates and an iterative least squares method for refinement has been described in
reference 14. The solution can be achieved under the assumptions that there is sufficient connectivity, a sufficient number of beacons nodes exist in the network, and
the ranging errors have a Gaussian distribution. The authors also present ideas for
optimizations using clusters or subtrees as well as distributed computations.
12.5 LOCALIZATION USING ANGLE ESTIMATION IN SENSOR
NETWORKS
Since estimation of distances is prone to errors, efforts have also been directed
toward localization using angle estimation in sensor networks. The basic idea of
angulation is described in Figure 12.1b, where it is assumed that the angles ∠NAP
and ∠NBP as well as the locations of reference nodes A and B are known to the sensor
node P.
Note that determination of the angles with respect to an absolute angular reference
(i.e., north) also requires the knowledge of the orientation of the node. To avoid that,
angulation may be performed by determining relative angles, that is, the angular spans
between nodes A, B, and C, as shown in Figure 12.6. Here, a sensor node located at
an unknown location P determines the angles α and β, sustained between reference
points A and B, and B and C, respectively. The sensor node then computes its location
(Xp , Yp ) as follows:
Xp = x2 + C cos(γ − η),
Yp = y2 − C sin(γ − η)
(12.5)
A
(x1,y1)
L
(x2,y2) B
S
C
M
P
(xp,yp)
C
(x3,y3)
Figure 12.6. Geometrical calculation of location coordinates from angle estimation.
354
LOCATION DISCOVERY IN SENSOR NETWORKS
where
−1
γ = tan−1
η = tan
C=
y1 − y2
x1 − x2
R sin(α) − sin(β) cos(A) − cos(β) sin(A)
sin(β) sin(A) − cos(β) cos(A) − R cos(α)
L sin(α + γ)
sin(α)
L sin(β)
M sin(α)
2
S − M 2 − L2
A = cos−1
2ML
S = (x1 − x3 )2 + (y1 − y3 )2
M = (x1 − x3 )2 + (y2 − y3 )2
L = (x1 − x2 )2 + (y1 − y2 )2
R=
(12.6)
Hence, with a mechanism that allows each node to determine the angles sustained
between reference nodes A and B, and B and C, resepctively, the network can selflocalize without the need for collaboration or centralized computations. A method for
doing this was proposed in reference 20 and a prototype implementation of the scheme
using the Crossbow mica2 wireless sensor nodes was presented in reference 18.
The principle of operation of this AOA-based localization scheme is presented next.
12.5.1 AOA Based Localization Using Rotating Directional Beacons
The principle of operation is somewhat similar to VOR stations that are used in aircraft
navigation. The system uses three specially constructed beacon generators to facilitate
angle determination at the sensor nodes. Each beacon generator transmits a directional
beacon signal that rotates with a fixed angular velocity. The sensor nodes calculate
the angles α and β by determining the times of arrival of the directional beams from
different beacon generators. It is assumed that the directional beacon signals have
very narrow beamwidths on the horizontal plane so as to enable accurate estimation
of times when the beams are pointed toward a sensor. They should have sufficiently
large vertical beam angles, so that the beams cover the entire sensor network area; that
is, they cover sensors located at different vertical angles with respect to the transmitter.
The beacon signal can be RF or optical, as long as it is possible to transmit it at a
narrow beam. A schematic illustration of the proposed implementation is depicted in
Figure 12.7a, which assumes the special case where the beacon generators (reference
points) are located at the vertices of a rectangle.
LOCALIZATION USING ANGLE ESTIMATION IN SENSOR NETWORKS
355
Figure 12.7. Illustration of (a) the system model for AOA-based localization and (b) the times
of arrival of beacon signals at a sensor node.
The main considerations required for successful implementation of this scheme
are as follows:
r Identical Angular Velocities. All three beacon signals must be rotating with
identical angular velocities. However, any arbitrary angular velocity may be
used, which can be measured at the sensor nodes by noting the time period of
rotations.
r Beacon Identification. The sensor nodes should be capable of identifying the
beacons from their signals. This can be implemented by various mechanisms,
such as using coding or modulation of the beacon signals.
r Phase Reference. The phase differences between the rotating beacon signals (θ1
between A and B and θ2 between B and C in Figure 12.7) must be known at
the sensors. To remove this problem, θ1 , θ2 , and θ3 may all be made zero by
appropriate initial configurations.
Each sensor will then receive the beacon signals periodically as shown in
Figure 12.7b. The required angle estimates α and β can then be obtained by measuring the times t1 , t2 , and t3 as follows:
α = ω(t2 − t1 ) − θ1
β = ω(t3 − t2 ) − θ2
(12.7)
Possible Sources of Error. The primary source of error in the above technique
is due to inaccurate angle estimation that can be caused by nonzero beamwidth of
the beacon signals. It is difficult to get very narrow beams of RF signals unless
large antenna arrays are used at the beacon generators. To avoid such expensive
infrastructure in case RF signals are used, the sensor nodes may be allowed to detect
the center of an incoming RF beam by detecting the time at which the received signal
356
LOCATION DISCOVERY IN SENSOR NETWORKS
strenght is maximum [19]. Alternatively, higher directionality of the beacon signals
can be achieved by using optical signals, such as a laser, as described in reference 18.
Errors will also be caused by nonidentical angular speeds and errors in estimating
the initial phase references θ1 , θ2 , and θ3 . However, these are relatively easy to correct
by appropriate infrastructure control mechanisms.
Experimental Implementation. To provide a deeper understanding of the issues
concerning implementing such a scheme, details of an experimental prototype development of this scheme are described in this section. The prototype was designed with
the goal of enabling Crossbow manufactured mica2 wireless sensor nodes (the Berkeley motes) to self-localize with minimum additional cost in a laboratory setting. Each
mica2 mote consists of the MPR410CB processor and radio platform that is equipped
with an Atmel ATmega128L processor, 128-kB program flash memory, 512-kB measurement flash, a 10-bit analog to digital converter (ADC), and 433-MHz radio interface. The motes were equipped with the MTS310 sensor board that has a photosensor
along with a number of other analog sensors such as microphone, accelerometer,
thermistor, and a magnetometer. The processor runs the Tinyos software operating
system developed by UC Berkeley, which supports large-scale self-configuring sensor
networking [21].
The beacon generators use optical lasers that are suitable for indoor use and
are also cost effective. The primary element in these modules is an Apinex 3-mW
650-nm semiconductor laser equipped with a diffraction grating for line generation
and a plastic lens that can be used to adjust the width and fan angle of the generated
optical line. In order to make the three beacon signals identifiable, a varying number of
laser line generators were used for each beacon—that is, a single laser line generator
for beacon-1, a double line for beacon-2, and a triple line for beacon-3.
Figure 12.8. Beacon generator assemblies for (a) beacon-1, (b) beacon-2, and (c) beacon-3.
The corresponding beacons signals are depicted in (d), (e), and (f), respectively.
LOCALIZATION USING ANGLE ESTIMATION IN SENSOR NETWORKS
357
The problem of maintaining identical angular rotations for all three beacon generator assemblies was solved by using stepper motors, all driven by a common controller.
The assembly of laser sources, gearbox for speed control, and stepper motor at each
beacon generator were mounted on a pole at a height of 6 feet with the axes of the lasers
pointing down at an approximate angle of 45◦ (see Figure 12.8). Since initial phase errors are critical for location accuracy and it is very difficult to align three independent
sources perfectly, a pilot sensor node is located at the farthest corner of the area for
estimating the angles θ1 and θ2 . Initially all sensor nodes are set to standby until the
pilot node determines these phase references. Once that is done, the pilot node broadcasts the estimates, which triggers the nodes in the network to start self localizing by
measuring the angles α and β at their locations. Each mote is programmed to identify
single, double, and triple beams, by counting optical signals detected within a small
window of time. Adequate measures are incorporated to reduce missed and false detections from overlapping or interfering signals. The details are shown in Figure 12.9.
/* Wait for pilot message and then start sampling and localization */
if packet received from pilot then
begin
read 1 and 2 from pilot message
start obtaining periodic samples s[i]
/* Begin beam counter window when first signal peak is detected*/
if ((s[i]>s[i-1]+8) AND (win_flag==0)) then
begin beam++; win_flag:=1; end
timer++;
/* Count beams within beam counter window */
if (win_flag==1) then
begin
win_time++;
if ((s[i],s[i-1]-8) AND (s[1-1]<s[i-2])) then
begin beacon++ end
if end of beam counter window then
begin
detected[beam]:=beacon; t[beam]:=timer;
win_flag:=0; beacon:=0; win_time:=0;
end
/* At the end of cycle, check for errors and localize */
if ((beam==4) AND (win_flag==0)) then
begin
if (detected[1]+detected[2]+detected[3]==6) then
begin
:=2 /(t[4]-t[1]);
if (detected[1]==1) then
begin
:= (t[2]-t[1])end
if (detected[1]==2) then
begin
1;
:=
(t[3]-t[2])-
2
;
:= (t[4]-t[3])- 1 ;
:= (t[2]-t[1])- 2 ;
end
if (detected[1]==3) then
begin
:= (t[3]-t[2])- 1 ;
:= (t[4]-t[3])- 2 ;
end
(x p ,y p ) locate( , ) /* apply equation (5) */
end
beam:=0; timer:=0
end
end
Figure 12.9. Algorithm for self-localization using angle estimation used at the sensor nodes.
358
LOCATION DISCOVERY IN SENSOR NETWORKS
10
9
8
7
6
5
4
3
2
1
0
0
1
2
3
4
5
6
7
8
9
10
Figure 12.10. Experimentally obtained locations of sensors that are placed in 8 × 8 grid
at separations of 1 ft. The angular speed of rotation of beacons was fixed at 9.5 rpm.
Experimental evaluation of the performance of the system shows that when used
in a 10-ft × 10-ft area, errors in localization is less than 4 inches and the maximum
time required for any sensor to localize is 45 seconds. Localization results of nodes
located 1 ft apart in the area are shown in Figure 12.10.
Although this scheme needs special considerations for designing the beacon generators, the attractive feature is that the sensors do not need any special hardware except
for being able to detect the beacon signal (i.e., photodetector), which is usually present
in most sensor nodes. The primary drawback of this scheme is its requirement for line
of sight. Within indoor environments, this can be mostly achieved for most locations
by mounting the beacons at the ceiling. For those nodes that do not have line of sight
from all three beacons, additional methods may be used, such as atomic or iterative
multilateration, to be performed after a sufficient number of nodes in the network
have self-localized.
12.6 OTHER LOCALIZATION TECHNIQUES
While the previous sections described the key approaches to localization in sensor
networks, the list of new ideas keep growing and it is difficult to capture all the
existing work in this rapidly developing topic. Here, samples of some other wellknown approaches are presented.
OTHER LOCALIZATION TECHNIQUES
359
12.6.1 Range-Free Position Estimation
Range-free techniques are those where location estimates are obtained without using
concrete distance or angle measurements from specific nodes. The principle here
is to determine an unknown location from the proximity to beacon nodes or nodes
with known locations. This avoids problems arising from errors in range or angle
estimation, and it typically requires less costly hardware and simpler calculations.
However, they generally have lower precision of localization than those using rangebased methods.
A popular example of range-free localization is the centroid method proposed in
reference 6. Here a node counts the number of beacon signals received from a set of
pre-positioned beacon nodes and achieves localization by obtaining the centroid of
the received beacon generators. Precision is dependent on the density and locations
of the beacon generators.
The DV-HOP solution uses locations of anchor nodes (which are special nodes
that broadcast control packets to enable localization), the hop counts from anchors,
and the average distance per hop for localization. It uses a mechanism that is similar
to classical distance vector routing. The anchor nodes broadcast beacon packets that
are flooded throughout the network. The beacon packets carry hop counts and the
locations of the corresponding anchor. Each receiving node maintains the minimum
counter value from each anchor, thereby allowing them to determine the shortest hop
distance to each anchor. The location is estimated using average hop distance from
anchors.
In reference 22, the authors present a Point-in-Triangulation (PIT) test, where
nodes use a set of signal strength measurements from neighboring beacon nodes
to determine the closest set of three nodes forming a triangle within which it is
located. This is repeated with different anchor combinations until all combinations
from nodes that are within range are exhausted. Localization is then performed using
the center of gravity of the intersection of all the triangles within which the node is
located.
A cluster-based distributed localization scheme is presented in reference 23. This
method avoids using distance or AOA measurements and long-range beacons by
utilizing the regularity of clusters in the network. The localization algorithm starts
with the development of regular-shaped clusters of nodes, each with a cluster-head
node. Such regular clusters can be formed using an algorithm like ACE [24]. Initially,
some of the cluster heads are location-aware anchor nodes (assuming that these nodes
have GPS or are manually deployed). The locations of other location-unaware cluster
heads are determined by a process of self-calibration, which utilizes the uniformity
of cluster shapes and the average edge length of clusters. The regularity of cluster
shapes is also utilized to refine early location estimates of cluster heads at a second
stage of the algorithm. When all cluster heads are calibrated, other follower nodes
can calibrate using the same range-free principle.
A localization scheme that uses RF connectivity and centralized computations
is presented in reference 25. The authors show that given a set of convex proximity
constraints and connectivity information under the constraints, fairly accurate location
360
LOCATION DISCOVERY IN SENSOR NETWORKS
estimates can be obtained using this scheme. This scheme targets large-scale networks
of small sensors such as those using the smart dust mote [10].
12.6.2 Probabilistic Position Estimation
Another approach to solving the uncertainties caused by errors in range or angle
estimation for localization is using a probabilitistic approach [26–29].
The work in reference 27 models range measurements by a set of density functions.
Nodes obtain initial range measurements using RSSI from a set of beacon nodes
that are location-aware, either using on-board GPS or from manual deployment. The
nodes then compute their position estimates that are represented as three-dimensional
density functions. This requires prior knowledge of the density function of range that
is obtained from collected sample values. Position estimates from different nodes
are then combined by intersection of density functions. When nodes improve their
position estimates from new information (i.e., the estimated density function becomes
sharper), they broadcast this information to other nodes.
A similar approach is applied in reference 28 to probabilistic localization using
AOA measurements. As usual, the error for AOA measurement is modeled probabilistically from offline experiments. The nodes utilize this information to refine position
estimates obtained from beacons.
In reference [26], the authors proposed a method applied to mobile sensors that is
based on Monte Carlo localization, which uses particle filter combined with probabilistic modeling. The basic idea is to first obtain a posterior distribution of possible
locations using a set of weighted samples. Nodes use a mobility model to refine the
probability distribution obtained at an initial step. In a second step, the nodes use
filtering to eliminate impossible locations based on new observations. This approach
has been found to work better with higher node mobility.
An algorithm that applies probabilistic modeling for localization using RSSI based
ranging from a mobile node is proposed in reference 29. Here, it is assumed that all
sensor nodes are static while a mobile beacon node travels through the network area
broadcasting beacon packets. Beacon packets carry updated location of the beacon
node. As the beacon packets are received at a sensor node, it obtains new position
estimates and combines them probabilistically to refine its position estimate.
12.7 CONCLUSION
Localization is a challenging problem to solve for successful implementation in small,
low-cost wireless sensor nodes. Although a significant amount of research has been
devoted to this problem, a single practical solution is hard to find. It is likely that
solutions would have to be application-specific, because sensors are used in a large
number of application scenarios and environments. Also, solutions for any application
may require multiple principles for addressing the needs for all nodes. For instance,
while some nodes may be able to self-localize, others may need to employ collaborative or centralized computations, depending on the situation at which it is in. Multiple
methods would also improve the robustness of the localization system.
BIBLIOGRAPHY
361
This chapter presents the basic principles of localization, the particular challenges
with respect to solving the problem in wireless sensor networks, and some existing
techniques to solve the problem. This topic is likely to generate more research ideas
in future because of its importance in almost all sensor network applications.
12.8 EXERCISES
1. State some of the different techniques that can be used to measure or estimate
distances from a wireless beacon generator or base station to a wireless node
for localization. Give examples of systems that use each technique. For each
technique, explain the technical challenges, if any, of their usage for localization
in wireless sensor networks.
2. The locations of three wireless bases stations are as follows: BS1 = (0, 0), BS2 =
(20, 5), and BS3 = (6, 25). The wireless system estimates that the distance from
a wireless node to BS1 is 15 units, that from BS2 is 7 units, and that from BS3 is
19.4 units. Determine the location (x, y) of the wireless node using triangulation.
Write all the equations and show your work. If you use a software tool such as
MATLAB, attach printouts of your code. Is it really necessary to get distances
from three base stations to estimate the location? Explain.
3. Given that in an angle-based localization scheme as shown in Figure 12.6, the
measured angle α can have errors up to 5◦ , determine an expression for the
corresponding error in the location estimate for a sensor node located at an
arbitrary location (x, y). Assume that the region of operation is a 10-ft × 10-ft
area and the beacon generators are located at the three corners as shown in
Figure 12.7a. Plot the error distribution within the square area.
4. Repeat the above problem when the beacon generators are located on a straight
line along the base of the square—that is, at locations (0, 0) (the left bottom
corner), (5, 0), and (10, 0).
5. How is the approach for probabilistic localization different from that of estimation of location coordinates under measurement errors? Explain by comparing an
existing probabilistic localization scheme based on ranging with that of atomic
multilateration.
BIBLIOGRAPHY
1. D. Estrin, R. Govindan, J. Heidemann, and S. Kumar. Next century challenges: Scalable
coordination in sensor networks. In ACM/IEEE International Conference on Mobile Computing and Networking (MOBICOM), August 1999, pp. 263–270.
2. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks:
A survey. Computer Networks, 38:393–422, 2002.
3. A. Rao, S. Ratnasamy, C. Papadimitriou, S. Shenker, and I. Stoica. Geographic routing without location information. In Proceedings of ACM MOBICOM, September 2003,
pp. 96–108.
362
LOCATION DISCOVERY IN SENSOR NETWORKS
4. B. Karp and H. T. Kung. GPSR: Greedy perimeter stateless routing for wireless networks. In
ACM/IEEE International Conference on Mobile Computing and Networking (MOBICOM),
August 2000, pp. 243–254.
5. A. D’Costa and A. M. Sayeed. Data versus decision fusion in wireless sensor networks. In
Proceedings of ICASSP), April 2003, pp. 832–835.
6. N. Bulusu, J. Heidemann, and D. Estrin. GPS-less low cost outdoor localization for
very small devices. IEEE Personal Communications Magazine, October 2000, pp. 28–
34.
7. T. S. Rappaport. Wireless Communications Principles and Practice. Prentice Hall PTR,
Englewood Cliffs, NJ, 2002.
8. C. D. McGillem and T. S. Rappaport. A beacon navigation method for autonomous vehicles.
IEEE Transactions on Vehicular Technology, 38(3):132–139, August 1989.
9. http://www.arundel.net/xplane/tml/vor.html.
10. B. Wameke, B. Atwood, and K. S. J. Pister. Smart dust mote forerunners. In Proceedings
of the International Conference on MEMS, January 2001, pp. 357–360.
11. N. Priyantha, A. Chakraborthy, and H. Balakrishnan. The cricket location support system.
In ACM/IEEE International Conference on Mobile Computing and Networking (MOBICOM), August 2000, pp. 32–43.
12. A. Harter, A. Hopper, P. Steggles, and A Ward. The anatomy of a context aware application.
In Proceedings of MOBICOM), August 1999, pp. 59–68.
13. P. Bahl and V. Padmanabhan. RADAR: An in-buidling RF-based user location and tracking
system. In Proceedings of INFOCOM, March 2000, pp. 775–784.
14. A. Savvides, C. C. Han, and M. B. Srivastava. Dynamic fine-grained localization in ad hoc
networks of sensors. In ACM/IEEE International Conference on Mobile Computing and
Networking (MOBICOM), July 2001, pp. 166–179.
15. D. Niculescu and Badri Nath. Ad hoc positioning system (APS). In Proceedings of GLOBECOM), November 2001, pp. 2926–2931.
16. K. Lorincz and M. Welsh. Motetrack: a robust decentralized approach to RF-based location tracking. In Proceedings of the International Workshop on Location and ContextAwareness, May 2005, pp. 63–82.
17. N. B. Priyantha, A. Miu, H. Balakrishna, and S. Teller. The cricket compass for contextaware mobile applications. In ACM/IEEE International Conference on Mobile Computing
and Networking (MOBICOM), July 2001, pp. 1–14.
18. A. Nasipuri and R. E. Najjar. Experimental evaluation of an angle based indoor localization
system. In IEEE International Workshop on Wireless Network Measurement (WiNMee),
April 2006, pp. 1–9.
19. D. Niculescu and B. Nath. VOR base stations for indoor 802.11 positioning. In ACM/IEEE
International Conference on Mobile Computing and Networking (MOBICOM), September
2004, pp. 58–69.
20. A. Nasipuri and K. Li. A directionality based location discovery scheme for wireless sensor
networks. In Proceedings of Workshop on Sensor Networks and Applications WSNA’02,
2002, pp. 105–111.
21. http://webs.cs.berkeley.edu/tos.
22. T. He, C. Huang, B. M. Blum, and J. A. Stankovic. Range-free localization schemes for
large scale sensor networks. In Proceedings of MOBICOM), August 2003, pp. 81–95.
BIBLIOGRAPHY
363
23. H. Chan, M. Luk, and A. Perrig. Using cluster information for sensor network localization.
In Proceedings of International Conference on Distributes Computing and Sensor Systems
(DCOSS), 2005, pp. 109–125.
24. H. Chan and A. Perrig. ACE: An emergent algorithm for highly uniform cluster formation.
In Proceedings of EWSN, 2004, pp. 154–171.
25. L. Doherty, K. S. J. Pister, and L. El-Ghaoui. Convex position estimation in sensor networks.
In Proceedings of INFOCOM, April 2001, pp. 1655–1663.
26. L. Hu and D. Evans. Localization for mobile sensor networks. In ACM/IEEE International
Conference on Mobile Computing and Networking (MOBICOM), September-October
2004, pp. 45–57.
27. R. Peng and M. L. Sichitiu. Probabilistic localization for wireless sensor networks. ACM
Mobile Computing and Communications Review, 11(1):53–64, 2007.
28. R. Peng and M. L. Sichitiu. Angle of arrival localization for wireless sensor networks. In
Proceedings of IEEE Conference on Sensor and Ad Hoc Communications and Networks),
September 2006, pp. 374–382.
29. M. L. Sichitiu and V. Ramadurai. Localization of wireless sensor networks with a mobile
beacon. In Proceedings of IEEE Conference on Mobile Ad Hoc and Sensor Systems MASS),
October 2004, pp. 174–183.
CHAPTER 13
QoS-Based Communication Protocols
in Wireless Sensor Networks
SERDAR VURAL, YUAN TIAN, and EYLEM EKICI
Department of Electrical and Computer Engineering, Ohio State University, Columbus, OH
43210
13.1 INTRODUCTION
Wireless sensor networks (WSN) have been one of the main research foci in wireless
networks area over the last decade [1]. With the evolution of the MEMS technology
and the availability of low cost communication and computation hardware, WSNs
have been transformed from conceptual paradigms to reality in this short period.
Prototype and deployed WSNs currently serve as enablers of several applications
such as environmental monitoring, urban safety, traffic monitoring, smart spaces,
and surveillance of hostile and inaccessible areas. The majority of the early research
efforts have focused on enabling technologies for WSNs, including development of
communication protocols, localization methods, and simple application-dependent
information processing.
Major research efforts in WSN area have resulted in many deployed testbeds
and other implementations. The driving motivation of delivering solutions that can
be implemented in short time resulted in systems that can provide only best effort
service. Minimization of energy consumption or achieving “high efficiency” (in its
versatile definition) has been the objective of many communication protocols designed
for WSNs. Their performance strictly depends on the configuration of the network
and the load it carries. While the boosting effect of the existing solutions on the
research community should be acknowledged, these solutions fall short of addressing
requirements of all WSN applications and deployment scenarios. More specifically,
mission critical and real-time applications suffer from unpredictable performance
levels when such communication protocols are used.
Mission critical and real-time applications require performance guarantees
from the system on which they are implemented. As an example, a real-time
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
365
366
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
intrusion detection application running on a WSN may require the event detection
decision to be made within a given delay bound [2]. Other applications may require
that the network deliver data packets to the sink with a given probability. Similarly,
the overall energy consumption of all communication events may be subject to energy consumption constraint. Applications running on video sensor networks [3–5],
which form a new and emerging class of WSNs, inherently require service guarantees from the communication network due to real-time nature of its multimedia source
content. These requirements are classified as quality of service (QoS) in other wired
and wireless networks, which we also adopt for WSN environments. We classify a
communication protocol as QoS-based protocol if it can guarantee one or more performance metrics to upper layers or the application. Under this classification, solutions
that simply minimize/maximize a performance metric (delay, energy consumption,
packet loss probability, network lifetime, etc.) without performance guarantees are
considered as non-QoS-based solutions. We should note that most of the existing
proposals for WSNs fall into non-QoS-based category.
In this chapter, we first contrast QoS provisioning in WSNs and other network
types and introduce a QoS provisioning framework for WSNs. Then we outline and
discuss proposed QoS-based communication protocols for WSNs. We also outline
methods that support QoS-based in-network processing along with communication
for WSNs. QoS-based capacity estimation methods are also discussed within the
proposed framework. We then conclude the chapter with future research directions.
13.2 QoS IN WIRELESS SENSOR NETWORKS
13.2.1 General Principles
QoS has been the target of many communication protocols for numerous network
types. In its broadest form, quality of service refers to the contract between the
service provider (i.e., the network) and customers (i.e., applications) [6]. In wired
networks, one of the main motivations for QoS solutions is the real-time multimedia
applications that need bandwidth, delay, and jitter guarantees. ATM networks [7] were
proposed to support such requirements from ground up. Although ATM networks are
not as widely in use as originally imagined, QoS support mechanisms proposed for
ATM networks still inspire new solutions. In IP networks, QoS support of individual
flows have been proposed to be handled through IntServ [8] mechanism, which has
not gained widespread acceptance due to its scalability problems.1 In cellular networks, the motivation for QoS support is also inherent to the primary application of
such networks: Voice (and recently) video calls are subject to stringent constraints
to be commercially viable. In these networks, QoS support are provided through
resource reservation mechanisms. To accomplish resource reservation, the following
steps are followed:
1 The DiffServ [9] architecture will be proposed later on to support differentiation of groups of flows rather
than individual flows to overcome the scalability problem. However, DiffServ mechanism does not provide
absolute performance guarantees and therefore cannot be classified as a QoS solution.
QoS IN WIRELESS SENSOR NETWORKS
367
r Available Resource Estimation. The first step in QoS support is the knowledge of
available resources. The estimation of available resources is performed using the
network state and the communication protocols employed in the network. The
network state is comprised of the network connectivity information, maximum
capacity of nodes and links, and allocated resources.
r Calculation of Required Resources. Given that the performance requirements
of applications are known, resources required to sustain the QoS expectations
are calculated in the network. Both performance metric conversion and the resource requirement estimation depend on the protocols used in the network. The
calculation also involves the selection of the resources in the network for the
information flow.
r Resource Allocation. Calculated resources are reserved in the network entities.
The reservation of such resources is performed via auxiliary protocols such as
RSVP [10] or as an integral part of the communication protocol.
r Resource Deallocation. When a session terminates, resources are returned to the
general pool. The deallocation can be done either explicitly or implicitly through
timeout mechanism.
13.2.2 QoS in Ad Hoc Networks
The above-outlined steps work well in networks where resources are separated from
each other with well-defined boundaries: In wired networks, link and node resources
are clearly separated from each other. As an example, two node-disjoint links can be
treated as independent resources even though the load on them may depend on each
other. Similarly, in cellular networks, point-to-point links between mobile stations
and base stations are separated from each other in time, frequency, code, space, or a
combination thereof. Hence, QoS provisioning in both types of networks can easily
follow the aforementioned steps.
In wireless ad hoc networks, the resource allocation strategy faces a roadblock
at a very fundamental level [11]: Estimation of available resources is a nontrivial
task even for simplest multihop ad hoc networks. Wireless resources are shared by
multiple nodes that do not have an inherent coordination infrastructure. The contention
resolution and resource bidding procedures usually propagate over long distances and
affect far-away nodes. The lack of isolation of resources and the dynamic nature of
the network make resource estimation and allocation very challenging tasks. This
fact, coupled with the weak motivation for QoS-demanding applications for ad hoc
networks, have limited the acceptance of QoS-based communication proposals for ad
hoc networks.
13.2.3 QoS in Wireless Sensor Networks
A WSN can be regarded as a special type of ad hoc network with very resourceconstrained nodes, lower mobility, and larger scale. With these additional constraints,
it is easy to dismiss QoS provisioning in WSNs as implausible. However, there is
one important difference between ad hoc networks and WSNs. WSNs are defined
368
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
by applications they are deployed for. A large set of WSN applications, including
security and surveillance, requires QoS guarantees from the network. Note that QoSbased applications were not integral parts of ad hoc networks and were proposed as
additional applications that could run in parallel with non-QoS applications. Hence,
QoS provisioning is a requirement, and not an optional feature, for WSNs.
Being multihop networks, WSNs potentially suffer under the same shortcomings
and problems as ad hoc networks if similar QoS provisioning mechanisms are adopted.
Furthermore, considering very limited resources in sensor nodes and the large scale
of WSNs, per flow resource reservation-based approaches are especially ill-suited
for WSNs. The two important differences between both network types suggests new
directions in QoS provisioning: First, the mobility of WSNs is very limited when compared with ad hoc networks. Resource availability in WSNs fluctuates as a function
of the offered load and not as a function of network connectivity over long periods of
time. Therefore, communication decisions do not need to be updated very frequently.
Second, the large scale of WSNs can easily be used as an advantage to eliminate
explicit resource allocation. Distribution of communication responsibility over larger
areas provides gains through diversity and allows local decisions to be made, leading
to end-to-end QoS guarantees.
QoS provisioning in WSNs is directly geared toward satisfying application requirements. In Figure 13.1, a generalized framework for information flow in a WSN
is depicted. The main components of this framework are the sink, source locality,
and relay nodes. The information flow starts with assignment of a particular task to
sensor nodes. Upon information retrieval through sensors, source and other nodes
nearby preprocess the information. The preprocessing may be simply forming data
packets with raw data, data aggregation, or completely processing data and forming
end results per application requirements. The information is then communicated via
the relay nodes to the sink. In return, sink may optionally give feedback to the source
locality and/or relay nodes. As will be presented in the next section, many of the
QoS-based communication methods form an almost open loop where the source does
not return any feedback to information sources or intermediate nodes.
Task Assignment / Feedback
Source Locality
Relay Nodes
Figure 13.1. Information flow in a WSN.
Sink
QoS-BASED MAC PROTOCOLS FOR WSNs
369
13.3 QoS-BASED MAC PROTOCOLS FOR WSNs
QoS provisioning in the MAC control layer deals mainly with the scheduling of packets on the wireless channel subject to local constraints. Since local constraints may
change based on the needs of individual flows, the decisions are generally very dynamic and must be computed rather fast. The three solutions outlined below consider
the requirements of real-time WSNs while scheduling medium access to contending
nodes.
13.3.1 QoS-Aware Medium Access Control Protocol [12]
A QoS-Aware Medium Access Control protocol (Q-MAC) is presented in reference [12]. Q-MAC assumes an environment of multihop WSNs where nodes may
generate packets with different priorities. The design objective of Q-MAC is to minimize energy consumption and provide QoS guarantees. Q-MAC is composed of intranode and inter-node QoS scheduling mechanisms. The intra-node QoS scheduling
scheme classifies outgoing packets according to their priorities, while the inter-node
QoS scheduling solution handles channel access with the objective of minimizing
energy consumption via reducing collision and idle listening.
The intra-node scheduling mechanism employs multiple First-In First-Out (FIFO)
queues with different priorities, among which an instant queue has the highest priority and its enqueued packets are always instantly served. The intra-node scheduling
mechanism is outlined in Figure 13.2. Self-generated and relayed packets are classified to different queues with several QoS metrics, such as content importance and
number of traveled hops. Data rate allocation between queues and serving packet
selection are achieved through the MAX-MIN fairness algorithm [13] and the GPS
algorithm [14], respectively.
After a packet is scheduled for transmission, the inter-node scheduling mechanism,
Power Conservation MACAW (PC-MACAW), is executed to achieve Loosely Prioritized Random Access (LPRA) between sensor nodes. In PC-MACAW, a successful
transmission consists of two periods: the contention period and the packet transmission period. In the contention period, a node sends out RTS after waiting for a certain
Intra-Node Scheduling:
1. WHILE TRUE
2.
IF new packet arrives
3.
Enqueued to different queues after classification
4.
WHILE instant queue is not empty
5.
Deliver the first packet in the instant queue
6.
IF there are queues not empty
7.
Select q among these queues with the MAX-MIN and GPS algorithms
8.
Deliver the first packet in q
Figure 13.2. The intra-node scheduling Mechanism of Q-MAC.
370
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
PC-MACAW:
1. /*Enter the contention period*/
2. Backoff for a random number of time slots (ranging in [1,CW])
3. IF channel is not clear during the backoff time
4.
IF the dropping threshold is not reached
5.
GOTO Step 2 with doubled CW
6.
ELSE
7.
Drop the packet
8. ELSE
9. Send RTS packet
10. WAIT CTS packet
11. IF the CTS packet is absent
12.
IF the dropping threshold is not reached
13.
GOTO Step 3 with doubled CW
14.
ELSE
15.
Drop the packet
16. ELSE
17.
/*Enter the packet transmission period*/
18.
Transmit the packet
19.
WAIT ACK packet
20.
IF the ACK packet is absent
21.
IF the dropping threshold is not reached
22.
GOTO Step 2 with doubled CW
23.
ELSE
24.
Drop the packet
Figure 13.3. The inter-node scheduling mechanism (PC-MACAW) of Q-MAC.
duration (contention time) and expects a CTS packet before accessing the channel.
The contention time is randomly generated with a contention window size CW, where
CW is determined by each node’s transmission urgency including packet criticality,
number of transmitted hops, residual energy, and queue’s proportional load. After
accessing the channel, the node enters the transmission period to send data packets
and waits for an ACK packet. In the case of collision, CW is doubled and the packet
is retransmitted. When the difference between the current time and when the packet is
generated exceeds a threshold, the packet is dropped. The PC-MACAW algorithm is
outlined in Figure 13.3.
Q-MAC presents a combined effort of intra-node and inter-node QoS scheduling
in WSNs. It is shown through simulations that Q-MAC provides the equivalent QoS
while consuming less energy in comparison with an existing mechanism, S-MAC.
However, complex scheduling mechanisms and relatively loosely defined QoS metrics
stand out as shortcomings of this proposal.
13.3.2 Coloring-Based Real-Time Communication Scheduling [15]
The Coloring-Based Real-Time Communication Scheduling (CoCo) solution is presented in reference [15]. CoCo is designed for multihop WSN environments that use
QoS-BASED MAC PROTOCOLS FOR WSNs
371
IEEE 802.11 MAC protocol, where all communication is unicast. It is assumed that
node locations are available at all times, and a central scheduler running CoCo is in
charge of communication scheduling. CoCo aims to schedule real-time communication avoiding collisions and minimizing the overall packet transmission time.
In CoCo, a set of messages waiting for transmission at various sensors are modeled
with a weighted, directed graph G = (V, E), where a vertex denotes a sensor node,
a directed edge from vertex vi to vj denotes a message to be sent from sensor vi to
vj , and the weight of an edge denotes the transmission time. The communication
problem is equivalent to assigning a color to each edge. Here, each color represents a
set of simultaneous communication during disjoint time periods, and the weight of a
color equals the maximum weight of the edges assigned with this color. CoCo aims
to find edge color assignment such that (i) no adjacent edges share the same color,
(ii) no two edges with the same color interfere with each other, and (iii) the overall
weight of used colors is minimized.
Since the optimal coloring problem is NP-complete, a coloring heuristic is presented in reference [15]. First, the edges of the vertex with the maximum degree are
assigned different colors. Once a color is assigned to an edge, it is removed from the
palettes of all adjacent edges, and its weight is updated. Then, the following steps
are repeated until all edges are colored: The edge with the smallest palette is chosen.
A color from the available palette is assigned to the edge such that no other edge
with that color interferes with the chosen edge. Then, the chosen color is removed
from the palettes of all uncolored adjacent edges. Three heuristics are presented for
selecting a color from an edge’s palette: The Random Color Selection Heuristic randomly picks a color from the palette that does not cause interference. The Least Used
Color (LUC) Heuristic chooses the color with the smallest number of colors. The
Minimal Weight Color (MWC) Heuristic first checks whether there are colors in the
palette whose weights are higher than the edge. If so, among these colors, the color
with the smallest weight is selected. Otherwise, the color with the maximum weight
is assigned from the palette.
CoCo aims to schedule a set of communication events with the minimum communication time in real-time WSNs. According to the simulations, MWC-based CoCo
provides performance superior to that of the other two-color selection heuristics, and
its performance is close to the optimal solution. The central computation requirement
limits the applicability of CoCo in large-scale sensor networks.
13.3.3 Reliability Maintenance Through Activity Management [16]
In reference [16], an activity management mechanism (AMM) is proposed to maintain
communication reliability in WSNs. WSNs are considered to work with underlying
IEEE 802.15.4 protocol operating in beacon-enabled slotted CSMA/CA mode. Sensors are organized with a star topology, where a node must be admitted by a coordinator
to participate in the network. The coordinator is aware of the number of nodes in the
network, packet arrival rates, and the desired reliability R. Here, the reliability R is defined as the number of packets delivered to the coordinator per unit time. AMM aims
to provide network reliability guarantee through sensor node activity management.
372
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
In AMM, the MAC layer exhaustively serves packets in First-Come First-Serve
(FCFS) manner. A sensor goes to sleep only after all packets in its buffer are transmitted. During the sleep period, newly arriving packets are enqueued waiting for
sensor wake-up to be delivered. In case of queue overflow, the packet at the head
of the queue is discarded. The sleep duration is geometrically distributed with the
parameter Psleep . A packet is transmitted starting with a random backoff countdown.
After the countdown, two Clear Channel Assessments (CCA) are executed by listening to the channel to make sure it is idle. If both CCAs pass, the packet transmission
starts and an ACK packet is expected. In the case where an ACK is not received,
the transmission is repeated and the countdown exponent is increased. Through
theoretical analysis of the throughput of the MAC mechanism above, the network
control equation is derived where the network reliability R is a function of several
MAC parameters, such as sensor sleep duration, transmission success probability,
and basic backoff period. Given desired network reliability R, the coordinator calculates the parameter Psleep and broadcasts it to all nodes to regulate their sleep times
accordingly.
AMM provides network reliability thorough activity management and analysis of
MAC parameters. It is shown through simulation that through proper activity management, the network reliability is robust against variation of network scale and packet
rates. As the author points out, AMM is computationally intensive, and distributed
activity management is left to future work.
13.4 QoS-BASED ROUTING PROTOCOLS FOR WSNs
QoS-based routing protocols for WSNs have been proposed in the literature mainly
to support two kinds of performance bounds, namely, delay and reliability. The protocols outlined in this section are implemented primarily in the network layer and, in
some instances, in the MAC sublayer. These solutions also differ in the management
of resources, where some rely on centralized computations while others utilized distributed methods. The common point in all these solutions is that all of them guarantee
at least one performance metric to be satisfied in the network.
13.4.1 Sequential Assignment Routing [17]
On-demand multihop routing algorithms such as AODV and TORA eliminate table
updates in high-mobility scenarios. However, they introduce high-energy cost during
route setup phase. Power-aware routing finds minimum metric paths on two different
metrics: minimum energy per packet and minimum cost per packet. The first metric
produces substantial energy savings, but performance degradation due to link/node
failure is not addressed. The second metric deals with failures by routing traffic away
from low-energy nodes at the expense of high path maintenance cost. The Sequential
Assignment Routing (SAR) [17] algorithm uses the idea of multiple paths while taking
parameters like energy resource, QoS on each path, and the priority of packets into
consideration.
QoS-BASED ROUTING PROTOCOLS FOR WSNs
373
In SAR, a table-driven multipath approach is used to improve energy efficiency
in a low-mobility sensor network. The failure protection is addressed by having at
least k-paths that have no common branches between a node and a sink. This is
called a k-disjoint structure. However, the disjoint property creates strong coupling
between routing tables, rendering localized recovery schemes ineffective. To reduce
this effect, the disjoint requirement is relaxed outside the 1-hop neighborhood of the
sink. Furthermore, localized path restoration procedures are used to decrease energy
cost in failure recovery. Multiple paths from each node to a sink are created by building
multiple trees, each rooted at the 1-hop neighborhood of the sink. Each node uses two
parameters to create routing paths:
r Energy resource that is estimated by maximum number of packets that can be
routed without energy depletion, assuming that the node has exclusive use of the
path.
r Additive QoS metrics where higher metric implies lower QoS.
Path selection is made by nodes that generate packets if no topology change occurs
while packets are being routed to their destinations. The energy cost and delay of links
are considered as additive QoS metrics. Packet priorities are used in a way that packets
with higher priorities use paths with lower latency. In short, for each packet, a weighted
QoS metric is computed as the product of a weight coefficient (the priority of the
packet) and the additive QoS metric. Hence, QoS is provided to each packet relative to
its priority level, where higher QoS is given to higher-priority level packets. The SAR
algorithm minimizes the average QoS metric throughout the lifetime of the network.
Periodic metric updates triggered at the sink node are used to account for possible
changes in the QoS on individual paths and the changes in energy resources. Simulations show that SAR performs better than a minimum metric algorithm that lowers
energy consumption without considering packet priorities. Furthermore, failure recovery is handled by local handshakes between upstream and downstream neighbors
in paths. SAR algorithm addresses low-mobility networks, and routes are established
at packet sources considering link costs and energy resources as a QoS parameter.
Packet priorities are taken into account to relay high-priority packets to popular paths
in terms of latency. However, the scheme requires resource-related topology information at packet sources that require frequent parameter updates by a common sink.
This incurs high overhead in WSNs with moderate or high mobility and in WSNs
carrying high data rates.
13.4.2 Energy-Aware Routing in Mobile and Wireless Ad Hoc
Networks [18]
Before focusing on energy-aware routing protocols for WSNs, it is worth focusing on
energy-saving routing protocol design for wireless ad hoc networks. These solutions
are also directly applicable to WSNs with limited number of nodes and where a
number of potentially mobile, high-capability nodes need to communicate possibly
374
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
over sensor nodes. A good example of modifying existing routing protocols to be
energy-aware has been presented in reference [18]. In this paper, two reactive ad hoc
routing protocols, namely DSR [19] and TORA [20], are modified to deliver QoS by
introducing energy-awareness.
DSR and TORA protocols involve a “route discovery phase” initiated when a
mobile source node needs to send data packets to the destination node but does not
have an active and valid route to it. Route discovery is performed using the control
packets called route request packets (RREQ). The source node broadcasts RREQs and
waits for a control packet called the route reply packet (RREP). Using the received
RREP, the source node updates its routing table, which is used to keep track of active
routes to individual destination nodes in the network. Intermediate nodes, which
forward RREQ and RREP packets between source and destination nodes, also use
RREP packets to update their routing tables. Although these protocols are effective
and robust, the broadcast of RREQ packets leads to unnecessary packet transmissions
and inefficient use of limited energy resources.
In reference [18], two modified protocols, EDSR and ETORA, based on the existing
DSR and TORA protocols, are introduced, respectively. Both EDSR and ETORA
involve an additional RREQ forwarding mechanism that does not exist in DSR and
TORA. In an intermediate node, this mechanism considers the current energy level of
the node, the energy level of the previous sender node, and the distance to the source
node when making routing decisions.
Distance estimation is accomplished using time stamps in RREQ packets. When
a node transmits an RREQ packet, it records the time of transmission in the RREQ.
Intermediate nodes that receive this RREQ calculate the time difference between
transmission and the reception time of the RREQ to estimate their distances to the
sender node. Besides transmission times, energy levels of the nodes right before
RREQ transmission are also recorded in RREQs.
The RREQ forwarding decision in reference [18] is based on cutoff circles that are
placed around each node in a network. Upon the reception of an RREQ packet, an
intermediate node calculates the diameter of its “cutoff” circle using its energy level,
the energy level of the previous node sending this RREQ packet, and its distance to
the previous node calculated by the time stamp in the RREQ packet. The receiving
node simply drops the packet, hence does not forward it, if its cutoff circle encircles
the previous node. In reference [18], ETORA and EDSR are shown to outperform
TORA and DSR, respectively, in terms of overall network throughput, the average
number of data packets received at destinations, average data transmission delay, and
energy consumption. The pseudocode of the proposed RREQ forwarding algorithm
is given in Figure 13.4.
13.4.3 Energy-Aware QoS Routing Protocol for WSNs [21]
The information delivery in video sensor networks requires end-to-end delay guarantees. The QoS-based routing protocol proposed in reference [21] aims to sustain paths
that can guarantee such delays for real-time traffic while supporting non-real-time
(best-effort) data flows, as well. The network architecture assumed in this proposal
QoS-BASED ROUTING PROTOCOLS FOR WSNs
375
Forward RREQ:
tp:
tr :
Ep :
Er :
dc :
l:
c:
1.
2.
3.
4.
5.
6.
7.
Time of transmission by the previous hop
Time of reception by the intermediate node (IN)
Energy level of previous node during RREQ transmission
Energy level of this IN when RREQ is received
Diameter of the cutoff circle of this IN
Distance to the previous hop
Speed of light
Calculate time difference ∆ t := tr − tp
Estimate distance l := ∆ t × c
Determine diameter of cutoff circle dc = 0 .4 · Ep + 0 .4 · Er + 0 .2l
IF dc ≤ l
Drop RREQ packet
ELSE
Forward RREQ packet
Figure 13.4. Energy-efficient RREQ forwarding algorithm.
involves a hierarchical organization: Sensor nodes are grouped into clusters, formed
based on criteria such as communication range, number and type of sensors, and geographical location. Clusters are assumed to be controlled by a single command node
and have their own gateway nodes, which act as cluster heads. Sensors in a cluster receive commands from and send readings to the cluster gateway node. Gateway nodes
of different clusters are able to communicate over long-haul communication links. All
sensors are assumed to be stationary. In addition to non-real-time data generated in
the network, the network also tries to recognize and track targets in individual clusters
using images and video feeds.
In this study, the aim is to find paths within a particular cluster under real-time
constraints, without explicitly addressing communication among gateways. Given a
cluster of sensor nodes, paths are computed centrally by gateways. To this end, a cost is
associated with each link. A link cost is a weighted sum of physical length of the link,
residual energy of the sender, expected lifetime of the sender assuming current power
consumption rate, and the estimated error rate of the link. Using these link costs, the
gateway computes k-shortest paths between individual nodes and the gateway itself.
However, this computation is not sufficient to satisfy the delay constraints of flows.
Based on precomputed k-paths between sources and the gateway, the gateway also
computes the expected delay on every link. Assuming perfect knowledge about the
demand of data sources and a fixed ratio r of resources used for real-time flows, a
queuing model is formed. This queuing model is used to compute the delay between
data sources and the gateway. The path computation algorithm aims to find a common
ratio r such that end-to-end delay requirements of all nodes are satisfied for at least
one of the k-paths associated with each data source. Once computed, the resource
ratio r is broadcast to all nodes along with the selected paths to route information in
the network.
376
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
This work has an interesting approach to the QoS provisioning in the network where
both real-time as well as non-real-time data flows are considered in the network.
The proposed solution is expected to perform well when the size of the cluster is
small. However, the link cost function helps minimize the energy consumption for
appropriate weights for different components of the cost. The authors do not comment
on the selection of the weight parameters. Furthermore, the assumption of knowing
the data rates of individual sensor nodes as well as the up-to-date state information
of relay nodes at a central location is far from being realistic. While this algorithm is
a nice attempt to solve a complex problem, it only does so at the expense of limiting
simplifications and assumptions of cluster-wide state information.
13.4.4 Energy-Aware Data-Centric Routing Algorithm for WSNs [22]
An energy aware and data-centric routing algorithm (EAD) for WSNs is proposed
in reference [22]. The algorithm aims to achieve two performance improving goals:
(i) elimination of redundant data by in-network processing and (ii) minimization of
overall network energy consumption by using a virtual backbone tree for data forwarding. The EAD algorithm is mainly focused on the construction and maintenance
of the forwarding backbone tree rooted at a single data sink with maximal number
leaf sensor nodes. The network operation is composed of two major phases, namely
the initialization and the data transmit phases. Initialization and data transmit phases
together form a “round.” The construction and following updates of the backbone tree
are performed during the initialization phase.
The construction of the forwarding tree is based on the following idea: The dominant part of the energy consumption in a sensor node is due to data transmission and
reception. To minimize the overall network energy consumption, some nodes should
turn their radios off (leaf nodes) while others should remain relaying packets (nonleaf nodes). Hence, to achieve minimal energy consumption, the focus of attention
is on the maximization of the leaf nodes. Since this is an NP-complete problem, the
EAD algorithm heuristically attempts to achieve this goal. The algorithm periodically
updates the distribution of the leaf and non-leaf nodes during the initialization phases
by considering a sensor node as a state machine.
The choice of being a non-leaf node depends on two mechanisms, namely neighboring broadcast scheduling and distributed competition among neighbors. These
mechanisms ensure that sensors with higher residual power have higher chance to
become a non-leaf node, hence conserving the local energy that eventually leads to
reduction in overall network energy consumption. Leaf and non-leaf nodes change
their states upon the reception of messages from neighbors indicating their parents,
energy levels, and distance to the sink node. Nodes sense the channel before transmitting these messages and also have waiting periods to avoid unnecessary local state
changes.
In EAD, data relaying and in-network data processing tasks are performed in nonleaf nodes. At each sensor, the local raw data is combined with partially processed
data delivered from sensors that are farther away from the sink. (The sensor nodes
keep records of their parent nodes and their child nodes, if any, through which they
QoS-BASED ROUTING PROTOCOLS FOR WSNs
377
determine their distance to the sink.) Non-leaf nodes summarize and forward the
aggregated data to their parents in the tree. To reduce the execution time of EAD processing, a topology-based algorithm is used which preprocesses the network topology
to ensure that all sensors are are spanned by the EAD tree even though a subset of
sensors participate in EAD execution. The outline of the two algorithms used in EAD
are given in Figures 13.5 and 13.6.
Receive_Control_Packet:
n: Current node
n.nodeId : Unique ID attribute of n
n.EAD_ type : Type attribute of n (0: undetermined, 1: leaf, 2: non-leaf )
n.EAD_ previous_type : Previous type attribute of n
n.EAD_ level : Current level attribute of n
n.EAD_ parent : Distance attribute of n to the previous hop
n.EAD has child : Boolean attribute of n indicating if it has children
sink:
P : Received packet
P.ead type : Type of the source of P
P.ead level : Level of the source of P
P.ead parent : Parent node of the source of P
P.source add: Source address of P
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
n.EAD previous type := n.EAD type
IF n.EAD type == 0
IF P.ead type == 2
n.EAD type := 1
n.EAD level := P.ead level + 1
n.EAD parent := P.source addr
ELSE IFP.ead type == 1
n.EAD type := 2
n.EAD level := P.ead level + 1
n.EAD parent := P.source addr
Call finalEADStatusUpdate function
Send control packet to 1-hop neighbors
ELSE IF n.EAD type == 1
IF P.ead parent == n.nodeId
n.EAD type := 2
Call finalEADStatusUpdate function
Send control packet to 1-hop neighbors
ELSE IF n.EAD type == 2
IF P.ead type == 2 AND P.ead parent == n.nodeId
n.EAD type := 2
Call finalEADStatusUpdate function
Send control packet to 1-hop neighbors
IF P.ead parent == nodeId
n.EAD has child := T RU E
n.EAD type := 2
Figure 13.5. Control packet reception procedure used in the EAD algorithm.
378
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
finalEADStatusUpdate:
n: Current node
P: Received packet
1. IF P.ead parent == nodeId
2.
n.EAD has child := T RU E
3. IF n.EAD has child == T RU E
4.
n.EAD type := 2
5. ELSE IF n.EAD previous type! = 0 AND n.EAD has child == FALSE
6.
n.EAD type := 1
Figure 13.6. EAD status update procedure used in the EAD algorithm.
The performance of the EAD algorithm is compared with the performance of a
simplified AODV without sensor mobility, and LEACH algorithms comparing total
number of active nodes, UDP packet throughput, and energy expenditure. The presented results illustrate that EAD outperforms the other two algorithms with respect
to these performance metrics. Furthermore, it is shown that there is a tradeoff between
system lifetime and system throughput when shorter or longer EAD refresh periods
are chosen. Small refresh intervals enable better throughput but require more energy.
Moreover, it is concluded that there is no tradeoff between the initialization and data
transmit phases of the EAD algorithm.
13.4.5 Reliable Information Forwarding Using Multiple Paths [23]
Data dissemination protocols which are not adaptive to channel error rates and do
not support information awareness either spend excessive amount of resources or
fail to deliver important information with sufficient reliability. Reliable information
forwarding using multiple paths (ReInForM) [23] is a protocol for WSNs to support
information awareness, such that the reliability of data transfer depends on the information content despite the presence of significant channel errors. To define the
desired reliability levels, ReInForM assigns different priority levels to data packets.
Depending on the priority level, multiple copies of the data packets are delivered
along multiple paths. Hence, ReInForM relies heavily on the existence of multiple
paths between a source and a destination, which is generally available in large-scale
WSNs. The simulations investigating the existence and number of edge-disjoint paths
show that a network slightly denser than a minimally connected graph is sufficient
to have as many edge-disjoint paths as the average node degree. The deviation in
the number of hops of these paths is found to be less than two hops, which suggests
that the paths have nearly identical lengths. Hence, data delivery on these paths has
comparable latency and efficient load balancing among multiple paths is possible.
Under ReInForM, the source node of a packet determines the importance of the
information in the packet and decides on a reliability level (rs ). Using the local channel
QoS-BASED ROUTING PROTOCOLS FOR WSNs
379
error information (es ) and the hop distance to the sink (hs ), the source computes the
number of paths, P(hs , rs , es ), required to deliver the packet at the chosen reliability
level. The neighbors of the source is divided into three subsets, Hs − ,Hs 0 , and Hs + ,
designating the neighborhoods at distances of hs − , hs 0 , and hs + hops to the sink,
where hs − < hs 0 < hs + and hs = hs 0 is the hop distance of the source to the sink.
The chosen total number of paths that the source is expected to create, P, is divided
into these three sets of neighborhood.
A random node in Hs − is chosen to be the default node, which always forwards
packets. This ensures that we have at least one path toward the sink. Other nodes in
Hs − , using their own local channel error rate e, hop distance to sink h, and reliability
r, compute their own P value. If this P value is larger than 1, then the node is chosen
to be a forwarding node. If the value is less than 1, then the probability that the node
is a forwarding node is simply this local P value. Eventually, a number of forwarding
nodes are chosen in set Hs − . If there are still more paths to be established (meaning
that the number of paths over the set Hs − is less than the total number of paths), then
additional paths are created by the nodes in the set Hs 0 in the same way. Paths over the
set Hs + are created only if there are still more paths needed besides the ones created
by Hs − and Hs 0 .
The nodes that decide not to forward a packet simply drop it. Packets carry minimal
state information to aid the forwarding decisions. The dynamic local states containing
hop distance to sink, reliability, and channel error rate are updated regularly at each
forwarding node. After receiving a packet from the source and updating the dynamic
states, the node effectively becomes a source. Using the local state information, the
node uses the same procedure to compute the path values for its own neighbors and
the process continues.
ReInForM is one of the leading examples of multipath routing protocols for WSNs.
The gains attained through local multipath forwarding mechanisms lend predictability
to applications in terms of reliability. While load balancing is argued to be a natural result of the protocol, the study does not present any the analysis of this issue.
Furthermore, not all alternative edge-disjoint paths are of equal length, potentially
leading to out-of-order delivery of packets and unpredictable delays.
13.4.6 SPEED Protocol [24]
In large-scale WSNs, the availability of state information of individual nodes cannot
always be assumed. Therefore, local decision-based solutions prove to be more flexible as well as feasible for such large-scale networks. SPEED protocol [24] is designed
to provide soft end-to-end deadline guarantees for real-time packets in WSN. It uses a
geographic forwarding mechanism such that each packet can be routed without global
topology information. Thus, it scales well in large sensor networks. More importantly,
it ensures a network wide speed of packet delivery for real-time guarantees. The network is assumed to be composed of nodes that have location information. Since the
protocol is based on a geographic routing algorithm, nodes are also assumed to gather
information about their neighbors’ locations.
380
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
j
delayi,j = 0.1sec
distj,k = 80m
dist i,j = 40m
distj,k = 100m
i
κ
dist i,k – distj,k = 20m
geographic progress
along the virtual direct line
Figure 13.7. Packet progress speed calculation under SPEED protocol.
The concept of soft real-time guarantees in the context of the SPEED protocol refers
to the fact that packets travel to their destination at a given propagation speed. For this,
each node maintains information about neighboring nodes such as geographic distance
and average delay to each neighbor. Using the distance and delay, each node evaluates
the packet progress speed of each neighbor node for a packet sent to a specific destination. An example of packet progress speed calculation is shown in Figure 13.7. Let a
packet in node i be destined to k, which is 100 m away. Let i forward the packet through
a neighbor j, which is 80 m away from k. If this forwarding takes, on the average,
0.1 s, then the packet progresses to its destination k at 20m
0.1s = 200 m/s. Under SPEED
protocol, a packet is forwarded through a neighboring node if and only if the progress
speed through that neighbor is higher than the specified lower-bound speed SetSpeed.
If each node can find a neighbor that can progress a packet with a speed higher
than SetSpeed, SetSpeed can be guaranteed in the entire network. However, if the
load carried in the network is too high, uniform speed guarantees cannot be provided
in the network. When a node cannot find any neighbor node whose speed is higher
than SetSpeed, it probabilistically drops packets to regulate the workload such that at
least one neighbor node with a speed higher than SetSpeed exists at all times. At the
same time, the node sends a back-pressure packet to the previous nodes to prevent
them from forwarding any further packets through this congested area. Hence, packet
delays can be upper-bounded at the expense of packet losses in the network to sustain
network-wide packet progress guarantees. The algorithms used in the SPEED protocol
are given in Figures 13.8 and 13.9.
With the SPEED protocol, a uniform packet progress speed is guaranteed in the
entire network. Furthermore, the protocol relies only on local information augmented
with limited scope feedback, which improves its scalability. However, the SPEED
protocol provides only one network-wide speed, which is not suitable for differentiating various flows with different deadlines. Furthermore, it does not provide any
guarantees in packet delivery: The fraction of packets lost or dropped in the network
QoS-BASED ROUTING PROTOCOLS FOR WSNs
Forward Packet:
i: Current node
D : Destination node
d(i, D ): Destination of node i to D
NS i : Neighborhood set of node i
FS i : Set of all nodes in NS i closer to D than i
HopDelay (i, j ): Estimated delay from i to j
Speed (i, j, D ): Estimated speed from i to j
p: Packet being processed
BPP : Back-pressure packet
US i : Upstream Node Set of node i
S setpoint : Speed threshold
FS i set 1 : Nodes in FS i with sufficient speed
FS i set 2 : Nodes in FS i with insufficient speed
en : Miss ratio of neighbor n
RR : Relay ratio
K : Proportionality gain
1. FSi := FS i set 1 := FS i,set2 := φ
2. FOR all k in NS i
3.
L next := d(i, D) − d(k, D)
4.
IF L next > 0
5.
Add k to F Si
6.
Speed (i, k, D) := (L − L next) /HopDelay (i, k )
7.
IF Speed ( i, k, D ) > Ssetpoint
8.
Add k to FS i,set 1
9.
ELSE
10.
Add k to FS i set 2
11. IF FS i == φ
12.
Drop p
13.
FOR all u in US i
14.
Send BPP to u
15. ELSE IF FS i set1! = φ
16.
Choose k in F S i,set1 with max(Speed (i, k, D))
17.
Forward p to k
18. ELSE
19.
R := 1− K × mean (e n)
20.
Generate random number RN in [0,1]
21.
IF RR < RN
22.
Drop p
23.
FOR all u in USi
24.
Send BPP to u
25.
ELSE
26.
Choose k in FS i with max(Speed (i, k, D ))
27.
Forward p to k
Figure 13.8. Stateless nongeographic forwarding algorithm of the SPEED protocol.
381
382
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
Process BPP:
i:
j:
D:
SenttoDelayj :
AvgSenttoDelay :
Current node (recipient of the BPP)
Node with congested downstream neighbors(sender of the BPP)
Destination node
Delay of downstream node j
Average of SenttoDelays of nodes in FS j
1. IF j ∈ FS i
2.
Drop BPP
3. ELSE
4.
Adjust SenttoDelay j with AvgSenttoDelay
5.
IF i is congested for all k in FS i
6.
Send BPP to all upstream nodes of i
Figure 13.9. Back-pressure packet handling algorithm of the SPEED protocol.
cannot be know ahead of time. This, in turn, renders the SPEED protocol not entirely
suitable for real-time applications for WSNs.
13.4.7 MMSPEED Protocol [25]
The two important shortcomings of the SPEED protocol [24] are the support of a
single propagation level and the lack of support for ensuring end-to-end reliability.
The MMSPEED protocol [25] addresses these two shortcomings. This is achieved
by creating multiple logical layers in the same physical network. In Figure 13.10a,
two logical delay layers that provide two different packet propagation speeds on the
same physical network are depicted. To provide delay guarantees, the idea of providing
network-wide speed guarantees is adopted [24] for each logical layer. The speed layers
are isolated from each other through prioritization in queuing and channel access.
Figure 13.10b shows how different reliability levels can be provided in the same
network. To guarantee different reliability levels, packets are routed over multiple
paths based on the requirements contained in headers and on local statistics on packet
loss.
Local forwarding decisions are made considering local statistics and required speed
and reliability levels contained in every packet’s header. Let a packet x be generated
by the source s. The source node s calculates the required speed S req (x) for x so
that x reaches its destination d by its deadline D (x), that is, S req (x) = D|s,d|
(x) , where
| s, d | is the distance between s and d. The source node s includes the required speed
S req (x) in x’s header. First, let us consider the delay QoS provisioning in a network
that supports L layers of propagation speed S1 , . . . , SL . At the source, s selects the
req(x) }.
minimum speed layer l larger than S req (x), that is, Sl = minL
j=1 {Sj | Sj ≥ S
Then x is forwarded to one of the neighbor nodes i that has a propagation speed of
d = |s,d|−|i,d| with respect to d, where d
Ss,i
s,i is the delay estimate from i to s, and
ds,i
d ≥ S . As the packet is forwarded in the network, it may be propagated slower
Ss,i
l
QoS-BASED ROUTING PROTOCOLS FOR WSNs
Logical high−speed network
(20km/s guaranteed uniform speed)
Logical low−speed network
(10km/s guaranteed uniform speed)
Physical network
383
20km/s progress speed
10km/s progress speed
source
destination
(a) Delay QoS Domain
loss
high reliability
loss
low reliability
Physical network
source
destination
(b) Reliability QoS Domain
Figure 13.10. Two QoS domains and corresponding layers implemented in the same network.
than anticipated due to random node placement and lack of coordination between
distant nodes. If an intermediate node finds that a packet cannot reach its destination
at a speed layer, the packet is pushed to a higher speed layer with a new required
speed S req . Hence, errors in local decisions are compensated as packets traverse the
network.
To provide reliability in reaching the destination, the packet loss rates to neighbors
are monitored and the total number of hops to the destination is estimated. The source
node s includes the reliability requirement Rreq (x) of packet x in the header. Consider
x being forwarded by node i. All nodes including i keep packet loss statistics for their
neighbors j, indicated by ei,j . The end-to-end reachability estimate Rdi,j from node
i to destination d over neighboring node j is calculated as Rdi,j = (1 − ei,j )K , where
|i,d|
⌉ is the estimate of hop distance from i to d. After determining its
K = ⌈ |i,d|−|j,d|
neighbors that can sustain the required
x, node i chooses a subset of
speed for packet
d
req
neighbors j1 , . . . , jm such that 1 − m
n=1 (1 − Ri,jn ) ≥ R (x). In other words, we
try to make sure that the probability of one copy of x to reach d is not smaller than
the original required reliability level. At the same time, we make sure that the packet
is propagated at the required speed. In each of the copies sent to these neighbors jn ,
the required reliability of packet x is updated as Rreq (x) = 1 − ei,jn . Hence, as the
number of paths x is relayed over increases, the individual reliability levels decrease
384
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
while preserving the total reliability level. Note that the probability of discarding
a packet to sustain a speed level increases as the required reliability for the packet
decreases.
At this point, we would like to emphasize the interactions between the network
layer and the MAC layer. The network layer makes decisions about the selection of forwarding nodes for each packet and maintains a prioritized queue for different speed
layers. On the other hand, the MAC layer implements mechanisms for prioritized
access to the channel according to speed levels and implements a multicast mechanism. The network layer makes its forwarding decisions based on the information
it obtains from the MAC layer. The MAC layer monitors the delay and packet loss
probabilities for each neighbor.
The MMSPEED protocol provides delay as well as reliability guarantees to realtime flows. With its dual objective nature, MMSPEED stands out as a multifaceted
protocol that cuts across MAC and network layers. These features come at the price
of higher complexity in monitoring the network as well as storage requirements.
Hence, added storage and processing requirements may potentially increase the cost
of individual nodes in the sensor network.
13.4.8 Dynamic Delay-Constrained Minimum Energy Dissemination
Protocol [26]
Some real-time applications for WSNs may require multiple sinks to obtain
sensory data from a single source. At the same time, the end-to-end data delivery
delay between a data source and sinks is required to be upper bounded. The Dynamic
Delay-Constrained Minimum Energy Dissemination (DEED) protocol [26] aims to
form and maintain multicast trees in the WSN that minimize energy consumption
while guaranteeing end-to-end delays. Since the number and locations of the sinks
may not be available, it is required to dynamically adapt the data dissemination tree
upon the arrival of new sinks and leaving of the existing ones. A single source node
in the WSN sends data, and arbitrarily located mobile/stationary sinks request these
data. Sensor nodes are assumed to be aware of their own geographic locations and
their immediate neighbors. Moreover, each sink has an upper bound for end-to-end
message delivery delay (UBED) from the source.
The dissemination tree (d-tree) is rooted at the data source node. New sensors are
added to the tree as new sinks request data from the source and register themselves
with the tree. Hence, the construction of the d-tree starts from the single source node,
and it continues as new sinks arrive. The DEED scheme deals with this construction
while minimizing the total energy consumption and meeting UBED requirements.
The structure of the d-tree has two main parts, namely, the static part and the mobile
part. The static part consists of the source, the sensors (relays) and the multihop
edges between the relays. The sensors that connect sinks to the d-tree are called
access relays (ARs). The mobile part of the d-tree is simply the set of one-to-one
connections between sinks and their corresponding access relays (AR).
The end-to-end delay in the DEED protocol (which is upper bounded by the UBED
Dm for each destination m) is composed of the end-to-AR delay upper bounded by
QoS-BASED ROUTING PROTOCOLS FOR WSNs
385
Pm and the delay between a sink and its access relay, upper bounded by δm , where
δm + Pm = Dm . The delay through multiple hops is the sum of queuing, transmission,
propagation, MAC, and retransmission delays. However, since the queuing, MAC,
and retransmission delays are unpredictable, DEED introduces a new parameter, the
average delay per distance q, which is obtained through tests like ping applications
and then given to sensor nodes. Furthermore, since the geometric distance of multihop
edges are nearly proportional to hop count in a sensor network, geometric distance
is used as a measure of delay along with the parameter q. The end-to-end delay is
computed as the sum of the edge delays along an end-to-end path.
The tree-update procedure when a new sink joins is also the mechanism of
d-tree construction. The packets for constructing the d-tree are forwarded by greedy
forwarding, whereas data packets are broadcast and only the nodes that cache the
addresses of the senders receive the packets. The procedure for d-tree construction
is as follows: When a sink m wants to join the d-tree, m locates its nearest neighbor
am and chooses am as its access relay. Then, the sink m sends a JOIN query over am
to the source. Upon the reception of the JOIN query, the source subscribes the sink
and its access relay. This procedure is called the subscription phase. Then, the source
initiates the gate-relay search procedure. A gate relay is the relay in the existing tree
where a separate branch of the tree is created to reach the access relay. The gate relay
is searched recursively over the relays of the d-tree starting from the source node.
Let r[i] be the relay that is checked to be a candidate gate relay, where r[0] is the
source sensor. Let H be the union of the set of children of the relay node r[i] in the
d-tree and the relay node r[i]. Furthermore, let si be the delay from the source to r[i].
The objective in the gate relay search is to minimize d(h, am ), which is the distance
between the access relay and a relay node in H. Moreover, the delay between the
source and the access relay over this gate relay should be lower than the end-to-AR
delay constraint Pm of the sink as follows:
(si + q[d(r[i], h) + d(h, am )]) < Pm
(13.1)
Here, h is the element of H with minimum d(h, am ). If h is found to be the relay
node itself (such that d(r[i], am ) is minimum d(h, am )) , then r[i] becomes the gate
relay. Otherwise, node h with minimum d(h, am ) is assigned to be the next relay node
r[i + 1] to run the same algorithm recursively and si+1 is updated accordingly.
The third step of the three construction aims to locally adjust the tree around the
gate relay (hence form a branch leading to the access relay) to produce an optimum
dissemination tree from source to the destination. A junction node J in the neighborhood of g is searched such that the total of the distances from gate relay g, its closest
child c, and the access relay am to the junction J is minimized. J should also satisfy
the delay constraints of the sink m, as well as all the sinks that node c has as its
descendants. If the delay constraints are not satisfied for all neighbors of g, then am
becomes a direct child of node g. Otherwise, the chosen junction J becomes a relay of
the new branch emerging from g toward am and J recursively runs the same algorithm
until the delay conditions are not met for a junction’s all neighbors. At this point, all
the relay nodes on the new branch are determined.
386
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
If the sink is mobile, the path leading to the sink and the distance between the
sink and the access relay change. Hence, the sink notifies its access relay (AR) of
its latest nearest-neighbor node in order to continue to communicate with the AR. If
the sink-to-AR distance is increased too much and the sink-AR delay constraint is
violated, then the sink needs to select another access relay.
DEED is an efficient algorithm to minimize energy consumption and meet
delay requirements. However, it does not propose service differentiation among flows
toward different sinks. Furthermore, decrease of reliability in data delivery due to
possible link errors in wireless channels is not addressed. Additional mechanisms to
address path changes due to link errors and congestion are needed. DEED adjusts the
delay/distance parameter to handle congestion, but this does not guarantee avoiding
highly congested local spots. Furthermore, these additional mechanisms must be local
and should not alter the overall tree structure.
13.4.9 QoS for Data Relaying in Hierarchical WSNs [27]
This study presented in reference [27] addresses the selection of one or more routes
from sensors to a base station. The routes are chosen such as to satisfy the delay
requirements. In this study, large heterogeneous WSNs are considered which are
organized in three tiers of hierarchy: a base station (BS), relay nodes (RN), and finally
sensor nodes acting as data sources. Relay nodes are placed such that connectivity
is maintained. Additional relay nodes are placed to improve energy consumption
and reduce interference. However, increasing the number of relay nodes increases
the end-to-end delay. Furthermore, relay nodes that are closer to the base station
consume more energy compared to others. To address these issues, a hybrid approach
is proposed that introduces relay gateways (RG) that receive data from relay nodes
and send them directly (in a single hop) to the base station. RGs are pre-deployed in
the network and are stationary.
The routing decisions at RN-RG and RN-RN communication level consider the
system lifetime as a constraint. System lifetime is defined as the time until at least
one RN or RG depletes its energy supply. The lifetime of a node is modeled by the
maximum amount of traffic it can handle, which is called the node capacity. The
routing from sensors to relay nodes is assumed to be handled by low level protocols.
Hence, this protocol only deals with efficient routing of data among relay nodes by
delivering it to one of the RGs while meeting the delay requirements. Two different
cases of selecting a relay path are proposed: multipath relaying with delay constraints
(MPD) and unconstrained multipath relaying (MP). Two different algorithms that are
both centralized and optimal are proposed to solve these problems.
In the MP problem, the aim is to minimize the end-to-end communication cost
while meeting the capacity constraints of RNs and RGs. However, this problem does
not deal with end-to-end delays. The WSN is modeled as a directed graph composed
of a set of vertices (composed of RNs and RGs) and a set of arcs which represent
the edges between vertices. The cost c of sending a packet over an arc is defined as
a function of individual arcs. Furthermore, the flow of data over an arc is defined
as a separate function x, which is used to model the routing decisions of every RN.
QoS-BASED COMPUTATION WITH COMMUNICATION SUPPORT
387
Hence
the total cost of communication over the selected arcs from source RN to RG,
x(a)c(a), is minimized, where a is an arc.
The capacity of a node i is represented as γ(i), whereas the total amount of data
forwarded to a relay from sensors is β(i), which is called the demand of i. The set
of all arcs entering a node i is δ(i−), and the set of all arcs leaving i is δ(i+). The
minimization of total cost is subject to x(δ(i+)) − x(δ(i−)) = β(i), meaning that the
demand of a relay node i is equal to the net flow into i, that is, x(δ(i−)) − tβ(i) ≤ γ(i),
where it is assumed that transmission requires t/(1 − t) more energy than reception.
The MP problem is modeled as a transshipment problem, which is a classical
problem in operations research. The transshipment problem can be solved in strongly
polynomial time. In the MPD problem, apart from minimizing the total cost and
meeting capacity constraints, the total number of intermediary nodes on a path should
not be larger than a given value to limit the total delay. Hence, packets that originate at
different RNs should be distinguished. The delay constraint can be formulated either
using flow functions or using feasible paths. In the former, relay nodes increment a
hop-count index that is assigned to individual flows. In the latter, the set of all feasible
directed paths P(r, k) originated from an RN r and ending at an RG k is defined.
The paths with lengths smaller than a threshold are chosen. However, the demand of
every node should be met and the capacity of all nodes should be observed. The MPD
with feasible paths formulation can be solved optimally using a linear program. The
advantage of feasible path formulation over the flow function formulation is that the
number of constraints is linear in the size of the graph. Column generation, which is
an implementation of the simplex algorithm for solving linear programs, is used to
find a solution to the MPD problem.
This work is based on a graph representation of a WSN, which is used to find
QoS-based paths in an hierarchical structure. However, methods to handle network
dynamics is not considered. Furthermore, since the algorithm is centralized, it is
assumed that every parameter, including the topology, link bandwidths, and link costs,
are known at a central location, which is not very practical. A distributed implementation is needed to ensure its practical applicability.
13.5 QoS-BASED COMPUTATION WITH COMMUNICATION
SUPPORT
The idea of reducing the communicated data volume through methods like data
aggregation has been recognized as a means to reduce energy consumption and prolong WSN lifetime. While majority of in-network processing proposals involve simple
operations, more complex algorithms have recently been proposed to process higher
data volumes such as in video sensor networks. The main idea behind such proposals is to leverage the collective processing power of individual sensor nodes to run
complex processing applications. To fully utilize the collective processing power of
sensor nodes, solutions from parallel processing literature have been adopted. The
proposed methods also have strong connections with the communication protocols:
The exchange of intermediate results occurs over shared wireless channels, which is
388
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
not considered in wired interconnected networks of processors. The resulting communication schedules determine channel access in sensor clusters, directly affecting
the MAC layer in WSNs.
13.5.1 Collaborative Resource Allocation Algorithm [28]
The Collaborative Resource Allocation (CoRAl) algorithm [28] aims to dynamically
allocate resources such as bandwidth and CPU time for multiple periodic applications
in WSNs. Subject to resource availability, CoRAl aims to adjust application sampling
frequencies to meet the temporal constraints and maximize network utility. CoRAl
is assumed to be executed in fully-connected single-hop WSNs, where all nodes
are synchronized and use Earliest Deadline First (EDF) as the scheduling algorithm.
Nodes also implement the implicit EDF algorithm as the underlying wireless network
MAC protocol. End-to-end applications are considered in reference [28] that are
composed of a chain of tasks already assigned to sensors and sequentially executed
in a pipelined manner.
CoRAl achieves its goals by iteratively executing the following steps until the
schedule converges: First, the task execution frequencies on each sensor are locally
optimized subject to application execution frequency upper bounds, whose initial values are set to be infinite. Then the execution frequency upper bound of each application
is reevaluated based on the updated task frequencies and bandwidth allocation.
In CoRAl, the wireless channel is modeled as a dummy node on which only communication can be executed, and the network bandwidth is allocated in the same manner as sensor CPU time allocation. The CoRAl algorithm is presented in Figure 13.11.
In each node, an extended version of the SLSS algorithm [29] is implemented to
CoRAl:
Ti :
f imax :
L:
mk:
f ild :
f ibn :
Application i
Maximum upper-bound frequency of application Ti
Number of Applications
Sensor node k
Frequency of leader task of application Ti
Frequency of bottleneck task of application Ti
1. Initialize maximum upper-bound frequency of each application Ti :
2.
f imax = + ∞ , i ∈ {1, ..., L }
3. WHILE schedule not converge
4.
FOR each sensor m k
5.
FOR each task of application Ti assigned on m k , i ∈ {1, ..., L }
6.
Locally optimize the task subject to f imax using the extended SLSS
7.
FOR each application Ti
8.
Reevaluate f imax with updated f ild and f ibn
Figure 13.11. The CoRAl algorithm.
QoS-BASED COMPUTATION WITH COMMUNICATION SUPPORT
389
compute locally optimal frequencies subject to node utility constraints. Different from
the original SLSS algorithm, the extended SLSS algorithm in reference [28] takes
each task’s application execution frequency upper-bound into consideration. After
each iteration of local optimization, the upper-bound frequency of each application is
calculated. Let the leader task ldi and bottleneck task bni of an application Ti be tasks
whose frequency fild and fibn are highest and lowest among all tasks of Ti , respectively. The frequency upper bound of Ti is updated as fimax = fibn + (fild − fibn )σ,
where σ is the factor that controls frequency convergence speed. The optimization
procedure terminates when the weighted difference between leader and bottleneck
frequencies converges.
CoRAl addresses online resource allocation among multiple applications. According to the simulation results, CoRAl provides performances comparable to the optimal solutions obtained by the nonlinear optimization tool of Matlab at a much higher
execution speed. However, in CoRAl, tasks of applications are assumed to be already assigned on sensors, and task mapping remains an open problem. Furthermore,
energy consumption is not explicitly considered in reference [28], which is a fundamental problem in WSNs.
13.5.2 EcoMapS Algorithm [30]
A task mapping and scheduling solution, EcoMapS, is presented in reference [30]
for energy-constrained applications in single-hop WSNs. It is assumed that networks
are composed by homogeneous sensors that can calculate and communicate simultaneously. EcoMapS aims to assign computation tasks and schedule communication
events with minimum application execution lengths subject to energy consumption
constraints. EcoMapS is composed of two phases: the Initialization Phase and the
Quick Recovery Phase. The Initialization Phase algorithm aims to minimize schedule
lengths subject to energy consumption constraints, while the Quick Recovery Phase
algorithm handles runtime sensor failures.
In the Initialization Phase, EcoMapS iteratively searches for the schedule with
an optimal number of computing sensors involved in computation that results in
the minimum schedule length under the energy consumption constraint. To exploit
the broadcast nature of wireless communication, a hypergraph representation of the
Directed Acyclic Graph (Hyper-DAG) is introduced. The Hyper-DAG representation of task dependency explicitly represents communication as well as computation
events: The edges between a task and its immediate successors in a DAG is replaced
with a net, which represents the communication task to send the result of a task to all
of its immediate successors in the DAG. The Hyper-DAG extension of the DAG in
Figure 13.12a is presented in Figure 13.12b, where ri s are the introduced nets. Similar
to CoRAl [28], EcoMapS also models the single-hop wireless channel as a virtual
node where only communication tasks can be executed. Based on the virtual node
model and Hyper-DAG, a communication scheduling algorithm is developed and embedded into the schedule search algorithm, E-CNPT. E-CNPT is a low-complexity
algorithm that first enqueues tasks according to the critical path of a Hyper-DAG, then
assigns the enqueued tasks to the node with minimum execution start time. In case
390
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
v1
v2
e25
e14
e15
v4
v3
e26
e16
e17
v5
e48
v7
v6
e68
e58
v1
v2
v3
r1
r2
r2
e 37
e78
v4
v5
v6
v7
r4
r5
r6
r7
v8
v8
(a)
(b)
Figure 13.12. (a) DAG and (b) hyper-DAG examples.
communication between sensors is necessary, the proposed communication scheduling algorithm is executed. The Quick Recovery Phase algorithm handles sensor failures by adaptively adjusting the previous schedule. If idle sensors exist, the tasks of
the failing sensor are migrated to an idle sensor. Otherwise, they are merged to a
sensor that has the most idle time.
EcoMapS is a task mapping and scheduling solution for WSNs that provides application energy consumption guarantee with minimum schedule lengths. According to
the simulation results, EcoMapS has superior performance over existing mechanisms
in terms of minimizing schedule lengths. Also, the alternative schedules generated
after sensor failures are shown to have satisfying performance with small recovery
latency. However, EcoMapS has no guarantee of application deadline constraints.
13.5.3 Energy-Balanced Task Allocation Algorithm [31]
An energy-balanced task allocation (EBTA) solution is presented in reference [31].
EbTA assumes single-hop clustered homogeneous WSNs with multiple wireless channels, where sensors are equipped with dynamic voltage scaling (DVS)-enabled processors. EBTA considers real-time applications composed by interdependent tasks.
The design objective of EBTA is to map and schedule application tasks to sensors
such that balanced energy consumption is minimized subject to deadline constraints.
In reference [31], applications are represented with directed acyclic graphs (DAGs)
and the scheduling problem is formulated as an integer linear programming (ILP)
problem. The exclusive wireless channel access feature is incorporated as additional
constraints in the ILP problem.
Because the formulated ILP problem is computationally costly, a three-phase
heuristic is proposed in reference [31] to provide a practical solution. In Phase 1,
tasks are grouped into clusters to minimize overall application execution time assuming infinite number of sensors. Each task first constitutes a cluster by itself. Then
all communication tasks are examined in a nonincreasing order of their data volume. For each communication event e(i, j) between computation task Ti and Tj , the
QoS-BASED COMPUTATION WITH COMMUNICATION SUPPORT
391
clusters containing Ti and Tj are merged if it leads to shorter application execution time. When evaluating application execution time, communication events are
scheduled to the channel with smallest available time using the First-Come FirstServe (FCFS) policy. In Phase 2, the task clusters from Phase 1 are assigned to
sensor nodes with the objective of minimizing the maximum energy expenditure
among all sensors. The task clusters from Phase 1 are first sorted in a nondecreasing
order of energy consumption and are stored in a queue . The clusters in are
then assigned to the sensor with the minimum normalized energy consumption (task
execution energy consumption normalized by sensor residue energy, norm energy
for short). Each time after a task cluster is assigned to a sensor, the norm-energy
of the sensor is updated. This procedure repeats until all task clusters are assigned.
Finally, a DVS heuristic is presented for Phase 3 to decrease energy consumption by
iteratively adjusting the CPU voltage level of each task. In each iteration, a critical
node that has the highest norm energy ε is selected. Among the tasks assigned on
the critical node, a task is selected such that, by decreasing its CPU supply voltage to the next level, ε is decreased the most. Each time when a task is adjusted, the
application schedule is iteratively adjusted accordingly to meet inter-task dependency
constraints.
EbTA is one of the first proposals that addresses task allocation in WSNs, where
both communication and computation tasks are considered. It is shown through simulations that the three-phase heuristic achieves longer lifetime compared with the
baseline without DVS. The performance of the three-phase heuristic is also found to
be comparable to that of the ILP-based approach via simulations.
13.5.4 RT-MapS Algorithm [32]
The RT-MapS algorithm [32] is proposed for single-hop clustered WSNs, which
are composed of homogeneous DVS sensors with finite number of voltage levels.
The design objective of RT-MapS is to provide application deadline guarantees with
the minimum energy consumption for WSNs applications. The RT-MapS algorithm
contains two phases, namely, Task Mapping and Scheduling (TMS) Phase and DVS
Phase. The flowchart of RT-MapS is shown in Figure 13.13. In the TMS phase,
computation and communication events are simultaneously assigned and scheduled
with the objective of minimizing energy consumption subject to deadline constraints.
To guarantee deadline constraints, sensors are scheduled with highest CPU speed in
the TMS phase. Schedules generated in the TMS phase are then further optimized in
the DVS phase by reducing CPU speed to decrease energy consumption. Similar to
EcoMapS [30], RT-MapS employs Hyper-DAGs to represent applications and utilizes
the virtual node model of wireless channels.
The RT-MapS solution is outlined with the pseudocode in Figure 13.14. Here, the
communication scheduling algorithm sequentially schedules communication tasks
on the virtual channel node to avoid packet collision. Broadcasting is also realized in the communication scheduling algorithm to conserve energy. The communication scheduling algorithm is embedded in the execution of the task mapping
and scheduling algorithms, H-CNPT and H-MinMin. H-CNPT is different from
392
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
Start
Convert DAG
to Hyper-DAG
Start to assign the
Task Mapping and
task with H-CNPT
Scheduling Phase
or H-MinMin
(Maximum CPU speed)
Communication
required by
dependency
constraint?
DVS Phase
Algorithm
Yes
NO
NO
Assign the task
with H-CNPT or
H-MinMin as
search engine
Execute
communication
scheduling
algorithm
End
YES
All tasks
scheduled?
Figure 13.13. Flowchart of RT-MapS.
E-CNPT [30]. Among schedules with different number of computing sensors, the
schedule satisfying the deadline constraint with the minimum energy consumption is
selected as the optimal solution. The H-MinMin algorithm is an extended version of
the Min–Min algorithm [33]. In the core of H-MinMin lies the fitness function that
RT-MapS:
N:
νi :
mk :
mk :
f ild :
f ibn :
1.
2.
3.
5.
6.
7.
8.
9.
10.
Number of sensor nodes in the cluster
Task i of the application
Sensor node k
Sensor node k
Frequency of leader task of application Ti
Frequency of bottleneck task of application Ti
Convert DAG to Hyper-DAG
FOR n = 0 to N
/* Schedule with n computing sensors/*
Assign tasks using H-MinMin or H-CNPT
IF communication is needed for an assignment of task νi on sensor m k
Execute the communication scheduling algorithm
Among these candidate schedules, find the optimal schedule H o :
H o has the smallest energy consumption subject to deadline constraints
Adjust the schedule H o using the DVS algorithm
Figure 13.14. The RT-MapS algorithm.
QoS-BASED CAPACITY ESTIMATION IN WSNs
393
H-MinMin:
L : Mappable task list
νi : Task i of the application
m k : Sensor node k
1. FOR α = 0 step 0.1 to 1
2.
Assign all entry-tasks
3.
Initialize the mappable-task list L
4.
WHILE L is not empty
5.
FOR task νi ∈ L
6.
Find sensor m k with the minimum fitness of νi
7.
Find (ν o , m o ) with the minimum fitness among these combinations
8.
Assign ν o to m o
9. Update L
10. Among all schedules with different values of α
11. Select the optimal schedule
Figure 13.15. The H-MinMin algorithm used in the RT-MapS algorithm.
combines schedule length and energy consumption resulting from assigning a task to
a sensor. In each iteration of task assignment, each “mappable” task whose immediate predecessors are already assigned is tentatively assigned to different sensors. The
suboptimal task-sensor combination with minimum fitness value is kept. Among all
suboptimal task–sensor combinations, the pair that gives the minimum fitness value
is chosen, and the task is assigned to the corresponding sensor. The pseudocode of
H-MinMin is presented in Figure 13.15. The DVS algorithm further reduces energy
consumption of the schedules generated in the TMS phase. In the DVS algorithm,
the communication tasks assigned on the channel are kept unchanged, and their start
time and finish time are taken as the upper and lower bounds to adjust the corresponding sensors’ speed during the time interval. During DVS adjustment procedure, CPU
speed is reduced in proportion to the CPU utility.
RT-MapS aims to provide application deadline guarantees with minimum energy
consumption. Due to the parallelism among sensors and exploiting the broadcast
feature of wireless communication, RT-MapS shows superior performance compared
with exiting mechanisms including EBTA [31], as presented in the simulation part of
reference [32]. However, as other outlined solutions, RT-MapS also fails to extend to
multihop WSN cluster. Furthermore, neither execution of concurrent application nor
interaction with neighboring clusters is considered, either.
13.6 QoS-BASED CAPACITY ESTIMATION IN WSNs
An important step in end-to-end QoS provisioning in any communication network is
the estimation of the network capacity to ensure that admitted flows can be sustained
394
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
in the network. Such estimations are also important to adjust the data injection rates
and other requirements of flows at sources. To provide accurate feedback, sinks must
be able to compute the capacity of the network as functions of QoS parameters and
the flow characteristics. In the literature, capacity modeling of wireless multihop
networks have been investigated with simple metrics. The line of work pioneered
by P. R. Kumar aims wireless network capacity in bit-meters per second [34]. This
approach has also stimulated similar analysis attempts with increasingly realistic
communication models [35–38]. The main shortcoming of these attempts lies in the
oversight of an overarching goal of communication networks: Delivery of information
to remote devices across a network under realistic conditions. Furthermore, combined
metrics do not represent individual metrics such as delay, throughput, and loss rates
explicitly.
Motivated by the lack of multimetric capacity modeling work for multihop networks, we have investigated the following problem [39]: Consider a dense sensor
network deployed in a rectangular area, where sources and destinations are located
along two opposite edges. Assuming a Rayleigh fading channel, what is the maximum
attainable throughput across the network given a delay bound for individual packets
and a maximum tolerable bit error rate? A corollary to this problem is the calculation
of the minimum delay attainable if a particular throughput is required. With the results
of this analysis, it is possible to estimate the total capacity of the network for a given
delay value.
Our analysis is based on a physical layer channel model that accounts for the effects
of the noise as well as the interference from other simultaneous transmissions. The bit
error rates are calculated as the probability of a signal to be below a given threshold
based on the Rayleigh channel model. With this comprehensive interference model,
we analyze the effect of packets in the same data flow (intraflow interference) and
other data flows (interflow interference).
As a first step, we analyzed a data stream that traverses a linear path. Based on a
discrete time model where each packet transmission lasts one time unit, we modeled
the behavior of the packets in the same flow. Our analysis confirmed that every packet
injected into the network slows down the progress of preceding packets. To solve
this problem, waiting times are introduced at the source while injecting packets into
the network, allowing earlier packets to move forward before another packet (i.e., an
interferer) is introduced. To achieve the lowest end-to-end delay, a packet must exist in
the network alone, which reduces the throughput to one packet per end-to-end delay.
Similarly, the maximum throughput can be achieved if packets are injected every time
unit, which causes the delay to go to infinity as the network density approaches infinity.
Intermediate combinations are achievable by adjusting the interpacket waiting times.
The two-dimensional case is modeled as parallel linear flows. In this case, every
transmitted packet experiences interflow as well as intraflow interference. To increase
the throughput, the distance between flows can be reduced. Small interflow separations
increases the interflow interference, causing packets to cover short distances every
transmission time, which in turn increases the end-to-end delay. If the flows are
separated by larger distances, then the overall throughput decreases along with the
end-to-end delay. The parameters that control this behavior are the interpacket delays,
CONCLUSIONS AND OPEN RESEARCH PROBLEMS
395
relative packet injection times between flows, and the interflow separations. Using
nonlinear optimization techniques, it is possible to compute the maximum throughput
for a given delay bound and parameters to achieve the desired performance.
This study [39] is merely a first step in this open research area, and it leaves out
many important issues such as crossing paths, single destination flows, and non-timesynchronized systems. Furthermore, protocol-dependent effects on capacity are also
left out for the sake of simplicity. We are hopeful that more accurate and comprehensive studies will follow this one that will address more realistic settings and that can
be directly integrated to admission control and QoS feedback mechanisms.
13.7 CONCLUSIONS AND OPEN RESEARCH PROBLEMS
QoS-based communication in WSNs is a very promising and emerging field to sustain the communication in next wave of truly real-time WSNs. Although there is a
significant amount of work that has been performed in the last five years in this area,
they are far from addressing all QoS problems in WSNs. While WSNs resemble ad
hoc networks and many more solutions for ad hoc networks do exist, it is important
to keep in mind that WSNs are fundamentally different than ad hoc networks and
have different requirements and constraint. Yet, it is equally important to consider the
lessons learned from earlier QoS-based communication protocol proposals to avoid
similar pitfalls.
Considering the available pool of solutions, several shortcomings in the development of QoS-based communication protocols stand out: Most of the solutions either
rely on hard-to-materialize assumptions or are very complex for implementation on
truly resource-constrained sensor nodes. Although the existing and widely used sensor nodes (such as Mica2 sensor nodes [40]) would not have any problems running
algorithms outlined in this chapter, their implementation on extremely small devices
is questionable. Therefore, simplification of the QoS support mechanisms is very
important. Similarly, implementation of select proposals on very simple devices
should also be undertaken to assess their value in real operation environments.
A majority of solutions consider only an isolated set of problems (such as only
routing, MAC, or processing) or very few QoS parameters (such as delay, reliability,
or energy). While the protocol design for WSNs is heavily influenced by applications,
protocols that can support multiple metrics are still few and far apart. Support of multiple QoS metrics is an important step to realize many real-time and mission-critical
applications. It is clear that such protocols would have to coordinate (if not merge)
several networking functions traditionally considered belonging to separate layers.
The resulting solutions must still maintain a simple nature for easy and ubiquitous
deployment on many platforms.
Another important aspect that has not been studied very extensively is the feedback
mechanisms. The solutions presented in this chapter operate in an open loop; no
feedback is obtained from end nodes to adjust the behavior of the flows or protocols.
Although some examples such as ESRT [41] use feedback mechanisms to sustain
specific reporting requirements, they are mostly application-specific solutions and do
396
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
not apply to generic settings. Furthermore, feedback mechanisms to adjust protocol
behavior would improve the application level QoS guarantees as well as improve the
total number of applications supported in a real-time WSN.
Finally, the QoS-based capacity estimation is a completely virgin field with very
limited work present. Capacity estimation of multihop wireless networks is a challenging topic in itself, and very few proposals can provide actual limits (and not only
scaling laws) even with very simplifying assumptions. Clearly, capacity estimation
has no low-hanging fruits. Significant and coordinated research efforts are required to
derive useful capacity bounds. These bounds should provide not only idealized limits,
but also useful estimates that can be used in combination with QoS-based protocols.
13.8 EXERCISES
1. One of the most important reasons for performance degradation in wireless
networks in general is the mismatch between solutions deployed together. Considering the protocols and solutions outlined in this chapter, identify
(a) Protocols of different layers that cannot be used in the same node, and
(b) Protocols not necessarily in different layers that cannot be run concurrently
in the same network.
2.
3.
4.
5.
6.
Elaborate on reasons for these incompatibilities. Expand your search to more
recent work published in the literature.
Considering the SPEED protocol, how would the selection of the SetSpeed parameter affect the system performance? Based on these observations, propose a
method to select the SetSpeed value for a given overall throughput requirement.
Repeat the exercise above for the MMSPEED protocol.
How should the RT-MapS algorithm be modified such that it can be used in
multihop clusters? Can the presented channel model be sufficient? If not, how
should it be updated?
The ESRT protocol aims to keep congestion under control by changing the
event reporting rates. Identify routing (and MAC if necessary) protocols that
are best suited to be used with ESRT. Discuss possible needs for modification
for seamless operability. Elaborate on the multihop feedback channel on ESRT
performance, and how routing (and MAC) protocols can help with potential
problems.
Discuss how energy-awareness can be incorporated into multi-path routing protocols in WSNs.
BIBLIOGRAPHY
1. I. F. Akyildiz, W. Su, Y. Sankarasubramaninam, and E. Cayirci. Wireless sensor networks:
A survey. Computer Networks Journal, 38(4):393–422, 2002.
BIBLIOGRAPHY
397
2. M. Younis, K. Akkaya, M. Eltoweissy, and A. Wadaa. On handling QoS traffic in wireless
sensor networks. In Proceedings of the 37th IEEE Annual Hawaii International Conference
on System Sciences (HICSS’04), January 2004, pp. 292–301.
3. M. Gerla and K. Xu. Multimedia streaming in large-scale sensor networks with mobile
swarms. ACM SIGMOD Record, December 2003.
4. J. Campbell, P. Gibbons, S. Nath, S. Seshan, and R. Sukthankar. IrisNet: An Internet-scale
architecture for multimedia sensors. In Proceedings of ACM Multimedia Conference, 2005.
5. W. Feng, B. Code, E. Kaiser, W. Feng, and M. L. Baillif. Panoptes: scalable low-power video
sensor networking technologies. ACM Transactions on Multimedia Computing, Communications, and Applications, January 2005.
6. S. Chen and K. Nahrstedt. An overview of quality of service routing for next-generation
high-speed networks: Problems and solutions. IEEE Network, 12(6):64–79, 1998.
7. M. de Prycker. Asynchronous Transfer Mode: Solutions for Broadband ISDN. 3rd edition,
Prentice-Hall, 1995.
8. R. Braden, D. Clark, and S. Shenker. Integrated services in the Internet architecture: An
overview. RFC 1633, June 1994.
9. S. Blake, D. Black, M. Carlson, E. Davies, Z. Wang, and W. Weiss. An architecture for
differentiated services. RFC 2475, December 1998.
10. P. White. RSVP and integrated services in the internet: A tutorial. IEEE Communications
Magazine, 35:100–106, 1997.
11. E. Ekici. QoS-based routing in wireless mobile networks. In Resource Management in
Wireless Networking. Kluwer, 2005.
12. Y. Liu, I. Elhanany, and H. Qi. An energy-efficient QoS-aware media access control
protocol for wireless sensor networks. In Proceedings of International Conference on
Mobile Ad Hoc Sensor Systems (MASS’05), November 2005, pp. 189–191.
13. D. Bertsekas and R. Gallager. Data Networks. Prentice-Hall, 1987.
14. A. Parekh and R. Gallager. A generalized processor sharing approach to flow control in
integrated services networks: The single-node case. In Proceedings of Joint Conference of
the IEEE Computer and Communications Societies (INFOCOM’92), 1992.
15. H. Li, P. Shenoy, and K. Ramamritham. Scheduling communication in real-time sensor
applications. In Proceedings of the 10th IEEE Real-Time and Embedded Technology and
Applications Symposium (RTAS’04), May 2004, pp. 10–18.
16. J. Misic, S. Shafi, and V. B. Misic. Maintaining reliability through activity management in
805.15.4 sensor networks. In Proceedings of 2nd International Conference on Quality of
Service in Heterogeneous Wired/Wireless Networks (QShine’05), August 2005.
17. K. Sohrabi, J. Gao, V. Ailawadhi, and G. J. Pottie. Protocols for self-organization of a
wireless sensor network. IEEE Personal Communications, October 2000, pp. 16–27.
18. A. Boukerche and H. Owens. Energy-aware routing protocol for mobile and wireless ad
hoc networks. In Proceedings of IEEE Conference on Local Networks, 2003, pp. 768–771.
19. D. B. Johnson, D. A. Maltz, Y. Hu, and J. G. Jetcheva. The dynamic source routing protocol
for mobile ad hoc networks (DSR). Internet Draft, draft-ietf-manet-dsr-07.txt, 2002.
20. V. D. Park and M. S. Corson. A highly adaptive distributed routing algorithm for mobile
wireless networks. In Proceedings of IEEE INFOCOM, 1997, pp. 1405–1413.
398
QoS-BASED COMMUNICATION PROTOCOLS IN WIRELESS SENSOR NETWORKS
21. K. Akkaya and M. Younis. An energy-aware QoS routing protocol for wireless sensor
networks. In Proceedings of the 23rd IEEE International Conference on Distributed
Computing Systems, 2003.
22. A. Boukerche, X. Cheng, and J. Li. A performance evaluation of a novel energy-aware
data-centering algorithm for wireless sensor networks. ACM/Springer Wireless Networks,
2005, pp. 619–636.
23. B. Deb, S. Bhatnagar, and B. Nath. ReInForm: Reliable information forwarding using
multiple paths in sensor networks. In Proceedings of IEEE International Conference on
Local Computer Networks, 2003, pp. 406–415.
24. T. He, J. A. Stankovic, C. Lu, and T. Abdelzaher. SPEED: A stateless protocol for realtime communication in sensor networks. In Proceedings of IEEE International Conference
on Distributed Computing Systems, 2003, pp. 46–55.
25. A. Felemban, C. G. Lee, and E. Ekici. MMSPEED: Multi-path multi-SPEED protocol for
QoS guarantee of reliability and timeliness in wireless sensor networks. IEEE Transactions
on Mobile Computing, 5(6):738–754, 2006.
26. H. S. Kim and T. F. Abdelzaher. Dynamic delay-constrained minimum-energy dissemination in wireless sensor networks. ACM Transactions on Embedded Computing Systems,
4(3):679–706, 2005.
27. R. Benkoczi, H. Hassanein, S. Akl, and S. Tai. QoS for data relaying in hierarchical wireless
sensor networks. In Proceedings of the 1st ACM International Workshop on Quality of
Service and Security in Wireless and Mobile Networks, 2005, pp. 47–54.
28. S. Giannecchini, M. Caccamo, and C.-S. Shih. Collaborative resource allocation in wireless sensor networks. In Proceedings of Euromicro Conference on Real-Time Systems
(ECRTS’04), June/July 2004, pp. 35–44.
29. D. Seto, J. P. Lehoczky, L. Sha, and K. G. Shin. On task schedulability in real-time control systems. In Proceedings of the 17th IEEE Real-Time Systems Symposium (RTSS’96),
December 1996.
30. Y. Tian, E. Ekici, and F. Özgüner. Energy constrained local task mapping and scheduling
in wireless sensor networks. In Proceedings of the 1st IEEE International Workshop on
Resource Provisioning and Management in Sensor Networks (RPMSN05), November 2005.
31. Y. Yu and V. K. Prasanna. Energy-balanced task allocation for collaborative processing
in wireless sensor networks. ACM/Kluwer Journal of Mobile Networks and Applications,
10(1–2):115–131, 2005.
32. Y. Tian, J. Boangoat, F. Özgüner, and E. Ekici. Real-time task mapping and scheduling
for collaborative in-network processing in DVS-enabled wireless sensor networks. In Proceedings of the IEEE International Parallel and Distributed Processing Symposium, April
2006.
33. S. Shivle, R. Castain, H. J. Siegel, A. A. Maciejewski, T. Banka, K. Chindam, S. Dussinger,
P. Pichumani, P. Satyasekaan, W. Saylor, D. Sendek, J. Sousa, J. Sridharan, P. Sugavanam,
and J. Velazco. Static mapping of subtasks in a heterogeneous ad hoc grid environment. In
Proceedings of IEEE Parallel and Distributed Processing Symposium, April 2004.
34. P. Gupta and P. R. Kumar. The capacity of wireless networks. IEEE Transactions on
Information Theory, 46(2):388–404, 2000.
35. F. Xue, L.-L. Xie, and P. R. Kumar. The transport capacity of wireless networks over fading
channels. IEEE Transactions on Information Theory, 51(3), 2005.
BIBLIOGRAPHY
399
36. G. Barrenechea, B. Beferull-Lozano, and M. Vetterli. Lattice sensor networks: Capacity
limits, optimal routing and robustness to failures. In Proceedings of 3rd International
Symposium on Information Processing in Sensor Networks, IPSN 2004, April 2004.
37. E. S. Sousa and J. A. Silvester. Optimum transmission ranges in a direct-sequence spreadspectrum multihop packet radio network. IEEE Journal on Selected Areas in Communications, 8(5):762–771, 1990.
38. M. Haenggi. Toward a circuit theory for sensor networks with fading channels. In Proceedings of the 2004 International Symposium on Circuits and Systems ISCAS ’04, Vol 4(IV),
May 2004, pp. 908–911.
39. A. Bader and E. Ekici. Throughput and delay optimization in interference-limited multihop networks. In Proceedings of 7th ACM International Symposium on Mobile Ad Hoc
Networking and Computing (MobiHoc 2006), May 2006.
40. Crossbow, Inc. MICA2DOT Data Sheet. http://www.xbow.com, 2006.
41. Y. Sankarasubramaniam, O. B. Akan, and I. F. Akyildiz. ESRT: Event-to-sink reliabile
transport in wireless sensor networks. In Proceedings of the 4th ACM International Symposium on Mobile Ad Hoc Networking and Computing (ACM MobiHoc 2003), 2003.
CHAPTER 14
Quality of Service in Wireless
Sensor Networks
GREGORY J. POTTIE and AMEESH PANDYA
Department of Electrical Engineering, UCLA, Los Angeles, CA 90095
14.1 INTRODUCTION
The purpose of a sensor network is to extract model parameters of the physical world
according to the fidelity requirements set by some end user(s), for applications such
as basic science, monitoring, or control. The nodes of the sensor network include
some combination of sensing, communications, and signal processing/storage capability, and they may additionally possess attributes such as mobility [1]. Early
sensor networks were typically arranged in a star topology with (a) very limited
processing on the nodes and (b) transport of essentially all data to a central fusion
point. Later, nodes with limited processing capabilities (e.g., microcontrollers) and
storage, along with simple sensors (e.g., geophones or infrared sensors), were developed [2]. There are now large families of nodes ranging in capabilities from small
nodes of this type (“smart dust”) [3] to DSP-based nodes with simple operating systems capable of forming ad hoc networks (“motes”) [4–6] through to nodes with
32-bit machines running Linux [7], advanced chemical sensors, cameras and 802.11
radios. In deploying nodes of all of these types for a variety of applications, including identification and localization of mobile targets and multimodal sensing for basic
science, it has been a challenge to determine in what manner nodes should cooperate
to achieve the basic user objectives. This includes questions such as the minimum
deployment density, what set of nodes should cooperate for detection/identification
of particular phenomena, and how they should cooperate to establish communications among groups performing data fusion and in transporting results to the end
user. There is not one best answer that applies to all situations, but rather careful
attention needs to be paid to the interplay of objectives, physical models, and network
resources.
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
401
402
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
The most important design criterion for any type of network is guaranteeing quality of service, QoS [8]. QoS has become an important issue in various kinds of data
networks, because some users are no longer satisfied with resource allocation based
on service provisioning. QoS measures include bandwidth, delay, and delivery guarantees. Different classes of traffic (e.g., voice, data, image, video, etc.) have different
bandwidth and delay requirements. Many issues of resource allocation for QoS provisioning are discussed in [9–11], although there remains a broad set of optimizations
based on user needs to explore.
Some basic sensor network problem types are depicted in Figure 14.1, illustrating a
number of QoS issues. The first category of problem is to extract model parameters of
some continuous phenomenon within some specified target region (e.g., specification
of a temperature field according to some limited spatial frequency range and with some
temporal resolution). Some sensors immediately outside the boundary region might
be required, while for smoother sections inside the study region some sensors may
not be needed. Even with a statistical model of the field, it may not be clear without
some preliminary survey what density of nodes is required to gather sufficient data
to meet user requirements for fidelity of modeling. An adaptive deployment or node
activation strategy might be needed, with the former implying delay and the latter
implying potentially excessive use of resources. The second problem category is to
determine parameters of some point source (or set of point sources) located within
the convex hull of the sensor network (e.g., location and target type). Here, typically a
small number of sensors in the immediate neighborhood of the source must cooperate,
with the resource optimization problem being to limit the number of nodes involved to
those needed to meet the fidelity requirement, subject to delay constraints. Resources
will be expended in disseminating queries about such events, collecting, processing,
and storing the data, and then transmitting it to the end user, with a number of tradeoffs
available between resource efficiency and latency. Similarly to the first problem,
without some knowledge of the source statistics, it will be difficult to determine the
appropriate node density and cooperative cluster sizes, and so adaptive procedures
may be needed (imposing a large initial delay). In the third problem category, sources
are located outside the network, with goals similar to the second problem category.
Figure 14.1. Three sensor network problems.
FIDELITY AND RESOURCE TRADEOFFS IN FLAT NETWORKS
403
Now, however, that a very large number of nodes will be excited at similar energy
levels; and so to avoid unnecessary expenditure of energy, means must be found to
limit the number of clusters that are involved in fusing information. Individual clusters
might, for example, perform beamforming to establish bearing angles and enhance
the SNR, with the longer-range transport being limited to likelihoods of target identity
and bearing angle, thus permitting fusion and localization.
In all of these cases, there is also potentially longer-range multihop transport of
data to the end user, along with distribution of queries from users to the network that
set policies for searching for and storing (fused and compressed) data. Latency in both
processes should be within the constraints set by the end user. The transport of both
data and queries should be reliable, and in many situations there will also be concerns
for the security of the transport. Additionally, there is little point in collecting sensor
data unless the sensors are calibrated and metadata relevant to the data interpretation
is also included. A variety of propagation models can be considered in these problems,
including distance loss laws and discontinuities caused by obstructions, while source
models can range from the very simple (low-order differential equations for fields,
Gaussian sources) to the more complicated, with varying levels of uncertainty in the
models.
The remainder of the chapter is as follows. In Section 14.2 we discuss some
background topics in network information theory relevant to the efficient collection,
compression, and reliable communication of sensor data. We then discuss how a QoS
perspective enables scalability in classical flat sensor networks, and we explore a number of practical approaches for high-fidelity data extraction in large-scale networks. In
Section 14.3 we discuss some of the implications of introducing mobile elements and
other forms of heterogeneity. In Section 14.4 we describe some of the data integrity
concerns in sensor networks, including both calibration and security issues, for both
flat and heterogeneous networks. In Section 14.5 we provide our conclusions.
14.2 FIDELITY AND RESOURCE TRADEOFFS IN FLAT NETWORKS
14.2.1 Network Information Theory
The sensor network problem is at a fundamental level governed by network information theory: It is a network rate distortion problem, subject to a number of practical
constraints. Sharp information theoretic results could therefore provide guidance on
design tradeoffs. Unfortunately, information theory has had relatively little impact
in multihop networking compared to its success in point-to-point communications.
There are several reasons, among them: It is difficult to compute the capacity of large
networks (or even some networks with as few as three nodes), there is no sourcechannel coding separation theorem for general networks, and the number of possible
interactions presents combinatorial difficulties. Additionally, the optimizations posed
by Shannon are not as complete a match to the practical issues in networking because
they are in point-to-point communications. Latency is less of a concern in single
404
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
links than when it gets aggregated in multiple links across a network. With multiple
users sharing a network, it is possible to assign different QoS values to flows, vastly
expanding the number of possible acceptable solutions. Consequently, the usual mode
of networking research is to (a) create a hierarchy of abstractions to manage the complexity and (b) then propose (and analyze) protocols within each layer to deal with
different QoS, physical resource, and traffic model scenarios. Corresponding to these
abstractions are particular physical components and algorithms [12]. This structure
enables many of the components to be reused even when the conditions or objectives change, minimizing the effort involved in redesigning the system. Therefore,
after reviewing progress in network information theory, we will consider how a QoS
perspective for the theory of sensor networks can enable fresh progress.
Information Theory for 1-Hop Neighborhoods. Neglecting cryptography,
there are two basic questions asked in classical information theory:
r Channel Capacity: For communications between a transmitter and receiver,
what is the maximum rate at which information can be conveyed given the
power, bandwidth and error rate constraints, and the noise in the channel?
r Source Coding: What is the rate of the minimum description of some random
process (discrete or continuous) such that we can perfectly reproduce the process
from our description (noiseless source coding) or can reproduce it with some
bounded distortion (rate distortion coding)?
Enormous progress has been made since Shannon first posed these questions in the
context of point-to-point communications systems [13, 14], and subsequently for
many-to-one or one-to-many systems. Channel coding systems have been devised
for the Gaussian channel and a number of its variants (e.g., the Rayleigh fading
channel) that get within a fraction of a decibel of the SNR corresponding to channel
capacity. The capacities of multiple access, broadcast, and multi-input multiple-output
(MIMO) channels have been characterized [15, 16]. In multiple cell communication
systems, users communicate directly with one or more base stations over a common
set of channel resources (bandwidth/time) while the base stations use an independent
(and low-cost) set of resources to communicate among themselves. One can assume
varying degrees of cooperation among the base stations in coordinating their own
communications and in assisting in separating the communications of the users, with,
for example, cellular capacity more than doubling in going from no cooperation to
having all base stations share information [17, 18]. There is also a broad set of practical
techniques for managing interference, assuming differing levels of ability to estimate
the (dynamic) channel [e.g., 19–21].
In the domain of source coding, practical lossless schemes have been devised
that get very close to the entropy limits for a single source. Additionally, Slepian–
Wolf coding enables separate lossless coding of multiple sources to the entropy
limit assuming modest statistical knowledge has been distributed [22]. In rate distortion (lossy) coding, considerable progress has been made for Gaussian sources. In
FIDELITY AND RESOURCE TRADEOFFS IN FLAT NETWORKS
405
Wyner–Ziv coding, a second source is presumed to supply side information about the
primary source (e.g., correlated sensor data) that can then be used to lower the overall
rate. Here the rate region is known for a single source with side information [23] and
for cases with multiple sources where the sources are conditionally independent, given
the side information [24]. In the Gaussian channel estimation officer (CEO) problem a
single fusion point is assumed, with the objective of minimizing the overall rates from
the (correlated) sources in the presence of Gaussian noise. While the rate region is
not known for all cases, bounds have been developed [25, 26]. For the m-helper problem, one main node makes observations and communicates with a fusion center on
some rate-constrained channel, along with m helpers that provide side information on
the source through a nonconstrained channel. The question is then, How much can the
helpers reduce the rate from the main node for a given distortion? A general solution
is presented in reference 27, which was then applied to problems that more closely
match the usual question of minimizing the total rate of all communications, such as
the separate coding and Gaussian CEO problems. This has resulted in tight upper and
lower bounds for general correlation relationships among the sensor readings [28].
Unicast Communications in Ad Hoc Networks. For unicast communication
in ad hoc networks, pairs of source and destination nodes exchange independent messages across a network. Nodes, all of which share the same communication resources,
may cooperate in transmitting, receiving, and relaying messages. The communications can cause interference to other communications within the network, degrading
its quality of service. The network is thus a combination of relay, interference, broadcast, and multiple access channels, the first two of which have unknown capacities
except in the simplest of situations. However, upper bounds on the scalability of the
network with respect to the capacity of a single link are known. For a network with
uniform source-destination traffic flows and decay of signal strength with distance
according to some rule such as d−k , for k > 2 the
arises from having an
√behavior
average path length (in near-neighbor hops) of n , and thus the need to carry
traffic of roughly that number of users. The overall capacity available per sourcedestination pair declines as the square root of the network size, for fixed transmission
bandwidth [29]. While antenna arrays, cooperative communications, and so on, all
lead to improved communications, they do not sufficiently change the asymptotics
[30–36], so that the network is not scalable under these assumptions on the traffic and
propagation model. Under the
of multihop transmission, it may be shown
√ assumption
that delay also scales as n [37].
In the random graph model, it is assumed that the quality of links between nodes
does not follow a simple geometric rule, a reasonable model for local links dominated
by multipath. The capacity depends strongly on the particular model chosen for link
quality, and can in the most favorable circumstances of high probabilities of long
links being present result in scalability, even assuming uniform source-destination
probabilities [38]. The bounds can be tightened if the source-destination distribution
is highly irregular [39, 40]. For example, for the model considered in reference 29,
if the traffic pattern is such that the average distance between source and destination
406
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
Physical Phenomenon
Cooperative
Communications
Intended Destination
Communication
gap
Cluster 1
(Transmitting)
Cluster 2
(Receiving)
Figure 14.2. Cooperative communication in sensor network.
nodes remains small as the network grows, the throughput per node was derived to
be (1) in reference 41. The mobility model also has a large impact on capacity; an
extended discussion is provided in Section 14.3.
Cooperative communication, while in itself insufficient for scalability, can lead
to significant improvements in the capacity of individual links and thus potentially
the reliability of finding a path through the network at sufficient rate to satisfy QoS
requirements. For example, in Figure 14.2, nodes are clustered in the network such that
there is a communications gap. The transmitting cluster senses the phenomenon and
the measured data needs to be transmitted to the destination. One likely scenario for
cooperation is to have two independent transmitters and two independent receivers.
For cooperation to be of much benefit, the distance between the two clusters should
be much larger than the distances within a cluster. The information theoretic aspects
of such a channel models are presented in references 42–46 with explicit derivation of
the data rates for each transmitting node and consideration of the effect of transmitter
and/or receiver cooperation on the rate region. The system consisting of three nodes
was considered in references 42, 45–47. The achievable rates for the channel model
with two cooperative transmitters and a receiver is derived in references 45 and 47.
The four-node scenario with two nodes acting purely as relays is considered in
reference 48. In references 49 and 50, channel with two cooperating transmitters
and noncooperating receivers is considered. Their derivation was based on outage
and diversity. The behavior for a fading channel is considered in reference 50 and
that of a nonfading channel but with a complicated transmitter cooperation scheme
involving dirty paper coding is presented in reference 49. In more recent work,
two cooperating receivers along with the two cooperating transmitters are considered in reference 31. However, the model did not consider the transmission of information from a transmitter to both the receivers. In contrast to this, the system
with two cooperating transmitters and two cooperating receivers is considered in
references 43 and 44, where the data stream from each transmitter is intended for
both the receivers. Apart from overcoming the gap in the sensor networks, the sensors
cooperate with each other to achieve more reliable and higher rate communications.
Power rather than bandwidth is the main constraint. Also, multiple sensors occupy
FIDELITY AND RESOURCE TRADEOFFS IN FLAT NETWORKS
407
the same channel; hence, standard multiplexing techniques such as TDMA, FDMA,
CDMA, and OFDM may not be readily employed. For many cooperative schemes, it
is assumed that a high level of synchronism can be maintained; if this cannot be done
(see Section 14.4), then considerably lower gains result from noncoherent combining.
Network Channel Coding. In the previous set of scenarios, it was assumed that
independent traffic streams would interfere, limiting capacity. In network channel
coding, the capability of combining information streams as information is redundantly
spread over the network is exploited [51]. There are two ways in which network coding
can improve network capacity [52]: (1) When there is one source to be multicast,
store and forward may fail to optimize bandwidth, and (2) when there are two or
more independent sources to be transmitted in a network, even for a unicast situation,
store and forward relaying may fail to optimize bandwidth. Considerable progress has
been made on practical schemes for the first situation [53], but less for the second.
In some examples, network coding makes the difference between network scalability
or not [54]. However, in other situations [55] it can be shown that classical routing
achieves the same performance as network coding.
Network Source Coding. We have previously addressed network source coding
for the classic instances in which data are to be collected in a single communication
hop, and now we focus on when it is aggregated across a multihop network. In data
fusion problems, information from multiple sensors, possibly of different types, is
combined to produce a decision variable [56, 57]. In classic data fusion problems,
reduction of the transmission rate between sensors or fusion points is not considered,
but clearly this is an important consideration in sensor networks [1, 58]. Remarkably,
while obviously suboptimal from the point of view of rate reduction, it may be shown
that the decision of a sensor to participate or not in contributing to fusion is sufficient
to enable network scalability whenever imperfect fidelity of source reconstruction
is permitted [59]. That is, while total network communications capacity increases
with the number of nodes, under a fidelity constraint the amount of information to
extract from the network eventually saturates. We will explore the consequences of
this observation more fully in Section 14.2.
Better energy efficiency can be achieved by jointly considering coding and routing
[60]. Here there are two basic choices to make, given that the optimization of network
resources (e.g., communications energy) is NP hard. One can perform a relatively
simple form of incremental compression (that is, with all data present at the node where
fusion occurs), but then face a difficult routing optimization problem. Alternatively,
one can perform (difficult) joint source coding, but can then route data using simple
methods such as shortest path tree (SPT). For lossless compression of discrete sources,
it has been shown that Slepian–Wolf coding followed by SPT is optimal. Either
approach can work in small neighborhoods, but becomes impractical over larger
regions. While it has been shown that at high resolution Wyner–Ziv coding resembles
Slepian–Wolf coding [61], there are still many open issues in rate-distortion coding
in spite of progress on practical coding methods (e.g., 62–66). Reference 67 presents
408
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
a two-step DPCM encoding process, where first local measurements are encoded and
then encoded outputs from other such encoders are combined. While producing higher
mean squared error than joint encoding, the problem is more stable. In the combined
problem of distributed coding with routing, the above scheme might be used in doing
incremental compression. For such problems, it may be shown that routing using trees
is suboptimal; and while the combined routing and fusion problem is NP hard (since
it embeds a Steiner tree problem), if source correlations are highly localized, then
simple heuristics can be quite effective [68, 69], considering both tree and non-tree
constructions. However, much remains to be done to quantify the precise conditions
under which such simplifications are effective and how much performance is lost.
Also of continuing interest is the problem of efficiently searching a sample volume,
whether for reliably detecting intrusions in a surveillance application or characterizing
some physical phenomenon. This may be done, for example, to learn a model as
a prelude to source coding (e.g., correlations to enable Slepian–Wolf or Wyner–
Ziv coding). The search may be conducted in many fashions: through activation of
dense sets of static sensor nodes, by multiscale multisensor search, or through the
use of mobile elements [70, 71]. Associated with each search procedure will be a
source coding/information routing problem. Another important consideration is the
construction of the physical model. Radically different conclusions on the efficacy
of sampling methods can result from differing assumptions on the underlying model.
For example, a smooth surface is far easier to deal with than one with discontinuities,
and assuming that a phenomenon is caused by some finite number of sources results in
different asymptotics than assuming it is due to high-order differential equations. As
cluttered and inhomogeneous environments are the norm rather than the exception,
devising information theoretic results for such domains is an interesting challenge.
14.2.2 Fidelity and Scalability
While ad hoc networks are generally not scalable, with particular conditions on
the source-destination (S-D) distribution, scalability of sensor networks is possible
[72]. The S-D distribution can be made sufficiently localized to enable scalability
through either (a) decision-making of the type to be outlined below or (b) inclusion of
additional communication layers, which make the S-D distribution appear local in
the lower layers. These higher layers may be either traditional telecommunications
infrastructure or mobile nodes, which can play the role of infrastructure with the cost
of additional delay.
We now outline how local decision-making is sufficient for scalability. The sensor
network problem is to extract information concerning some physical phenomenon
(point source or field) to within some fidelity, given nodes with some constraint
on resources (e.g., bandwidth). As the density of nodes increases, the possibilities
for spatial re-use of frequencies improve and so the information volume that can
be carried increases. However, the information that must be conveyed will saturate
according to the fidelity threshold, if only nodes have a mechanism for determining
which ones will be involved in some form of local fusion and which ones will report
nothing. For example, in an identification problem, as the density of nodes increases,
FIDELITY AND RESOURCE TRADEOFFS IN FLAT NETWORKS
409
eventually there will be a node so close to the source that it alone will be able to
make the decision. All that is required then is the capability for the higher SNR
nodes to suppress activity by nodes that detect the phenomenon at lower fidelity but
at SNRs at levels where they might be inclined to engage in cooperative detection.
A simple relay strategy will then suffice to get the information out of the network.
Notice that the keys to the scalability are: separation of function between source
coding and long-range communication (allowing different densities for the logical
functions of communications and detection), local decision-making, and a fidelity
criterion. Without the last two, the amount of information to convey would outstrip
communications capability for certain source models.
Now clearly neither the communications relay strategy nor the local decision rules
are optimal in information theoretic senses; they are merely sufficient to ensure scalability. However, this result is highly suggestive of what an optimal strategy might
look like under a fidelity (quality of service) constraint. Once a sufficient number of
nodes are identified that (with appropriate network source coding) achieve mutual
information between observations and source phenomenon above some threshold,
then no more nodes need be involved. Long-range transport of data need not further
consider interactions of source and channel coding. If the network is to be scalable,
local communications must dominate, with longer-range communications becoming
much less important (e.g., not requiring heroic cooperation strategies). Cooperation
in relaying the data may be more of a secondary matter of achieving reliable communications over local gaps in the network. Given the local interactions dominate, it is
in this domain that combined source coding and routing/channel coding are of most
interest.
Put another way, while there is no general source-channel separation theorem,
when source models, deployment scenarios, and QoS (fidelity) objectives are specified, the above sensor network scalability results imply that there can be separation
on particular scales. For example, if the objective is to identify and track a number of
objects that enter a study region, and the signals decay strongly with distance, then
only the nodes that are in close proximity to each object will have sufficient mutual
information to be included in the decision-making. While some performance loss
may result in separating source and channel coding to the point of data fusion among
this group, beyond this the problem reduces to a pure network coding (or routing)
problem, with no further consideration of source coding. Furthermore, in the limit
of very strong decay of signals with distance, the relative SNRs of the best and next
best sensor will in likelihood be quite large so that even optimal fusion buys little
performance gain, allowing separation of source and channel coding even at the local
level. This suggests a two-scale design approach: For local regions, a solution specific to the underlying physical model is devised, which may involve some combined
source coding/routing/channel coding scheme, while for long-range transport only
communications considerations apply.
The corresponding optimization problem can be conceived of as a two-step process.
The first step is the selection of a cost function that will control the set of nodes that
will be locally cooperating, given a cooperation protocol. One may then select the cooperation protocol from a set of candidates. For a significant set of practical scenarios
410
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
the number of nodes involved and their time of involvement are the most important
resource questions. For example, in short-range radio communications, power is consumed simply from the fact of a radio being on, with transmission power being a
secondary issue. The power and bandwidth requirements linearly depend upon their
duty cycle. Similarly, the time spent processing is related to the amount of data being
collected, again a function of the duty cycle of involvement. The amount of processing (or communications) required is often an exponential function of the number of
nodes involved, again suggesting that minimization of the number of nodes involved
(while meeting the fidelity constraint) is a primary consideration.
This type of cost function also leads to a reasonable set of optimization problems,
as follows. The scaling laws for the number of nodes to be involved in fusing information about some source event to the desired fidelity can be relatively easily obtained in
many circumstances. In particular, this is true for both maximal ratio (MR) combining
and beamforming problems. For example, if the fidelity requirement is that the fused
sensor data must achieve an SNR greater than some threshold in detecting some point
source, and if AWGN is assumed and there is a second power source propagation loss,
then assuming that coherent combining is possible every time the radius of the cooperation region is doubled, the expected SNR improves 6 dB (i.e., it scales with area,
assuming a uniform distribution of nodes) [1, Chapter 5]. The relationship between
coverage and density may similarly be easily obtained for this set of assumptions.
For other models and cooperation protocols, the details of the tradeoff will vary, but
the above serves as a good prototypical scenario for assessing the suitability of particular cost functions and also conveniently serves as a lower bound on achievable
performance. Assuming that there is some maximum cost in terms of the number
of nodes involved in the fusion activity, one would obtain a family of curves of the
type illustrated in Figure 14.3 for different fidelity requirements. Below density ρl
we hit the ceiling nmax on the maximum number n of nodes permitted, resulting in an
inability to meet the fidelity requirements, while above density ρh only the minimum
number of nodes nmin to carry out the task are needed (e.g., one in detection, four for
3-D localization, etc.).
The cost function must include the resource cost for the network as a whole
in having a higher density (that is, the total number of nodes in the network, N),
which reflects the equipment and maintenance costs. This will drive the solution
Figure 14.3. Number of cooperating nodes versus node density.
FIDELITY IN HETEROGENEOUS NETWORKS
411
toward lower densities, while the operational costs for each detection event will drive
it toward higher densities. Thus, for a given cooperation protocol source model,
fidelity requirement there will be an optimal deployment density for at least some
cost functions in this family. A broad set of such problems remains to be more fully
explored.
14.3 FIDELITY IN HETEROGENEOUS NETWORKS
The basic set of assumptions regarding network configuration in the previous section
might be termed “sensor network classic:” a dense set of nodes of the same type
(sensor, radios, processors) usually with highly limited resources (especially energy)
but which are inexpensive and can be deployed in large numbers to form a multihop
network. Apart from the challenge of data extraction at minimal resource cost, other
difficult problems for networks of this type that have attracted the attention of the
research community will be discussed in Section 14.4. For the remainder of this
chapter we will also consider heterogeneous networks, including the possibility of
mobile elements and hierarchical networks that have nodes with greater capabilities
(e.g., sensing or radio range, more storage, position-location ability, reliable energy
supply, etc.). In this section we explore the implications for the fidelity problem, first
with mobility and then with hierarchical networks.
14.3.1 Mobility
In MANETs, mobility is one of the major contributors to overhead from the protocol
stack, thereby reducing communication efficiency [73–77]. More recently, however,
mobility has been found to increase the capacity of wireless ad hoc networks [78–81].
Here, the networks have a mixture of static and mobile nodes. The type of mobility
that mobile agents can possess falls into one of the following three categories [43]:
r Random Mobility. The nodes are assumed to move in an arbitrarily random pattern typically modeled as uniform Brownian motion for analytical convenience
[78–81]. This model has also been used in data mule work [82, 83].
r Predictable Mobility. This model assumes that the pattern of mobility of the
mobile nodes is known, and this knowledge can be exploited to route data [84–
86]. The mobile agents are not moving for the purpose of data transfer and hence
their paths may not coincide with the routing requirements.
r Controlled Mobility. Here the mobility pattern of the mobiles is completely
under the control of the network. Prototypes of such networks have been produced [87–90]. Controlled mobility could also be implemented by providing
infrastructure to move the nodes. This infrastructure can then serve additional
purposes such as logistical support to the network [158].
There are fundamental differences in the throughput and delay properties when the
mobility is controlled as opposed to when the mobility is random or predictable, as
412
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
we now detail. In reference 81, it is assumed that all nodes are randomly mobile
within a unit disc area. The data traffic pattern is assumed to be random as in
reference 40. Data travels over only two wireless hops, from the source node to a
mobile node that acts as a relay and then from the relay to the destination. With this
model the throughput was found to be (1). A similar result was obtained for mobility constrained to one dimension in [79]. The delay for the above scenarios was
found in reference 80, and algorithms were discussed for trading delay and capacity.
Delay D(n) is related to throughput T(n) by D(n) = (nT (n)) for the wireless network scenario of [29]. For the model in reference 81 when
√ the nodes are randomly
mobile with average velocity v(n), the delay scales as ( n/v(n). As noted in reference 80, three important features that influence the throughput and delay in ad hoc
networks are the number of hops, the transmission range, and the node mobility and
velocity. In reference 80, schemes are proposed that exploit these three features to
different degrees to obtain different points on the throughput-delay curve in an optimal way, enabling achievement of the results in references 29 and 81 as limiting cases
of a larger achievable region. Another scenario for a network with mobile nodes was
considered in reference 78. The network consisted of n static nodes, which acted as
sources and destinations for data. However, the network also had m randomly mobile
nodes which were used as relays. For this model, using the routing scheme proposed
in reference 78, the throughout is (m/n log3 n) with an average delay of 2d/v where
v is the velocity of the mobile nodes.
Now consider the network scenario as in reference 81 where all nodes are mobile
except that their motion is controlled rather than random. Here a naı̈ve communication strategy is for each source to move to its destination and communicate at
almost zero range. Hence interference among simultaneous transmissions is zero,
and each sender–receiver pair can utilize the full available bandwidth W. The pernode throughput is W with constant delay. The delay depends on the traveling time
of mobile nodes to reach their destinations, which is constant as network area is constant. This compares to the worst-case delay in reference 81, which is infinite. Clearly,
controlled mobility has the potential to yield fundamentally different throughput and
delay limits compared to those achieved with random mobility in references 78 and
81. Further details, including mobile routing protocols, are presented in reference 43.
We note that, of course, hybrid routing strategies are possible, with some combination
of multihopping in the static network within close neighborhoods of mobile nodes,
in which case the mobile nodes serve the role of a higher level of communication
infrastructure.
In a network where most nodes are static, a reasonable way to view the mobile nodes
is as a form of infrastructure for the support of the static elements. One consequence
of providing long-range data transport is that the mobile nodes can help save energy
in the static nodes, since these must no longer relay data from other nodes over long
multihop wireless routes. The extra energy overhead of mobility may not be a major
concern because the mobile nodes can periodically recharge themselves.
The reliability of transmission across a network also depends strongly on the
mobility assumptions. One of the common problems arising in ad hoc networks
FIDELITY IN HETEROGENEOUS NETWORKS
413
is that of having disjoint networks. Mobile agents can be used to connect such
sparse and disjoint networks. Also, by using mobile agents with longer-range
radios, the number of wireless hops traveled by a data packet is reduced. This,
in turn, reduces the possibility of packet error, thereby helping enhance goodput
performance, and it reduces delays due to retransmission. The reliability analysis
in reference 37 considered a number of possible metrics for network dependability, and it posed a number of graph theoretic optimizations to model the behavior
of heterogeneous mobile ad hoc networks. Given that the optimizations are NPhard, simulations were used to compare protocols for a number of different network
configurations.
Mobile components can provide additional benefits. It was shown in reference 91
that the time synchronization error increases with an increasing number of hops between two nodes. Using the mobiles for time synchronization reduces the hop distance
between nodes, and hence much finer time synchronization is possible than in a multihop case. Controlled mobility also helps improve the performance of localization
systems [92]. Mobile components can support other system activities such as delivering required resources [93, 94]. Further details on some of these uses of mobility
are provided in Section 14.4.
Other Forms of Hierarchy. Much of the literature on sensor networking is devoted
to optimizations that consider variously computation, storage, energy, and communications bandwidth to be the scarce resources. Yet often logistical tasks are also
limiting, including the planning for deployment, maintenance in the field, and time
spent securing the network. This is part of the reason why, in practice, most telecommunication networks are hierarchical. Specialization of components allows limiting
the demands placed on different levels of the hierarchy, simplifying design of each
task and permitting different components/algorithms to be employed. A particular
advantage is that it is easier to compose, debug, and re-use software in such settings,
in that the design choice has been made to provide the higher levels of the network
with far greater resources. When considering the large volumes of code required to
get large networks to do anything useful, along with the frequency of updates required
as application demands change, the challenge of having resource-constrained nodes
perform such tasks is obvious. With hierarchical networks, one can have the advantage
of re-using code and the potentially abundant bandwidth for the upper levels (e.g., not
having to reinvent the capabilities of the TCP/IP stack or SQL databases), together
with the capabilities of covering local regions with dense collections of nodes with
deployments matched to the physical phenomenon of interest. To date, the largest
deployments of sensor networks have also been hierarchical.
A basic principle of design in such situations is to take maximum advantage of
the additional resources provided by the higher network layers. This can include some
combination of low-latency long-range high-bandwidth data transport, reliable time
and position information, high-capacity storage, longer-range and/or well-calibrated
sensors, and so on. For example, in the Tenet protocol suite [95] the asymmetry of
414
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
resources is exploited to enable simplified retasking and debugging of the network,
thereby off-loading (for example) computation of new routes to the master nodes in
the network. Clearly also with such a topology the data in the lower level need go only
a few hops to reach a base station, thereby mitigating the reliability and scalability
problems.
14.4 DATA INTEGRITY IN SENSOR NETWORKS
In this section we consider the question of how to trust the data that the network
produces, given the problem of component drift and miscalibration and also malicious attack upon the network. These are referred to collectively as data integrity
and are clearly prime QoS concerns. Data integrity must be maintained at multiple
levels [96]. Errors can be introduced at the level of the physical events, the sensors, and middleware services (e.g., location, synchronization, data aggregation) in
addition to the usual network consideration of communications losses. Confusion
can arise in the events being monitored either due to malicious activity or the occurrence of some confounding set of events that are beyond the model assumed.
Sensors can be faulty or can be captured. Services may yield less accuracy than
promised (or needed) due to simplistic modeling or failure/capture of nodes. The
solutions to these problems may be categorized according to how trust is embedded
in the network. One may place trust in some physical model (for example, resulting
from a sequence of experiments with carefully calibrated instruments), some set of
trusted nodes (e.g., a base station or cluster heads), or the relative reliability of redundant deployments of nodes compared to sparse ones. Usually some combination
of these trusted elements is required to provide a robust solution, although different
elements are emphasized. For example, while intuitively one would expect that deployments of large numbers of nodes should enable greater collective reliability than
the reliability of the individual nodes, this rests on particular assumptions regarding
both the physical process under study and the corruptive processes. In particular,
it is generally assumed that the process is oversampled, so that correlated readings
are available to provide consistency checks, and that the corruptive processes are
independent (in location and/or time). The quality of the data integrity assurances
depends on how reliable these models are, notwithstanding the analysis of protocols designed assuming particular models. In the following we consider a number
of specific problems in ensuring data integrity in sensor networks: calibration, location, synchronization, secure key distribution, reputation systems, and reliable query
systems.
14.4.1 Calibration of Sensors
Calibration of sensors can in certain circumstances be quite easy. If the resolution
requirements are loose and the sensor is stable for the environmental conditions under
which it is deployed, a one-time calibration (comparison to a trusted standard) is
sufficient. Temperature sensors deployed indoors fall into this category. However,
DATA INTEGRITY IN SENSOR NETWORKS
415
usually the situation is more difficult, in that some of the following uncertainties can
apply:
r The coupling to the medium may be unknown, causing the laboratory calibration
to be largely irrelevant; and the medium may be dynamic, making a one-time
field calibration ineffectual.
r There may be confounding physical effects (e.g., sunlight heating up the package).
r The sensor may be subject to biofouling (a particular problem for biochemical
sensors), or otherwise become degraded by the environment.
r The electronic components can be subject to drift (e.g., droop in battery supply
voltage).
r There can be hard faults (e.g., stuck-at values).
These uncertainties are often poorly modeled. While manufacturers will often
design sensors to deal with some of the largest confounding effects, at other times the
calibration curves provided are fictional with respect to field deployments [97, 98].
There are two basic ways to deal with these problems: redundancy of deployment
(including multiplicity of sensing modes) and an explicit maintenance/calibration
schedule. We deal with them in turn.
First consider the limit of a very large numbers of sensors, such that the phenomenon of interest is greatly oversampled with respect to the spatial and temporal
fidelity requirements. Suppose the sensors are all initially perfectly calibrated and in
working order. If sensor failures are uniformly distributed over some spatiotemporal
window, then consensus-based mechanisms can be used to identify outliers [99] and,
over time, exclude those nodes that are farthest from the group consensus of the remaining set. Supposing for now that these mechanisms are perfect, absent sensor drift
the performance of the remaining ensemble will meet fidelity requirements until such
time as their density is above the target, and the nodes will be able to reliably report
their performance with respect to this target to enable decisions such as adding nodes
to the network. Hence, the essential requirements are (a) ability to recognize the quality of individual readings with respect to the ensemble and (b) ability to recognize the
quality of the decision of the ensemble. Now suppose that all the measurements are
subject to independent drifts. With redundant deployment of sensors, either the law
of large numbers or the Cramer–Rao bound can be invoked to argue that the consensus measurement will be more reliable than that of individual sensors. Unfortunately,
over time, this consensus will deteriorate (e.g., as a result of the random walk of the
combined noise and drift); thus, to have the same network lifetime as when there are
random faults alone, an even greater level of redundancy is needed. Note that here
as well what is fundamentally required to give reliability guarantees is some means
for measuring the quality of the decision of the ensemble. Examples of such measures include (a) the variance of the measurements used to compute the ensemble and
(b) its history over a sequence of measurements.
416
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
In the limit that the number of sensors goes to infinity, relatively simple mechanisms
can be used to provide reliable decisions. Suppose that the objective is to determine
the location of some source using coherent acoustic beamforming, and also suppose
that the nodes have been synchronized using radios. The nodes can learn their own
positions if sources at known locations and times emit sounds; this may be done
simultaneously or sequentially. Then for sources at unknown locations, a random
selection of a small number of nodes close to the source can compute the apparent
position. The results for many such groups of nodes can be averaged, with groups that
produce results far from the consensus excluded. Two basic approaches are to employ
thresholds based upon (a) absolute distance of measurement values or (b) distance
normalized by the average spread of the remaining members of the (local) ensemble.
The choice of decision metric depends on the objectives, the prior knowledge of
the source phenomenon and sensor behavior (e.g., models), and the dynamics of the
physical situation (e.g., deterministic or rapidly and randomly changing).
With finite numbers of nodes, matters become more complicated in that it is more
difficult to construct reliable reference ensembles; errors and drift in a small number
of sensors can significantly alter results. With only two sensors, for example, it is
impossible to tell which is more reliable, absent some side information such as a
reliable model of the source phenomenon. When there are more sensors, clearly there
will be greater certainty in the consensus, but statistical tests designed to gauge the
reliability of the consensus or to make exclusion decisions will themselves be subject
to considerable uncertainty. With modest redundancy, one must therefore fall back
upon some source of trust beyond the primary sensors, such as the physical model,
reference events, complementary sensors, or auditing by trusted nodes. Many possible
combinations of these can be explored at varying levels of redundancy of deployment
to produce effective protocols.
For example, trusted mobile elements may be employed to periodically calibrate
nodes (e.g., a person or robot carrying trusted equipment). A variety of strategies are
possible:
r Generation of standardized physical events at known locations and times (e.g.,
beacon signals)
r Measurement of physical phenomena in close proximity to node(s) being calibrated
r Replacement of fielded nodes with freshly calibrated ones
r Hybrid combinations of the above
The mobile patrols can be according to some fixed schedule, responsive to reported node failures or other indications from the network that performance has become unacceptable, or adaptive, using actual drift/faults to adjust the schedule of
visits. That is, the patrol/audit period depends on the perceived drift and fault models.
Additionally, the density at which standardized events are generated or the proximity within which calibration is required depends on model of the source phenomenon. If the events to be detected are rare, the second strategy might not be
DATA INTEGRITY IN SENSOR NETWORKS
417
sufficient, and so the more aggressive tactic of generating events on a schedule may be
needed.
In all these cases the conclusions made based on the data between calibration events
are tentative, and thus use of the data might need to wait until the next calibration
event. If the calibration takes place frequently enough for data to be usually within the
desired fidelity bounds, then fewer such post-facto adjustments are needed. Having
operated for some time in the field, one may also build up a better model of sensor
drift and use this to predict offsets so that the objective of calibration is then to
correct the model prediction error. With a good model, this may be considerably
less frequently than otherwise. Similarly, one may employ redundant deployment to
reduce the frequency of calibration events.
14.4.2 Reliable Sensor Location
Sensor self-localization or localization of sources may be attempted using a wide
variety of algorithms. Most simply, sensor nodes can be told their locations, obtained
using standard survey techniques, which themselves rely upon knowing some combination of orientation and the positions of landmarks or beacons such as in GPS
[157]. With dense deployments of nodes, physical phenomena may be located by the
expedient of identifying which set of nodes has signal strength above some threshold
and then selecting either (a) the node with the strongest signal as the location or (b)
some weighted combination of the nodes above the threshold. These techniques have
low spatial fidelity, being on the order of the node separation. While it may be argued
that if the propagation environment is known, then a weighted combination of signal
strengths should be highly accurate, in practice most environments in which we would
employ sensor networks do not have such simple propagation laws, or else we would
simply have deployed a very small number of trusted long-range sensors. A better
estimate of source position may be obtained by coherent beamforming [100, 101], in
which either (a) several clusters of nodes form bearing angles toward the source and
then a least squares estimate is made of the intersection point or (b) a single cluster
directly estimates position, also usually with some type of least squares estimate. The
latter technique is accurate only when the source lies within the convex hull of the
nodes performing the sensing, while the former can be successful at considerable
range. Since a wavefront will typically only be coherent within some local region,
this implies that multiple clusters are usually needed for distant sources. Note that the
price paid for the superior capabilities of coherent beamforming is that the nodes must
be tightly synchronized, with absolute error among the clocks in the cluster being a
small fraction of the period of oscillation of the waves in question.
Large collections of nodes can locate themselves with respect to a relative coordinate system without outside assistance, but require some landmarks or known
positions to establish absolute location with respect to absolute coordinates. A noncoherent approach to (approximate) localization of nodes begins with certain nodes
with known positions recording the number of communication hops to other nodes
in the network. If these are at the periphery of the network, then nodes can estimate
position with respect to these fixed nodes, assuming that the line of propagation to
418
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
each is straight and that the propagation losses are constant. A least squares estimate
can be used to deal with inconsistencies of the estimates, and refinements can be made
based upon actual propagation losses between links [102, 103]. However, in reality
a relatively high density of landmarks will typically be required due to the typically
large variations in near-ground propagation.
If nodes are synchronized, then time distance of arrival techniques can be used
to systematically determine the actual distances of nodes from each other, and then
geometric relations can be used to snap together the links in a consistent fashion,
again using least squares techniques to resolve the imperfections of the distance
estimates. Such multilateration techniques can construct a relative coordinate system
or, given some landmarks, an absolute coordinate system [104]. Note, however, that
the greater the number of links that are built outward from known coordinates, the
greater the chance that errors will accumulate in the position estimates. The errors
can be controlled by sprinkling nodes with known (surveyed) positions within the
network volume.
The above techniques all assume that nodes are working and are reliably reporting
information. Implicitly, errors are small and random and the propagation conditions
are reasonably uniform, at least within neighborhoods. Algorithms become complicated in practice because these conditions are seldom met: Some nodes are faulty,
obstructions produce multipath and complicate range estimates, and the system may
even be subject to malicious attack. The first level of defense, as in calibration, is to
deploy a redundant set of nodes and then perform statistical tests that seek to identify
and discard outliers. The next level, as already indicated, is to explicitly include more
nodes whose position is trusted. In the SeRLoc and HiRLoc protocols [68, 69], multiple secure nodes produce beacons that are used by lower-cost sensors to passively
estimate their positions, eliminating interactions among insecure nodes and thus many
attacks. In the SPINE protocol [105], multiple beacons are also employed, and certain
facts regarding propagation are used to verify distances; in particular, it is shown that
attackers may be able to lie about position to extend distances, but cannot produce
viable attacks that shorten distances.
One might accomplish a similar aim with mobile nodes with accurate location
capability passing through the network in one sweep, in a position calibration exercise.
Mobile elements can also provide timing beacons successively to different parts of
the network to enable accurate distance estimates, while sets of reliable elements
can better estimate the propagation model parameters for use in refining position
estimates. Some of the benefits possible with controlled mobility are described in
reference 92. Turning things around, static elements, once having determined their
positions, can greatly assist mobile elements in estimating their own positions as they
traverse a network.
14.4.3 Reliable Synchronization
As noted in the prior section, synchronization is usually a requirement for reliable
position location, and of course the time that events take place is often of interest on its
own. The level of synchronization required very much depends on the applications. In
DATA INTEGRITY IN SENSOR NETWORKS
419
practical terms, one may distinguish among synchronism mediated by the operating
system of the sensor node as compared to synchronism for physical layer communications. Very different types of algorithms are required; and in practice, multiple
layers of synchronism are often present in sensor networks.
Explicit synchronism algorithms are required because clocks that are initially
aligned will not stay that way. Even if they are physically identical, small temperature differences will be manifested as frequency offsets, and thermal noise will cause
jitter. Radios typically present the most stringent synchronism requirements in sensor
networks since the carrier frequency and phase must usually be aligned, using some
combination of frequency and phase-locked loops [106], wherein the receiver locks
to the reference of the transmitter. Motion of either the communicating parties or
strong reflecting media will also affect timing loops; in general, high channel dynamics are also problematic in that it is difficult to achieve high-precision estimates
before the situation changes. The data bearing signals themselves are used to establish
symbol synchronism, while preambles of packets or frames can be used to establish
frame synchronism. It is this level of alignment that usually forms the basis of more
coarse levels of network synchronism, as may be required for time stamping data,
duty-cycling nodes to conserve energy, or performing coherent combining of low
frequency signals such as acoustic or seismic waves.
Synchronism within the communications range of a single master clock is relatively
straightforward, with the clocks of all the subordinate nodes “slaved” to it, offset by
some delay. The propagation delay is important only for applications such as coherent
combining of radio signals; for other applications it will be a negligible fraction of
the total timing offset produced by lower-cost procedures. Chief among the delays
are (a) the time taken to interrupt a processor to record a timing event and (b) delays
experienced when there is queuing of packets over congested links. These delays will
then accumulate in a multihop network as timing references propagate outward from
the master clock. In the NTP protocol used in wired networks [107], queuing delays are
mitigated by sending large numbers of packets and taking the early arrivals and returns
as being more reliably indicative of the actual delay than other statistical quantities.
The RBS protocol [108] shares this feature and additionally is designed to reduce
software delays that are typical when a processor controls a radio. In small networks
it produces sufficient accuracy to enable acoustic beamforming [109], but in larger
networks it suffers from the problem of accumulation of random error as more links
are traversed. This error can be mitigated to some extent through explicit estimation
of clock skews; but even so, the error variance will generally grow linearly with the
number of hops since the problem is well-modeled as a random walk. Interestingly,
this growth in synchronization error with network size is not fundamental but rather
an algorithmic artifact. It has been shown that there exist distributed algorithms for
which the synchronization error variance is O(1) as network size grows [110]. Quite
a number of protocols have been proposed to establish synchronism in networks with
resource-constrained nodes including TPSN [91], LTS [111], Mini/Tiny Sync [112],
and RATS [113].
If there is GPS at every node, synchronism to a high level is quite easy, but there are
many physical situations in which GPS is either unavailable (indoors, mountainous
420
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
terrain, heavy tree cover, etc.) or simply too expensive in cost or energy consumption.
Nonetheless, if timing beacons are available, they can greatly simplify network synchronism, and to achieve phase alignment of radios there is no practical alternative.
Just as a set of nodes with known locations can anchor network location algorithms,
a network of nodes slaved to a common timing base can similarly stabilize network
synchronism. Thus, even if not all nodes have GPS or lock with some other timing
beacon, the remaining nodes can benefit through limitation of the number of hops
over which clock synchronism is propagated.
Again as in the calibration and position location problems, provision must be made
for misbehaving nodes, with explicit estimation of outliers. This is only possible with
redundant sources of timing information (e.g., communication of time by multiple
neighbors), or the presence of trusted elements such as timing beacons [114]. The
RANSAC [115] and LMS [116] algorithms have been proposed for detection of
outliers, whether due to node failures or malicious attack. One may additionally
exploit certain physical facts concerning propagation time, which can be exploited
to detect inconsistencies in timing for individual links [96]. Secure routing protocols
may then be used as the basis for secure multihop synchronism.
14.4.4 Key Distribution in Sensor Networks
Another component of data integrity is some security in the communications system.
Many solutions that apply to MANETs also apply to sensor networks, with the caveat
that some of the nodes in the network may have low computational and storage
capabilities. Since the nodes are potentially deployed for long periods of time, they
may be captured and subverted. This can result in insertion of bad data, interception
of good data, and failure to forward information. While it is unrealistic to expect that
full recovery from some events is possible, what is desired is that the damage should
be detectable and limited in scope in its effects on the network. Since the managers of
the network may be in a position to add new nodes to the network over time (e.g., to
recover from regions of failures/subversion), it is desirable therefore for the security
protocol to be able to accommodate variable numbers of nodes, added at different
times.
A basic requirement for any security protocol is some mechanism for distribution of
cryptographic keys. An ideal protocol would provide strong encryption, scalability,
low vulnerability to the capture of small numbers of nodes, adaptability to adding
new nodes, low computational cost, and low communications overhead. Obviously,
these are difficult to all satisfy, and so protocols have been proposed using a variety
of approaches. TinySec, among its various security features, provides symmetric
keys and is designed specifically for resource constrained motes [117]. Keys may
be deterministically preassigned, as in the SPINS [118] and LEAP [119] protocols.
This has the virtue of perfect security in that capture of one node does not impact the
security of other links, but relies upon sharing of keys with one secure base station
and requires considerable work prior to deployment. Such strategies are thus suitable
for modest-sized stable deployments, but are not well-suited to large-scale systems or
rapidly changing topologies. Random key distribution schemes [120–122] yield more
DATA INTEGRITY IN SENSOR NETWORKS
421
flexibility but are vulnerable to the capture of a relatively small number of nodes in the
network. Post-deployment derivation of keys as in the LEAP, Secure Pebblenet [123],
and PIKE [124] protocols uses a shared secret in combination with pseudorandom
number generators implemented in software which take as parameters some physical
attribute of the node (e.g., location or ID number) to create the keys. Since nodes
erase the shared secret after creation of their key, the network is not vulnerable to
node capture, but nodes are also unable to generate new keys in the future.
Recently, trusted platform modules (TPM) have appeared, which provide strong
security primitives in a tamper-resistant and compact hardware module. While perhaps
at the moment too expensive for the lowest-cost nodes, the presence of such trusted
devices in a subset of nodes in the network can be used to dramatically improve
performance. In reference 96, such modules are exploited to produce a scalable and
adaptable protocol, with strong security. This is a particular instance of algorithms that
rely upon centralized (or multiple) trusted authorities [125–128]. In general, where
there are sources of trust available, be they gateways or TPMs, it is very much to the
advantage of the network designer to make use of them, so that the many difficulties
inherent in fully distributed key generation are mitigated. In this, the situation is
parallel to that of calibration; the more sources of trust available, the easier it is to
construct reliable procedures.
14.4.5 Reputation Systems
So far we have discussed means for making lower-level services related to data integrity more robust. One may also attack the problem at higher levels such as aggregation of decisions from nodes. Clearly, a system in which data integrity is assured
at multiple levels will in general be more robust than one that relies upon just one
technique. Here as well, protocols may be categorized according to the sources of
trust invoked, namely, trusted authorities versus some combination of redundant deployment and trust in models.
Aggregation of decisions or measurements can be very negatively impacted by
nodes that have failed or been subverted. In reputation systems the idea is to rate
the reliability of current information based upon some combination of past reliability
and plausibility of the current information given the group consensus. Since trusted
authorities may be assigned a higher reputation, such systems can be designed for both
flat and hierarchical networks. Similarly, systems can also be designed to deal with
fusion of information from multiple sensing modes. Information from nodes/sensors
with poor reputations can then either receive a low weight or be ignored.
Reputation in sensor network decisions is different in some respects from other
domains such as e-commerce and pure communication networks [129–135]. To begin
with, for scalability of the protocol there will not necessarily be a single repository of
reputation information, although of course there can be some hierarchy in reputation
storage. Second, unlike MANETs, the relative stability of at least the static portions
of the network enables time to build up information on the performance of near
neighbors, with whom most interactions will likely take place (e.g., for the reasons
articulated in Section 14.2). Third, in ad hoc networks, reputation is often used as a
422
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
means to incentivize rational nodes to pass on data, but faulty nodes will not behave
rationally. Finally, for reputation to be checked with respect to sensor data, some
redundant deployment will be required, making implicit use of at least the correlation
of data from nodes in close proximity. In this the situation is very similar to our
discussion of the benefits of redundancy with respect to calibration, localization, and
synchronization, in that trust is placed in some combination of the physical model
and the opinions of other nodes in order to compute some parameter from the data.
Reputation systems are, however, somewhat more general because the decisions to
be made can come from multiple sensor streams, and reputation scores can to some
extent be abstracted away from the particular physical models.
Part of the commonality with the calibration problem is the need for an outlier
detection protocol. With strong prior information concerning the model of the physical
system, model-based outlier detection can be employed in sensor networks [87, 136,
137]. The reputation rating of data points is made proportional to its conformance to
this model. The model could be deterministic or statistical, or even constructed using
adaptive techniques in a phase where all nodes are trusted (e.g., following calibration).
With only partial knowledge of the physical situation, one may apply simple criteria
such as rejection of values outside particular ranges, or sequences which change too
quickly, or sequences that change too little (indicating stuck-at faults).
Apart from the difficulty of dealing with dynamic environments, model-based
techniques suffer from the difficulty of modeling the fault behavior of nodes. An
alternative is to use consensus-based techniques, which use much weaker assumptions concerning models. In particular, it is assumed that neighboring nodes have
correlated readings and that nonfaulty nodes are the large majority of those still permitted to participate in decisions. To plan a sensor node density sufficient for the
first of these conditions to be met of course requires some knowledge of the physical
model. In operation, readings are compared to the consensus of neighboring nodes in
order to compute a reputation score, rather than being compared to the model predictions. There are two basic consensus criteria: (a) outlier designation based on distance
from the consensus or (b) density-based outlier detection. In distance-based mechanisms, the absolute difference in sensor readings from the (reputation-weighted)
average of the remaining nodes is used as the basis for outlier rejection [138, 139].
To set the outlier detection thresholds, some knowledge of the statistical model of
the system (phenomenon and sensor behavior) is thus required. Density-based outlier rejection [140, 141] operates similarly, except that now the data can be clustered into groups with differing spreads, for which different distance thresholds then
apply.
Any outlier detection scheme [142] can be used to produce the reputation score.
Having accomplished this, there are then several possibilities for how these scores will
be accumulated and used. One can assign scores from any transaction as being binary
(satisfactory/unsatisfactory) or give more levels. A particularly convenient framework
is to accumulate reputation metrics using a beta distribution [96, 143] since only two
parameters are needed to characterize the entire probability distribution. Reputations
can age over time, so that more weight is given to recent behaviors. Formulation of
reputation as being related to the probability of yielding a useful input is a flexible
DATA INTEGRITY IN SENSOR NETWORKS
423
framework that allows for fully trusted nodes to be included (e.g., a trusted central
or mobile authority), and for fusion of information collected by very different types
of sensor. It is also amenable to using results from decision theory [143–144] to deal
with various malicious attacks [96].
14.4.6 Efficient Query Systems
Sensor networks are fundamentally a means to actively answer queries about the
physical world. They differ from traditional databases in some important respects.
For example, streaming data from sensors presents a problem of scaling. Even without energy and communication bandwidth constraints, a network of thousands of
sensors will quickly overrun whatever fixed storage media are supplied (and especially if the bandwidth is large). For scalability it is required that queries that direct
the collection, storage, and transport of data should specify not only geographic
and temporal scope, but also spatial, temporal, and accuracy fidelity requirements.
An expression of priority or interest is also needed because network resources are
finite. On the other hand, problems of concurrency, integrity, security, and efficient
access to perform logical inferences remain, and so it is desirable to re-use as much
of the scaffolding from traditional databases as possible, particularly when multiple
applications are intended to use the same system.
A useful abstraction for such problems is diffusion of interests. A user indicates
interests in data with particular attributes, with multiple such interests interacting
within the network to determine the flow of information. At the bottom (physical)
layer they might include sensor type, spatial resolution (or location), and temporal
resolution (or time). Higher-level queries such as a request for a map of the temperature gradients in some particular region must then be translated into a set of requests
for information from particular nodes, a strategy for communicating the information,
and selection of the means for performing the signal processing, all consonant with
resource constraints. Interests interact in increasing the likelihood of particular information being stored, processed, and communicated, but also in creating possibilities
for congestion and other resource contention. These interests thus compete with each
other and against the inhibition built into the network for usage of its resources. Examples of protocols that fairly directly implement this abstraction to establish routes in a
resource-efficient fashion for both queries and responses are push-directed diffusion
[145] (oriented toward situations where there are many data sources and sinks) and
two-phase pull diffusion [146] (few sinks and many data sources).
In general, which routing mechanism to employ depends on (a) how data are
to be stored and used and (b) constraints such as latency for query dissemination
and network reply. Very different traffic volumes result, depending upon the choices
made. In external storage systems, data are automatically sent to a sink whenever
some standing conditions have been met (generically, an event has occurred); data
transport clearly dominates the traffic in this case. In local storage systems, data are
only sent to a sink in response to a particular request. In this case, there can be many
queries. In rendezvous systems, data regarding events and queries meet at intermediate
points, which may shift in order to balance the load, as for example in data-centric
424
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
storage (DCS) [147]. This allows trading storage requirements against query response
time, as well as trading query against data traffic, at the cost of requiring geographic
information for routing.
Clearly, establishment of routes is but part of the problem. There are possibilities for data fusion and aggregation along the way [148, 149], data resolution may
be reduced if resources are deficient [e.g., 150], and consideration must be given to
the timeliness of response to queries and the reliability of the data storage in the
network (e.g., replication to deal with node failure) [147]. An example of a system that enables network operators to balance some of this concerns is TinyDB,
a SQL-like utility developed for operation on networks of (resource-constrained)
sensor nodes running TinyOS [151]. It has a number of features that allow datadriven applications to be developed and deployed with far less effort than attempting
to use fully custom code, including meta-data management, high-level query support, network topology management (including routing tables), support for multiple
queries with different fidelity criteria, and adaptation to nodes entering/leaving the
network.
Hierarchy can be used to considerable benefit. For example, in the aggregration
system proposed in reference 148, a base station with additional resources is required
for computing the routing trees and aggregation strategy. Nodes with longer-range
communication links and additional storage are natural locations for rendevous points,
allowing for faster response to queries and greater fidelity in the data records. More
generally, with powerful higher-level nodes, standard communication and database
might be re-used, with novel algorithms required only for control of the information
flows to and from the lower-level nodes. On the other hand, rapid network topology
changes will, as for routing in MANETs, result in considerably more overhead if
response times are to be kept reasonable.
14.5 CONCLUSIONS
Sensor networks come in a wide variety of forms, from flat networks of highly
resource-constrained nodes to hierarchical networks with mobile elements and powerful communications, processing, and storage capability. The networks themselves
may be designed for one special purpose, or as a tool to answer a broad variety of
questions about the physical world. Which resource constraints apply and the scope
of the network objectives lead to vastly different optimization problems; and thus in
considering QoS requirements for sensor networks, the assumptions concerning the
technology, physical model, and usage must be made explicit.
In this chapter, we have attempted to show that while classical network information theory problems are at best a partial fit to problems of resource optimization
in sensor networks, a QoS perspective at once makes many of the problems both
more relevant and tractable. Mobility in network elements at the same time expands
the richness of the problem set and gives rise to simple solutions to many otherwise
difficult problems in the domains of data communication and integrity. This is a yet
another manifestation of the truism that for robust/simple design there must be at
EXERCISES
425
least some quantity available in abundance. If everything is constrained, then models
must be highly accurate and the system tends to be both overoptimized and brittle.
With a little headroom (e.g., additional trusted beacons, sources of security, communications paths, computational cycles) the system can be more easily adapted to deal
with uncertainties in models and objectives. Much of the sensor network community
continues to be engaged in the project of exploring how having an abundance of nodes
(redundancy with respect to the physical phenomena) can lead to robust and flexible
design, while there is an emerging trend to look at how hierarchical overlays of more
capable nodes (including mobile ones) can simplify design and expand capabilities.
While considerable progress has been made in both domains, a voluminous design
space remains to be explored.
One such example is presented below. Consider mobile ad hoc networks, in which
nodes must balance a variety of tasks including sensing and communications. Nodes
might thus change location or trajectory for sensing purposes, with constraints on
disruption to the QoS of the network. For example, a sensor node may be required
to have its position change for improved source identification without disrupting
the preexisting communications. Similarly, in an overlay network, a node might
have to move to replace a failed backbone node. This gives rise to following two
problems:
1. Maximization of non-communication application QoS (such as node motion to
facilitate sensing) with communication QoS constraints (packet delay, etc.).
2. Maximizing throughput for the newly formed links with communication QoS
and link capacities constraints at the new position of the node.
One solution to these problems is the formulation of a geometric program [43,
59]. Because the mobile radio channel is fast-changing and the number of user
nodes is large, a fast and robust decision-making algorithm is needed that accommodates a large number of variables for dynamic resource allocation to be feasible [9, 11]. A substantial literature is available on resource allocation problems that
use techniques such as game theory, cross-layer optimization, and so on [152–156].
Many other problems involving a mix of controls, sensing, and networking can be
formulated.
14.6 EXERCISES
1. Mixed Deployments of Fixed and Mobile Nodes: Discuss the design tradeoffs
among deploying fixed and mobile nodes with costs nf and nm respectively,
where the objective is to reproduce a field (such as temperature) to within some
specified spatiotemporal fidelity. How is the choice affected by prior knowledge
of the model parameters of the field? With high initial uncertainty in the model,
why might a sequence of deployments be desirable?
426
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
2. Sensor Network Asymptotics: Various results on the scalability of the network with respect to communications capacity, reconstruction under fidelity
constraints, and synchronization error have been presented. What are the benefits of asymptotic analysis in the design of practical networks, and what are the
limitations?
3. Small-group Cooperation: With small groups of nodes, it is often possible
to formulate QoS optimization problems that while having poor scalability are
nonetheless tractable due to the small numbers. Given that centralized computation is possible in such cases, why are heuristics and distributed algorithms still
often preferred solutions? (Hint: Consider the various sources of uncertainties
in an optimization problem such as localization of a source in an obstructed
environment.)
4. Data Integrity: Suppose a set of static sensor nodes is used to track a slowly
varying phenomenon with known spatial correlation function. The sensors drift
according to a random walk with known rate. A number of higher-cost mobile
nodes are available that are fully reliable, but limited to some maximum velocity.
Discuss the design tradeoffs for creating a hybrid system of static and mobile
nodes to study the phenomenon, where spatiotemporal fidelity requirements
must be met.
5. Sources of Trust: In order to trust an experiment, one must have confidence
in a combination of the apparatus, experimental procedure, and the model of
the physical phenomenon (and failure modes of the apparatus). The experiment
will typically be designed to answer some small set of questions well. Suppose that one has available a trusted experimental procedure for determining
seismic response at one particular spatial scale, but the procedure is expensive,
resulting in low numbers of nodes being feasible. Suppose now that lower-cost
accelerometers are available that provide lower resolution. How might these be
used in combination with the trusted procedure, possibly in an iterated fashion, to extend a physical model, either in spatial density or to cover larger
scales?
6. Design Trends: The great project of the scientific revolution is the extension
of mathematical models to explain physical phenomena. Sensor networks of
different varieties provide new tools for this enterprise. Choose an application
domain. How do current technological trends support the future use of sensor
networks in this domain, and what are the principal design barriers to overcome
in order for there to be widespread adoption of the technology? Do the trends
favor homogeneous or hierarchical networks? What nontechnological barriers
must also be overcome, and how do these affect the choices of technology?
BIBLIOGRAPHY
1. G. J. Pottie and W. J. Kaiser. Principles of Embedded Networked Systems Design,
Cambridge University Press, New York, 2005.
BIBLIOGRAPHY
427
2. K. Bult et al. Wireless integrated microsensors. In Solid-State Sensor and Actuator Workshop, Hilton Head Island, SC, June, 1996, pp. 205–210.
3. B. Warneke, M. Last, B. Liebowitz, and K. Pister. Smart dust: Communicating with a
cubic-millimeter computer. Computer, 34(1):44–51, 2001.
4. G. J. Pottie and L. P. Clare. Wireless integrated network sensors: Towards low-cost and
robust self-organizing security networks. In Proceedings of SPIE, Boston, November 3–5,
1998, Vol. 3577, pp. 86–95.
5. J. Agre, L. P. Clare, and G. J. Pottie. Development platform for self-organizing sensor
networks. In Aerosense’99, Orlando, FL, April 1999.
6. Mica wireless measurement system. Datasheet, Crossbow Technology Inc. http://www.
xbow.com/Products/Product pdf files/Wireless pdf/MICA.pdf.
7. “Stargate: Xscale processor platform,” http://www.xbow.com/Products/Product pdf files/
Wireless pdf/6020-0049-01 C Stargate.pdf. Data Sheet.
8. E. Mingozzi. Scheduling algorithms for guaranteeing QoS in wireless Networks. In Proceedings of MTM Workshop, February 2000.
9. M. Chiang, D. O’Neill, D. Julian, and S. Boyd. Resource allocation for QoS provisioning
in wireless ad hoc networks. IEEE Globecom, 5:2911–2915, 2001.
10. D. Julian, M. Chiang, D. O’Neill, and S. Boyd. QoS and fairness constrained convex
optimization of resource allocation for wireless cellular and ad hoc networks. In IEEE
INFOCOM, June 2002, pp. 477–486.
11. D. Julian, D. O’Neill, and M. Chiang. Robust and QoS constrained optimization of power
control in wireless cellular networks. In IEEE Vehicular Technology Conference, October
2001, pp. 1932–1936.
12. M. Schwartz. Telecommunication Networks: Protocols, Modeling, and Analysis. Addison
Wesley, Reading, MA, 1987.
13. C. E. Shannon. A mathematical theory of communication. Bell System Technical Journal,
27 (July and October): 379–423 and 623–656, 1948.
14. C. E. Shannon. Communication in the presence of noise. Proceedings IRE, 37:10–21,
1949.
15. T. M. Cover and J. A. Thomas. Elements of Information Theory. John Wiley & Sons,
New York, 1991.
16. I. E. Teletar. Capacity of multi-antenna Gaussian channels. European Transactions on
Telecommunications, 10:585–596, 1999.
17. S. V. Hanly and P. A. Whiting. Information-theoretic capacity of multireceiver networks.
Telecommunication Systems, 1:1–42, 1993.
18. A. Vanelli-Coralli, R. Padovani, J. Hou, and J. Smee. Capacity of cell clusters with coordinated processing. In Information Theory and Its Applications Workshop (ITA ’06),
San Diego, February 6–10, 2006.
19. S. Verdu. Multiuser Detection. In Advances in Statistical Signal Processing, 2:369–409,
1993.
20. A. Duel-Hallen. Decorrelating decision-feedback multi-user detector for synchronous
code-division multiple access channel. IEEE Transactions and Communications, 41:285–
290, 1993.
21. U. Madhow and M. L. Honig. MMSE interference suppression for direct-sequence spreadspectrum CDMA, IEEE Transactions on Communications, 42:3178–3188, 1994.
428
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
22. D. Slepian and J. K. Wolf. Noiseless coding of correlated information sources. IEEE
Transactions on Information Theory, 19:471–480, 1973.
23. A. D. Wyner and J. Ziv. The rate-distortion function for source coding with side information at the decoder. IEEE Transaction on Information Theory, 22:1–11, 1976.
24. M. Gastpar. The Wyner–Ziv problem with multiple sources. IEEE Transactions on Information Theory, 50(11):2762–2768, 2004.
25. T. Berger, Z. Zhang, and H. Vishwanathan. The CEO problem. IEEE Transactions on
Information Theory, 42:887–902, 1996.
26. H. Vishwanathan and T. Berger. The quadratic Gaussian CEO problem. IEEE Transactions
on Information Theory, 43:1549–1559, 1997.
27. A. Pandya, A. Kansal, G. J. Pottie, and M. Srivastava. Lossy source coding of multiple
Gaussian sources: m-helper problem. In Proceedings of Information Theory Workshop
2004, San Antonio, TX, October 24–29, 2004.
28. A. Pandya, A. Kansal, G. J. Pottie, and M. B. Srivastava. Fidelity and resource sensitive
data gathering. In 42nd Allerton Conference, Allerton, IL, September 2004.
29. P. Gupta and P. R. Kumar. The capacity of wireless networks. IEEE Transactions on
Information Theory, 46:388–404, 2000.
30. P. Gupta and P. R. Kumar. Towards an information theory of large networks: An achievable
rate region. IEEE Transactions on Information Theory, 49, 1877–1894, 2003.
31. M. Gastpar and M. Vetterli. On the capacity of wireless networks: The relay case. In
Proceedings of the 21st INFOCOM, New York, June 2002, pp. 1577–1586.
32. N. Jindal, U. Mitra, and A. Goldsmith. In Capacity of ad-hoc networks with node cooperation. In Proceedings of IEEE International Symposium Information Theory, Adelaide,
June 2004, p. 271.
33. M. Uppal, Z. Liu, V. Stankovic, A. Host-Madsen, and Z. Xiong. Capacity bounds and code
designs for cooperative diversity. In Information Theory and Its Applications Workshop
(ITA ’06), San Diego, February 6–10, 2006.
34. C. T. K. Ng and A. J. Goldsmith. Capacity and cooperation in wireless networks. In
Information Theory and Its Applications Workshop (ITA ’06), San Diego, February 6–10,
2006.
35. A. Host-Madsen. On the achievable rate for receiver cooperation in ad-hoc networks.
In Proceedings of IEEE International Symposium on Information Theory, Adelaide,
2004, p. 272.
36. A. Host-Madsen and J. Zhang. Capacity bounds and power allocation for wireless relay
channel. IEEE Transactions on Information Theory, 51(6):2020–2040, 2005.
37. M. Ahmed. Decentralized Information Processing in Wireless Peer to Peer Networks.
Ph.D. dissertation, Department of Electrical Engineering, UCLA, 2002.
38. R. Gowaikar, B. Hochwald, and B. Hassibi. Throughput scaling in random wireless
networks. In Information Theory and Its Applications Workshop (ITA ’06), San Diego,
February 6–10, 2006.
39. A. Sriniva and E. Modiano. Finding minimum energy disjoint paths in wireless ad hoc
networks. In ACM Wireless Networks, November 11:201–417, 2005.
40. M. Neely and E. Modiano. Capacity and delay tradeoffs for ad hoc mobile networks. In
IEEE Transactions on Information Theory, 51(6):1917–1937, 2005.
BIBLIOGRAPHY
429
41. J. Li, C. Blake, D. De Couto, H. Lee, and R. Morris. Capacity of ad hoc wireless networks.
In 7th Annual International Conference on Mobile Computing and Networking, ACM
Press, New York, 2001, pp. 61–69.
42. Y. S. Tu, Cooperative communications among wireless sensor networks. Ph.D. Dissertation, Department of Electrical Engineering, UCLA, 2003.
43. A. Pandya. Fundamental limits of scalable sensor networks. Ph.D. Dissertation, Department of Electrical Engineering, UCLA, 2004.
44. A. Pandya and G. Pottie. Bounds on achievable rates for cooperative channel coding. In
38th Annual Asilomar Conference on Signals, Systems, and Computers, Pacific Grove,
CA, November 2004.
45. A. Sendonaris, E. Erkip, and B. Aazhang. User cooperation diversity—Part I: system
description. IEEE Transactions on Communications, 51(11):1927–1938, 2003.
46. A. Sendonaris, E. Erkip, and B. Aazhang. User cooperation diversity—Part II: Implementation aspects and performance analysis. IEEE Transactions on Communications,
51(11):1939–1948, 2003.
47. K. Yazdi, H. Gamal, and P. Schniter. On the design of cooperative transmission schemes.
In Allerton Communications, Computing, and Control, October 2003.
48. U. Mitra and A. Sabharwal. Complexity constrained sensor networks: Achievable rates for
two relay networks and generalizations. In Third International Symposium on Information
Processing in Sensor Networks, ACM Press, New York, 2004, pp. 301–310.
49. A. Host-Madsen. A new achievable rate for cooperative diversity based on generalized
writing on dirty paper. In IEEE International Symposium on Information Theory, June
2003, p. 317.
50. J. Laneman, D. Tse, and G. Wornell. Cooperative diversity in wireless networks: Efficient
protocols and outage behavior. IEEE Transactions on Information Theory, 50(12):3062–
3080, 2004.
51. R. Ahlsede, N. Cai, S.-Y. R. Li, and R. W. Yeung. Network information flow. IEEE
Transactions on Information Theory, 46:1204–1216, 2000.
52. R. W. Yeung and N. Cai. Network coding, algebraic coding, and network error correction. In Information Theory and its Applications Workshop (ITA ’06), San Diego,
February 6–10, 2006.
53. R. Koetter and M. Medard. An algebraic approach to network coding. IEEE/ACM Transactions on Networking, 11:782–795, 2003.
54. A. G. Dimakis, D. Petrovic, and K. Ramchandran. From dumb wireless networks to smart
wireless networks using network coding. In Information Theory and Its Applications
Workshop (ITA ’06), San Diego, February 6–10, 2006.
55. S. Savari and G. Kramer. Capacity bounds for relay networks. In Information Theory and
Its Applications Workshop (ITA ’06), San Diego, February 6–10, 2006.
56. J. W. Fisher, T. Darrell, W. T. Freeman, and P. Viola. Learning joint statistical models for
audio-visual fusion and segregation. In T. K. Leen, T. G. Dietterich, and V. Tresp, editors,
Advances in Neural Information Processing Systems, Vol. 13, MIT Press: Cambridge,
MA, 2001, pp. 772–778.
57. M. Ahmed and G. Pottie. Fusion in the context of information theory. In S. R. Iyengar
and R. R. Brooks, editors, Distributed Sensor Networks, CRC Press, Boca Raton, FL,
2005.
430
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
58. G. J. Pottie and W. J. Kaiser. Wireless integrated network sensors. Communications of
ACM, 43:51–58, 2000.
59. A. Pandya and G. J. Pottie. On scalability and source/channel coding decoupling in large
scale sensor networks. CENS Technical Report 17, 2003.
60. R. Cristescu, B. Beferull-Lozano, and M. Vetterli. On network correlated data gathering.
In IEEE Infocom, 2004, pp. 2571–2582.
61. R. Zamir and T. Berger. Multiterminal source coding with high resolution. IEEE Transactions on Information Theory, 45(1):106–117, 1999.
62. J. Garcia-Frias and W. Zhong. LDPC codes for compression of multi-terminal source
with hidden Markov correlation. IEEE Communication Letters, 7(3):115–117, 2003.
63. S. S. Pradhan, J. Kusuma, and K. Ramchandran. Distributed compression in a dense
microsensor network. IEEE Signal Processing Magazine, 19:51–60, 2002.
64. S. S. Pradhan and K. Ramchandran. Distributed source coding using syndromes (DISCUS): Design and construction. IEEE Transactions on Information Theory, 49:626–643,
2003.
65. Z. Xiong, A. D. Liveris, and S. Cheng. Distributed source coding for sensor networks.
IEEE Signal Processing Magazine, 21:80–94, 2004.
66. R. Zamir, S. Shamai, and U. Erez. Nested linear/lattice codes for structured multiterminal
binning. IEEE Transactions on Information Theory, 48(6):1250–1276, 2002.
67. H. Luo, Y. Tong, and G. Pottie. A two-stage DPCM scheme for wireless sensor networks.
In Proceedings on IEEE Intlernational Conference on Acoustics, Speech and Signal Processing (ICASSP ’05), Philadelphia, Vol. 3, 2005, pp. 661–664.
68. L. Lazos and R. Poovendran. SeRLoc: robust localization for wireless sensor networks.
ACM Transactions on Sensor Networks (TOSN), 1:73–100, 2005.
69. L. Lazos and R. Poovendran. HiRLoc: High resolution localization for wireless sensor
networks. IEEE Journal on Selected Areas in Communications (JSAC), 24(2):233–246,
2006.
70. M. Rahimi, M. Hansen, W. J. Kaiser, G. S. Sukhatme and D. Estrin. Adaptive sampling
for environmental field estimation using robotic sensors. In IEEE/RSJ International Conference on Intelligent Robots and Systems, August 2005, pp. 3692–3698.
71. X. Kong, R. Pon, W. Kaiser, and G. Pottie. Environmental sampling with multiscale
sensing. In ICASSP 2006, Toulouse, Vol. 4, May 2006, pp. 873–876.
72. A. Pandya and G. Pottie. QoS in ad hoc networks. IEEE Vehicular Technology Conference,
October 2003.
73. S. Baatz, M. Frank, R. Gopffarth, D. Kassatkine, P. Martini, M. Scetelig, and A. Vilavaara.
Handoff support for mobility with IP over Bluetooth. In 25th Annual Conference on Local
Computer Networks (LCN00), Tampa, FL, November 2000.
74. A. Campbell, J. Gomez, C. Wan, and S. Kim. Cellular IP. http://comet.ctr.
columbia.edu/cellularip/pub/draft-ietf-mobileipcellularip-00.txt. Internet draft, January
2000.
75. J. Kempf and P. Yegani. Openran: A new architecture for mobile wireless internet radio
access networks. IEEE Communications Magazine, 40:118–123, 2002.
76. C. Lin and M. Gerla. Adaptive clustering for mobile wireless networks. IEEE Journal of
Selected Areas in Communications, 15(7):1265–1275, 1997.
77. C. Perkins. editor, In Ad Hoc Networking, Addison Wesley, Boston, 2001, pp. 1–28.
BIBLIOGRAPHY
431
78. N. Bansal and Z. Liu. Capacity, delay and mobility in wireless ad-hoc networks. In IEEE
INFOCOM 2003, San Francisco, CA, April 1–3, 2003, pp. 1553–1563.
79. S. Diggavi, M. Grossglauser, and D. Tse. Even one-dimensional mobility increases ad
hoc wireless capacity. In Proceedings of ISIT, Lausanne, Switzerland, June 2002, p. 352.
80. A. Gamal, J. Mammen, B. Prabhakar, and D. Shah. Throughput-delay tradeoff in wireless
networks. In IEEE INFOCOM 2004, Hong Kong, Vol. 1, March 7–11, 2004, pp. 464–
475.
81. M. Grossglauer and D. Tse. Mobility increases the capacity of wireless of ad hoc wireless
networks. IEEE/ACM Transactions on Networking, 10, pp. 477–486, 2002.
82. S. Jain, R. Shah, G. Borriello, W. Brunette, and S. Roy. Exploiting mobility for energy
efficient data collection in sensor networks. In Modeling and Optimization in Mobile, Ad
Hoc and Wireless Networks (WiOpt), 2004.
83. R. Shah, S. Roy, S. Jain, and W. Brunette. Datamules: Modelling a three tiered architecture
for sparse sensor networks. In First IEEE International Workshop on Sensor Network
Protocols and Applications (SNPA), May 2003.
84. A. Chakrabarty, A. Sabharwal, and B. Aazhang. Using predictable observer mobility
for power efficient design of a sensor network. In Second International Workshop on
Information Processing in Sensor Networks (IPSN), Palo Alto, CA, April 2003, pp. 129–
145.
85. S. Jain, K. Fall, and R. Patra. Routing in a delay tolerant network. In SIGCOMM, 2004,
pp. 145–157.
86. A. Pentland, R. Fletcher, and A. Hasson. Daknet: Rethinking connectivity in developing
nations. IEEE Transactions on Computer, 37(1):78–83, 2004.
87. F. Koushanfar, M. Potkonjak, and A. Sangiovanni-Vincentelli. On-line fault detection of sensor measurements. In IEEE International Conference on Sensors, October
2003, pp. 974–980.
88. A. Kansal, A. Somasundara, D. Jea, M. Srivastava, and D. Estrin. Intelligent fluid infrastructure for embedded networks. In Second International Conference on Mobile Systems,
Applications, and Services (MobiSys), June 2004.
89. W. Zhao and M. Ammar. Message ferrying: Proactive routing in highly partitioned wireless ad hoc networks. In Ninth IEEE Workshop on Future Trends of Distributed Computing
Systems (FTDCS), May 2003.
90. W. Zhao, M. Ammar, and E. Zegura. A message ferrying approach for data delivery in
sparse mobile ad hoc networks. In Fifth ACM International Symposium on Mobile Ad
Hoc Networking and Computing (MobiHoc), 2004, pp. 187–198.
91. J. Elson, L. Girod, and D. Estrin. Fine-grained network time synchronization using reference broadcasts. SIGOPS Operating System Review, 36(1):147–163, 2002.
92. J. Scott and M. Hazas. User-friendly surveying techniques for location-aware systems. In
Proceedings ubicomp, published in LNCS Vol. 2864, Oct. 2003, pp.45–54.
93. A. LaMarca, W. Brunette, D. Koizumi, M. Lease, S. Sigurdsson, K. Sikorski, D. Fox, and
G. Borriello. Plantcare: An investigation in practical ubiquitous systems. In Ubicomp,
September 2002.
94. M. Rahimi, H. Shah, G. Sukhatme, J. Heidemann, and D. Estrin. Studying the feasibility
of energy harvesting in a mobile sensor network. In IEEE International Conference on
Robotics and Automation, 2003, pp. 19–24.
432
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
95. R. Govindan, E. Kohler, D. Estrin, F. Bian, K. Chintalapudi, O. Gnawali, S. Randwala,
R. Gummadi, and T. Sathopoulos. Tenet: An architecture for tiered embedded networks. Center for Embedded Networked Sensing Technical Report 56, November
2005.
96. Saurabh Ganeriwal. Trustworthy sensor networks. Ph.D. dissertation, Department of Electrical Engineering, UCLA, 2006.
97. P. Buonadonna, D. Gay, J. Hellerstein, W. Hong, and S. Madden. TASK: Sensor network in a box. Intel Research Berkeley Technical Report IRB-TR-04-021, January
2005.
98. Y. H. Hu and B. Benson. Sensor network data quality assurance. Technical Report,
University of Wisconsin, Madison, 2005.
99. V. Barnett and T. Lewis. Outliers in Statistical Data, Wiley, 1994.
100. J. C. Chen, K. Yao, and R. E. Hudson. Acoustic source localization and beamforming: Theory and practice. Eurasian Journal on Applied Signal Processing, 4:359–370,
2003.
101. J. C. Chen, K. Yao, T. L. Tung, C. W. Reed, and D. Chen. Source localization and
tracking of a wideband source using a random distributed beamforming sensors array.
International Journal of High Performance Computing Applications, 16:259–272, 2002.
102. N. Bulusu. Self-Configuring Localization Systems. Ph.D. dissertation. Computer Science
Department, UCLA, 2001.
103. D. Niculescu and B. Nath. Position and orientation in ad hoc networks. Elsevier Journal
of Ad Hoc Networks, 2(2):133–151, 2004.
104. A. Savvides. Design, implementation and analysis of ad-hoc localization methods. Ph.D.
Dissertation, Electrical Engineering Department, UCLA, 2003.
105. S. Capkun and J. P. Hubaux. Secure positioning in wireless networks. IEEE Journal on
Selected Areas in Communications, 24(2):221–232, 2006.
106. J. G. Proakis. Digital Communications, McGraw-Hill, New York, 2001.
107. http://www.ntp.org.
108. J. Elson, Time synchronization in wireless sensor networks, Ph.D. dissertation, Computer
Science Deptartment, UCLA, 2003.
109. J. C. Chen, L. Yip, J. Elson, H. Wang, D. Maniezzo, R. E. Hudson, K. Yao, and D. Estrin.
Coherent acoustic array processing and localization on wireless sensor network. In Proceedings of the IEEE, Vol. 91, no. 8, August 2003, pp. 1154–1162.
110. A. Giridhar and P. R. Kumar. Distributed clock synchronizatin over wirelss networks:
Algorithms and analysis. In Proceedings of the 45th IEEE Conference on Decision and
Control, San Diego, December 13–15, 2006, pp. 4915–4920.
111. J. van Greunen and J. Rabaey. Lightweight time synchronization for sensor networks.
In 2nd ACM International Conference on Wireless Sensor Networks and Applications
(WSNA), 2003, pp. 11–19.
112. M. Sichitiu and C. Veeraraittiphan. Simple, accurate time synchronization for wireless
sensor networks. In IEEE Wireless Communication and Networking Conference (WCNC),
2003, pp. 1266–1273.
113. S. Ganeriwal, D. Ganesan, H. Sim, V. Tsiatsis, and M. B. Srivastava. Estimating clock uncertainty for efficient duty-cycling in sensor networks. In ACM Conference on Networked
Sensor Systems (Sensys), 2003, pp. 138–149.
BIBLIOGRAPHY
433
114. K. Sun, P. Ning, and C. Wang. Secure and resilient clock synchronization in wireless
sensor networks. IEEE Journal on Selected Areas in Communications, 24(2):395–408,
2006.
115. J. P. Rousseeuw and M. A. Leroy, Robust Regression and Outlier Detection, John Wiley
& Sons, New York, 1987.
116. M. Manzo, T. Roosta, and S. Sastry. Time synchronization attacks in sensor networks.
In ACM Workshop on Security of Ad Hoc Sensor Networks (SASN), 2005, pp. 107–
116.
117. C. Karlof, N. Sastry, and D. Wagner. Tinysec: A link layer security architecture for wireless
sensor networks. In ACM Sensys ’04 2004, pp. 162–175.
118. A. Perrig, R. Szewczyk, V. Wen, D. Culler, and J. D. Tygar. SPINS: Security protocols
for sensor networks. In 7th ACM Conference on Mobile Computing and Networking
(Mobicom), 2001, pp. 189–199.
119. S. Zhu, S. Setia, and S. Jajodia. LEAP: Efficient security mechanisms for large-scale
distributed sensor networks. In 10th ACM Conference on Computer and Communication
Security, ACM Press, New York, 2003, pp. 62–72.
120. L. Eschenauer and V. D. Gligor. A key-management schemes for distributed sensor networks. In ACM Conference on Computer and Communications Security, ACM Press,
New York, 2002, pp. 41–47.
121. H. Chan, A. Perrig, and D. Song. Random key predistribution schemes for sensor networks. IEEE Symposium on Research in Security and Privacy, IEEE Computer Society,
May 2003, p. 197.
122. D. Liu and P. Ning. Efficient distribution of key chain commitments for broadcast authentication in distributed sensor networks. In 10th Annual Network and Distributed System
Security Symposium, 2003.
123. S. Basagni, K. Herrin, D. Bruschi, and E. Rosti. Secure pebblenets. In ACM International
Symposium on Mobile Ad Hoc Networking and Computing (MOBIHOC), ACM Press,
New York, 2001, pp. 156–163.
124. R. Anderson, H. Chan, and A. Perrig. Key infection: Smart trust for smart dust. IEEE
Conference on Network Protocols, Berlin, Germany, October 2004.
125. W. Du, J. Deng, Y. S. Han, S. Chen, and P. K. Varshney. A key management scheme for
wireless sensor networks using deployment knowledge. IEEE Conference on Computer
Communications (Infocom) 2004, pp. 586–597.
126. W. Du, J. Deng, Y. S. Han, and P. K. Varshney. A pairwise key predistribution schemes for
wirelss sensor networks. In ACM Conference on Computer and Communications Security,
October 2003, pp. 42–51.
127. J. Hwang and Y. Kim. Revisiting random key pre-distribution for sensor networks. In ACM Workshop on Security of Ad Hoc and Sensor Networks (SASN)
2004, pp. 43–52.
128. D. Liu and P. Ning. Establishing pairwise keys in distributed sensor networks. In ACM
Conference on Computer Communications Security, October 2003, pp. 52–61.
129. M. Blaze, J. Feigenbaum, J. Ioannidis, and A. Keromytics. The keynote trust management
system. In IETF RFC 2704, 1999.
130. M. Blaze, J. Feigenbaum, and J. Lacy. Decentralized trust management. In IEEE Conference on Security and Privacy, 1996, pp. 164–173.
434
QUALITY OF SERVICE IN WIRELESS SENSOR NETWORKS
131. S. Buchegger and J. Y. LeBoudec. Performance analysis of the CONFIDANT protocol. In
ACM International Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc)
2002, pp. 226–236.
132. P. Michiardi and R. Molva. Core: A COllaborative REputation mechanism to enforce
node cooperation in mobile ad hoc networks. In Communication and Multimedia Security,
September 2002.
133. P. Resnick, K. Kuwabara, R. Zeckhauser, and E. Friedman. Reputation systems: facilitating trust in e-commerce systems. Communications of the ACM 43(12):45–48, 2000.
134. P. Resnick and R. Zeckhauser. Trust among strangers in Internet transactions: Empirical analysis of eBay’s reputation system. Working paper for the NBER Workshop on
Empirical Studies of Electronic Commerce, 2000.
135. L. Xiong and L. Liu. A reputation-based trust model for peer-to-peer ecommerce communities. In IEEE Conference on e-commerce, 2003.
136. F. Koushanfar and M. Potkonjak. Markov-cain based models for missing and faulty
data in mica2 sensor networks. In IEEE International Conference on Sensors, October
2005, pp. 576–579.
137. F. Koushanfar, M. Potkonjak and A. Sangiovanni-Vincentelli. Fault tolerance in wireless ad hoc sensor networks. In IEEE International Conference on Sensors, Oct. 2003,
pp. 974–980.
138. E. M. Knorr and R. T. Ng. Algorithms for mining distance-based outliers in large datasets.
In 24th International Conference on Very Large Databases, 1998, pp. 392–403.
139. E. M. Knorr and R. T. Ng. Finding intensional knowledge of distance-based outliers. In
25th International Conference on Very Large Databases, 1999, pp. 211–222.
140. M. M. Breunig, H. P. Kriegel, R. T. Ng, and J. Sander. LOF: Identifying density-based
local outliers. In ACM SIGMOD Conference 2000, pp. 93–104.
141. S. Papdimitriou, H. Kitawaa, P. B. Gibbons, and C. Faloutsos. LOCI: Fast outlier detection
using the local correlation integral. In Proceedings of 19th International Conference on
Data Engineering, March 2003, pp. 315–326.
142. V. Barnett and T. Lewis. Outliers in Statistical Data, John Wiley & Sons, New York,
1994.
143. A. Josang and R. Ismail. The beta reputation system. In 15th Bled Electronic Commerce
Conference, June 2002.
144. G. Shafer. A Mathematical Theory of Evidence, Princeton University Press, Princeton,
NJ, 1976.
145. J. Heidemann, F. Silva, and D. Estrin. Matching data dissemination algorithms to application requirements. In Proceedings of ACM SenSys ’03, Los Angeles, November 5–7,
2003.
146. C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed diffusion: A scalable and robust
communication paradigm for sensor networks. In Proceedings of Mobicom 2000, Boston,
MA, August, 2000.
147. S. Shenker, S. Ratnasamy, B. Karp, R. Govindan, and D. Estrin. Data-centric storage in
sensor networks. SIGCOMM Computer Communication Review, 33(1):137–142, 2003.
148. S. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong. TAG: A tiny aggregation
service for ad-hoc sensor networks. SIGOPS Operating Systems Review, 36(SI):131–146,
2002.
BIBLIOGRAPHY
435
149. S. Madden, R. Szewczyk, M. J. Franklin, and D. Culler. Supporting aggregate queries over
ad-hoc wireless sensor networks. In 4th IEEE Workshop on Mobile Computing Systems
and Applications, 2002, pp. 49–58.
150. D. Ganesan, B. Greenstein, D. Perelyubskiy, D. Estrin, and J. Heidemann. An evaluation
of multi-resolution storage in resource-constrained sensor networks. Proceedings ACM
Sensys’03, Los Angeles, November 5–7, 2003.
151. S. Madden, J. Hellerstein, and W. Hong. TinyDB: In-Network Query Processing in TinyOS.
UC Berkeley EECS Department, 2003. http://telegraph.cs.berkeley.edu/tinydb.
152. Y. Xue, B. Li, and K. Nahrstedt. Optimal resource allocation in wireless ad hoc networks:
A price-based approach. IEEE Transactions on Mobile Computing, 5(4):347–364, 2006.
153. P. Marbach and R. Berry. Downlink resource allocation and pricing for wireless networks.
In INFOCOM, 2002, 1470–1479.
154. S. Cui, R. Madan, A. Goldsmith, and S. Lall. Cross-layer energy minimization in Sensor
Networks, Proceedings: Allerton Conference on Communications, Control, and Computing, Monticello, IL, September 2004, pp. 1891–1900.
155. S. Shakkottai, T. S. Rappaport, and P. C. Karlsson. Cross-layer design for wireless networks. IEEE Communications Magazine, 41:74–80, 2003.
156. V. Kawadia and P. R. Kumar. A cautionary perspective on cross-layer design. IEEE
Transactions on Wireless Communications, 12(1):3–11, 2005.
157. B. Hofmann-Wellenhof, H. Lictenegger, and J. Collins GPS: Theory and Practice,
Springer-Verlag, Berlin, 1997.
158. W. Kaiser, G. Pottie, M. Srivastava, G. Sukhatme, J. Villasenor, and D. Estrin. Networked
infomechanical systems (NIMS) for ambient intelligience. In W. Weber, J. M. Rabaey,
and E. Aarts, editors, In Ambient Intelligence, Springer, New York, 2005.
159. H. Luo and G. J. Pottie. Routing explicit side information ofr data compression in wireless
sensor networks. In DCOSS 2005, Marina del Rey, CA, June 30–July 1, 2005.
160. H. Luo and G. J. Pottie. Balanced aggregation trees for routing correlated data in wireless
sensor networks. In IEEE ISIT 2005, Adelaide Australia, September 4–9 2005, 14–18.
CHAPTER 15
Energy-Efficient Algorithms
in Wireless Sensor Networks
AZZEDINE BOUKERCHE
School of Information Technology and Engineering, University of Ottawa, Ottawa, Ontario K1N
6N5, Canada
SOTIRIS NIKOLETSEAS
Department of Computer Engineering and Informatics, University of Patras, Patras, Greece;
and Computer Technology Institute (CTI), Patras 26500, Greece
15.1 INTRODUCTION
Recent dramatic developments in microelectromechanical (MEMS) systems, wireless
communications, and digital electronics have already led to the development of smallsized, low-power, low-cost sensor devices. Such extremely small devices integrate
sensing, data processing, and wireless communication capabilities. Current devices
have a size at the cubic centimeter scale, a CPU running at 4 MHz, some memory,
and a wireless communication capability at a 4-kbps rate. Also, they are equipped
with a small but effective operating system and are able to switch between “sleeping”
and “awake” modes to save energy. Pioneering groups (like the “Smart Dust” Project
at Berkeley, the “Wireless Integrated Network Sensors” Project at UCLA, and the
“Ultra low Wireless Sensor” Project at MIT) pursue further important goals, like
a total volume of a few cubic millimeters and extremely low energy consumption,
by using alternative technologies, based on radio frequency (RF) or optical (laser)
transmission.
Examining each such device individually might appear to have small utility; however, the effective distributed coordination of large numbers of such devices may lead
to the efficient accomplishment of large sensing tasks. Large numbers of sensor nodes
can be deployed in areas of interest (such as inaccessible terrains or disaster places)
and use self-organization and collaborative methods to form a sensor network.
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
437
438
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
Their wide range of applications is based on the possible use of various sensor
types (i.e., thermal, visual, seismic, acoustic, radar, magnetic, etc.) in order to monitor a wide variety of conditions (e.g., temperature, object presence and movement,
humidity, pressure, noise levels, etc.). Thus, sensor networks can be used for continuous sensing, event detection, location sensing, and as microsensing. Hence, sensor
networks have important applications, including (a) military (like forces and equipment monitoring; battlefield surveillance; targeting; nuclear, biological, and chemical
attack detection), (b) environmental applications (such as fire detection, flood detection, precision agriculture), (c) health applications (like telemonitoring of human
physiological data), and (d) home applications (e.g., smart environments and home
automation). For an excellent survey of wireless sensor networks see, references 1–3.
15.1.1 Critical Challenges
The efficient and robust realization of such large, highly dynamic, complex, nonconventional networking environments is a challenging algorithmic and technological task. Features including the huge number of sensor devices involved, the severe
power, computational, and memory limitations, their dense deployment, and frequent
failures pose new design, analysis, and implementation aspects that are essentially
different with respect to not only distributed computing and systems approaches but
also ad-hoc networking techniques.
We emphasize the following characteristic differences between sensor networks
and ad hoc networks:
r The number of sensor particles in a sensor network is extremely large compared
to that in a typical ad hoc network.
r Sensor networks are typically prone to faults.
r Because of faults as well as energy limitations, sensor nodes may (permanently
or temporarily) join or leave the network. This leads to highly dynamic network
topology changes.
r The density of deployed devices in sensor networks is much higher than in
ad hoc networks.
r The limitations in energy, computational power, and memory are much more
severe in sensor networks.
Because of the above rather unique characteristics of sensor networks, efficient
and robust distributed protocols and algorithms should exhibit the following critical
properties:
Scalability. Distributed protocols for sensor networks should be highly scalable, in
the sense that they should operate efficiently in extremely large networks composed
of huge numbers of nodes. This feature calls for an urgent need to prove by analytical
means and also validate (by large-scale simulations) certain efficiency and robustness
(and their tradeoffs) guarantees for asymptotic network sizes.
INTRODUCTION
439
Efficiency. Because of the severe energy limitations of sensor networks and also
because of their time-critical application scenarios, protocols for sensor networks
should be efficient, with respect to both energy and time.
Fault-Tolerance. Sensor particles are prone to several types of faults and unavailabilities and may become inoperative (permanently or temporarily). Various reasons
for such faults include (a) physical damage during either the deployment or the operation phase and (b) permanent (or temporary) cease of operation in the case of power
exhaustion (or energy saving schemes, respectively). The sensor network should be
able to continue its proper operation for as long as possible despite the fact that certain
nodes in it may fail.
15.1.2 The Energy Efficiency Challenge
Since one of the most severe limitations of sensor devices is their limited energy
supply, one of the most crucial goals in designing efficient protocols for wireless sensor
networks is minimizing the energy consumption in the network. This goal has various
aspects, including (a) minimizing the total energy spent in the network, (b) minimizing
the number (or the range) of data transmissions, (c) combining energy efficiency and
fault-tolerance, by allowing redundant data transmissions which, however, should
be optimized to not spend too much energy, (d) maximizing the number of “alive”
particles over time, thus prolonging the system’s lifetime, and (e) balancing the energy
dissipation among the sensors in the network, in order to avoid the early depletion of
certain sensors and thus the breakdown of the network.
We note that it is very difficult to achieve all the above goals at the same time. There
even exist tradeoffs between some of the goals above. Furthermore, the importance
and priority of each of these goals may depend on the particular application. Thus,
it is important to have a variety of protocols, each of which may possibly focus on
some of the energy efficiency goals above (while still performing well with respect
to the rest of the goals).
In this chapter, we present four energy-efficient protocols:
r The Local Target Protocol (LTP), which performs a local optimization trying to
minimize the number of data transmissions.
r The Probabilistic Forwarding Protocol (PFR), which creates redundant data
transmissions that are probabilistically optimized, into perform a tradeoff between energy efficiency and fault-tolerance.
r The Energy-Balanced Protocol (EBP), which focuses on guaranteeing the same
per sensor energy dissipation, in order to prolong the lifetime of the network.
r The Variable Transmission Range Protocol (VTRP), which dynamically adapts
the transmission range to bypass sensors that tend to be overused, in order to
avoid possible obstacles.
Because of the complex nature of a sensor network (that integrates various aspects of
communication and computing), protocols, algorithmic solutions, and design schemes
440
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
for all layers of the networking infrastructure are needed. Far from being exhaustive,
we mention the need for frequency management solutions at the physical layer, for
Medium Access Control (MAC) protocols to cope with multihop transmissions at
the data link layer. The interested reader may use the excellent survey by Akyildiz
et al. [4] for a detailed discussion of design aspects of all layers of the networking
infrastructure.
We focus in this chapter on algorithms and protocols for the network layer [references 23–32]. We believe that a complementary use of rigorous analysis and largescale simulations is needed to fully investigate the performance of data propagation
protocols in wireless sensor networks. In particular, asymptotic analysis may lead to
provable efficiency and robustness guarantees toward the desired scalability of protocols for sensor networks that have extremely large size. On the other hand, simulation
allows us to investigate the effect of a great number of detailed technical specifications
of real devices, a task that is difficult (if possible at all) for analytic techniques which,
by their nature, use abstraction and model simplicity.
15.2 LTP: A HOP-BY-HOP DATA PROPAGATION PROTOCOL
15.2.1 The Model
The LTP Protocol was introduced by Chatzigiannakis, Nikoletseas, and Spirakis [5].
The authors adopt a two-dimensional (plane) framework: A smart dust cloud (a set
of particles) is spread in an area (for a graphical presentation, see Figure 15.1).
Let d (usually measured in numbers of particles/m2 ) be the density of particles
in the area. Let R be the maximum (radio/laser) transmission range of each grain
particle.
A receiving wall W is defined to be an infinite line in the smart-dust plane. Any
particle transmission within range R from the wall W is received by W. W is assumed
to have very strong computing power, able to collect and analyze received data, and
has a constant power supply and so it has no energy constraints. The wall represents in
fact the authorities (the fixed control center) to whom the realization of a crucial event
should be reported. The wall notion generalizes that of the sink and may correspond
to multiple (and/or moving) sinks. Note that a wall of appropriately big (finite) length
suffices.
Control center
Sensor field
Sensor nodes
Figure 15.1. A smart dust cloud.
LTP: A HOP-BY-HOP DATA PROPAGATION PROTOCOL
441
The notion of multiple sinks that may be static or moving has also been studied
in reference 6, where Triantafilloy et al. introduce “NanoPeer Words,” which are
merging notions from Peer-to-Peer Computing and Smart Dust.
Furthermore, there is a setup phase of the smart dust network, during which the
smart cloud is dropped in the terrain of interest; when using special control messages
(which are very short, inexpensive, and transmitted only once), each smart dust particle
is provided with the direction of W. By assuming that each smart dust particle has
individually a sense of direction and by using these control messages, each particle
is aware of the general location of W.
15.2.2 The Protocol
Let d(pi , pj ) be the distance (along the corresponding vertical lines toward W) of
particles pi , and pj , and let d(pi , W) the (vertical) distance of pi from W. Let info(E)
be the information about the realization of the crucial event E to be propagated. Let
p be the particle sensing the event and starting the execution of the protocol. In this
protocol, each particle p′ that has received info(E) does the following:
r Search Phase. It uses a periodic low-energy directional broadcast in order to
discover a particle nearer to W than itself (i.e., a particle p′′ where d(p′′ , W) <
d(p′ , W)).
r Direct Transmission Phase. Then, p′ sends info(E) to p′′ .
r Backtrack Phase. If consecutive repetitions of the search phase fail to discover a
particle nearer to W, then p′ sends info(E) to the particle from which it originally
received the information.
Note that one can estimate an a priori upper bound on the number of repetitions of
the search phase needed, by calculating the probability of success of each search phase
as a function of various parameters (such as density, search angle, and transmission
range). This bound can be used to decide when to backtrack.
For a graphical representation see Figures 15.2 and 15.3.
a
-a
p'
Beacon circle
W
Figure 15.2. Example of the search phase.
442
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
p3
p2
p1
a2
a1
a0
p0
W
Figure 15.3. Example of a transmission.
15.2.3 Theoretical Analysis
Reference 5 first provides some basic definitions.
Definition 15.2.1. Let hopt (p, W) be the (optimal) number of “hops” (direct, vertical
to W transmissions) needed to reach the wall, in the ideal case in which particles
always exist in pairwise distances R on the vertical line from p to W. Let be a
smart dust propagation protocol, using a transmission path of length L(, p, W)
to send information about event E to wall W. Let h(, p, W) be the actual number
of hops (transmissions) taken to reach W. The “hops” efficiency of protocol is the
ratio
Ch =
h(, p, W)
hopt (p, W)
Clearly, the number of hops (transmissions) needed characterizes the energy consumption and the time needed to propagate the information E to the wall. Remark that
hopt = ⌈ d(p,RW ) ⌉, where d(p, W) is the (vertical) distance of p from the wall W.
In the case where the protocol is randomized, or in the case where the distribution
of the particles in the cloud is a random distribution, the number of hops h and the
efficiency ratio Ch are random variables and one wishes to study their expected values.
The reason behind these definitions is that when p (or any intermediate particle
in the information propagation to W) “looks around” for a particle as near to W as
possible to pass its information about E, it may not get any particle in the perfect
direction of the line vertical to W. This difficulty comes mainly from three causes:
(a) Due to the initial spreading of particles of the cloud in the area and because
particles do not move, there might not be any particle in that direction. (b) Particles
of sufficient remaining battery power may not be currently available in the right
direction. (c) Particles may temporarily “sleep” (i.e., not listen to transmissions) in
order to save battery power.
Note that any given distribution of particles in the smart dust cloud may not allow
the ideal optimal number of hops to be achieved at all. In fact, the least possible
LTP: A HOP-BY-HOP DATA PROPAGATION PROTOCOL
443
number of hops depends on the input (the positions of the grain particles). Reference 5,
however, compares the efficiency of protocols to the ideal case. A comparison with the
best achievable number of hops in each input case will of course give better efficiency
ratios for protocols.
To enable a first step toward a rigorous analysis of smart dust protocols, reference 5
makes the following simplifying assumption: The search phase always finds a p′′ (of
sufficiently high battery) in the semicircle of center the particle p′ currently possessing
the information about the event and radius R, in the direction toward W. Note that this
assumption on always finding a particle can be relaxed in the following ways: (a) By
repetitions of the search phase until a particle is found. This makes sense if at least one
particle exists but was sleeping during the failed searches. (b) By considering, instead
of just the semicircle, a cyclic sector defined by circles of radiuses R − R, R and
also taking into account the density of the smart cloud. (c) If the protocol during a
search phase ultimately fails to find a particle toward the wall, it may backtrack.
Reference 5 also assumes that the position of p′′ is uniform in the arc of angle 2a
around the direct line from p′ vertical to W. Each data transmission (one hop) takes
constant time t (so the “hops” and time efficiency of our protocols coincide in this
case). It is also assumed that each target selection is stochastically independent of the
others, in the sense that it is always drawn uniformly randomly in the arc (−α, α).
The above assumptions may not be very realistic in practice; however, they can
be relaxed and in any case allow us to perform a first effort toward providing some
concrete analytical results.
Lemma 1 [5]. The expected “hops efficiency” of the local target protocol in the
a-uniform case is
E(Ch ) ≃
α
sin α
for large hopt . Also
1 ≤ E(Ch ) ≤
for 0 ≤ α ≤
π
≃ 1.57
2
π
2.
Proof. Due to the protocol, a sequence of points is generated, p0 = p, p1 ,
p2 , . . . , ph−1 , ph , where ph−1 is a particle within W’s range and ph is part of
the wall. Let αi be the (positive or negative) angle of pi with respect to pi−1 ’s vertical
line to W. It is
h−1
i=1
d(pi−1 , pi ) ≤ d(p, W) ≤
h
i=1
d(pi−1 , pi )
444
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
Since the (vertical) progress toward W is then i = d(pi−1 , pi ) = R cos αi , we get
h−1
i=1
cos αi ≤ hopt ≤
h
cos αi
i=1
From Wald’s equation for the expectation of a sum of a random number of independent
random variables (see reference 7), then
E(h − 1) · E(cos αi ) ≤ E(hopt ) = hopt ≤ E(h) · E(cos αi )
Now, ∀i, E(cos αi ) =
α
1
−α cos x 2α dx
=
sin α
α .
Thus
α
E(h)
α
1
≤
+
= E(Ch ) ≤
sin α
hopt
sin α
hopt
Assuming large values for hopt (i.e., events happening far away from the wall, which
is the most interesting case in practice since the detection and propagation difficulty
increases with distance), we have (since for 0 ≤ α ≤ π2 it is 1 ≤ sinα α ≤ π2 ) and
the result follows.
15.2.4 Local Optimization: The Min-two Uniform Targets
Protocol (M2TP)
Reference 5 further assumes that the search phase always returns two points p′′ , and
p′′′ , each uniform in (−α, α), and that a modified protocol M2TP selects the best of
the two points, with respect to the local (vertical) progress. This is in fact an optimized
version of the Local Target Protocol.
In a similar way as in the proof of the previous lemma, the authors prove the
following result:
Lemma 2 [5]. The expected “hops efficiency” of the “min-two uniform targets”
protocol in the a-uniform case is
E(Ch ) ≃
α2
2(1 − cos α)
for 0 ≤ α ≤ π2 and for large h.
Now remark that
lim E(Ch ) = lim
α→0
α→0
2α
=1
2 sin a
PFR—A PROBABILISTIC FORWARDING PROTOCOL
445
and
limπ E(Ch ) =
α→ 2
π2
(π/2)2
=
≃ 1.24
2(1 − 0)
8
Thus, reference 5 proves the following:
Lemma 3 [5]. The expected “hops” efficiency of the min-two uniform targets protocol is
1 ≤ E(Ch ) ≤
π2
≃ 1.24
8
for large h and for 0 ≤ α ≤ π2 .
Remark that, with respect to the expected hops efficiency of the local target protocol, the min-two uniform targets protocol achieves, because of the one additional
search, a relative gain that is (π/2 − π2 /8)/(π/2) ≃ 21.5%. Reference 5 also experimentally investigates the further gain of additional (i.e., m > 2) searches.
15.3 PFR—A PROBABILISTIC FORWARDING PROTOCOL
The LTP protocol, as shown in the previous section, manages to be very efficient
by always selecting exactly one next-hop particle, with respect to some optimization
criterion. Thus, it tries to minimize the number of data transmissions. LTP is indeed
very successful in the case of dense and robust networks, since in such networks a next
hop particle is very likely to be discovered. In sparse or faulty networks, however,
the LTP protocol may behave poorly, because of many backtracks due to frequent
failure to find a next hop particle. To combine energy efficiency and fault-tolerance,
the Probabilistic Forwarding Protocol (PFR) has been introduced. The tradeoffs in
the performance of the two protocols implied above are shown and discussed in great
detail in reference 8.
15.3.1 The Model
The PFR protocol was introduced by Chatzigiannakis et al. [9]. They assume the case
where particles are randomly deployed in a given area of interest. Such a placement
may occur, for example, when throwing sensors from an airplane over an area.
As a special case, they consider the network being a lattice (or grid) deployment
of sensors. This grid placement of grain particles is motivated by certain applications,
where it is possible to have a pre-deployed sensor network, where sensors are put
(possibly by a human or a robot) in a way that they form a two-dimensional lattice.
Note indeed that such sensor networks, deployed in a structured way, might be useful
in precise agriculture, where humans or robots may want to deploy the sensors in
a lattice structure to monitor, in a rather homogeneous and uniform way, certain
446
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
conditions in the spatial area of interest. Certainly, exact terrain monitoring in military
applications may also need some sort of a grid-like-shaped sensor network. Note
also that Akyildiz et al. in a recent state-of-the-art survey [1] do not exclude the
pre-deployment possibility. Also, reference 10 explicitly refers to the lattice case.
Moreover, as the authors of reference 10 state in an extended version of their work
([11]), they consider, for reasons of “analytic tractability,” a square grid topology.
Let N be the number of deployed grain particles. There is a single point in the
network area which we call the sink S, and it represents a control center to which data
should be propagated.
We assume that each grain particle has the following abilities:
1. It can estimate the direction of a received transmission (e.g., via the technology
of direction-sensing antennae).
2. It can estimate the distance from a nearby particle that did the transmission
(e.g., via estimation of the attenuation of the received signal).
3. It knows the direction toward the sink S. This can be implemented during a setup
phase, where the (very powerful in energy) sink broadcasts the information
about itself to all particles.
4. All particles have a common coordinates system.
Notice that GPS information is not needed for this protocol. Also, there is no
need to know the global structure of the network.
15.3.2 The Protocol
The PFR protocol is inspired by the probabilistic multipath design choice for the
Directed Diffusion paradigm mentioned in reference 10. Its basic idea of the protocol
(introduced in reference 9) lies in probabilistically favoring transmissions toward the
sink within a thin zone of particles around the line connecting the particle sensing the
event E and the sink (see Figure 15.4). Note that transmission along this line is energy
optimal. However, it is not always possible to achieve this optimality, basically because
certain sensors on this direct line might be inactive, either permanently (because
their energy has been exhausted) or temporarily (because these sensors might enter
a sleeping mode to save energy). Further reasons include (a) physical damage of
sensors, (b) deliberate removal of some of them (possibly by an adversary in military
applications), (c) changes in the position of the sensors due to a variety of reasons
(weather conditions, human interaction, etc). and (d) physical obstacles blocking
communication.
The protocol evolves in two phases:
Phase 1: The “Front” Creation Phase. Initially the protocol builds (by using a
limited, in terms of rounds, flooding) a sufficiently large “front” of particles,
in order to guarantee the survivability of the data propagation process. During
PFR—A PROBABILISTIC FORWARDING PROTOCOL
447
E
Particles that
participate in
forwarding path
S
Figure 15.4. Thin zone of particles.
this phase, each particle having received the data to be propagated, deterministically forwards them toward
√ the sink. In particular, and for a sufficiently large
number of steps s = 180 2, each particle broadcasts the information to all
its neighbors, toward the sink. Remark that to implement this phase, and in
particular to count the number of
√steps, we use a counter in each message. This
counter needs at most ⌈log 180 2⌉ bits.
Phase 2: The Probabilistic Forwarding Phase. During this phase, each particle P
possessing the information under propagation calculates an angle φ by calling
the subprotocol “φ-calculation” (see description below) and broadcasts info(E )
to all its neighbors with probability IPfwd (or it does not propagate any data with
probability 1 − IPfwd ) defined as follows:
IPfwd =
1
φ
π
if φ ≥ φthreshold
otherwise
where φ is the angle defined by the line EP and the line PS and φthreshold = 134◦
(the selection reasons of this φthreshold will become evident in Section 15.3.4).
In both phases, if a particle has already broadcast info(E ) and receives it again,
it ignores it. Also the PFR protocol is presented for a single event tracing. Thus no
multiple paths arise and packet sizes do not increase with time.
Note that when φ = π, then P lies on the line ES and vice versa (and always
transmits).
If the density of particles is appropriately large, then for a line ES there is (with
high probability) a sequence of points “closely surrounding ES” whose angles φ are
larger than φthreshold , and thus successive points are within transmission range. All
such points broadcast and thus essentially they follow the line ES (see Figure 15.4).
448
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
P prev
P
φ
S
E
Figure 15.5. Angle φ calculation example.
The φ-Calculation Subprotocol (see Figure 15.5). Let Pprev be the particle
that transmitted info(E) to P.
1. When Pprev broadcasts info(E), it also attaches the info |EPprev | and the direc−−−→
tion Pprev E.
2. P estimates the direction and length of line segment Pprev P, as described in the
model.
−
→
3. P now computes angle (EP
prev P), and computes |EP| and the direction of PE
(this will be used in further transmission from P).
4. P also computes angle (P
prev PE) and by subtracting it from (P
prev PS) it
finds φ.
Notice the following:
i. The direction and distance from activated sensors to E is inductively propagated
(i.e., P becomes Pprev in the next phase).
ii. The protocol needs only messages of length bounded by log A, where A is
some measure of the size of the network area, since (because of (i) above)
there is no cumulative effect on message lengths.
Essentially, the protocol captures the intuitive, deterministic idea “if my distance
from ES is small, then send, else do not send.” Chatzigiannakis et al. [9] have chosen
to enhance this idea by random decisions (above a threshold) to allow some local
flooding to happen with small probability and thus to cope with local sensor failures.
15.3.3 Properties of PFR
Any protocol solving the data propagation problem must satisfy the following three
properties:
r Correctness. must guarantee that data arrive to the position S, given that the
whole network exists and is operational.
PFR—A PROBABILISTIC FORWARDING PROTOCOL
Particles
449
E
Lattice
Dissection
S
Figure 15.6. A lattice dissection G.
r Robustness. must guarantee that data arrive at enough points in a small interval around S, in cases where part of the network has become inoperative.
r Efficiency. If activates k particles during its operation, then should have a
small ratio of the number of activated over the total number of particles r = Nk .
Thus r is an energy efficiency measure of .
Reference 9 shows that this is indeed the case for PFR.
Consider a partition of the network area into small squares of a fictitious grid G
(see Figure 15.6). Let the length of the side of each square be l. Let the number of
squares be q. The area covered is bounded by ql2 . Assuming that we randomly throw
in the area at least αq log q = N particles (where α > 0 a suitable constant), then the
probability that a particular square is avoided tends to 0. So with very high probability
(tending to 1), all squares get particles.
Reference 9 conditions all the analysis on this event, call it F , of at least one
particle in each square.
15.3.4 The Correctness of PFR
Without loss of generality, we assume that each square of the fictitious lattice G has
side length 1.
In reference 9 the authors prove the correctness of the PFR protocol, by using a
geometric analysis. We sketch their proof below.
Consider any square intersecting the ES line. By the occupancy argument above,
there is with high probability a particle in this square. Clearly, the worst case is when
the particle is located in one of the corners of (since the two corners located most
far away from the ES line have the smallest φ-angle among all positions in ).
By some geometric calculations, reference 9 finally proves that the angle φ of this
particle is φ > 134◦ . But the initial square (i.e., that containing E) always broadcasts,
and any intermediate intersecting square will be notified (by induction) and thus will
450
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
broadcast because of the argument above. Thus the sink will be reached if the whole
network is operational.
Lemma 4 [9]. PFR succeeds with probability 1 in sending the information from E
to S given the event F .
15.3.5 The Energy Efficiency of PFR
Reference 9 considers the fictitious lattice G of the network area and let the event F
hold. There is (at least) one particle inside each square. Now join all nearby particles
of each particle to it, thus forming a new graph G′ which is “lattice-shaped,” but its
elementary “boxes” may not be orthogonal and may have varied length. When G′ s
squares become smaller and smaller, then G′ will look like G. Thus, for reasons of
analytic tractability, Chatzigiannakis et al. [9] assume that particles form a lattice
(see Figure 15.7). They also assume length l = 1 in each square, for normalization
purposes. Notice, however, that when l → 0, then “G′ → G” and thus all results in
this section hold for any random deployment “in the limit.”
The analysis of the energy efficiency considers particles that are active but are as
far as possible from ES. Thus the approximation is suitable for remote particles.
Reference 9 estimates an upper bound on the number of particles in an n × n
(i.e., N = n × n) lattice. If k is this number, then r = nk2 (0 < r ≤ 1) is the “energy
efficiency ratio” of PFR.
More specifically, Chatzigiennakis et al. [9] prove the (very satisfactory) result
below. They consider the area around the ES line, whose particles participate in
the propagation process. The number of active particles is thus, roughly speaking,
captured by the size of this area, which in turn is equal to |ES| times the maximum
distance from |ES| (where maximum is over all active particles).
This maximum distance is clearly a random variable. To calculate the expectation and variance of this variable, the authors in reference 9 basically “upper bound”
the stochastic process of the distance from ES by a random walk on the line, and
subsequently they “upper bound” this random walk by a well-known stochastic process (i.e., the “discouraged arrivals” birth and death Markovian process; see, e.g.,
reference 12). Thus they prove the following:
E
Sensor
Particles
Sink
Figure 15.7. A lattice sensor network.
EBP: THE ENERGY BALANCE PROTOCOL
451
Theorem 15.3.1 [9]. The energy efficiency of the PFR protocol is (( nn0 )2 ), where
√
n0 = |ES| and n = N, where N is the number of particles in the network. For
n0 = |ES| = o(n), this is o(1).
15.3.6 The Robustness of PFR
To prove the following robustness result, the authors in reference 9 consider particles
“very near” to the ES line. Clearly, such particles have large φ-angles (i.e., φ > 134◦ ).
Thus, even in the case where some of these particles are not operating, the probability that none of those operating transmits (during the probabilistic phase 2) is very
small. Thus, reference 9 proves the following.
Lemma 5 [9]. PFR manages to propagate the crucial data across lines parallel to ES,
and of constant distance, with fixed nonzero probability (not depending on n, |ES|).
15.4 THE ENERGY BALANCE PROBLEM
In order to save energy and keep the network functional for as long as possible,
various approaches, including hop-by-hop transmission techniques [5, 10, 11], as
well as clustering techniques [13] and alternating power-saving modes [14], have
been proposed.
All such techniques do not explicitly take care of the possible overuse of certain
sensors in the network. As an example, note that in hop-by-hop transmissions toward
the sink, the sensors lying closer to the sink tend to be utilized exhaustively (since all
data pass through them). Thus, these sensors may die out very early, thus resulting
in network collapse, although there may be still significant amounts of energy in the
other sensors of the network. Similarly, in clustering techniques the cluster heads that
are located far away with respect to the sink tend to spend a lot of energy.
In this chapter, we present two protocols trying to balance energy dissipation
among the sensors in the network: (a) the EPB (Energy Balance) protocol, introduced in reference 15, which probabilistically chooses between either propagating
data one hop toward the sink or sending directly to the sink. The first choice is more
energy-efficient, while the latter bypasses the critical (close to the sink) sectors. The
appropriate probability for each choice in order to achieve energy balance is calculated in reference 15. (b) VTRP (Variable Transmission Range Protocol), proposed in
reference 16, which implicitly contributes to energy balance by appropriately adapting (increasing) the transmission range, thus bypassing critical sensors and avoiding
possible obstacles.
15.5 EBP: THE ENERGY BALANCE PROTOCOL
15.5.1 The Model and the Problem
We assume that crucial events, which should be reported to a control center, occur in
the network area. Furthermore, we assume that these events are happening at random
452
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
uniform positions. Let N be their total number in a certain period (i.e., during the
execution of the protocol).
The sensors are spread in the network area randomly uniformly, so their number in
a certain area is proportional to the area’s size. There is a single point in the network,
which we call the “sink” S, that represents the fixed control center, to which data about
event realization should be reported. The sink is very powerful, in terms of both energy
and computing power. Sensors can be aware of the direction (and position) of the sink,
as well as of their distance to the sink. Such information can be easily obtained during
a setup phase, by having the (powerful) sink broadcast control messages to the entire
network area. We assume that the transmission range of sensors can vary with time
(in fact, for each sensor our protocol may use only two different ranges: R and i · R,
where i is a measure of the sensor’s distance to the sink). The sensors do not move.
We virtually “cover” the network area by a cycle sector of angle φ (see Figure 15.8).
The cycle sector is divided into n ring sectors or “slices.” The first slice has radius
R (i.e., the sensors’ transmission range). Slice i (2 ≤ i ≤ n) is defined by two cycles
sectors, one of radius i · R and the other of radius (i − 1) · R. Taking a sufficiently
large angle φ and/or by taking multiple sectors, we can cover the whole area.
As far as energy dissipation is concerned, we assume that the energy spent at a
sensor when transmitting data is proportional to the square of the transmitting distance.
Our protocol’s performance analysis can, however, be extended to any energy cost
model. Note that the energy dissipation for receiving is not always negligible with
respect to the energy when sending, such as in the case when transmissions are too
short and/or radio electronics energy is high (see reference [13]). In the analysis,
we only count (for simplicity) energy spent during transmissions. Since, however,
in our protocol (see next section) there is one receipt for each transmission, it is
clear that even when energy during receipt is more or less the same as energy during
transmissions, the analysis can be extended easily to the full case (counting both
transmissions and receipts).
R
n
2
1
ϕ
Figure 15.8. Sensor network with n ring sectors, angle φ, and ring “width” R.
EBP: THE ENERGY BALANCE PROTOCOL
453
Definition 15.5.1. The area between two consecutive cycle sectors is called a ring
sector (or “slice”). Let Ti (1 ≤ i ≤ n) be the ith ring sector of the network.
T1 stands for the ring sector with center the sink and radius equal to R.
Definition 15.5.2. Let Si be the area size of the ring sector Ti of the network (1 ≤
i ≤ n).
We wish to solve the “energy balanced data propagation problem”,—that is, propagate data to the sink in such a way that the “average” energy dissipation in each sensor
is at each time the same. The average energy dissipation per sensor is taken to be the
fraction of the total energy spent by sensors in a ring sector over the number of sensors in that sector. Because of our assumption that the number of sensors in an area is
proportional to the area size, the average energy dissipation per sensor is calculated
by dividing the total energy spent in a sector by the sector size.
Here we do not study medium access aspects, assuming the existence of underlying
MAC protocol.
15.5.2 The Protocol
We assume that each event is sensed by only one sensor. This assumption is not
restrictive since we may consider multiple sensing and propagation of an event by
various sensors as sensing and propagation of many different events. A sensor sensing
an event generates a data message that should be eventually delivered to the sink. On
each ring sector, Ti , a number of events occur and a corresponding number of messages
(one for each event) is generated.
Randomization is used to achieve some “load balancing” by evenly spreading the
“load” (energy dissipation). In particular, on ring sector Ti each event is propagated
to Ti−1 (i.e., the “next” sector towards the sink) with probability pi , while with probability 1 − pi it is propagated directly to the sink S. Each message in Ti is handled
stochastically independently of the other events’ messages.
The choice of probability pi for Ti is made so that the average energy consumption
per area unit (and thus per sensor) is the same for the whole network. There is a tradeoff
from choosing pi : If pi increases, then transmissions tend to happen locally, thus
energy consumption is low; however, sensors closer to the sink tend to be overused
since all data passes through them. On the other hand, if pi decreases, there are
distant transmissions (thus a lot of energy is consumed); however, particles closer
to the sink are bypassed. Calculating the appropriate probability pi for each Ti and
solving the problem of energy balance is very important since it combines efficient
data propagation with increased network’s lifetime.
By using an underlying subprotocol [5, 10], we can guarantee that only one “next
hop” sensor receives the transmitted message. Note also that data messages are of
fixed size; that is, no further information is added to a message on its route toward
the sink.
Our protocol is (a) distributed, since each sensor chooses propagation probability
independently of other sensors, and (b) uses only local information, in the sense that
454
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
pi depends only on i, that is, a parameter related to the distance from the sink. Note
that distance from the sink information for each sensor can be easily obtained—that
is, during a setup phase where the sink broadcasts control messages to the network.
Several techniques (including signal attenuation evaluation) can be used to estimate
each sensor’s distance for the sink. (c) The protocol is very simple, since it just uses
a random choice based only on parameter i.
15.5.3 Basic Definitions—Preliminaries
We aim at calculating probability pi for each i in order to ensure the energy balance
property. Using simple geometry, one can easily prove the following lemmas.
Lemma 6. The area size, S1 , of the ring sector T1 is S1 =
φ
2
· R2 .
Lemma 7. The relation between the area size of the ring sector Ti and that of T1 is
Si = (2i − 1) · S1 .
Definition 15.5.3. Let λi the probability that an event will occur on the ring sector
Ti .
There are n ring sectors in the network.
Lemma 8. Assuming a random uniform generation of events in the network area, the
probability λi of an event occurring on the ring sector Ti (1 ≤ i ≤ n) is
λi =
(2i − 1)
n2
Let us now consider sector Ti .
Definition 15.5.4. An area Ti “handles” an event generated in ring sector j if either
the message was generated in the area Ti (i.e. j = i) or the message was propagated
to Ti from the ring sector Ti+1 .
Definition 15.5.5. Let hi be the number of the messages that are “handled” by the
area Ti .
We now define energy ǫij spent for message j when sector i handles it.
Definition 15.5.6. Let ǫij a random variable which measures the energy that dissipates
the sector Ti so as to handle the message j. For ǫij we have that
ǫij =
cR2
with probability pi
c(iR)2
with probability 1 − pi
EBP: THE ENERGY BALANCE PROTOCOL
455
where cR2 is the energy dissipation for sending a message j from Ti to its adjacent
ring sector Ti−1 and c is a constant.
Thus, the expected energy dissipation in sector i for handling a message is
E[ǫi,j ] = cR2 · [i2 − pi (i2 − 1)]
(15.1)
Note: The expected energy above is the same for all messages; we use j just for
counting purposes.
Definition 15.5.7. Let Ei the total energy spent by sensors in Ti . Clearly,
Ei =
hi
(15.2)
ǫij
j=1
Energy balance is defined as follows:
Definition 15.5.8. The network is energy balanced if the average per sensor energy
dissipation is the same for all sectors, i.e. when
E[Ei ]
E[Ej ]
=
Si
Sj
i, j = 1, . . . , n
(15.3)
15.5.4 The General Solution
We next provide a lemma useful in the estimation of the total energy dissipation in a
sector.
Lemma 9. The expected total energy dissipation in sector i is
E[Ei ] = E[hi ] · E[ǫik ]
Proof.
E[Ei ] = E
=
=
N
n=0
N
n=0
hi
k=1
E
E
ǫik
hi
k=0
h
i
k=1
(ǫik ∩ hi = n)
ǫik hi = n · IP{hi = n}
456
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
Furthermore,
E
hi
k=1
n
n
ǫik hi = n = E
ǫik =
E[ǫik ]
k=1
k=1
= n · E[ǫik ]
Thus, we have from the above that
E[Ei ] =
N
n=0
E
hi
k=1
ǫik hi = n · IP{hi = n}
= E[ǫik ] · E[hi ]
Definition 15.5.9. Let gi be the number of the messages that are generated in the
area Ti .
Note that messages are generated in an area only when events occur in this area.
Definition 15.5.10. Let fi be the number of the messages that are forwarded to the
area Ti .
We note that messages are forwarded to a ring sector (say i) only because of an
event generation at a sector j > i and successive one-hop propagations from sector j
to sector i.
We notice the following important relation:
hi = gi + fi
(15.4)
which means that the number of messages that area Ti handles equals the number of
the messages that are generated in Ti , plus the number of messages that are forwarded
to it. Because of event generation according to a probability distribution and also
because of the probabilistic nature of message propagation in, all three quantities
above are random variables. By linearity of expectation, we get the following lemma.
Lemma 10. E[hi ] = E[gi ] + E[fi ].
We establish a relationship between E[fi ] and E[hi+1 ].
Lemma 11. E[fi ] = pi+1 · E[hi+1 ].
EBP: THE ENERGY BALANCE PROTOCOL
457
Proof. Let δi,j an indicator random variable that is equal to 1 if area Ti forwards the
message j to the area Ti−1 and 0 otherwise. Thus
δi,j =
1
0
with probability pi
with probability 1 − pi
Clearly, δi,j depends only on i, but we add j for counting purposes. Obviously,
E[δi,j ] = pi . It is
fi =
hi+1
δi+1,j
j=0
Similarly to the proof of Lemma 9, we get
E[fi ] =
N
n=0
hi+1
E
δi+1,j hi+1 = n · IP{hi+1 = n}
j=0
and the proof is completed.
Recall that, according to Definition 15.5.8, to achieve the same on the average
energy dissipation per area unit (and thus per sensor) in the network area, the following equality should hold:
hi
hj
ǫik
ǫjk
k=1
k=1
E
= E
Si
Sj
∀i, j ∈ {1, . . . , n}
(15.5)
that is, the average energy consumption per sensor should be equal in any two ring
sectors. By induction, it suffices to guarantee this for any two adjacent sectors. In
what follows, we guarantee the above balance property, requiring a certain recurrence
relation to hold. This recurrence basically relates three successive terms of the E[fi ]
sequence (the E[gi ] terms depend only on i and on input parameters).
Theorem 15.5.1. To achieve energy balance in the network, the following recurrency
equation should hold:
ai+1 E[fi+1 ] − (di + ai )E[fi ] + di−1 E[fi−1 ]
= ai E[gi ] − ai+1 E[gi+1 ]
458
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
where
ai =
i2
2i−1 ,
(i+1)2 −1
2i+1
di =
Proof. For the case j = i + 1 of Eq. (15.5) and using Lemmas 7 and 9, we have
E[hi+1 ]E[ǫi+1,j ]
E[hi ]E[ǫi,j ]
=
⇔
Si
Si+1
E[hi ][i2 − pi (i2 − 1)]
E[hi+1 ]{(i + 1)2 − pi+1 [(i + 1)2 − 1]}
=
(2i − 1)
(2i + 1)
⇔
i2 − 1
(i + 1)2 − 1
(i + 1)2
i2
E[hi ] − pi E[hi ]
=
E[hi+1 ] − pi+1 E[hi+1 ]
2i − 1
2i − 1
2i + 1
2i + 1
Let ai , and di be as defined in the theorem statement above. By Lemma 11 we know
that pi E[hi ] = E[fi−1 ], and by Lemma 10 it is E[hi ] = E[gi ] + E[fi ]; thus the last
equation becomes
ai+1 E[fi+1 ] − (di + ai )E[fi ] + di−1 E[fi−1 ]
= ai E[gi ] − ai+1 E[gi+1 ]
To solve the above recurrency we must compute E[gi ].
Lemma 12. If N is the total number of events that are generated in the network, the
mean value of gi is given by the following relationship:
E[gi ] = N · λi
Proof. Because the position of each event is independent of other events and because
for each sector i, probability λi is the same, clearly gi is binomial with parameters N
and λi .
In order to have a simpler recurrence involving only two (successive in fact) terms
of the E[fi ] sequence, we will transform the recurrency relation of Theorem 15.5.1.
into the following (easier to solve) relation:
Lemma 13. The recurrency relation
ti − ti−1 = ai · E[fi ] − ai+1 · E[fi+1 ]
t0 = a1 · E[f1 ]
has as a solution the function
for
i = 1, . . . , n − 1 and
EBP: THE ENERGY BALANCE PROTOCOL
ti =
i
j=1
459
aj E[gj ] − aj+1 E[gj+1 ] + a1 · E[f1 ]
Proof. The proof is done by induction on i. For i = 0, it is obviously true. Let it be
true for i − 1. For i we have
ti = ti−1 + ai · E[gi ] − ai+1 · E[gi+1 ]
By the induction hypothesis, we get the solution
ti =
i
j=1
aj E[gj ] − aj+1 E[gj+1 ] + a1 · E[f1 ]
Now the recurrency relation of Theorem 15.5.1. is simplified:
ai+1 · E[fi+1 ] − di · E[fi ] = ti ,
i = 1, . . . , n − 1
Thus, we get a recurrence for sequence E[fi ] involving only two successive terms
of the sequence.
Theorem 15.5.2. The recurrency relation
ai+1 E[fi+1 ] − di E[fi ] = ti ,
i = 1, . . . , n − 1
where ti is defined in Lemma 13, has the following solution:
E[fn−i ] = −
i
k=1
i−1
j=k an−j
i
j=k dn−j
· tn−k
The proof is rather complex, so we omit it here. The interested reader may find it
in reference 15.
The full expression for E[fi ] can be expressed by substituting i with n − i, thus
E[fi ] = −
n−i
k=1
n−i+1
j=k
n−i
an−j
j=k dn−j
·
n−k
j=1
(aj E[gj ] − aj+1 E[gj+1 ]) + a1 · E[f1 ]
where i−1
ai = 1.
i
We note that all the parameters of the recurrency solution above are expressed
as a function of E[f1 ] and i. So as to compute them, we firstly compute the value
of E[f1 ]. Then we can compute all the other parameters by replacing the already
computed E[f1 ].
Now, the calculation of the probabilities pi is quite easy.
460
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
Theorem 15.5.3. The energy balance property is achieved if any ring sector (say Ti )
propagates each message it handles with probability pi to the next ring sector, Ti−1 ,
and with probability 1 − pi it propagates the message directly to the sink. The value
of each pi is given by the following relation:
pi =
E[fi−1 ]
E[gi ] + E[fi ]
where the values of E[fi ] and E[gi ] are obtained from Theorem 15.5.2. and Lemma
12, respectively.
Proof. From Eq. (15.11) we know that E[fi−1 ] = pi E[hi ] and also by Lemma 10 we
know that E[hi ] = E[gi ] + E[fi ].
Remark. Note that, interestingly, pi ’s are independent of the number N of the events
that occur in the network, since pi ’s depend only on i and the number of ring sectors
n (which is broadcast to sectors by the sink). Thus the protocol assumes only local
information.
We note that the analysis above allows the exact derivation of probabilities pi ’s
as a function of i and n which (although complicated and not obviously leading to a
closed form) can be easily calculated by the sensors in the network by carrying out
very simple calculations.
The authors of reference 15 also prove the correctness of the protocol.
Theorem 15.5.4. Given that the energy is each time on the average the same in all
network sensors, each message will finally get to the sink.
15.5.5 A Closed Form
Under specific assumptions (that we discuss and motivate), we can make the calculation of probabilities pi simpler. Combining Lemmas 10 and 12, we have that
E[hi ] = λi N + E[fi ] =
2i − 1
N + E[fi ]
n2
(15.6)
By the corresponding relation for E[hi−1 ], it must be
2i−1
N+E[fi ]
n2
(2i−1) φ2 R2
[pi + (1 − pi )i2 ]R2 =
2i−3
N+E[fi−1 ]
n2
(2i−3) φ2 R2
[pi−1 + (1 − pi−1 )(i − 1)2 ]R2
But 2i − 1 ≃ 2i − 3 and φ2 , R2 cancel. Dividing by N, we get
1+
E[fi ]
N
pi + (1 − pi )i2 ≃ 1 +
E[fi−1 ]
N
pi−1 + (1 − pi−1 )(i − 1)2
VTRP: THE VARIABLE TRANSMISSION RANGE PROTOCOL
461
If E[fi ] ≃ E[fi−1 ], then the previous relation becomes
pi + (1 − pi )i2 = pi−1 + (1 − pi−1 )(i − 1)2
(15.7)
In reference 15 we show how to solve the above recurrence.
Theorem 15.5.5. If E[fi ] ≃ E[fi−1 ], 3 ≤ i ≤ n, then the one-hop forwarding probability, guaranteeing energy balance, is
pi = 1 −
3x
(i + 1)(i − 1)
where p2 = x ∈ (0, 1) a free parameter and p1 = 0.
We note that the assumption E[fi ] ≃ E[fi−1 ] is quite reasonable and wellmotivated. We provide the following intuitive explanation of why this happens. Note
indeed that the area sizes of adjacent sectors (and thus the number of events generated
in such sections) are more or less the same, especially when i increases. Furthermore,
the probability pi of forwarding to the adjacent (towards the sink) sector increases
very fast with i and becomes 1 in most sectors of the network (in the middle territory).
15.6 VTRP: THE VARIABLE TRANSMISSION RANGE PROTOCOL
15.6.1 Introduction
The VTRP protocol has been proposed by T. Antoniou et al. [16]. They study the
problem of multiple event detection and propagation—that is, the local sensing of
a series of crucial events and the energy-efficient and fault-tolerant propagation of
data reporting the realization of these events to a (fixed or mobile) control center. The
control center could in fact be some human authorities responsible for taking action
upon the realization of the crucial event. We use the term sink for this control center.
We note that this problem generalizes the single event propagation problem (with
respect to references 5, 8, and 15) and poses new challenges for designing efficient
and fault-tolerant data propagation protocols. The new protocol we present here can
also be used for the more general problem of data propagation in sensor networks
(see reference 10).
The basic innovation in our protocol is to vary the range of data transmissions. This
feature aims at better performance, compared to typical fixed transmission range data
propagation, in some rather frequently occurring situations such as the following:
(a) The case of low densities of sensor particles. In such networks, fixed-range
protocols may trap in backtracking actions when no particles toward the sink
are found. Our protocol, by increasing the transmission range, may find such
particles and avoid extensive backtracking.
462
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
(b) Because of the possibility to increase transmission range, VTRP performs
better in cases of obstacles or faulty/sleeping sensors. Also, it bypasses certain
critical sensors (like those close to the sink) that tend to be overused, thereby
prolonging the network lifetime.
To demonstrate the above properties of VTRP, we compare it to a typical fixed-range
protocol: the Local Target Protocol ( LTP).
The ability of LTP to propagate information regarding the realization of a crucial
event to the control center depends on the particle density of the network. The experiments conducted in reference 5 indicate that for low particle densities, LTP fails to
propagate the messages to the control center (while for high particle densities the
failure rate drops very fast to zero; that is, the messages are almost always reported
correctly). The new protocol that we propose in this chapter successfully overcomes
this problem by increasing the transmission range of the particles that fail to locate an
active neighboring particle toward the sink. In fact, the experiments conducted in this
chapter (see Section 15.6.7) demonstrate the superiority of VTRP over LTP even for
sensor networks with very low particle densities.
Further note that this is the first time that the LTP protocol is evaluated under the
setting of multiple events. Our findings indicate that LTP has a fundamental design
flaw in this case, because the success of the propagation process heavily depends on
the lifetime of the particles that are located around the control center. As soon as these
particles exhaust their power supplies, the whole network becomes inoperable. The
new protocol that we present here successfully overcomes this problem by adjusting
the transmission range of the particles as soon as the particles closer to the control
center “die”. Our experiments indicate that VTRP increases the ability of the network
to report multiple events up to 100%, compared to LTP.
We propose four different mechanisms for varying the transmission range of the
particles that aim at different types of smart dust networks regarding particles densities and energy-saving criteria. Our experimental results show that VTRP can
be easily modified to further improve its performance. Actually, VTRPp (where
range is increased aggressively) and VTRPr (which randomizes between the various
range change functions toward a better average case performance) successfully propagate about 50% more events than the “basic” VTRP and almost 200% more events
than the original LTP protocol.
15.6.2 The Model
The model considered here is an extension of the one presented in Section 15.2.1. In
our model here (Figure 15.9), we assume that the transmission range (R) can vary (i.e.,
by setting the transmission power at appropriate levels) while the transmission angle
(let it be α) is fixed and cannot change throughout the operation of the network (since
this would require a modification or movement of the antenna used). Note that the
protocols we study in this work can operate even under the broadcast communication
mode (i.e., α = 2π). The laser possibility is added for reducing energy dissipation in
long-distance transmissions.
VTRP: THE VARIABLE TRANSMISSION RANGE PROTOCOL
α
463
S
-α
R
p'
beacon circle
Figure 15.9. Directed transmission of angle α.
Each particle can be in one of three different modes at any given time, regarding the
energy consumption. These modes are (a) transmission of a message, (b) reception
of a message, and (c) sensing of events.
For the case of transmitting and receiving a message, we assume the following
simple model where the radio dissipates Eelec to run the transmitter and receiver
circuitry and ǫamp for the transmit amplifier to achieve acceptable SNR (signal-tonoise ratio). We also assume an r 2 energy consumption due to channel transmission
at distance r. Thus to transmit a k-bit message at distance r in our model, the radio
expends
ET (k, r) = ET −elec (k) + ET −amp (k, r)
ET (k, r) = Eelec · k + ǫamp · k · r 2
and to receive this message, the radio expends
ER (k) = ER−elec (k)
ER (k, r) = Eelec · k
where ET −elec , ER−elec stand for the energy consumed by the transmitter’s and
receiver’s electronics, respectively.
Concluding, there are three different kinds of energy dissipation:
r ET : Energy dissipation for transmission.
r ER : Energy dissipation for receiving.
r Eidle : Energy dissipation for idle state.
For the idle state, we assume that the energy consumed for the circuity is constant
for each time unit and equals Eelec (the time unit is 1 second).
We note that in our simulations we explicitly measure the above energy costs. We
feel that our model, although simple, depicts accurately enough the technological
464
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
specifications of real smart dust systems. Similar models are being used by other
researchers in order to study sensor networks (see reference 17). In contrast to references 10 and 18, our model is weaker in the sense that no geolocation abilities are
assumed (e.g., a GPS device) for the smart dust particles, leading to more generic and
thus stronger results. In reference 19 a thorough comparative study and description
of smart dust systems is given, from the technological point of view.
15.6.3 The Problem
Assume the realization of a series of K crucial events Ei , with each event being
sensed by a single particle pi (i = 1, 2, . . . , K). Then the multiple event propagation
problem P is the following:
How can each particle pi (i = 1, 2, . . . , K), via cooperation with the rest of the grain
particles, in an efficient (mainly with respect to energy and time) and fault-tolerant way,
propagate information info(Ei ) reporting realization of event Ei to the sink S?
We note that this problem is a generalization of the single event propagation
problem, which is more difficult to cope with because of the severe energy restrictions
of the particles.
Certainly, because of the dense deployment of sensor particles close to each other,
communication between two particles is much more energy-efficient than direct transmission to the sink. Furthermore, short-range hop-by-hop transmissions can effectively overcome some of the signal propagation effects in long-distance transmissions
and may help to smoothly adjust propagation around obstacles. Finally, the low-energy
transmission in multihop communication may enhance security, protecting from
undesired discovery of the data propagation operation.
On the other hand, long-range transmissions require the participation of few
particles and therefore reduce the overhead on particle resources and provide better network response times. Furthermore, long-range communication permits the
deployment of clustering and other efficient techniques, developed for ad hoc wireless networks. In particular, a clustering scheme enables cluster heads to reduce the
amount of transmitted data by aggregating information.
The above suggest that many diverse approaches exist to the solution of the multiple
event propagation problem P. Further to choosing between long or short transmissions, certain additional tradeoffs are introduced by choosing between fixed or varying
transmission range. In particular, we wish to focus on the following important properties:
(a) Obstacle Avoidance: This may be achieved by increasing transmission range
when an obstacle is encountered.
(b) Fault Tolerance: Increasing range may reach active sensors when the current
range does not succeed, because of either faulty or “sleeping” sensors close
to sensor which is currently transmitting or in the case of very low network
densities.
VTRP: THE VARIABLE TRANSMISSION RANGE PROTOCOL
465
(c) Network Longevity: An interesting aspect of the problem under investigation is
the lifetime of particles, since it affects the ability of the network to propagate
data to the sink, because available routes are reduced as more particles consume
their energy resources and “die.” Varying transmission range may bypass the
sensors lying close to the sink, which tend to be overused in the case of fixed
range transmissions, since all data pass through them in this case. The same
holds also in the case of a geographical concentration of event generation.
15.6.4 The Variable Transmission Range Protocol (VTRP)
In this protocol, each particle p′ that has received info(E) from p (via, possibly, other
particles) does the following:
Phase 1: The Search Phase. It uses a periodic low-energy broadcast of a
beacon in order to discover a particle nearer to S than itself. Among the particles returned, p′ selects a unique particle p′′ that is “best” with respect to
progress toward the sink. More specifically, the particle p′′E that among all particles found achieves the bigger progress on the p′ S line should be selected
(see Figure 15.2).
Phase 2: The Direct Transmission Phase. Then, p′ sends info(E) to p′′ and sends
a success message to p (i.e., to the particle from which it originally received
the information).
Phase 3: The Transmission Range Variation Phase. If the search phase fails
to discover a particle nearer to S, p′ enters the transmission range variation
phase. More specifically, each particle maintains a local counter τ, with initial
value τ = 0. Every time the search phase fails, this counter is increased by
1. Thus τ is an indication of the number of failures to locate an active particle. Based on τ, the particle modifies its transmission range R according to a
change function F(τ). We here consider four different functions for varying
the transmission range:
(a) Constant Progress. This choice is more suitable in the case where the network is comprised of a large number of particles and thus, a small increment
of the transmission range will probably suffice to locate an active particle.
Based on this assumption, the change function is defined as follows:
F(τ) = Rnew = Rinit + c · τ,
(i.e., c = 10)
where c is a constant set to a small value
This is considered as the “basis” VTRP and is denoted as VTRPc .
(b) Multiplicative Progress. In this case, the transmission range of the particle
is increased mode drastically. We call this variation of our protocol VTRPm .
F(τ) = Rnew = Rinit + Rinit · m · τ ,
a small value (i.e., m = 3)
where m is a constant set to
466
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
This drastic change has bigger probability of finding an active particle;
however, it leads to higher energy consumption.
(c) Power Progress. In this case, the transmission range of the particle is increased even faster using the following scheme:
√
(τ+1)
F(τ) = Rnew = Rinit + Rinit
We call this protocol VTRPp .
(d) Random Progress. When the density of the network is not known in advance, we use randomization to avoid bad behavior due to the worst-case
input distributions for each choice above (i.e., small modifications to the
transmission range in VTRPc in case of low densities and big modifications resulting from VTRPp in high particle densities). We call this
variation VTRPr which is defined as follows:
F(0) = Rinit
F(τ) = F(τ − 1) + Rinit · r ,
where r ∈ (0, 8], a random value
At any given time there could be more than one event being propagated toward
the sink. In order to avoid repeated transmissions and infinite loops, each particle
is provided with a limited “cache memory.” In this cache, the particle registers the
event IDs for each distinct event it has “heard of.” Each event ID’s uniqueness is
guaranteed, by choosing it to be a concatenation of the source particle ID and the
timestamp of the sensed event. Upon the receival of a message, a particle checks
whether the pertinent event is enlisted in its cache. If the event that it is not in the
particle’s cache, it is registered and then the particle proceeds to the proper actions
defined by the VTRP protocol. However, if the event was already seen, the message
is dropped and no further action is taken.
Presumably, a relatively small amount of memory (e.g., up to 2 MB) would be
adequate for such purpose. Note that in the future the particle cache could enforce a
policy of limited lifetime for each of its contents, thus reducing the space requirements
to a minimum. Data aggregation also poses a challenge for further study and efficiency
assessment.
15.6.5 Implementation Details
To implement the protocols presented in the previous sections, we have used simDust
[16], which operates in Linux using C++ and the LEDA [20] algorithmic and data
structures library. An interesting feature for our simulator is its ability to experiment
with very large networks of thousands of nodes. In fact, two major reasons for creating
this simulator (a) the complexity of extending existing networks simulators (b) and
their (in cases of large instances) time-consuming execution. simDust enables the
protocol designer to implement the protocol using just C++ and avoids complicate
procedures that involve the use of more than one programming language. Additionally,
simDust generates all the necessary statistics at the end of each simulation based on
VTRP: THE VARIABLE TRANSMISSION RANGE PROTOCOL
467
a wide variety of metrics that are implemented (such as delivery percentage, energy
consumption, delivery delay, longevity, etc.)
The key points in simDust’s implementation are the following:
Operation in Rounds. A basic concept used in the simulator is that its operation
is divided into discrete rounds. One round represents a time interval in which
a particle can transmit or receive a message and process it according to the
protocol that is being simulated.
MAC Layer Assumptions. simDust leaves transmission collisions to be handled
by lower MAC layer protocols and does not take them into account. It is our
intention to consider them in future versions of this simulator.
Energy Assumptions. We have included a detailed energy dissipation scheme for
both protocols implemented. In particular, we have assumed that a particle consumes a standard amount of energy Eelec per round while being awake. Furthermore, in each transmission, energy consumption is proportional to the square of
the transmission distance. For each receive, a node is credited with an amount of
energy that practically reflects the power needed to run the transceiver circuit,
namely, Eelec . Finally, a particle can switch to the sleep state to save energy.
No energy consumption virtually takes place while the particle remains in the
sleep mode, since it keeps its transceiver and its sensors shut down.
Size of Messages: Regarding the communication cost in terms of the bits transmitted per message, we assume that information messages require 1 Kbyte,
plus a 40-bits header, containing a 32-bit identifier for the sender particle and
an 8-bit code that determines the message type.
15.6.6 Efficiency Measures
On each execution of the experiment, let K be the total number of crucial events
(E1 , E2 , . . . , EK ) and let k be the number of events that were successfully reported to
the sink S. Then, we define the success rate as follows:
Definition 15.6.1. The success rate, IPs , is the fraction of the number of events successfully propagated to the sink over the total number of events, that is, IPs = k/K.
Another crucial efficiency measure of our comparative evaluation of the two protocols is the average available energy of each particle in the network over time:
Definition 15.6.2. Let Ei be the available energy for the particle i. Then Etot = ni Ei
is the total energy available in the smart dust network, where n is the number of the
total particles dropped. Note that Ei and Etot vary with time.
Clearly, the less energy a protocol consumes, the better, but we have to notice that
the comparison, in order to be fair, should be done in cases where the other parameters
of efficiency should be similar (i.e., satisfy certain quality of service guarantees).
468
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
Finally, we consider as a measure of efficiency of the two protocols the number of
alive particles, capturing the network survivability in each case. As in the case of the
energy, the more particles that are alive, the better. A source of crucial information
is also the particular manner in which particles die over time, such as (a) the geographical distribution of the nodes that die out earlier and (b) the evolution of energy
consumption in critical sensors such as those lying close to control center.
Definition 15.6.3. Let hA (for “alive”) be the number of “alive” sensor particles
participating in the sensor network.
15.6.7 Experimental Results
We start our experimentation by evaluating the effect of the particle density on the
performance of the new protocol VTRP when compared to the already existing one,
LTP. We generate a variety of sensor fields in a 2000-m by 2000-m square; in these
fields, we drop n ∈ [1000, 8000] particles uniformly distributed on the smart-dust
plane. In each execution, we generate a single event by randomly selecting a particle
in the network. The results of this experiments are shown in Figure 15.10.
It is evident that the effect of particle density has significant impact on the performance of LTP. We observe that for low densities (i.e., n ≤ 2000) the protocol
almost always fails to report the event to S, while when n ≥ 5000 the success rate
increases approaching very fast one. This can be justified by taking into account the
average degree of each particle for various network sizes n. Note that similar obser-
1
0.9
Success Rate
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1000 2000 3000 4000 5000 6000 7000 8000
Total Number of Particles
LTP
VTRP
Figure 15.10. Success rate (IPs ) for LTP and VTRP for various particle densities (n ∈
[1000, 8000]).
VTRP: THE VARIABLE TRANSMISSION RANGE PROTOCOL
469
vations for LTP have been made in reference 5. On the other hand, the mechanism of
VTRP that increases the transmission range of the particles successfully overcomes
these problems. Even for the cases of very low particle densities, VTRP manages to
propagate the information reporting the realization of the event to the sink, with high
probability.
We continue our experimentation by investigating the performance of the protocols
in the case of multiple events. For this set of experiments we drop n = 5000 particles
uniformly distributed in a 2000-m by 2000-m square field. Then in each simulation
round, we generate one event at a random location in the sensor field that is sensed
by only one particle (given that this particle has enough power to sense it); that is,
we use a high event generation rate. This is repeated until a total of 9000 events are
generated. Note that this is the first time that the LTP protocol is evaluated under the
setting of multiple events.
Figure 15.11 depicts the success rate of the two protocols as the multiple events
are generated. Clearly, VTRP achieves better results than LTP and in fact manages
to propagate almost two times more events. The superiority of VTRP is explained
by the fact that in LTP the particles that are closer to S will always participate in the
propagation of the messages. The continuous transmissions of messages will eventually exhaust the power of this small group of (highly critical) particles, rendering
the rest of the network useless (although there are still energy supplies available)
since no further events can be reported to S. VTRP overcomes this problem by
activating the Transmission Range Variation Phase. As soon as the particles close to
S “die,” the neighboring nodes will sense it (during the Search Phase) and adjust their
transmission range appropriately, thus bypassing them and reaching the sink directly.
1
0.9
0.8
Success Rate
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1000 2000 3000 4000 5000 6000 7000 8000 9000
Number of Events
LTP
VTRP
Figure 15.11. Success rate (IPs ) for LTP and VTRP for multiple events (n = 5000).
470
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
2000
2000
1500
1500
1000
1000
500
500
0
0
500
1000
1500
2000
0
0
500
t=1
2000
1500
1500
1000
1000
500
500
0
500
1000
t = 5000
1500
2000
1500
2000
t = 2500
2000
0
1000
1500
2000
0
0
500
1000
t = 7500
Figure 15.12. Snapshots of the network showing alive particles when executing LTP at different
time instances (n = 5000).
This is clearly seen in Figure 15.12, where snapshots of the network are taken for
different time instances. As soon as some particles around S “die,” LTP fails to deliver
the remaining events.
Essentially, VTRP manages the energy of the network in a more efficient way. By
examining Figure 15.13, we observe that VTRP ends up using slightly more energy
than LTP in order to propagate more events to the control center. In fact, VTRP
will force the particles to spend more energy so that their transmissions manage to
reach S even if this will exhaust their power supplies. Again, this is clearly seen in
Figure 15.14, where snapshots of the network are taken for different time instances.
To get a more complete view on how each protocol manages the energy resources
of the particles, Figures 15.15 and 15.16 show the number of alive particles based
on their distance from the sink. In these figures we have grouped the particles in 32
sets based on the division of the diagonal line connecting (0, 0) with (2000, 2000)
in 32 sectors. We observe that for different time instances, the total number of alive
particles that are close to the sink (for sections 1–10) drops as the time increases while
the particles further away almost always remain active until the end of the experiment.
Observe how VTRP forces the particles close to S to sacrifice their battery supplies
in order to propagate more messages.
VTRP: THE VARIABLE TRANSMISSION RANGE PROTOCOL
471
5000
4500
Total Energy (J)
4000
3500
3000
2500
2000
1500
1000
500
0
0
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Simulation Time (rounds)
LTP
VTRP
Figure 15.13. Total energy (Etot ) for LTP and VTRP for multiple events (n = 5000).
2000
2000
1500
1500
1000
1000
500
500
0
0
500
1000
1500
2000
0
0
500
t=1
2000
2000
1500
1500
1000
1000
500
500
0
0
500
1000
t = 5000
1000
1500
2000
1500
2000
t = 2500
1500
2000
0
0
500
1000
t = 7500
Figure 15.14. Snapshots of the network showing alive particles when executing VTRP at
different time instances (n = 5000).
472
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
400
Alive Particles
350
300
250
200
150
100
50
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
Distance from Sink (section)
t=1
t=2000
t=4000
t=6000
t=8000
Figure 15.15. Alive particles (hA ) for LTP at different time instances (n = 5000).
In the last set of experiments we evaluate the performance of the four different
functions for varying the transmission range of the particles when phase 3 is activated.
We use a similar setting as in the previous experiments; that is, the field size is 2000 m
by 2000 m, and we deploy n = 5000 sensors and generate 9000 events. The result of
this set of experiments are shown in Figures 15.17–15.22.
400
Alive Particles
350
300
250
200
150
100
50
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
Distance from Sink (section)
t=1
t=2000
t=4000
t=6000
t=8000
Figure 15.16. Alive particles (hA ) for VTRP at different time instances (n = 5000).
VTRP: THE VARIABLE TRANSMISSION RANGE PROTOCOL
473
1
0.9
0.8
Success Rate
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
1000 2000 3000 4000 5000 6000 7000 8000 9000
Number of Events
VTRPc
VTRPm
VTRPp
VTRPr
Figure 15.17. Success rate (IPs ) for the VTRP variations for multiple events (n = 5000).
The results indicate that the constant progress seems to be the least efficient function regarding the success rate metric (Figure 15.17), while for the other three functions the achieved success rate seems to be at similar levels. In fact, this is also the
case for the total energy consumption (Figure 15.18). The constant progress function
seems to be the most conservative; however; as in the case of LTP, it actually implies
that VTRPC just fails to reach the sink.
5000
4500
Total Energy (J)
4000
3500
3000
2500
2000
1500
1000
500
0
0
1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Simulation Time (rounds)
VTRPc
VTRPm
VTRPp
VTRPr
Figure 15.18. Total energy (Etot ) for the VTRP variations for multiple events (n = 5000).
474
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
350
Alive Particles
300
250
200
150
100
50
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
Distance from Sink (section)
t=1
t=2000
t=4000
t=6000
t=8000
Figure 15.19. Alive particles (hA ) for VTRPc at different time instances (n = 5000).
A possible explanation to this behavior of VTRPC is the way the protocol modifies
the transmission range by making small, constant steps. At the early stages of the
network’s operation, when only a small number of particles have “died,” these small
steps suffice to reach the sink. However, as the distance of the closest still-active
350
Alive Particles
300
250
200
150
100
50
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
Distance from Sink (section)
t=1
t=2000
t=4000
t=6000
t=8000
Figure 15.20. Alive particles (hA ) for VTRPm at different time instances (n = 5000).
VTRP: THE VARIABLE TRANSMISSION RANGE PROTOCOL
475
350
Alive Particles
300
250
200
150
100
50
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
Distance from Sink (section)
t=1
t=2000
t=4000
t=6000
t=8000
Figure 15.21. Alive particles (hA ) for VTRPp at different time instances (n = 5000).
particle to S increases (see Figure 15.14), the strategy of making small steps becomes
inefficient. The series of small increments in the transmission range and failed searches
waste the power sources of the particles and eventually cause the “death” of the particle
before the information reaches the sink.
350
Alive Particles
300
250
200
150
100
50
0
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32
Distance from Sink (section)
t=1
t=2000
t=4000
t=6000
t=8000
Figure 15.22. Alive particles (hA ) for VTRPr at different time instances (n = 5000).
476
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
15.7 SOME RECENT RELEVANT WORK
Nikoletseas et al. [21] propose extended versions of two data propagation protocols: the Sleep–Awake Probabilistic Forwarding Protocol ( SW-PFR) and the Hierarchical Threshold-Sensitive Energy-Efficient Network Protocol (HTEEN). These
nontrivial extensions aim at improving the performance of the original protocols by
(a) introducing sleep–awake periods in the PFR protocol to save energy and (b)
introducing a hierarchy of clustering in the TEEN protocol to better cope with large
networks areas. The authors have implemented the two protocols and performed an
extensive experimental comparison (using simulation) of various important measures
of their performance with a focus on energy consumption. They investigate in detail
the relative advantages and disadvantages of each protocol and discuss and explain
their behavior. As a result of the above, they propose and discuss a possible hybrid
combination of the two protocols toward optimizing certain goals.
Recently, Boukerche et al. [22] proposed a novel and efficient energy-aware distributed heuristic, which they refer to as EAD, to build a special rooted broadcast tree
with many leaves that is used to facilitate data-centric routing in wireless microsensor
networks. The EAD algorithm makes no assumption on local network topology and
is based on residual power. It makes use of a neighboring broadcast scheduling and
distributed competition among neighboring nodess.
EAD basically computes a tree with many leaves. With the transceivers of all
leaf nodes being turned off, the network lifetime can be greatly extended. Boukerche
et al. [22] implement an EAD scheme and present an extensive simulation experiments
to study the its performance. The experimental results indicate clearly that the EAD
scheme outperforms previous schemes, such as LEACH among other protocols.
ACKNOWLEDGMENTS
The work of Azzedine Boukerche has been supported by the Canada Research Chair
Program, Canada Foundation Innovation Grant and Ontario Distinguished Researcher
Award #201722.
The work of Sotiris Nikoletseas has been supported by the IST/FET programme
of the European Union under contract numbers IST-1999-14186 (ALCOM-FT), IST2001-33116 (FLAGS) and IST-2001-33135 (CRESCCO).
BIBLIOGRAPHY
1. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. Wireless sensor networks:
A survey. Journal of Computer Networks, 38:393–422, 2002.
2. D. Estrin, R. Govindan, J. Heidemann, and S. Kumar. Next century challenges: Scalable
coordination in sensor networks. In Proceedings of the 5th ACM/IEEE International Conference on Mobile Computing—MOBICOM, 1999.
BIBLIOGRAPHY
477
3. J. M. Kahn, R. H. Katz, and K. S. J. Pister. Next century challenges: Mobile networking
for smart dust. In Proceedings of the 5th ACM/IEEE International Conference on Mobile
Computing, September 1999, pp. 271–278.
4. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cayirci. A survey on sensor networks.
IEEE Communications Magazine, pp. 102–114, August 2002.
5. I. Chatzigiannakis, S. Nikoletseas, and P. Spirakis. Smart dust protocols for local detection and propagation. Distinguished paper. In Proceedings of the 2nd ACM Workshop on
Principles of Mobile Computing—POMC ’2002, pp. 44–51.
6. P. Triantafilloy, N. Ntarmos, S. Nikoletseas, and P. Spirakis. NanoPeer networks and P2P
worlds. In Proceedings of the 3rd IEEE International Conference on Peer-to-Peer Computing, 2003. IEEE CD-ROM.
7. S. M. Ross. Stochastic Processes, 2nd edition, John Wiley & Sons, New York, 1995.
8. I. Chatzigiannakis, T. Dimitriou, M. Mavronicolas, S. Nikoletseas, and P. Spirakis. A
comparative study of protocols for efficient data propagation in smart dust networks. In
Proceedings of the International Conference on Parallel and Distributed Computing —
EUPOPAR 2003, pp. 436–440.
9. I. Chatzigiannakis, T. Dimitriou, S. Nikoletseas, and P. Spirakis. A probabilistic algorithm
for efficient and robust data propagation in smart dust networks. In Proceedings of the 5th
European Wireless Conference on Mobile and Wireless Systems Beyond 3G (EW 2004),
2004.
10. C. Intanagonwiwat, R. Govindan, and D. Estrin. Directed diffusion: A scalable and
robust communication paradigm for sensor networks. In Proceedings of the 6th ACM/IEEE
International Conference on Mobile Computing—MOBICOM, 2000.
11. C. Intanagonwiwat, R. Govindan, D. Estrin, J. Heidemann, and F. Silva. Directed diffusion
for wireless sensor networking. Extended version of reference [10].
12. L. Kleinrock. Queueing Systems, Theory, Vol. I, John Wiley & Sons, New York, 1975,
p. 100.
13. W. R. Heinzelman, A. Chandrakasan, and H. Balakrishnan. Energy-efficient communication protocol for wireless microsensor networks. In Proceedings of the 33rd Hawaii
International Conference on System Sciences—HICSS, 2000.
14. C. Schurgers, V. Tsiatsis, S. Ganeriwal, and M. Srivastava. Topology management for
sensor networks: Exploiting latency and density. In Proceedings of MOBICOM, 2002.
15. H. Euthimiou, S. Nikoletseas, and J. Rolim. Energy balanced data propagation in wireless
sensor networks. In Proceedings of the 4th International Workshop on Algorithms for
Wireless, Mobile, Ad-Hoc and Sensor Networks (WMAN ’04), IPDPS, 2004.
16. T. Antoniou, A. Boukerche, I. Chatzigiannakis, G. Mylonas, and S. Nikoletseas. A new
energy efficient and fault-tolerant protocol for data propagation in smart dust networks
using varying transmission range. In Proceedings of the 37th Annual ACM/IEEE Simulation
Symposium (ANSS ’04), 2004.
17. A. Manjeshwar and D. P. Agrawal. TEEN: A routing protocol for enhanced efficiency in
wireless sensor networks. In Proceedings of the 2nd International Workshop on Parallel
and Distributed Computing Issues in Wireless Networks and Mobile Computing, satellite
workshop of 16th Annual International Parallel & Distributed Processing Symposium—
IPDPS, 2002.
18. B. Karp. Geographic routing for wireless networks. Ph.D. dissertation, Harvard University,
Cambridge, MA, 2000.
478
ENERGY-EFFICIENT ALGORITHMS IN WIRELESS SENSOR NETWORKS
19. S.E.A. Hollar. COTS dust. M.Sc. Thesis in Engineering, Department of Mechanical Engineering, University of California, Berkeley, 2000.
20. K. Mehlhorn and S. N´’aher. LEDA: A Platform for Combinatorial and Geometric Computing, Cambridge University Press, New York, 1999.
21. S. Nikoletseas, I. Chatzigiannakis, H. Euthimiou, A. Kinalis, A. Antoniou, and
G. Mylonas. Energy efficient protocols for sensing multiple events in smart dust networks.
In Proceedings of the 37th Annual ACM/IEEE Simulation Symposium (ANSS ’04), 2004.
22. A. Boukerche, X. Cheng, and J. Linus. Energy-aware data-centric routing in microsensor
networks. In Proceedings of ACM Modeling, Analysis and Simulation of Wireless and
Mobile Systems, September 2003, pp. 42–49.
23. I. Chatzigiannakis and S. Nikoletseas. A sleep–awake protocol for information propagation in smart dust networks. In Proceedings of the 3rd Workshop on Mobile and Ad-Hoc
Networks (WMAN)—IPDPS Workshops, IEEE Press, New York, 2003, p. 225.
24. B. Chen, K. Jamieson, H. Balakrishnan, and R. Morris: SPAN: An energy efficient coordination algorithm for topology maintenance in ad-hoc wireless networks. In Proceedings
of the ACM/IEEE International Conference on Mobile Computing—MOBICOM, 2001.
25. W. R. Heinzelman, J. Kulik, and H. Balakrishnan. Adaptive protocols for information dissemination in wireless sensor networks. In Proceedings of the 5th ACM/IEEE International
Conference on Mobile Computing—MOBICOM, 1999.
26. C. Intanagonwiwat, D. Estrin, R. Govindan, and J. Heidemann. Impact of network density on data aggregation in wireless sensor networks. Technical Report 01-750, Computer
Science Department, University of Southern California, November 2001.
27. µ-Adaptive Multi-domain Power Aware Sensors: http://www-mtl.mit.edu/
research/icsystems/uamps, April, 2001.
28. C. E. Perkins: Ad Hoc Networking. Addison-Wesley, Boston, 2001.
29. TinyOS: A Component-Based OS for the Network Sensor Regime. http://webs.cs.
berkeley.edu/tos/, October, 2002.
30. W. Ye, J. Heidemann, and D. Estrin. An energy-efficient MAC protocol for wireless sensor networks. In Proceedings of the 12th IEEE International Conference on Computer
Networks—INFOCOM, 2002.
31. Wireless Integrated Sensor Networks: http:/www.janet.ucla.edu/WINS/, April, 2001.
32. Y. Xu, J. Heidemann, and D. Estrin. Geography-informed energy conservation for adhoc routing. In Proceedings of the ACM/IEEE International Conference on Mobile
Computing—MOBICOM, 2001.
CHAPTER 16
Security Issues and
Countermeasures in Wireless
Sensor Networks
TANVEER ZIA and ALBERT Y. ZOMAYA
School of Information Technologies, The University of Sydney, Sydney, NSW 2006, Australia
16.1 INTRODUCTION
Sensor networks pose unique security challenges because of their inherent limitations
in communication and computing. The deployment nature of sensor networks makes
them more vulnerable to various attacks. Sensor networks are deployed in applications where they have physical interactions with the environment, people, and other
objects, making them more vulnerable to security threats. We envision that sensor networks would be deployed in mission critical applications such as battlefield, security
of key land marks, building and bridges, measuring traffic flow, habitat monitoring,
and farming. Inherent limitations of sensor networks can be categorized as node and
network limitations. The privacy and security issues in sensor networks raise rich
research questions. Dense deployment of sensor networks in an unattended environment makes sensor nodes vulnerable to potential attacks. Attackers can capture the
sensor nodes and compromise the network to accept malicious nodes as legitimate
nodes. Once within the network, attackers can wage a variety of attacks. We expect
Moore’s law to be applied to drive down the cost of sensor nodes instead of improving
its resources and performance.
Hardware and software improvements will address these issues to some extent,
but complete secure sensor networks require deployment of countermeasures such as
secure key management, secure routing, and lightweight encryption techniques. This
chapter provides an overview of security issues known so far in sensor networks,
along with the countermeasures against these issues followed by key management
schemes and secure routing protocols. The key management and routing protocols
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
479
480
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
discussed in this chapter are the resilient solutions toward the attacks identified so
far; however, research toward a complete secure sensor network is still in its infancy
stages.
16.2 LIMITATIONS IN SENSOR NETWORKS
The following sections list the inherent limitations in sensor networks which make
the design of security procedures more complicated.
16.2.1 Node Limitations
A typical sensor node processor is of 4–8 MHz, having 4 Kbyte of RAM, 128 Kbyte
flash and ideally 916 MHz of radio frequency. Heterogeneous nature of sensor nodes
is an additional limitation that prevents one security solution. Due to the deployment
nature, sensor nodes would be deployed in environments where they would be highly
prone to physical vandalism.
16.2.2 Network Limitations
Besides node limitations, sensor networks bring all the limitations of a mobile ad hoc
network where they lack physical infrastructure, and they rely on insecure wireless
media.
16.2.3 Physical Limitations
Sensor networks’ deployment nature in public and hostile environments in many
applications makes them highly vulnerable to capture and vandalism. Security sensor
nodes with tamper-proof material increases the node cost.
16.3 SENSOR NETWORKS AND MANETS
Wireless sensor networks characteristics are common with Mobile Ad-hoc Networks
(MANETs). Both have limitations in terms of memory, power, and computational
capabilities, and they rely on wireless communication via radio frequency (RF). Both
include data collection and data aggregation to make it available for processing. In
both networks, some nodes are statically deployed while most have a high level
of mobility after deployment. However, sensor networks differ from MANETs in
many areas: Sensor networks are very densely deployed (thousands as compared to
hundreds in MANETs), they are dynamic [1, 2] (i.e., they allow addition or deletion
of nodes after deployment to extend the network or to eliminate failed nodes without
physical contact), and they are highly susceptible to capture and manipulation by an
adversary.
SECURITY IN SENSOR NETWORKS
481
16.4 SECURITY IN SENSOR NETWORKS
Security goals in sensor networks depend on the need to know what we are going to
protect. We determine four security goals in sensor networks: confidentiality, integrity,
authentication, and availability (CIAA).
r Confidentiality. This is the ability to conceal a message from a passive attacker,
where the message communicated on sensor networks remains confidential.
r Integrity. This is the ability to confirm that the message has not been tampered
with, altered, or changed while it was on the network.
r Authentication. If the messages are from the node it claims to be from, we need
to determine the reliability of the message’s origin.
r Availability. If a node has the ability to use the resources, the network is available
for the messages to move on.
16.4.1 Security Classes
Pfleeger [3] has identified four classes of security in computing systems. We integrate
these four threat classes in sensor networks. In computing systems the major assets
are hardware, software, and data. While in sensor networks, our goal is to protect the
network itself, the nodes, and communication among the sensor nodes. There are four
classes of threats which exploit the vulnerability of our security goals. Figure 16.1
shows these four threat classes:
1. In an interruption, a communication link in sensor networks becomes lost or unavailable. Examples of this sort of threat are node capture, message corruption,
addition of malicious code, and so on.
2. An interception means a sensor net has been compromised by an adversary
where the attacker gains unauthorized access to a sensor node or to data in it.
An example of this type of attack is node capture.
3. Modification means that an unauthorized party not only accesses the data but
tampers with it—for example, modifying the data packets being transmitted,
causing a denial of service attack; flooding the network with bogus data; and
so on.
4. In fabrication, an adversary adds false data, making the whole network unreliable.
16.4.2 Security Threats in Sensor Networks
Having built a foundation of security threats in computing, the following sections list
the possible security threats in sensor networks identified by Undercoffer et al. [4].
Passive Information Gathering. An adversary with powerful resources collecting information from sensor networks if information is not encrypted.
482
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
Interruption
Modification
Interception
Fabrication
Figure 16.1. Pfleeger’s four classes of systems security threats.
Node Subversion. Capture of a node may reveal its information including disclosure of cryptographic keys, hence compromising the whole sensor net.
False Node. Addition of a malicious node by an adversary to inject the malicious
data, false node would be computationally robust to lure other nodes to send data
to it.
Node Malfunction. A malfunctioning node will generate inaccurate data that
would jeopardize the integrity of a sensor net, especially when that node is a data
aggregating node—for example, a cluster leader.
Node Outage. What happens when a cluster leader stops functioning? Sensor net
protocols should be robust enough to mitigate the effects of node outages by providing
an alternate route.
Message Corruption. When contents of a message are modified by an attacker, it
compromises the message integrity.
Traffic Analysis. Even if the message transfer is encrypted in sensor networks, its
still leaves the high probability of analysis of communication patterns and sensor activities revealing enough information to enable an adversary to cause more malicious
harm to sensor networks.
16.4.3 Security Attacks in Sensor Networks
Chris Karlof et al. [5] have presented detailed attacks in sensor networks which are
described in the following section. Table 16.1 lists these attacks.
SECURITY IN SENSOR NETWORKS
483
TABLE 16.1. Security Attacks in Wireless Sensor Networks
Spoofed, altered, or
replayed routing
information
Selective forwarding
Sinkhole attacks
Sybil attacks
Wormhole attacks
Hello floods
Create routing loop, attract or repel network traffic, extend or
shorten source routes, generate false error messages, etc.
Either in-path or beneath path by deliberate jamming, allows
to control which information is forwarded. A malicious
node acts like a black hole and refuses to forward every
packet it receives.
Attracting traffic to a specific node—for example, to prepare
selective forwarding.
A single node presents multiple identities, allows us to reduce
the effectiveness of fault-tolerant schemes such as
distributed storage and multipath.
Tunneling of messages over alternative low-latency links to
confuse the routing protocol, thereby creating sinkholes,
etc.
An attacker sends or replays a routing protocols hello packets
with more energy.
Routing Loops. In sensor networks, routing loops attack the information exchanged between nodes. False error messages are generated when an attacker alters
and replays the routing information. Routing loops attract or repel the network traffic
and increase node-to-node latency.
Selective Forwarding. Selective forwarding is a way to influence the network
traffic by believing that all the participating nodes in network are reliable to forward
the message. In selective forwarding attack, malicious nodes simply drop certain
messages instead of forwarding every message. Once a malicious node cherry picks
on the messages, it reduces the latency and deceives the neighboring nodes that they
are on a shorter route. Effectiveness of this attack depends on two factors. First is the
location of the malicious node. the closer it is to the base station, the more traffic it will
attract. Second is the percentage of messages it drops. When a selective forwarder
drops more messages and forwards less, it retains its energy level, thus remaining
powerful to trick the neighboring nodes.
Sinkhole Attacks. In sinkhole attacks, an adversary attracts the traffic to a compromised node. The simplest way of creating a sinkhole is to place a malicious node
where it can attract most of the traffic, possibly closer to the base station or malicious
node itself, which deceptively acts as a base station. One reason for sinkhole attacks
is to make selective forwarding possible to attract the traffic toward a compromised
node. The nature of sensor networks where all the traffic flows toward one base station
makes this type of attack more susceptible.
484
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
Sybil Attacks. This is, a type of attack where a node creates multiple illegitimate
identities in sensor networks either by fabricating or stealing the identities of legitimate
nodes. Sybil attacks can be used against routing algorithms and topology maintenance;
it reduces the effectiveness of fault-tolerant schemes such as distributed storage and
dispersity. Another malicious factor is geographic routing where a Sybil node can
appear at more than one place simultaneously.
Wormholes. In wormhole attacks, an adversary positioned closer to the base station
can completely disrupt the traffic by tunneling messages over a low-latency link. Here
an adversary convinces the nodes which are multihop away that they are closer to the
base station. This creates a sinkhole because an adversary on the other side of the
sinkhole provides a better route to the base station.
Hello Flood Attacks. This involves broadcasting a message with stronger transmission power and pretending that the HELLO message is coming from the base
station. Message receiving nodes assume that the HELLO message sending node is
the closest one and they try to send all their messages through this node. In this type
of attack, all nodes will be responding to HELLO floods and wasting the energies.
The real base station will also be broadcasting the similar messages but will have only
a few nodes responding to it.
DoS Attacks. Denial of service attacks occur at a physical level, causing radio
jamming, interfering with the network protocol, battery exhaustion, and so on.
16.4.4 Layering-Based Security Approach
Application Layer. Data are collected and managed at application layer therefore
it is important to ensure the reliability of data. Wagner [6] has presented a resilient aggregation scheme that is applicable to a cluster-based network where a cluster leader
acts as an aggregator in sensor networks. However, this technique is applicable if
the aggregating node is in the range with all the source nodes and there is no intervening aggregator between the aggregator and source nodes. In the hierarchical
clustering approach, a communication channel between the aggregator and base station has potentially limited bandwidth because the cluster leader as an aggregator
itself is a sensor node [6, 7]. To prove the validity of the aggregation, cluster leaders
use the cryptographic techniques to ensure the data reliability. We will discuss the
cryptography in key management section.
Network Layer. The network layer is responsible for routing of messages from
node to node, node to cluster leader, cluster leaders to cluster leaders, cluster leaders
to the base station, and vice versa.
Routing protocols in sensor networks are of two types: (1) ID-based protocols, in
which packets are routed to the destination based on the IDs specified in the packets,
SECURITY IN SENSOR NETWORKS
485
and (2) data-centric protocols [8] in which packets contain attributes that specify the
type of data being provided. Law and Havinga [7] have described Karlof and Wagner
[5] routing attacks in sensor networks as below:
1. Packets are dropped completely, or selectively.
2. The network is flooded with global broadcasts.
3. Some sensor nodes in the network are misguided into believing that nodes are
either multiple hops away or do not exist at all in the neighbors.
4. A significant proportion of the traffic is tunneled from one place in the network
to another distant place of the network depriving other parts of the network that
under normal circumstances would have received the traffic themselves.
5. Sometimes traffic is lured to a particular node or a small group of nodes, depriving other parts of the network that normally would have received the traffic
themselves.
Security of routing protocols depends on the location of nodes and the encryption
techniques.
Data Link Layer. The data link layer does the error detection and correction, as
well as encoding of data. The link layer is vulnerable to jamming and DoS attacks.
TinySec [9] has introduced link layer encryption, which depends on a key management
scheme. However, an attacker having better energy efficiency can still wage an attack.
Protocols like LMAC [10] have better anti-jamming properties, which are viable
countermeasures at this layer.
Physical Layer. The physical layer emphasizes the transmission media between
sending and receiving nodes; the data rate, signal strength, and frequency types are
also addressed in this layer. Ideally, the FHSS frequency hopping spread spectrum is
used in sensor networks.
Table 16.2 summarizes the attacks and countermeasures in a layering model in
sensor networks.
TABLE 16.2. Layering Approach in Sensor Network Attacks and Countermeasures
Attack Types
Application layer
Network layer
Data link layer
Physical layer
Subversion and malicious
nodes
Wormholes, sinkholes, Sybil,
routing loops
Link layer jamming
DoS attacks, Radio jamming,
node capture
Countermeasures
Malicious node detection and
isolation
Key management, secure
routing
Link layer encryption
Adaptive antennas, spread
spectrum
486
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
16.5 CRYPTOGRAPHY IN SENSOR NETWORKS
To achieve security in any communication model, it is important to encrypt messages
among the sensor nodes on an agreed key management scheme. In traditional networks
we use the complex key management schemes such as Public and Diffie–Hellman
[11] keying schemes. However, providing a secure key management scheme in sensor
networks is difficult due to ad hoc, dynamic topology and resource limitations.
The general key distribution problem refers to the task of distributing secret keys
between communicating parties to provide security properties such as secrecy and
authentication.
In sensor networks, key distribution is usually combined with initial communication establishment to bootstrap a secure communication infrastructure from a collection of deployed sensor nodes.
To bootstrap the security in sensor networks, nodes must be able to establish a
secure node-to-node communication such as the following:
r Additional legitimate nodes deployed at a later time should be able to form secure
connections with already-deployed nodes.
r Unauthorized nodes should not be able to gain entry into the network, either
through packet injection or masquerading as legitimate node.
r The scheme must work without prior knowledge of which nodes will come into
the communication range of each other after deployment.
r The computational and storage requirement of the key management scheme must
be low, and the scheme should be robust to DoS attacks from out-of-network
source.
Figure16.2 shows a keying process to establish secure communication between
node A and node B.
16.5.1 Asymmetric Cryptography
Asymmetric cryptography, also known as public key, is the most commonly used
keying method in computing. However, the following three factors make public
keying not feasible for sensor networks:
Figure 16.2. Secure channel between nodes A and B.
KEY MANAGEMENT SCHEMES
487
First, asymmetric cryptography requires involves extensive hardware and software and mathematical functions that are beyond the processing power of tiny sensor
nodes. In order to deploy asymmetric cryptography on sensor nodes, it is necessary to
have dedicated cryptographic hardware that would either increase the nodes cost or
embed mathematical functions in software, which would be much slower than using
symmetric keying. Second, asymmetric cryptography involves intensive computation
that may take up to minutes of processing for sensor node to complete one signature generation, making nodes vulnerable to battery exhaustion and denial of service
attacks where they are continuously requested to generate signatures. Third, there is
no resistance against node captures.
16.5.2 Symmetric Cryptography
In symmetric cryptography, every node in a sensor network shares a unique
symmetric key with every other node in the network. According to Chan and Perrig
[12], using symmetric cryptography will have n nodes in a sensor network. Every node
stores n – 1 keys, one of each of the other nodes in the network. After deployment,
nodes must perform key discovery to verify identity of the node that they are communicating with. Symmetric keying in sensor networks has two benefits. First, any
node captured reveals no information about the communication, providing stronger
resilience against node capture attack. Second, due to the broadcast nature of sensor
networks, compromised keys can be revoked. For example, a captured node’s entire
n – 1 keys can be broadcasted, and a node hearing this broadcast will stop using the
keys.
16.6 KEY MANAGEMENT SCHEMES
This section discusses various key management schemes introduced in sensor networks, along with the secure triple key management scheme we have introduced.
16.6.1 Master-Key-Based Key Predistribution Scheme
A Master-key-based predistribution scheme is based on a single master key that is predeployed in all the sensor nodes. Each sensor node uses part of the memory to store the
master keys. A pair of sensor nodes exchanges random nonce values. The pair uses a
master key to establish the session keys. Compromise of master key makes the whole
sensor net insecure. Dutertre et al. [13] have presented a lightweight key management
system where they have proposed more than one master key on an assumption that
sensor networks are deployed in various phases. Each sensor node stores a group
authentication key bk1 and a key generation key bk2. If two sensors S1 and S2 are
from the same phase, they authenticate each other by using the authentication keys
bk1. They exchange random nonce values RN1 and RN2 and establish the session key.
If the nodes are from two different phases, a sensor node from an old phase stores a
random nonce and a secret key for each new phase and authenticates itself with that
488
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
secret key. Once authenticated, both nodes generate a pairwise key to establish the
secure communication.
16.6.2 Basic Random Key Pre-distribution Scheme
A probabilistic key pre-distribution scheme is presented by Eschenauer and Gligor [1]
where each sensor node receives a random subset of keys from a large key pool before
deployment. To agree on a key for communication, two nodes find one common key
within their subsets and use that key as their shared key.
The random key pre-distribution scheme denotes m as the number of distinct
cryptographic keys that can be stored in the key ring of a sensor node. In this scheme,
sensor nodes pick a random pool of keys Q out of the total possible key space in an
initialization phase. Nodes store m number of keys randomly selected from the pool
Q. This set of m keys is called a node’s key ring. The number of keys in the key pool
Q is chosen such that two random subsets of m in Q will share at least one key with
some probability p. In this scheme all nodes use the same key pool Q. That means
security of the network would be weak if keys from pool Q are compromised. Basic
random key pre-distribution contains five offline steps:
i. Generation of a large pool of P keys and their key identifiers
ii. Random drawing of k keys out of P without replacement to establish the key
ring of a sensor
iii. Loading of the key ring into the memory of each sensor
iv. Saving of the key identifiers of a key ring and associated sensor identifier on
a trusted controller node
v. For each node, loading the ith controller node with the key shared with that
node
The key pre-distribution phase ensures that only a small number of keys need to
be placed on each sensor node’s key ring, ensuring that any two nodes share at least
one key with a chosen probability.
In the shared-key discovery phase, every node discovers its neighbors in wireless
communication range with which it shares the key. Each node broadcasts in plain text
the list of identifiers of the keys on their key ring. Shared-key discovery takes place
during sensor network initialization, and it establishes the topology of the sensor array
by the routing layer of the sensor network. A link exists between two sensor nodes
only if they share a key; and if a link exists between two nodes, all communication
on that link is secured by link encryption.
During the path-key establishment phase a path key is assigned to a set of selected
pairs of sensor nodes in wireless communication range that do not share a key but are
connected by two or more links at the end of the shared-key discovery phase.
Revocation is used to eliminate the key ring of nodes which have been compromised. To execute revocation, a controller node (which is comparatively more
powerful and mobile in terms of range) broadcasts a single revocation message
KEY MANAGEMENT SCHEMES
489
containing a signed list of K key identifiers for the key ring to be revoked. To sign
the list of key identifiers, the controller generates a signature key Ke and unicasts it
to each node by encrypting it with a key Kci . After obtaining the signature key, each
node verifies the signature of the signed list of key identifiers, locates those identifiers
in its key ring, and removes the corresponding keys. Once the keys are removed from
key rings, some links may disappear and the affected nodes need to reconfigure those
links by restarting the shared-key discovery and possibly path-key establishment.
Re-keying is described as self-revocation of a key by a node if its life is expired.
After removal of the expired key, affected nodes restart the shared-key discovery and
path-key establishment process.
Resiliency to Sensor Node Capture. There are two levels of threats posed by
the node capture. The first level of threat is manipulation of sensor’s data when
an adversary injects the fake data in sensor network. Detecting this nature of attack
requires offline data correlation analysis and data anomaly detection by collection and
processing nodes. The second level of threat occurs when physically a sensor node
is in control of the adversary. This level of active threat includes data manipulation
of captured node and other nodes on the network. Countermeasures to node capture
threats are tamper-detection technologies [14] to shield the sensors in such a way that
vandalism of nodes causes the erasure of a sensor’s key ring, eventually disabling the
sensor’s operation.
16.6.3 Extended Random Key Pre-distribution Scheme
Chan et al. [15] extended the idea of Eschenauer and Gligor [1] and developed three
key pre-distribution schemes; q-composite, multipath reinforcement, and randompairwise keys schemes to over come the weakness of basic random key predistribution scheme.
In the q-composite scheme, q common keys (q > 1) are needed, instead of just one.
In this scheme the key pool size S is reduced and multiple keys are used to establish
communications instead of just one.
To compute the key pool size, let p(i) be the probability that any two nodes have
exactly i keys in common. Let Pconnect be the probability of any two nodes sharing sufficient keys from a secure connection. P connect = 1−(probability that the two nodes
share insufficient keys to form a connection), hence P connect = 1 − (p(0) + p(1)
+ . . . + p(q − 1)).
For a given key ring size m, minimum key overlap q, and minimum connection
probability p, the largest |S| is chosen such that P connect ≥ p.
To initialize the key setup, a set of S of random keys out of the total key space is
picked. For each node, m random keys are selected from S (where m is the number of
keys each node can carry in its key ring) and store them into the node’s key ring. In
the key setup phase, each node discovers all common keys it possesses with each of
its neighbors.
q-composite key scheme strengthens the network resilience against node capture
when the number of nodes captured is small. However, if the large number of nodes
490
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
have been captured, the q-composite keys scheme tends to reveal a larger fraction of
network to the adversary.
In the multipath key reinforcement scheme, Chan et al. [15] present a method
to strengthen the security of an established link key by establishing the link key
through multiple paths. Assuming that key setup has been completed, there are now
many secure links formed through the common keys in the various nodes key rings.
Suppose node A has a secure link to node B after key setup. This link is secured
using a single key k from the key pool S. k may be residing in the key ring memory
of some other nodes elsewhere in the network. If any of those nodes are captured,
the security of the link between nodes A and B is jeopardized. To address this, the
communication key is updated to a random value after key setup. However, key update cannot be coordinated using the direct link between nodes A and B since if
the adversary has been recording all key-setup traffic, it could decrypt the key update message after it obtained k and still obtain the new communication key. In this
approach, key update is coordinated over multiple independent paths. Assume that
enough routing information can be exchanged such that node A knows all disjoint
paths to node B created during initial key setup that are h hops or less. The more paths
between two nodes A and B, the more security multipath key reinforcement provides
from the link between A and B. However, for any given path, the probability that
the adversary can eavesdrop on the path increases with the length of the path since
if any one link on the path is insecure, then the entire path is made insecure. Furthermore, it is increasingly expensive in terms of communication overhead to find
multiple disjoint paths that are very long. To address this issue, a 2-hop multipath
key reinforcement scheme is presented where paths of only 2 links are considered.
The effectiveness of 2-hop multipath key reinforcement is evaluated by simulating
the random deployment of 10,000 sensor nodes on a square planar field. Successfully implementing multipath key reinforcement on the key management scheme
of Eschenauer and Gligor [1] enables it to outperform the q-composite scheme for
q ≥ 2 even when the q-composite scheme is supplemented by key reinforcement.
However, compounding both schemes compounds their weaknesses: The smaller key
pool size of the q-composite keys scheme undermines the effectiveness of multipath
key reinforcement by making it easier to build up a critically larger collection of
keys.
The third mechanism introduced in this scheme is a random pairwise scheme where
node-to-node authentication is established. The random pairwise scheme introduces
the following properties:
r
r
r
r
r
Resilience against node capture
Node-to-node identity authentication
Distributed node revocation
Resistence to node replication and generation
Comparable scalability
In random key pool distribution schemes like q-composite and multipath, keys
can be issued multiple times out of the key pool, and the node-to-node authentication
KEY MANAGEMENT SCHEMES
491
is not possible. Pairwise key distribution assigns a unique key to each pair of
nodes.
Recall that the size of each node’s key rings is m key, and the probability of any two
nodes being able to communicate securely is p. The random pairwise keys scheme
proceeds as follows:
i. In the pre-deployment initialization phase, a total of (n = m/p) unique node
identities are generated. The actual size of the network may be smaller than n.
Unused node identities will be used if additional nodes are added to the network
in the future. Each node identity is matched up with m other randomly selected
distinct node IDs, and a pairwise key is generated for each pair of nodes. The
key is stored in both nodes key rings, along with the ID of the other node that
also knows the key.
ii. In the post-deployment key-setup phase, each node first broadcasts its node ID
to its intermediate neighbors. By searching for each other’s IDs in their key
rings, the neighboring nodes can tell if they share a common pairwise key for
communication.
16.6.4 Multiple Space Key Pre-distribution Scheme
Du et al. [16, 17] propose a key pre-distribution scheme that is based on Blom [18]
key pre-distribution method introduced in 1985. Blom’s key pre-distribution method
was not for sensor networks; however, it allows any pair of nodes in a network to be
able to find a pairwise secret key. As long as no more than X nodes are compromised,
the network is perfectly secure (this is called the X-secure property).
In this scheme, the concept is taken from graph theory where we draw an edge
between two nodes if and only if they can find a secret key between themselves. The
hypothesis here is that by requiring the graph to be only connected, each sensor node
needs to carry less key information.
During the key pre-distribution phase, key information is assigned to each node,
such that after deployment, neighboring sensor nodes can find a secret key between
them. In Du et al. [16] the key pre-distribution phase generates G and D matrices
followed by selection of a key space; then in the key agreement phase, after deployment, each node discovers whether it shares any key space with its neighbors. To
achieve this, each node broadcasts a message containing the node’s ID, the indices
of key spaces it carries, and the seed of the column of G it carries. In the previous
two schemes, an adversary only needs to compromise less than 100 nodes in order
to compromise 10% of the rest of the secure links, whereas in this scheme the adversary needs to compromise 500 nodes, thus lowering the payoff to the adversary of
smaller-scale network breaches.
16.6.5 Key Management Scheme Using Deployment Knowledge
Du et al. [19] make an extension on the key management scheme developed by
Eschenauer and Gligor [1]. Their scheme describes the deployment knowledge where
492
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
sensors are prearranged in a sequence before deployment. Once dropped from a plane
or helicopter, sensors have better probability to recognize the prearranged sequence of
neighbor sensors. This prior deployment knowledge is useful for key pre-distribution.
When neighbor sensors are known, key pre-distribution becomes trivial; for each node
n we just need to generate a pairwise key between n and each of its neighboring nodes,
and we save these keys in n’s memory. This guarantees that each node can establish
a secure channel with each of its neighbors after deployment.
Deployment knowledge in this scheme is modeled using probability density functions (pdf). When the pdf is uniform, no information can be gained regarding where a
node is more likely to reside. Du et al. [19] focus on nonuniform pdf functions. Since
the distribution is different from uniform distribution, we can assume that a sensor is
more likely to be deployed in certain areas.
16.6.6 Trusted Base Station as Key Distribution Center
A base station is used as a trusted third party to distribute the keys to provide link
keys to sensor nodes such as Kerberos [20]. In this keying mechanism a base station
becomes the single point of vulnerability if compromised. Due to extensive communication with the base station, nodes closer to the base station will lose their battery
early. This approach assumes some level of reliable communication mode between
the node and the base station before any key scheme is deployed.
16.6.7 A Secure Triple Key Management Scheme
Our secure triple-key management scheme [21] consists of three keys: two predeployed keys in all nodes and one in-network generated cluster key for a cluster
to address the hierarchical nature of sensor network.
Otherwise, the packet is dropped by the cluster leader. The node builds the message
using the fields below:
Kn
MAC
ID
TS
S
message
Level 2
Cluster Leader to Next-Hop Cluster Leader Key Calculation. A cluster
leader aggregates the messages received from its nodes and forwards it to a next-level
cluster leader; or if the cluster leader is one hop closer to the base station, it directly
sends the message to the base station. The receiving cluster leader checks its routing
table and constructs the following packet to be sent to the next-level cluster leader or
base station. The cluster leader adds its own IDCLn , its network, and its cluster key in
incoming packet and rebuilds the packet as follows:
{IDCLn , Kn , [IDsn , Kn, TS, MAC, S(Aggr message)]}
KEY MANAGEMENT SCHEMES
Kn
MAC
ID
TS
S
Aggr message
493
Level 1
Here ID is the ID of the receiving cluster leader which wraps the message and sends it
to the next-hop cluster leader or to the base station if directly connected. The next-hop
cluster leader receives the packet and checks the ID. If the ID embedded in the packet
is the same as it holds, it updates the ID for the next hop and broadcasts it; otherwise
the packet is discarded. Aggr message refers to the message aggregated by the cluster
leader.
Cluster Leader to Base Station Key Calculation. A base station receives the
packet from its directly connected cluster leader; it checks the ID of the sending
cluster leader, and it verifies the authentication and integrity of the packet through
MAC. The cluster leader directly connected with base station adds its own ID along
with the packet received from the sending cluster leader. The packet contains the
following fields:
{IDCL2 [IDCL4 , Kn , [IDs10 , Kn, TS, MAC, S (Aggr message)]]}
Figure 16.3 illustrate the hierarchal structure of our secure triple-key management
scheme where we have shown sensor nodes S1..S11, cluster leaders CL1..CL4, and
base station communication using the triple-key management scheme.
Kn (Network Key). This is generated by the base station, pre-deployed in each
sensor node, and shared by the entire sensor network. Nodes use this key to
encrypt the data and pass onto next hop.
Ks (Sensor Key). This is generated by the base station, pre-deployed in each sensor
node, and shared by the entire sensor network. A base station uses this key to
decrypt and process the data, and a cluster leader uses this key to decrypt the
data and send it to base station.
Kc (Cluster Key). This is generated by the cluster leader, and it is shared by the
nodes in that particular cluster. Nodes from a cluster use this key to decrypt the
data and forward it to the cluster leader.
The secure triple-key management scheme is a much resilient solution against
many of sensor networks attacks. The next section describes how triple keys are used
to send messages from base station to node, nodes to cluster leader, cluster to cluster
leader and base station.
Base Station to Node Key Calculation. The base station uses Kn to encrypt
and broadcast data. When a sensor node receives the message, it decrypts it by using
its Ks . In Figure 16.3, the base station uses Kn1..nn to broadcast the message. This
494
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
BS
{IDCL2[IDCL4, Kn, [IDs10, Kn, TS, MAC, S (Aggr message)]]}
CL1
CL2
CL3
{IDCL4, Kn, [IDs10, Kn, TS, MAC, S (Aggr message)]}
S1
S2
S3
S4
S5
S6
S7
S8
CL4
{IDs10, Kn, TS, MAC, S (message)}
S9
S10
S11
Figure 16.3. A secure triple-key management.
process is as follows: The base station encrypts its own ID, a current timestamp TS,
and its Kn as a private key.
The base station generates a random seed S and assumes itself at level 0. The packet
contains the following fields:
Kn
MAC
ID
TS
S
message
Level 0
The sensor node decrypts the message received from the base station using Ks .
MAC is a message authentication code for a message m.
Nodes to Cluster Leader Key Calculation. When a node sends a message to
cluster leader, it constructs the message as follows
{IDsn , Ks , TS, MAC, S (message)}
SECURE ROUTING
495
The cluster leader checks the ID from the packet, checks if the ID in the packet
matches the ID it holds, and verifies the authentication and integrity of the packet
through MAC.
The cluster key Kc is used for message decryption with in a cluster where a cluster
leader needs to decrypt the message sent by a node and prepare that message to
forward the next hop.
This secure triple-key management scheme provides added resilience toward susceptible attacks on sensor networks by keeping in mind the resource-starved nature
of sensor nodes.
16.7 SECURE ROUTING
Secure routing in sensor networks is challenging due to the unique characteristics
sensor networks have compared to wired and wireless ad hoc networks. Traditional
IP-based routing is not a viable solution due to a relatively large number of sensor
nodes because the overhead of IP maintenance is very high. Following are the issues
that need to be kept in mind while designing a secure routing protocol in sensor
networks.
1. Sensor nodes are self-organizing due to the ad hoc deployment, and nodes are
left unattended after the deployment.
2. In sensor networks, most of the time flow of data would be from nodes to cluster
leader and base station.
3. Careful route management due to nodes limitations.
4. Frequent changes in network topology due to the dynamic nature of sensor
networks.
5. Sensor networks are application-specific and data-centric.
6. Secure location of sensor nodes because Global Posting Systems (GPS) are not
suitable for sensor networks.
There are very few routing protocols proposed to address the secure routing issues
in sensor networks. In the following sections we present some of these solutions
discussed by Saraogi [22].
16.7.1 SPINS: Security Protocols for Sensor Networks
SPINS is a suite of security building blocks proposed by Perig et al. [23]. It is optimized for resource constrained environments and wireless communication. SPINS
has two secure building blocks: SNEP and µTESLA (the micro version of TESLA).
SNEP provides data confidentiality, two-party data authentication, and data freshness. µTESLA provides authenticated broadcast for severely resource-constrained
environments. All cryptographic primitives (i.e., encryption, message authentication
code (MAC), hash, random number generator) are constructed out of a single block
496
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
cipher for code reuse. This, along with the symmetric cryptographic primitives used,
reduces the overhead on the resource constrained sensor network. In a broadcast
medium such as a sensor network, data authentication through a symmetric mechanism cannot be applied because all the receivers know the key. µTESLA constructs
authenticated broadcast from symmetric primitives, but introduces asymmetry with
delayed key disclosure and one-way function key chains.
SNEP. SNEP uses encryption to achieve confidentiality and uses message authentication code (MAC) to achieve two-party authentication and data integrity. Apart
from confidentiality, another important security property is semantic security, which
ensures that an eavesdropper has no information about the plaintext, even if it sees
multiple encryptions of the same plaintext [24]. The basic technique to achieve this
is randomization: Before encrypting the message with a chaining encryption function (i.e., DESCBC), the sender precedes the message with a random bit string (also
called the initialization vector). This prevents the attacker from inferring the plaintext
of encrypted messages if it knows plaintext–ciphertext pairs encrypted with the same
key. To avoid adding the additional transmission overhead of these extra bits, SNEP
uses a shared counter between the sender and the receiver for the block cipher in
counter mode (CTR). The communicating parties share the counter and increment it
after each block. SNEP offers the following unique features:
Semantic Security. Since the counter value is incremented after each message, the
same message is encrypted differently each time. The counter value is long enough
that it never repeats within the lifetime of the node.
Data Authentication. If the MAC verifies correctly, a receiver can be assured that
the message originated from the claimed sender.
Replay Protection. The counter value in the MAC prevents replaying old messages. Note that if the counter were not present in the MAC, an adversary could easily
replay messages.
Data Freshness. If the message verified correctly, a receiver knows that the message must have been sent after the previous message it received correctly (that had a
lower counter value). This enforces a message ordering and yields weak freshness.
Low Communication Overhead. The counter state is kept at each end point and
does not need to be sent in each message.
µTESLA. Most of the proposals for authenticated broadcast are impractical for sensor networks, because they rely on asymmetric digital signatures for the authentication. The TESLA protocol provides efficient authenticated broadcast [25, 26];
SECURE ROUTING
497
however, it is not designed for limited computing environments. µTESLA solves the
following three inadequacies of TESLA in sensor networks:
1. TESLA authenticates the initial packet with a digital signature, which is too
expensive for sensor nodes. µTESLA uses only symmetric mechanisms.
2. Disclosing a key in each packet requires too much energy for sending and
receiving. µTESLA discloses the key once per epoch.
3. Sensor node is unable to store the one-way key chain making TESLA infeasible
for sensor networks.
µTESLA restricts the number of authenticated senders. µTESLA uses symmetric
authentication but introduces asymmetry through a delayed disclosure of the symmetric keys, which results in an efficient broadcast authentication scheme. For the base
station to broadcast authenticated information to the nodes, µTESLA requires that
the base station and nodes are loosely time synchronized, and each node knows an
upper bound on the maximum synchronization error. To send an authenticated packet,
the base station simply computes a MAC on the packet with a key that is secret at
that point in time. When a node gets a packet, it can verify that the corresponding
MAC key was not yet disclosed by the base station (based on its loosely synchronized
clock, its maximum synchronization error, and the time schedule at which keys are
disclosed). Since a receiving node is assured that the MAC key is known only by the
base station, it assures that no adversary could have altered the packet in transit. The
node stores the packet in a buffer. At the time of key disclosure the base station broadcasts the verification key to all receivers. When a node receives the disclosed key, it
can easily verify the correctness of the key. If the key is correct, the node can now
use it to authenticate the packet stored in its buffer. Each MAC key is a key of a key
chain, generated by a public one-way function F. To generate the one-way key chain,
the sender chooses the last key Kn of the chain randomly, and repeatedly applies F
to compute all other keys: Ki = F (Ki + 1). Each node can easily perform time synchronization and retrieve an authenticated key of the key chain for the commitment
in a secure and authenticated manner, using the SNEP building block.
SPINS leaves some questions such as security of compromised nodes, DoS issues,
network traffic analysis issues. SPINS assumes the static network topology ignoring
the ad hoc and mobile nature of sensor nodes.
16.7.2 INSENS: Intrusion-Tolerant Routing in Wireless Sensor
Networks
INSENS [26, 27] tolerates intrusion by bypassing the malicious nodes instead of
detecting the intrusion. Even if a malicious node exists in the network, INSENS
mitigates the impact of that intrusion. INSENS design is based on the following three
principles:
1. To prevent DoS-style attacks, the type of communication is constrained; that
is, individual nodes are not allowed to broadcast. Only the base station does
498
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
broadcasting. Authentication of base station is used via one-way hashes so that
individual nodes cannot spoof the base station and thereby flood the network.
This feature exploits redundancy to tolerate intrusions without any need for
detecting the nodes where intrusions have occurred. INSENS operates correctly
in the presence of undetected intruders.
2. To prevent advertisement of false routing data, control routing information
must be authenticated. Due to localized intrusions, the base station might not
receive the entire topology discovery and control information and will farm
an incorrect network. INSENS performs all complex computation at the base
station and minimizes the overheads from sensor nodes in building routing
tables. INSENS also addresses other issues of sensor nodes limitations such as
memory, computation, and bandwidth.
3. INSENS addresses the resource limitations by using symmetric key cryptography for confidentiality and authentication between base station and nodes. To
address the issues of compromise nodes, redundant multipath routing is built
in INSENS. Even if a node is compromised, a secondary path will exist to forward the packet to the right destination. This feature limits the damage done
by adversaries by limiting flooding and using right authentication mechanisms
such as symmetric key cryptography.
INSENS does not rely on conventional anomaly-based intrusion detection techniques; instead it uses an intrusion-tolerant mechanism that reduces the harm caused
by presence of a small number of undetected intruders in the network by incorporating
redundancy in routing. INSENS is comprised of two phases: (1) a route discovery
phase and (2) a data forwarding phase
The route discovery phase builds the topology of the network and creates forwarding tables on various nodes. Route discovery is further divided into three
rounds:
(a) A Request Message. This is broadcasted by the base station to all the sensor
nodes.
(b) A Feedback Message. Each sensor node sends its neighborhood topology information back to the base station.
(c) Routing Update Message. The base station authenticates the neighborhood
information, constructs the topology of the network, computes the forwarding
table for each sensor node, and sends the tables to each sensor node.
The data forwarding phase enables data forwarding from each sensor node to the
base station.
16.7.3 TinySec: A Link Layer Security Architecture
TinySec [9] is a lightweight, generic security architecture that can be integrated
into sensor network applications. It is incorporated into the official TinyOS release.
SECURE ROUTING
499
Ganesan et al. [8] reason why link layer security is ideal for sensor networks. Sensor
networks use in-network processing such as aggregation and duplicate elimination
[28, 29] to reduce traffic and save energy. Since in-network processing requires the
intermediate nodes to access, modify, and suppress the contents of messages, end-toend security mechanisms between each sensor node and the base station cannot be
used to guarantee the authenticity, integrity, and confidentiality of messages.
End-to-end security mechanisms are also vulnerable to certain denial of service
attacks. If message integrity is only checked at the final destination, the network
may route packets injected by an adversary many hops before they are detected. This
kind of attack will waste energy and bandwidth. Link layer security architecture can
detect unauthorized packets when they are first injected into the network. TinySec
provides the basic security properties of message authentication and integrity (using
MAC), message confidentiality (through encryption), semantic security (through an
initialization vector), and replay protection. TinySec supports two different security
options: authenticated encryption (TinySec-AE) and authentication only (TinySecAuth). With authenticated encryption, TinySec encrypts the data payload and authenticates the packet with a MAC. The MAC is computed over the encrypted data and the
packet header. In authentication-only mode, TinySec authenticates the entire packet
with a MAC, but the data payload is not encrypted.
TinySec uses an 8-byte initialization vector (IV) and cipher block chaining (CBC)
[30]. The structure of the IV is dst||AM||l||src||ctr, where dst is the destination address
of the receiver, AM is the active message (AM) handler type, l is the length of the data
payload, src is the source address of the sender, and ctr is a 16-bit counter. The counter
starts at 0 and the sender increases it by 1 after each message sent. A stream cipher
uses a key K and IV as a seed and stretches it into a large pseudorandom keystream
GK(IV). The keystream is then xored against the message: C=(IV, GK(IV) xor P).
The fastest stream ciphers are faster than the fastest block ciphers, which might make
them look tempting in a resource-constrained environment. However, stream ciphers
have a failure mode: If the same IV is ever used to encrypt two different packets,
then it is often possible to recover both plaintexts. Guaranteeing that IVs are never
re-used requires IVs to be fairly long, say at least 8 bytes. Since an 8-byte overhead in
a 30-byte packet is unacceptable in the resource-constrained sensor network, TinySec
uses block cipher.
Using a block cipher for encryption has an additional advantage. Since the most
efficient message authentication code (MAC) algorithms use a block cipher, the nodes
will need to implement a block cipher in any event. Using this block cipher for
encryption conserves code space. The advantage of using CBC is that it degrades
gracefully in the presence of repeated IVs. If we encrypt two plaintexts P1 and P2
with the same IV under CBC mode, then the cipher texts will leak the length (in blocks)
of the longest shared prefix of P1 and P2 and will leak nothing more. For instance, if
the first block of P1 is different from the first block of P2, as will typically be the case,
then the cryptanalyst learns nothing apart from this fact. CBC mode is provably secure
when IVs do not repeat. However, CBC mode was designed to be used with a random
IV, and it has a separate leakage issue when used with a counter as the IV (note that
the TinySec IV has a 16-bit counter). To fix this issue, TinySec pre-encrypts the IV.
500
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
The developers of TinySec give reasons behind their choice of cipher in reference 9.
Initially they found AES and Triple-DES to be slow for sensor networks. They found
RC5 and Skipjack to be most appropriate for software implementation on embedded
microcontrollers. Although RC5 was slightly faster, it is patented. Also, for good
performance, RC5 requires the key schedule to be pre-computed, which uses 104
extra bytes of RAM per key. Because of these drawbacks, the default block cipher in
TinySec is Skipjack.
TinySec always authenticates messages, but encryption is optional. TinySec uses a
cipher block chaining construction, CBC-MAC, for computing and verifying MACs.
CBC-MAC is efficient and fast, and the fact that it also relies on a block cipher minimizes the number of cryptographic primitives that we must implement in the limited
memory available. However, the standard CBC-MAC construction is not secure for
variably sized messages. Adversaries can forge a MAC for certain messages.
TinySec creates confusion because of its three characteristics, such as no TinySec,
TinySec-Auth, and TinySec-AE. Also, TinySec assumes a message length of 8 bytes
or more, and it does not address the smaller messages. TinySec does not provide a
secure routing mechanism while our framework addresses these issues.
16.8 SUMMARY
Sensor networks have become promising future to many applications. In the absence
of adequate security, deployment of sensor networks is vulnerable to a variety of
attacks. A sensor node’s limitations and nature of wireless communication pose unique
security challenges. Researchers have introduced secure communication using secure
key management and secure routing techniques to address the unique security needs
in sensor networks. Current research in sensor network security is mostly built on
a trusted environment [31]; however, several research challenges remain before we
can trust on sensor networks. In this chapter we have listed the prominent research
efforts to meet the security goals of confidentiality, integrity, authentication, and
availability in sensor networks. We have discussed cryptographic needs of tiny sensors
and emphasized that only cryptography is not enough to secure sensor networks.
Security issues such as presence of a malicious node and compromise of a legitimate
node raise concern to look beyond cryptography. On the basis of our observation, we
motivate the need for a security framework to address the secure key management,
secure routing, malicious node detection, and secure location in sensor networks.
16.9 EXERCISES
1. Why is security different in wireless sensor networks? What makes sensor
networks more vulnerable compare to wireless networks and mobile ad hoc
networks?
2. List and discuss at least five security threats in wireless sensor networks.
3. Explain the difference between sinkhole and wormhole attacks.
BIBLIOGRAPHY
501
4. Discuss the attacks and countermeasures in application, network, data link, and
physical layers in sensor networks.
5. Summarize basic and random key pre-distribution schemes. How is multiple
space key pre-distribution scheme different from basic and random key predistribution schemes?
6. What is SPINS? Discuss the features of SNEP and µTESLA.
7. What is TinySec? What are the two security options offered by TinySec?
BIBLIOGRAPHY
1. L. Eschenauer and V. Gligor. A key-management scheme for distributed sensor networks.
In Proceedings of the 9th ACM conference on Computer and Communication Security
2002, Washington, DC.
2. A. Boukerche. Handbook of Algorithms for Wireless Networking and Mobile Computing,
CRC/Chapman Hall, Boca Raton, FL, 2006.
3. C. P. Fleeger. Security in Computing, 3rd edition, Prentice-Hall, Upper Saddle River, NJ,
2003.
4. J. Undercoffer, S. Avancha, A. Joshi, and J. Pinkston. Security for sensor networks. In
CADIP Research Symposium, 2002.
5. C. Karlof and D. Wagner. Secure routing in wireless sensor networks: Attacks and countermeasures. Elsevier’s Ad Hoc Networks Journal, Special Issue on Sensor Network Applications and Protocols, 1(2–3):293–315, 2003.
6. D. Wagner, Resilient aggregation in sensor networks. In Proceedings of the 2nd ACM
Workshop on Security of Ad Hoc and Sensor Networks, ACM Press, New York, 2004, pp.
78–87.
7. Y. W. Law and P. J. M. Havinga. How to secure sensor network. In Proceedings of the 2005
International Conference on Sensor Networks and Information Processing, 5–8, December
2005, pp. 89–95.
8. D. Ganesan, A. Cerpa, Y. Yu, and D. Estrin. Networking issues in wireless sensor networks,
Journal of Parallel and Distributed Computing (JPDC), Special issue on Frontiers in
Distributed Sensor Networks, 64, 2004.
9. C. Karlof, N. Shastry , and D. Wagner. TinySec: A link layer security architecture for
wireless sensor networks, SenSys ’04, November 3–5, 2004, Baltimore. MD.
10. L. V. Hoesel and P. Havinga. A lightweight medium access protocol (LMAC) for wireless sensor networks: Reducing preamble transmissions and transceiver state switches. In
Proceedings of INSS, June 2004.
11. W. Diffie and M. E. Hellman. New directions in cryptography. IEEE Transactions on
Information Theory, 22:644–654, 1976.
12. H. Chan and A. Perrig. Security and privacy in sensor networks. IEEE Journal of Computing, 36(10):103–105, 2003.
13. B. Dutertre, S. Cheung, and J. Levy. Lightweight key management in wireless sensor
networks by leveraging initial trust, SDL Technical Report SRI-SDL-0402, April 6 2004.
14. F. Stajano. Security for Ubiquitous Computing, John Wiley & Sons, New York, 2002.
502
SECURITY ISSUES AND COUNTERMEASURES IN WIRELESS SENSOR NETWORKS
15. H. Chan, A. Perrig, and D. Song. Random key predistribution schemes for sensor networks.
In Proceedings of the IEEE Symposium on Security and Privacy, Oakland, CA.
16. W. Du, J. Deng, Y. S. Han, and P. K. Varshney. A pairwise key pre-distribution scheme for
wireless sensor networks. In ACM CCS, 2003.
17. W. Du, J. Deng, Y. S. Han, S. Chen, and P. K. Varshney. A key management scheme for
wireless sensor networks using deployment knowledge. IEEE InfoCom, 2004.
18. R. Blom. An optimal class of symmetric key generation systems. Advances in cryptology.
In Proceedings of EUROCRYPT 84, 1985.
19. W. Du, J. Deng, Y. S. Han, and P. K. Varshney. A key predistribution scheme for sensor
networks using deployment knowledge. IEEE Transactions on Dependable and Secure
Computing, 3(1): pp. 62–77, 2006.
20. B. C. Neuman and T. Tso. Kerberos: An authentication service for computer networks.
IEEE Communications, 32(9):33–38, 1994.
21. T. Zia and A. Zomaya. A secure triple-key management scheme for wireless sensor networks. In IEEE Infocom Workshop 2006, Barcelona, Spain.
22. M. Saraogi. Security in wireless sensor networks, Department of Computer Science,
University of Tennessee, Knoxville, TN, 2005.
23. A. Perrig, R. Szewczyk, V. Wen, D. Cullar, and J. Tygar. SPINS: Security protocols for
sensor networks. In Proceedings of the 7th Annual ACM/IEEE International Conference
on Mobile Computing and Networking (MobiCom), pp. 189–199.
24. F. Stajano. Security for Ubiquitous Computing, John Wiley & Sons, New York, 2002.
25. A. Perrig , R. Canetti, J. D. Tygar, and D. Song. Efficient authentication and signing of
multicast streams over lossy channels. In IEEE Symposium on Security and Privacy, 2000.
26. J. Deng, R. Han, and S. Mishra. INSENS: Intrusion-tolerant routing in wireless sensor
networks. University of Colorado, Department of Computer Science, Technical Report
CU-CS-939-02, 2003.
27. J. Deng, R. Han, and S. Mishra. A performance evaluation of intrusion-tolerant routing in
wireless sensor networks. In IPSN 2003, LNCS 2634, Springer-Verlag, Berlin, pp. 349–364.
2003.
28. S. R. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong. TAG: A tiny aggregation
service for ad-hoc sensor networks. In The Fifth Symposium on Operating Systems Design
and Implementation (OSDI 2002), 2002.
29. S. R. Madden, R. Szewczyk, M. J. Franklin, and D. Culler. Supporting aggregate queries
over ad-hoc wireless sensor networks. In Workshop on Mobile Computing and Systems
Applications, 2002.
30. M. Bellare, A. Desai, E. Jokipii, and P. Rogaway. A concrete security treatment of symmetric encryption: Analysis of the DES modes of operation. In Proceedings of 38th Annual
Symposium on Foundations of Computer Science (FOCS ’97), 1997.
31. E. Shi and A. Perrig. Designing secure sensor networks, Journal of IEEE Wireless Communications, 11(6):38–43, 2004.
CHAPTER 17
A Taxonomy of Secure Time
Synchronization Algorithms for
Wireless Sensor Networks
AZZEDINE BOUKERCHE
School of Information Technology and Engineering, University of Ottawa, Ottawa, Ontario
K1N 6N5, Canada
DAMLA TURGUT
School of Electrical Engineering and Computer Science, University of Central Florida,
Orlando, FL 32816-2362
17.1 INTRODUCTION
Wireless sensor networks naturally sense the desired event or phenomena in the realworld environment, and they communicate the sensed data to the global processing
unit via the intermediate sensor nodes or gateway nodes for processing to draw relevant
conclusions. Most of the time, the data received from multiple sensors are aggregated
before reaching the final processing unit. In order to carry out the tasks discussed
above, the physical time of the sensor nodes has to be synchronized with each other.
The distributed wireless sensor networks heavily depend on the time synchronization
for various reasons such as determining location and proximity of the deployed sensor
nodes, intra-network coordination among different sensor nodes, temporal message
ordering, security, time division multiplexing in wireless communication, improving
energy-efficiency of sensor nodes by scheduling the sleep times of the sensor nodes,
and so on [1].
In the computer synchronization history, Lamport’s work [2] is considered a
pioneering approach that emphasizes the need of virtual clocks in computer systems
in which causality is more important than the absolute time. Even though the total
ordering of the events was the focus of Lamport’s method, this work has influenced
the way the sensor networks have emerged today.
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
503
504
A TAXONOMY OF SECURE TIME SYNCHRONIZATION ALGORITHMS
Most computer devices contain an internal clock, usually designed to be synchronized with the exact real-world time at the specific location of the computer,
although many functionalities depend on the clock even on desktop computers—for
instance, the scheduled Friday afternoon virus checks, or the popular “make” program,
which determines whether a file needs to be recompiled comparing the timestamp of
the source and object files. However, in practice, a desktop computer can function
correctly even if its internal clock is minutes or even years away from the correct
time.
Let us first define the ways in which the clocks of two nodes A and B might
be out of sync. Let us note the clock of a node X with a function CX (t), which
returns the reading of the clock at real time t. The first type of difference is the offset:
δAB = CA (t) − CB (t). That is, the two clocks are identical, except that the clock of
node A is early (if δAB > 0). Now, we might fix this by setting the clock of node
A back; however, that would create a problem, because the same time slice would
appear twice for node A. This creates major problems for a number of protocols. It
is better to set the clock of node B forward; however, most protocols simply require
the nodes to keep track of their offsets without actually changing the internal clock.
The second type of synchronization difference is the clock skew; that is, one of
the clocks is running faster than the other. This can be expressed as a difference in
the derivatives of the clock function with respect to the time: ηA B = ∂C∂tA (t) − ∂C∂tB (t) .
While this appears to be a more difficult problem, if a node is aware of its clock skew,
it can very easily account for it.
Neither clock offset nor clock skew requires periodic synchronizations. If we know
the offset and the skew of a node’s clock, we can calculate the time difference at any
moment in time. However, the frequency of the clocks can change randomly because of
environmental conditions such as temperature difference or the aging of the hardware,
a condition called drift error. The
drift error
appears as a nonzero second derivative
2
2
in one or both clocks λA B = ∂ C∂tA2 (t) − ∂ C∂tB2 (t) . The clocks of sensor nodes usually
accumulate several seconds of drift error per day; because the drift is not predictable,
it needs to be solved using clock synchronization.
However, for sensor networks, the correct synchronization of the clocks is frequently a necessary component of the ability of the sensor network to function correctly. Unsynchronized clocks can yield invalid observations, can create uncovered
areas and timeslots, and in the worst case can disable the communication architecture
of the network.
Let us consider several examples. The individual nodes of the sensor network are
sending their timestamped observations to the sink.
In Figure 17.1, an intruder is sensed consecutively by sensors S1 and S2 , and their
reports are sent to the sink. Based on the reports (intruder, S1 , t1 ) and (intruder, S2 , t2 ),
knowing the locations of the sensors S1 and S2 , and noticing that t1 < t2 and
t2 − t1 < 1 second, the sink can correctly infer that the observations refer to the
same intruder who is moving from left to right (in certain cases, this inference can
be performed through in-network processing). However, this inference is valid only
under the assumption that the clocks of the two sensors are synchronized at the level
of tenths of seconds. Let us explain this further.
INTRODUCTION
505
location of the intruder
at universal time t 1
location of the intruder
at universal time t 1
Intruder
Intruder
sensor 1
sensor 2
(i, t1(1))
(i, t2 (2))
Sink
Figure 17.1. Intruder movement detected by sensors 1 and 2 at times t1(1) according to clock 1
and t2(2) according to clock 2.
If clocks 1 and 2 are synchronized—that is, they have the same offset δ(1) = δ(2)
(1)
(2)
compared to a universal time t—then t2 − t1 = t2 + δ(2) − t1 − δ(1) = t2 − t1 > 0.
So, t2 − t1 indicates the correct order of the arrival to the sensors. However, if there
is a large offset between these two, then we have δ(1) − δ(2) ≪ 0; that is, the clock of
(1)
(2)
δ(1) is early. We might have a situation where t2 − t1 = (t2 − t1 ) + δ(2) − δ(1) < 0;
that is, the sink will infer incorrectly that the intruder is moving from right to left. This
is case where if the clock of the sensor S1 is two seconds late, the inference would be
that the intruder moves from right to the left. As a drift of several seconds per day is
a normal occurrence for the internal oscillators of the devices, we cannot rely on the
initial setting of the clocks at deployment time. The clocks need to be synchronized
periodically in the field.
Notice that the faster the intruder moves, the smaller (t2 − t1 ) and the more accurate
synchronization is needed in order to make the correct inferences.
Our second example concerns the wake-up time of the sensors. Sensors have
limited power resources. To extend the lifetime of a deployed network, the sensor
nodes are frequently selectively put to sleep. The idea of the method is that the set
of currently active nodes at any given moment in time covers the area to be surveyed
and form a connected network. If an attacker can modify the internal clock of certain
sensor nodes, such that these nodes, for instance, do not wake up in time, certain
areas might not be surveilled by the sensors for a certain amount of time, allowing an
intruder to operate unreported.
Finally, Time Division Multiple Access (TDMA)-based channel sharing protocols
rely on the participating nodes to transmit at well-defined timeslots. Relatively small
time drifts in the clock of the individual nodes can make the transmission intrude
on the adjacent time slot, causing a collision. Repeated collisions can significantly
disrupt the network. A detailed survey on clock synchronization protocols can be
found in references 3 and 4.
506
A TAXONOMY OF SECURE TIME SYNCHRONIZATION ALGORITHMS
The rest of the chapter is organized as follows. The challenges and design issues of
time synchronization protocols in sensor networks are summarized in Section 17.2. In
Section 17.3, we survey some of the time synchronization protocol including possible
attacks and proposed countermeasures against these attacks. The types of attackers
and attacks are presented in Section 17.4. Section 17.5 gives detailed discussion about
the approaches for secure time synchronization. We conclude in Section 17.6.
17.2 TIME SYNCHRONIZATION IN SENSOR NETWORKS
17.2.1 Challenges
Different applications will have different synchronization requirements, and almost all the synchronization methods rely on message exchange between nodes.
The nondeterminism in the network operations can cause delays in the message
delivery and, in turn, contributes to the error in synchronization. Let us see how the
source of a message’s latency is decomposed in Figure 17.2. According to references
5 and 6, the time synchronization schemes have four basic packet delay components:
send time, access time, propogation time, and receive time.
r Send Time. This is the time it takes to construct a message at the sender including
the overhead of the operating system and the time to transfer the message to the
network interface.
r Access Time. This is the delay encountered at the medium access control (MAC)
layer prior to accessing the transmission channel due to contention, collisions,
and so on. This delay is dependent on the MAC protocol in place. For instance,
carrier sensing multiple access (CSMA) [7] requires every node to sense the
carrier before transmitting, and it does not start a transmission if the medium is
busy. CSMA/CA consists of both carrier sensing and a collision avoidance; the
IEEE 802.11 standard [8] is the best-known instance of CSMA/CA.
r Propogation Time. This is the time required in propagation of the message from
sender to the receiver. The propagation time varies, depending on the location of
the sender and the receiver. For instance, if they are one-hop neighboring nodes in
an ad hoc network, the propagation time equals to the physical propagation time
Sender
Send
Time
Access
Time
Propagation
Receiver
Propagation
Time
Receive
Time
Figure 17.2. Packet delay components.
TIME SYNCHRONIZATION IN SENSOR NETWORKS
507
of the message traveling in the media. On the other hand, the propagation time
can be much larger if we add switching and queueing delays into the formula.
r Receive Time. This is the time for the receiver node to process the message and
acknowledge the host the arrival of the message. Depending on the level in which
the arrival time was timestamped, the receive time may or may not include the
overhead of the transfering of the message from the network interface to the
host.
The challenge is not only the existence of packet delay but also the difficulty of
prediction of the time needed on each delay component even though Figure 17.2
shows each component of equal length.
17.2.2 Design Issues
Here, we present a set of metrics for evaluating time synchronization protocols in
sensor networks. Naturally, there are tradeoffs between these metrics; therefore, it is
difficult to find a protocol that can satisfy them all.
r Availability and Scope. All the nodes in the network can be synchronized based
on a global time, or a set of nodes locally in close proximity to each other can
be synchronized based on a local time. Due to mainly the energy and bandwidth
constraints in large-scale sensor networks, global synchronization is not only
difficult to implement but can also be costly. Only a few applications would
require global synchronization. Completeness of coverage within the specific
region of nodes determines the availability requirement.
r Cost and Size. Since most sensor nodes considered are small and low-cost devices, the secure synchronization algorithms must be designed to meet the cost
and size requirement. For example, it is not a viable solution to attach a global
positioning system (GPS) device to a sensor node.
r Efficiency. Due to the low weight and small size of the sensor nodes used in many
sensor network applications, they possess limited resources such as energy, memory, computation power, and so on. The security mechanisms developed within
the time synchronization protocols must adapt to the limited computation power
and memory of these sensor nodes since most security protocols are complex
and require extensive resources to run efficiently. Additionally, communication
capabilities are also limited due to these resource constraints. GPS or Universal
Time (UTC) are generally used to synchronize the network to an accurate time
in traditional protocols such as Network Time Protoocol (NTP) [9]. Using GPS
requires high energy consumption; therefore, it is not considered a viable solution. Reducing the energy consumption can be achieved by transmitting over
sequences of hops of short distances rather than a single one-hop long path. The
secure time synchronization algorithms should also take into consideration the
time needed to synchronize the nodes.
508
A TAXONOMY OF SECURE TIME SYNCHRONIZATION ALGORITHMS
r Infrastructure. Sensor networks may have to be deployed in a random fashion
to remote or dangerous regions. In this case, the sensor nodes are expected to
self-organize themselves into a network since they cannot rely on an existing
infrastructure such as NTP [9] in which the precise time is available to any node
in the network from only a few hops away. NTP achieves this by providing
the reference time at various points in the network. The task of the secure time
synchronization becomes complex due to the many factors such as scalability
issues, mobility, and of course the lack of infrastructure to rely on.
r Lifetime. The duration of the synchronization can be one of the two forms: (i)
almost instantenous or (ii) persistent, in which case the synchronization lasts as
long as the lifetime of the network itself. The nearly instantenous synchronization
can be appropriate for applications in which an immediate action needs to be
triggered based on the nodes’ detection times of a specific event.
r Network Dynamics. Regardless of the type of deployment (random versus preengineered), sensor nodes face a high degree of network dynamics such as frequent changes in the network topology due to mobility, network partitioning, or
energy depletion of the nodes. Designing secure time synchronization protocol
becomes an even more complex task when one needs to consider ways to ensure
the running of the network operations smoothly under these dynamic changes.
r Precision. The level of precision or accuracy required generally depends on
both the application and the objective of the synchronization. There are three
basic types of synchronization methods. The first one relies on the ordering
of messages and events; it is considered the simplest. The next method allows
the nodes to keep track of the drift and offset with respect to their neighboring
nodes. This is the most common type of synchronization encountered in the time
synchronization protocol applications. The most strict and complex one is the
global synchronization in which all the nodes are synchronized according to a
global time across the network. This method is the least used one since it is the
most difficult to implement and also precise time synchronization is not always
essential.
17.3 SECURE TIME SYNCHRONIZATION
These examples show us that time synchronization is vital for the correct operation of
a sensor network. Because relatively small time drifts can cause significant disruption,
we cannot rely on the precision of the hardware; we need to use external synchronization protocols. Furthermore, it was found that because relatively small changes
in the clocks can disturb the operation of the sensor network or even cause it to make
erroneous inferences about the observed event, the time synchronization protocols are
a convenient target for malicious attackers. Most time synchronization protocols were
not designed with security in mind. Recently, however, several research groups performed an analysis of various vulnerabilities and proposed countermeasures against
them.
SECURE TIME SYNCHRONIZATION
509
In the following, we survey some of the time synchronization protocols, discuss
their benefits, and outline possible attacks and proposed countermeasures.
17.3.1 Reference Broadcast Synchronization
A pioneering approach called post facto synchronization was proposed by Elson
and Estrin [10]. This is a low-power method of synchronization of clocks when the
timestamps must be accurate for desired events. The nodes’ clocks are are normally
unsynchronized. When a phenomenon is sensed, each node records the time of sensing
according to its own local clock. Shortly after, a beacon node sends a synchronization
message to the nodes within its transmission range. The nodes receiving this message
can adjust their timestamps of the sensed phenomena to the time of the receipt of
this synchronization message. The post facto synchronization forms the basis for the
reference broadcast synchronization method.
A sender-to-receiver synchronization method is used in most time synchronization
protocols. In this method, the sender transmits the timestamp information and the
receiver synchronizes.
The Reference Broadcast Synchronization (RBS) [11] protocol differs from the
sender-to-receiver synchronization since it uses receiver to receiver synchronization
method. Basically, the RBS is based on a synchronization signal broadcasted by
an external unit. The receivers record their local time when they receive this reference message, and then they exchange this information among themselves (see
Figure 17.3). The recording of a message is not 100% exact, because of hazards such
as the propagation time of the message or the processing time of the packet at the
lower protocol layers. To improve the precision, a number of reference messages can
be broadcasted, and the nodes exchange the arrival times for each message and then
find the best approximation using a least squares fit.
The keypoint of the RBS is that it uses a broadcast technique within a wireless
medium to minimize latency and nondeterminism issues related to latency in the time
synchronization protocol. On the other hand, the most notable drawback of the RBS
is its requirement of a network with a physical broadcast channel [11].
Let us consider a scenario shown in Figure 17.4 a where a group of nodes lie in
the range of multiple broadcast instead of a range of a single broadcast. Both sender
Sender
NIC
Receiver
Time
Sender
NIC
Receiver 1
Critical Path
Receiver 2
Critical Path
Figure 17.3. A critical path analysis for traditional time synchronization protocols (left) and
RBS (right). (Adapted from reference 11).
510
A TAXONOMY OF SECURE TIME SYNCHRONIZATION ALGORITHMS
1
1
2
B
A
A
3
5
2
5
4
3
2
5
3
4
7
8
9
10
11
6
6
C
B
7
7
4
1
6
8
9
D
10
(a)
11
(b)
(c)
Figure 17.4. Simple (a) versus complex (b) multihop network topology with the corresponding
logical topology (c) of the complex topology. (Adapted from reference 11).
nodes A and B send a synchronization message. According to the transmission ranges,
nodes 1–4 will be able to synchronize with node A and nodes 4–7 can synchronize
with node B; however, nodes A and B cannot hear each other’s synchronization message. Among the nodes, node 4 is the only one that can hear the synchronization
messages from both nodes A and B. This means that node 4 can correlate the clocks
in node A’s neighborhood to the clocks in node B’s neighborhood or vice versa. In
Figure 17.4 b, we have 3-hop network topology and the corresponding logical topology in Figure 17.4 c. The dotted lines—that is, the graph edges—are drawn between
two nodes whose clocks are known to each other—for instance, node pairs (1–4),
(7–9),(8–9), and so on. Note that there are two links between nodes 8 and 9, meaning
that they have two receptions in common since they are located in the overlapping
region of C and D.
Possible Attacks on RBS. In RBS, upon receiving a broadcast signal, two nodes
exchange their local clock time. An attack can happen if one of the receiver nodes is
compromised with an incorrect time. The compromised node then can send the incorrect time information to its neighbor, causing the uncompromised node calculating
an incorrect offset.
The multihop version of RBS can face attacks as well. If a compromised node
is located in any of the overlapping regions, it can inject an incorrect value into the
clock conversion process, affecting multiple regions at once. This miscalculation in
the clock conversion can be propagated across the network.
17.3.2 Time Synchronization Protocol Sensor Networks (TPSN)
TPSN [12] has two phases: level discovery and synchronization. In the level discovery
phase, a spanning tree is created for the sensor network where each node is assigned a
level. The root of the tree is usually a base station and assigned level 0. It is assumed
511
SECURE TIME SYNCHRONIZATION
T2
node B
T3
node A
T1
T4
Figure 17.5. Two-way message exchange between pair of nodes. T2 and T3 are measured in
node B’s clock while T1 and T4 are measured in node A’s clock. (Adapted from reference 12.)
that a node at level n can communicate with a node or a set of nodes at level n − 1.
In the synchronization phase, the child nodes are synchronized to the parent. The
synchronization is initiated by the child node, which sends a synchronization packet
at time t1 . This is received by the parent at t2 , and an acknowledgment packet is sent
in response at time t3 . The values of t2 and t3 will be included in the acknowledgment
packet. This packet is received by the child node at time t4 (see Figure 17.5). Knowing
these four time values, the child node can calculate its clock offset relative to the parent
node as well as propagation delay as follows:
t =
(t2 − t1 ) − (t4 − t3 )
2
(17.1)
d=
(t2 − t1 ) + (t4 − t3 )
2
(17.2)
The decomposition of the packet delay is depicted in Figure 17.6 similar to
references 5 and 6 as shown in Figure 17.2. The transmission and reception time
are more detailed here. The transmission time is the time it takes to transmit a packet
on a bit-by-bit basis at the physical layer via the wireless link [12]. The reception
Sender
Send
Access
Propagation
Transmission
Receiver
Reception
Receive
Figure 17.6. Decomposition of packet delay over a wireless link in TPSN. (Adapted from
reference 12).
512
A TAXONOMY OF SECURE TIME SYNCHRONIZATION ALGORITHMS
time is the time required to receive and forward all the bits to the link layer. Both
transmission time and reception time are generally considered deterministic; however,
variations can occur, depending on the underlying hardware structure.
The nodes are assumed to have unique ids, and each node has knowledge of its
neighboring nodes. TPSN takes advantage of only the symmetric links for pairwise
synchronization between nodes even though the network may also have asymmetric
links. This can be considered one of the drawbacks of the protocols. Other drawbacks
may include its limited suitability for applications serving highly mobile networks
and its lack of support for multihop communication. On the other hand, TPSN is
scalable and its computation overhead is less than some other protocols such as NTP
[9]. The root selection and the tree construction mechanisms need to be reinvoked
when topology changes occur due to node failures or other factors.
Possible Attacks on TPSN. A compromised node cannot cause any problem by
requesting a time synchronization message because the message will reach to the
parent node only. On the other hand, the compromised node can send erroneous time
information to its children.
Naturally, the child node is relying on the parent for its clock synchronization;
by providing incorrect values for t2 and t3 , the parent can set an arbitrary offset on
its child node. What is more, this incorrect offset will then be propagated down the
tree. Therefore, the number of nodes whose synchronization can be affected by the
compromised node depends on the location of the compromised node on the tree. One
way for a malicious attacker to compromise a larger number of nodes is to reposition
itself in a higher location on the tree or to answer queries instead of the proper parent.
This is surprisingly easy to do in the original algorithm.
Yet another type of attack can surface when the compromised node misinforms
its level in the tree, basically announcing a lower level than its current level. The
compromised node can also attempt to trick other nodes at its level in requesting synchronization updates from itself. Furthermore, the compromised node can disconnect
a number of nodes from being included in the tree by simply not participating in the
level discovery phase.
17.3.3 Flooding Time Synchronization Protocol (FTSP)
In FTSP [13], the nodes are participating in a process in which a root node is elected.
The root is the origin of the time synchronization messages. If a node does not hear a
time synchronization message for a while, it declares itself the new root. The protocol
requires that if at a later time the node receives a time synchronization message from
a node with a lower id than itself, it gives up its root status. When a node receives a
time synchronization message from the root, it adjusts its clock and broadcasts its own
time to its neighbors. In the message broadcast, the preamble bytes are transmitted
first, followed by sync bytes, message descriptor, the actual message, and finally crc
bytes (see Figure 17.7).
SECURE TIME SYNCHRONIZATION
sender
preamble
synch
data
513
crc
propagation delay
receiver
preamble
synch
data
crc
byte alignment
Figure 17.7. Data packets transmitted over the radio channel. (Adapted from reference 13.)
Possible Attacks on FTSP. The FTSP protocol is more robust for node failures
than the TPSN protocol, because there is no need to maintain a tree structure that
is notoriously vulnerable to single-point failures: The failure of a single node can
disconnect the whole subtree. The weak point of the FTSP protocol is the election
process. Any node can declare itself a root, and the protocol relies on the node to step
back if a lower id root appears. A compromised node can easily masquerade as a root;
and by declaring a very low id, it can actually dislodge the existing legitimate root.
Then, by sending a synchronization message with a fake timestamp, it can make the
nodes synchronize to an incorrect time.
17.3.4 Countermeasures for the Attacks
In this section, we describe the possible countermeasures for time synchronization
attacks in both single-hop and multihop networks [14]. In single-hop networks, transmission range of each node reaches to every other node in the network. Let us consider
a specific case of a single-hop network where there is a base station and its transmission range covers all the nodes in the network. The challenge for this type of
networks is to preclude the malicious node(s) from compromising the base station
and immediately start injecting the network with invalid timing information. In this
case, the message from the base station must be authenticated in order to carry out the
correct sequence of time synchronization methods. This can be achieved by utilizing
a broadcast authentication scheme such as µTESLA [15]. Another approach can be
the use of different private keys between the sender and receiver nodes.
In the multihop networks, many nodes may need to communicate with each other
via intermediate nodes due to the limited transmission ranges. It makes sense for nodes
to receive the global time of their immediate neighbors instead of neighbors several
hops away. In that case, we need to take into consideration the incurred network delays
between these nodes and their far-away neighbors. An approximation approach can
be used find an upper bound on the error produced by the malicious node. Introduction
of redundancy to the network is another viable approach. For instance, both FTSP
and TPSN compute the offset and skew of their clock based on the timing information
obtained from only one neighboring node. This redundancy approach can be easily
514
A TAXONOMY OF SECURE TIME SYNCHRONIZATION ALGORITHMS
applied to FTSP where a set of nodes can be used for the synchronization computations. The nodes can take the median of these received multiple updates. The private
keys can be set up between the nodes and their neighbors at the beginning such that a
malicious node can be precluded from injecting erroneous updates into the network.
Furthermore, if a node becomes aware of one of its neighbors’ update value being
substantially different than the updates from its other neighbors, the node can refrain
itself from including the updates from the suspicious neighbor into its computations
of clock skew and offset. This approach of containment works, however, under the
assumptions that the nodes have a sufficient number of sources for the updates. For
instance, in TPSN, children of the nodes nearby the malicious or compromised node
must find another parent, which may not always be feasible. Essentially, maintaining
multiple trees comes with a price whereas in FTSP, no such cost exists. Utilizing the
LS linear regression by each node in the network to compute the skew of its clock
can further enhance security in time synchronization protocols. Algorithms such as
RANSAC [16] can be used for this purpose.
17.4 ATTACKERS AND ATTACKS
The possible attacks against time synchronization protocols depend on the nature and
capabilities of the attackers. We will first identify three different types of attackers,
outline the types of attacks they are capable of, and then discuss the various types of
defenses proposed against these types of attacks.
We will describe the system with the characters regularly used in the description of
cryptographic protocols. We assume that Alice and Bob (and, potentially, additional
nodes Carol, Dave, and so on) are engaging in a time synchronization process.
The malicious outsider Malory is a wireless device inserted in the range of the
nodes of the sensor network, which has the ability to send and receive packets.
We assume that the attacker can eavesdrop on any ongoing transmission; we can
also assume that the attacker can eavesdrop on any ongoing transmission and that
the attacker can transmit messages which are physically indistinguishable from the
other nodes messages. However, this type of attacker does not have access to keys
or other confidential information, other than what it can infer from eavesdropping on
transmissions.
Attacker with Jamming and Replay Ability. Jimmy is an attacker which has
the ability to jam the message, record it, and possibly replay it at later time. This
type of attack is called a pulse delay attack. Although, in principle, the existence of
jamming can be detected, it requires significant resources; by default, most nodes are
not prepared for it.
A compromised node is a node that was taken over by the attacker. We will call
this Zach (for zombie). One example of this is the physical capture of the node by
an attacker, although a node can, in principle, be compromised with purely software
methods. Compromised nodes have access to all the keys and other information of the
original node, and they represent the most difficult type of attacker to defend against.
APPROACHES FOR SECURE TIME SYNCHRONIZATION
515
The challenge of secure time synchronization is to defend against all three types
of attackers. An attacker is considered successful if it succeeds in making the nodes
calculate an incorrect offset. By default, all three time synchronization protocols we
described are vulnerable to all three types of attacker.
17.5 APPROACHES FOR SECURE TIME SYNCHRONIZATION
Malicious outsiders can affect all types of protocol. The primary defense against a
malicious outsider is cryptographic techniques of the authentication of messages. If
the sender and the receiver is sharing a key KB , they can use it to sign the messages.
To prevent an attacker from capturing a valid message and inserting a copy of it
later in a different synchronization round, the sender sends a random nonce in the
initial message, which then needs to be signed by the synchronization partner. While
the attacker can still replay the same message in the same synchronization round,
by simply considering only the first arrived message, the receiver can ignore the
malicious outsider.
Ganeriwal et. al. [17, 18] proposed a series of secure time synchronization protocols
based on this idea. The protocols are adapted to pairwise single-hop and multihop
synchronization and for group synchronization. The protocols can also detect the
existence of a pulse delay attack by calculating the end-to-end delay d of the message.
If the delay is larger than a predetermined threshold d∗, the protocols assume that an
attack is in progress and abort the synchronization. We should note that key exchange
is a major problem for these types of algorithm, due to the ways in which sensor
nodes are deployed, which does not always permit the exchange of the keys in a
secure environment.
Notice that cryptographic methods are not feasible against a compromised node,
which has all the keys and knowledge to correctly answer all the challenges and appropriately sign its messages. If the time synchronization protocol happens only between
Alice and Zach, it is impossible for Alice to detect or mitigate the attack. However,
for protocols with a larger number of participants, we can use the redundancy in the
synchronization messages to identify the malicious participants or messages. We note
that delay attacks can be performed by either Zach or Jimmy, but not Mallory.
Song et al. [19] propose a method for making time synchronization protocols
resilient to delay attacks based on techniques of outlier detection. The essential assumption behind this method is that the synchronization signals received from compromised nodes will be “much different from one another.” Thus, messages coming
from compromised nodes can be identified using statistical techniques as outliers and
then eliminated from the package, and the synchronization can be performed with the
remaining nodes.
The authors propose two alternative methods. One of them uses the generalized
extreme studentized deviate (GESD), a generalization of the well-known Grubb’s test
from statistics. GESD can identify multiple outliers in a sample drawn from a normal
distribution. GESD requires as one of its outputs the estimated number of malicious
nodes.
516
A TAXONOMY OF SECURE TIME SYNCHRONIZATION ALGORITHMS
A somewhat simpler approach is based on a delay threshold. At the setup time
of the system, the nodes determine the maximum amount of time offsets they will
tolerate, based on information about the typical drift rate of the nodes. A received
offset that is higher than this value will be considered to come from a malicious node
and discarded.
As an observation, naturally, Zach, the compromised node, would have exact
knowledge about the thresholds used (but not Jimmy). Therefore, Zach has the possibility to remain undetected, by setting the delay such that it will put it just below the
threshold (or, in the GESD case, such that it will not be identified as an outlier), but
still have a distorting effect on the time synchronization process. Thus, the statistical
techniques can only reduce, but not necessarily eliminate, the effect of delay attacks
by nodes with insider knowledge.
Sun et al. [20] propose a statistical method for secure and resilient clock synchronization in the presence of compromised nodes. The techniques are applied for both
level-based clock synchronization, where a hierarchical structure of nodes is developed which determines which node is synchronized with whom, and diffusion-based
clock synchronization, which does not use such a structure and simply relies on the
reachability information of the network. Naturally, the level-based approach allows
for a more disciplined control of the synchronization flow, and thus a higher accuracy, whereas the diffusion method has the advantage that it can be applied to dynamic
sensor networks with mobile nodes.
Furthermore, the authors consider both (a) the case with a single source of synchronization information and (b) the case with multiple sources. For instance, in the
single-source case, the goal is assumed to be to find the clock offset δiS from the node
to the source. The technique assumes that at every level, a normal node collects 2t + 1
candidate source clock differences from its 2t + 1 neighbors and chooses the median
of them. Thus, the node can tolerate up to t compromised nodes, while retaining correct synchronization. Similar considerations apply to the diffusion-based approach. In
the case of multiple sources, the node can receive synchronization information from
2s + 1 sources, synchronized to the same external standard (such as a GPS signal)
and tolerate up to s compromised sources by selecting the median.
Note that this approach uses the whole redundancy of the system to defend against
an external attack: Out of 2t + 1 recorded offsets, the method will pick a single one,
the offsets median. Approaches that assume a benign environment usually select the
mean of these measurements, therefore improving accuracy; however, the mean is
vulnerable to even a single malicious node.
In addition, this approach requires a unique pairwise-key-based authentication
of the nodes. Otherwise, the malicious node could impersonate multiple nodes (the
so-called Sybil attack).
A significantly improved version of this technique was presented in
reference 21. In the approach called TinySeRSync, time synchronization is performed
in two phases. While we will call them Phase I and Phase II for convenience, these two
processes are taking place asynchronously in the sensor network. In the first phase,
single-hop pairwise synchronization is performed. The main feature of the pairwise
synchronization process is that it relies on a hardware-enhanced authenticated MAC
517
APPROACHES FOR SECURE TIME SYNCHRONIZATION
TABLE 17.1. A Summary of the Various Techniques Proposed for Secure Time
Synchronization
Approach
Ganeriwal et al. [17]
Song et al. [19]
Sun et al. [20]
TinySeRSync, Sun et al. [21]
Manzo et al. [14]
Protects
Against
Jimmy, Mallory
Zach, Jimmy, Mallory
Zach
Zach, Mallory
Zach
Uses
Crytographic
Techniques?
Uses
Statistical
Techniques?
Yes
No
No
Yes
Yes
Yes
Yes
Yes
Yes
Yes
layer timestamping. The hardware is programmed to add a timestamp authenticated
with a message integrity code (MIC) to every MAC packet transmitted. This is especially challenging for the newer radios, such as the ones on the newer-generation
MICAz motes, where the time required to authenticate the timestamp can interfere with the transmission rate of the radio. The authors propose a prediction-based
approach where the authenticated timestamp includes a prediction of the time required
to calculate the MIC.
Through these techniques (see Table 17.1 for summary), the nodes achieve a sufficiently good local-level synchronization, which is exploited in the second phase. The
second phase implements a global synchronization using the µTESLA broadcast authentication protocol. This protocol relies on the loose time synchronization between
the nodes, and it uses a unidirectional keychain. Messages received need to be stored
by the receiver and will be authenticated only after several timeslots. This prevents an
attacker from forging messages, but it opens the doors for a denial-of-service attack.
The messages received need to be buffered for future authentication; and because the
memory of sensor nodes is limited, Mallory can create fake messages, which will not
pass the authentication test, but will fill the buffer, preventing the node from receiving legitimate messages. To prevent denial-of-service attacks, the authors propose a
modified version of µTESLA. To reduce the timeslots when the adversary nodes can
flood the node with messages based on captured keys (which the receiver node needs
to store for future authentication), TinySeRSync uses an implementation with very
short delays r (made possible by the good local synchronization achieved in phase I).
However, such short delays would require the generation of a large number of keys;
the implementation utilized short intervals r used for message broadcasting, alternated
with long intervals R used for broadcasting the disclosed keys.
The global synchronization in TinySeRSync still relies on the selection of the
median from the 2t + 1 candidate offsets, therefore tolerating the presence of at most
t compromised nodes.
Notice that the approach presented in references 20 and 21 uses the median rather
than the mean as a choice of the estimated time offset, thus obtaining a high protection
against malicious nodes (provided that the technique is coupled with cryptographic
defenses). However, it sacrifices the ability to improve precision through multiple
independent observations.
518
A TAXONOMY OF SECURE TIME SYNCHRONIZATION ALGORITHMS
We can attack the general problem of finding the best estimation of the time offset
δbest from a set of candidate offsets {δ1 , ..., δn } by applying the principles of robust
estimation. We note that the individual offset measurements δi might have natural
noise, but some of them might be a result of a malicious attack. For any estimation
method, the breakdown point is the smallest number of contaminated values that
can move the estimate arbitrarily far from the correct value. Unfortunately, the most
frequently used estimators, the average and the least squares estimators, have a very
low breakdown point; a single malicious value can modify the estimate arbitrarily
far. Manzo et al. [14] propose the use of the least mean squares (LMS) estimator for
a more robust modeling. The generalized extreme studentized deviate (GESD) used
by reference 19 is another example of the application of the techniques of robust
estimation.
17.6 CONCLUSIONS
Among the many challenges in designing and employing wireless sensor networks
is the clock synchronization between the sensor nodes. Agreeing on a common time
is needed and even required by many of the sensor applications to carry out the
sensing, communication, and processing of the sensed data. The time synchronization
protocols in traditional wired networks cannot simply be re-used in the wireless
sensor networks domain due to the inherent characteristics and limited resources of
these networks. Therefore, several time synchronization protocols have been proposed
recently; however, most of them do not consider security aspect during the design
stages. There are only handful of protocols where the security has been taken into the
consideration.
In this chapter, we reviewed the three most common secure time synchronization
protocols: (i) Reference Broadcast Synchronization (RBS), (ii) Time Synchronization
Protocol Sensor Networks (TPSN), and (iii) Flooding Time Synchronization Protocol
(FTSP). We then evaluated these algorithms based on factors such as their countermeasures against various attacks and the types of techniques used (cryptographic
versus statistical).
BIBLIOGRAPHY
1. K. Römer. Time synchronization and localization in sensor networks. Ph.D. thesis, Swiss
Federal Institute of Technology Zurich (ETH Zurich), Zurich, Switzerland, 2005.
2. L. Lamport. Time, clocks, and ordering of events in a distributed system. Communications
of ACM, 21(4):558–565, July 1978.
3. B. Sundararaman, U. Buy, and A. D. Kshemkalyani. Clock synchronization for wireless
sensor networks: A survey. Ad Hoc Networks, 3(3):281–323, 2005.
4. F. Sivrikaya and B. Yener. Time synchronization in sensor networks: A survey. IEEE
Network Magazine’s special issue on Ad Hoc Networking: Data Communications & Topology Control, 18(4):45–50, 2004.
BIBLIOGRAPHY
519
5. H. Kopetz and W. Ochsenreiter. Clock synchronization in distributed real-time systems.
IEEE Transactions on Computers, C-36(8):933–939, 1987.
6. H. Kopetz and W. Schwabl. Global time in distributed real-time systems. Technical report
15/89, Technische Universität Wien, 1989.
7. L. Kleinrock and F. A. Tobagi. Packet switching in radio channels: Part I—carrier sense
multiple-access modes and their throughput-delay characeteristics. IEEE Transactions on
Communications, COM-23(12):1400–1416, 1975.
8. IEEE std 802.11b-1999. part 11: Wireless LAN medium access control (MAC) and physical
layer (PHY) specifications, August 1999.
9. D. L. Mills. Internet time synchronization: The network time protocol. IEEE Transactions
on Communications, 39(10):1482–1483, 1991.
10. J. Elson and D. Estrin. Time synchronization for wireless sensor networks, In Parallel and
Distributed Processing Symposium (IPDPS), April 2001, pp. 1–6.
11. J. Elson, L. Girod, and D. Estrin. Fine-grained network time synchronization using reference broadcasts. In Proceedings of the Fifth Symposium on Operating Systems Design and
Implementation (OSDI 2002), December 2002, pp. 147–163.
12. S. Ganeriwal, R. Kumar, and M. B. Srivastava. Timing-sync protocol for sensor networks.
In ACM Proceedings of the 1st International Conference on Embedded Networked Sensor
Systems (SenSyS 2003), November 2003, pp. 138–149.
13. M. Maróti, B. Kusy, G. Simon, and Á. Lédeczi. The flooding time synchronization protocol.
In ACM Proceedings of the 2nd International Conference on Embedded Networked Sensor
Systems (SenSyS 2004), November 2004, pp. 39–49.
14. M. Manzo, T. Roosta, and S. Sastry. Time synchronization attacks in sensor networks.
In The Third ACM Workshop on Security of Ad Hoc and Sensor Networks (SASN 2005),
November 2005, pp. 107–116.
15. A. Perrig, R. Szewczyk, V. Wen, D. Culler, and J. D. Tygar. Spins: Security protocols
for sensor networks. In International Conference on Mobile Computing and Networking,
July 2001, pp. 189–199.
16. M. A. Fischler and R. C. Bolles. Random sample consensus: A paradigm for model fitting
with applications to image analysis and automated cartography. Communications of ACM,
24(6):381–395, 1981.
17. S. Ganeriwal, S. Capkun, C.-C. Han, and M. B. Srivastava. Secure time synchronization service for sensor networks. In Wireless Security Workshop (WISE), September 2005,
pp. 97–106.
18. S. Ganeriwal, S. Capkun, and M. B. Srivastava. Secure time synchronization in sensor
networks. In ACM Transactions on Information and Systems Security, March 2006.
19. H. Song, S. Zhu, and G. Cao. Attack-resilient time synchronization for wireless sensor
networks. Ad Hoc Networks Journal, 5(1):112–125, 2007.
20. K. Sun, P. Ning, and C. Wang. Secure and resilient clock synchronization in wireless sensor
networks. IEEE Journal on Selected Areas in Communications, 24(2):395–408, 2006.
21. K. Sun, P. Ning, C. Wang, A. Liu, and Y. Zhou, TinySeRSync: Secure and resilient time
synchronization in wireless sensor networks. In Proceedings of the 13th ACM Conference
on Computer and Communications Security (CCS ’06), November 2006, pp. 36–42.
CHAPTER 18
Secure Localization Systems:
Protocols and Techniques in Wireless
Sensor Networks
AZZEDINE BOUKERCHE
School of Information Technology and Engineering, University of Ottawa, Ottawa, Ontario K1N
6N5, Canada
HORACIO A. B. F. OLIVEIRA
School of Information Technology and Engineering, University of Ottawa, Ottawa, Ontario K1N
6N5, Canada; Federal University of Minas Gerais, Brazil; and Federal University of Amazonas,
Brazil
EDUARDO F. NAKAMURA
Federal University of Minas Gerais, Brazil; and FUCAPI—Analysis, Research, and Technology
Innovation Center, Brazil
ANTONIO A. F. LOUREIRO
Department of Computer Sciences, Federal University of Minas Gerais, Belo Horizonte, Brazil
18.1 INTRODUCTION
In Chapter 1, wireless sensor networks (WSNs) [1–8] are defined has a network composed of a large number of sensor nodes that cooperate among themselves to monitor
an area of interest. This type of network has become popular due to its wide applicability, including many different areas such as the environmental, medical, industrial, and
military fields. For military applications, WSNs show a number of desirable characteristics such as (a) being autonomous systems able to be deployed in remote—possibly
hostile—environments and (b) their ability to perform tasks such as battlefield surveillance or enemy tracking as well as security monitoring of military facilities. However,
one of the main challenges in these critical applications, as shown in Chapter 16, is the
security issue. We need to avoid attacks on networks located in hostile environments
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
521
522
SECURE LOCALIZATION SYSTEMS: PROTOCOLS AND TECHNIQUES
as well as attacks to sensor networks that monitor the security of military facilities in
order to guarantee the reliability of the provided data and requested tasks.
As mentioned in Chapter 11, a WSN is able to monitor a number of physical
properties such as temperature, humidity, pressure, ambient light, and movement.
However, all these gathered data—and, thus, the sensor nodes themselves—need
to be localized in space in order to identify the events’ location. This positioning
task is the main goal of the localization systems [6, 8–18]. Since positioning data
are used not only to locate events but also as base information for routing [19–22],
density control, tracking, and a number of other protocols, the localization systems are
considered as key parts of WSNs. By performing key essential functions, localization
systems can be used as a security target that could compromise the whole functioning
of a WSN, which could lead to incorrect military plans and decision-making, among
other problems. Thus, these systems should be secured in order to protect the whole
network and avoid the consequences of having hostile nodes and untrusted data traffic
affecting the performance and function of the WSN. In this chapter, we show how
current localization systems are vulnerable to these security attacks and how proposed
techniques can be used to prevent or make these attacks difficult to perform in WSNs.
As we will see, for each proposed technique, a set of network resources (e.g.,
nodes’ energy, media access, processor, and memory consumption) are required in
order to implement and maintain the provided security. While in most networks this is
not a problem, in WSNs this issue becomes a little bit more complicated due to their
resource limitations. In this case, as we show in this chapter, we need to decide on the
required level of security, which is application-dependent, and how many resources
can be spent in providing these levels of security. Depending on this cost–benefit
analysis, we can decide which solution or what security techniques will be used to
secure the WSN.
The remainder of this chapter is organized as follows. In the next section, we briefly
present an overview and definition of a secure localization system. Section 18.3
identifies the vulnerabilities to which localization systems are exposed, while in
Section 18.4 we show some techniques used to eliminate these vulnerabilities. Finally, Section 18.5 presents our conclusions and future directions for secure localization systems.
18.2 PROBLEM STATEMENT
Before defining secure localization systems, we will first take a look at some general
concepts and definitions used in normal localization systems. From the viewpoint
of localization systems, we have basically two types of nodes: regular nodes and
beacons. Regular nodes, also known as unknown, free, or dumb nodes, refer to the
nodes in the network that have no knowledge of their position and have no special
hardware to acquire this information. The beacon nodes, also known as landmarks,
anchors, or locators, are the nodes that do not need a localization system in order to
estimate their physical positions; in fact, they form the base of these systems. Their
position is obtained by manual placement or by external means such as GPS.
PROBLEM STATEMENT
523
Figure 18.1. The division of localization systems into three distinct components.
Therefore, in a localization system, we want to solve the following problem: Given
a multihop network and a set of beacon nodes with their known positions, we want
to find the position (e.g., latitude, longitude) of regular nodes based on available
information. For more information about the localization systems in WSNs, the reader
can refer to Chapter 11, in which the localization systems are divided into three distinct
Components (Figure 18.1):
1. Distance/Angle Estimation. This component is responsible for estimating information regarding the distances and/or angles between two nodes. Known
techniques used in this component include received signal strength indicator
(RSSI), time [difference] of arrival (ToA/TDoA), number of hops, or the angle
of arrival (AoA).
2. Position Computation. This component is responsible for computing a node’s
position based on available information about the distances/angles and positions of reference nodes. Some techniques used to compute a position include
trilateration, multilateration, or triangulation.
3. Localization Algorithm. This is the main component of a localization system. It
determines how the available information will be manipulated in order to allow
most or all of the nodes of the WSN to estimate their positions. It is a distributed
and usually multihop algorithm. Some known algorithms include the ad hoc
positioning system (APS) [12], and the directed position estimation (DPE) [6].
To be deployed in hostile environments, WSNs need a secure localization system,
in which we need to solve the localization problem but must also be aware
that we are in the presence of compromised nodes—malicious nodes or network nodes that have been corrupted by a malicious code—and/or in the presence of a compromised environment, where hostiles can change the environment’s
characteristics and may also have physical access to nodes.
524
SECURE LOCALIZATION SYSTEMS: PROTOCOLS AND TECHNIQUES
18.3 FRAGILITY OF CURRENT LOCALIZATION SCHEMAS
A number of attacks and hostile techniques can used to compromise a localization
system. Also, when used in WSNs, these systems become more vulnerable due to
a number of resources saving techniques. Thus, as a result, localization systems in
WSNs can be attacked in a number of different ways. In the last section, the localization systems were divided into different components and we also showed how these
components are strongly connected. Any small misbehavior in any of these components can greatly affect the whole functioning of a localization system. For instance, a
malicious erroneous distance estimation can cause a position miscomputation, which
will be propagated to the localization algorithm and will probably cause a major localization error for the sensor nodes. Due to this strong relationship between them,
any of these components can be used to attack a localization system, making these
systems very fragile and hard to secure. In the next sections, we discuss each of the
localization systems’ components, showing to which type of attack each is vulnerable.
18.3.1 Attacks on Distance/Angle Estimation
As mentioned before, distance estimations can be made based on signal strength,
time of arrival, or hop count analysis. In the first case, an easy way to generate erroneous estimations is by making a compromised node send a packet
with a greater or reduced transmission power in order to make neighboring nodes
think it is nearer or farther away than it really is. In the second case, the
transmission time of a packet can be delayed, causing problems to both ToA- and
TDoA-based systems. Hop-count-based distance estimations can be confused by
compromised nodes that advertise wrongly computed hop counts. In fact, since it is
a multihop algorithm, hop count estimations can also be affected by attacks to the
localization algorithm (Section 18.3.3).
When the nodes are secured but they are located in a compromised environment, both signal strength and time of arrival techniques can be targeted by
changing the physical medium—for example, by introducing noise, obstacles, or even
smoke. Also, angle of arrival-based systems could be compromised by deploying
magnets in the sensor field.
18.3.2 Attacks on Position Computation
Positions computation are highly dependent on the received positions from the reference node as well as on the estimated distances. To compute a position, a node needs
at least three known positions and three distance estimations. Any attack on distance
estimations, as shown in the previous section, has the main goal of affecting the position computation. However, some attacks can affect the position computation directly
by advertising wrong known positions. A wrong advertised position can lead to an
erroneous position computation even when the distance is correctly estimated. In this
case, a compromised node can not only send its own packet with a wrong position but
also send additional packets as if it were different nodes in different locations. In a
FRAGILITY OF CURRENT LOCALIZATION SCHEMAS
525
compromised environment, a GPS signal can be jammed, making it erroneous or even
not possible for beacon nodes to estimate their positions using their GPS receivers.
18.3.3 Attacks on the Localization Algorithm
While attacks on distance estimation and position computation components are attacks
specific only to localization systems, the third component of the localization systems,
the localization algorithm, shares the same kind of vulnerabilities associated with
other distributed systems. It happens because the localization algorithm is a distributed
and usually multihop algorithm executed by all nodes of the WSN. Some of the attacks
in which distributed systems are usually vulnerable include the Sybil, the replay, and
the wormhole attack.
r Sybil Attacks. In this type of attack, a malicious node makes it appear that it
is a set of different nodes and starts sending erroneous information. This erroneous information can be distance estimations, positions, number of hops, or
nonexisting nodes or beacons. Figure 18.2a illustrates this attack when node 6
claims to be also nodes 12 to 15.
r Replay Attacks. In a replay attack, a compromised node stores a received packet
(from a beacon node, for instance) and then resends the same packet later. Since
it is a copy of the original packet, neighboring nodes wrongly deduce that the
malicious node is the node that sent the original packet (Figure 18.2b). In this
case, since the distance estimation will be done based on the compromised node
while the position in the packet will be based on the original node, the position
computation will be affected. Both signal strength and time-based distance estimations are affected, since the packet sent by the compromised node will have
a different signal strength and different propagation time.
r Wormhole Attacks. In this case, the information received by one malicious node
on one side of the network is sent and replicated by another malicious node on
the other side of the network. The multihop path between these two attackers
is a wormhole in the sense that packets arriving on one side are transported
and received on the other side of the network, appearing as if it came from a
Figure 18.2. Attacks to the localization algorithms: (a) Sybil, (b) replay, and (c) wormhole.
526
SECURE LOCALIZATION SYSTEMS: PROTOCOLS AND TECHNIQUES
neighboring node. This attack is illustrated in Figure 18.2c. This type of attack
can greatly disrupt an insecure localization system by putting totally different
and erroneous reference points in the position computations.
18.3.4 Examples in Current Insecure Solutions
In the following, we cite some known proposed localization systems for WSNs and
we show how they can be vulnerable to a number of security attacks.
r Ad Hoc Positioning System (APS) [12]. This is a distributed, multihop localization system that computes the average size of a hop and uses this information
in the multilateration process. In this localization system, compromised beacon
nodes can report wrong positions and a wrong average hop size, causing an erroneous position computation for the nodes. Compromised regular nodes can
perform a Sybil attack to act like beacons, and they can also perform a replay
attack or a wormhole attack. In these last two cases, the number of hops will be
incorrect, which will cause a wrong estimation of the average size of a hop and
result also in an erroneous position computation for the nodes.
r Recursive Position Estimation (RPE) [8]. In RPE, estimated positions are
broadcasted to help other nodes estimate their position as well. A similar algorithm is the Directed Position Estimation (DPE) [6], which uses a recursion
with a direction. This kind of algorithm is more subject to security problems,
because the miscomputation or the hostile advertisement of a single wrong position will be used in the computation of another position that will in turn be used
to compute another position, until all nodes have an incorrect estimation. As an
example, a hostile node can perform a replay attack by replicating a received
position. Neighboring nodes will estimate their distance to the hostile node while
using the position in the replicated packet, which will start an error propagation
that will increase at each localization round.
r Localization with a Mobile Beacon (MBL) [23]. in the MBL, once the nodes
are deployed, the mobile beacon travels through the sensor field broadcasting
messages that contain its current coordinates. When a regular node receives more
than three messages from the mobile beacon, it computes its position, using
a probabilistic approach, based on the received coordinates and on the RSSI
distance estimations. A good aspect of this algorithm is that it is not vulnerable
to some of the attacks performed in distributed algorithms such as the wormhole
attack. It happens because network nodes do not need to exchange packets among
them. They only need to receive packets from the mobile beacon. On the other
hand, a compromised node can start sending packets as if it were the mobile
beacon (Sybil or replay) and thereby mislead surrounding nodes. Also, a new
hostile mobile beacon can be introduced in the network or the current mobile
beacon can be compromised in order to attack the whole network.
r Point-In-Triangulation Test. The APIT algorithm [7] uses triangles formed by
three beacon nodes, and a node decides if it is inside or outside these triangles
TECHNIQUES FOR A SECURE LOCALIZATION
527
by comparing its signal strength measurements with the measurements of its
neighbors. In the presence of a wormhole, a node can always think that it is
outside the triangles, since it will receive packets from distant nodes that seem
to be coming from neighbors.
18.4 TECHNIQUES FOR A SECURE LOCALIZATION
A number of security techniques have been proposed in recent years to enforce the localization systems and provide the secure positioning of nodes in hostile and military
applications of WSNs. Most of these solutions achieve security by using cryptography, detecting and blocking compromised nodes or information, making statistical
decisions, or filtering the positions used in computations.
18.4.1 Security through Cryptography
Cryptography is the first line of defense in most secure protocols. Since most security
attacks are performed by a malicious node trying to pass as an entity it is not or by
changing the values in data packets, these problems can be easily solved through
cryptography by using authentication and/or message integrity checks. When using
cryptography techniques, it is also possible to provide position Confidentiality, preventing malicious nodes and entities from gathering network information that could
be used in other attacks or to search for uncovered areas in the sensor field (since the
intruder would have access to the nodes’ positions).
The good news is that cryptography can be used to protect against externally
deployed hostile nodes that could execute any of the attacks cited. However, in the
presence of compromised nodes in the local network, the attackers gain access to
locally stored keys and passwords, which compromises the whole functioning of
any cryptographic system. For this reason, most secure localization algorithms use
noncryptographic security techniques, as will be shown in the next sections, and
rely on cryptography as a second line of defense. This is the case in the HiRLoc [24],
SeRLoc [25], and ROPE [26] localization algorithms, in which efficient cryptography
is used to secure beacon transmissions. In SPINE [27], cryptography can be used to
make an authenticated distance estimation, while it is used in reference 28 to assist
the detection of malicious beacon nodes. In most cases, it is assumed that network
nodes can establish pairwise secret keys.
Due to the limited availability of processor and memory resources in sensor nodes,
cryptography is avoided in some works. However, it is a gray area, since, in most cases,
if we need secure localization, we will also need secure media access, secure data
routing and transfer, secure time synchronization, and a lot of other secure algorithms,
all of which could take advantage of the same stored keys and implemented cryptographic algorithms. In these cases, cryptography can be a very plausible solution.
Also, cryptography techniques can be used to provide a layer of security for all three
localization system components (Distance/Angle Estimation, Position Computation,
and Localization Algorithm), explained in Section 18.3, by providing authentication
528
SECURE LOCALIZATION SYSTEMS: PROTOCOLS AND TECHNIQUES
and integrity checks of the exchanged packets. However, it does not protect against
compromised environment where it is possible to introduce obstacles, smoke, and so
on, in the sensor field. In this last case, misbehavior detection, as shown in the next
section, could be a possible solution.
18.4.2 Misbehavior Detection and Block
There are a number of cases where it is not possible to guarantee the security of the
nodes and communications such as when the cryptography is compromised or when
the environment is compromised. In these cases, the next layer of defense would be
to make the nodes themselves detect and block the compromised factors. One way to
make this happen and to defend against a possible attack is to observe the behavior of
the nodes over time and decide whether to trust them. These techniques can be used
mainly to protect the Position Computation component, since information gathered
from untrusted nodes can simply be ignored when computing the position of the nodes.
For instance, in Liu et al. [28] it is proposed a set of techniques for detecting
malicious beacon nodes. One technique detects malicious beacon nodes by comparing
the distance estimated by using the location information provided by these beacon
nodes and the distance estimated by means of the signal (e.g., RSSI, TDoA, AoA).
Another technique evaluates the round-trip time (RTT) between two neighbors, based
on the observation that the replay of a (malicious) beacon signal introduces extra delay.
Then, the base station (or sink node) uses such information about malicious beacons to
reason about the suspiciousness of each beacon node, and then it filters out malicious
beacon nodes accordingly.
Srinivasan et al. [29] extends the techniques proposed by Liu et al. [28] by using
a continuous scale and a reputation- and trust-based mechanism. The result is the
Distributed Reputation-based Beacon Trust System (DRBTS), which is a distributed
security protocol for excluding malicious beacon nodes. In DRBTS, each beacon node
monitors its neighborhood for suspicious beacon nodes and provides information by
maintaining and exchanging a Neighbor-Reputation Table, in such a way that other
sensor nodes can choose trustworthy beacons based on a voting approach.
18.4.3 Robust Position Computation
Robust position computation can be used in order to increase the security of a localization system. This technique can be used when both cryptography and misbehavior
detection is compromised. In this case, a way to deal with malicious nodes is to accept
that they will be present in the network and propose robust position computations that
can still work in the presence of bogus information. This is done mostly by using
statistical and outlier filtering techniques. In these cases, it is assumed that the benign
nodes outnumber the malicious ones. These techniques are used to protect against
(or be robust to) attacks on the Position Computation and Distance/Angle Estimation
components.
A technique proposed by Li et al. [30] uses the principle behind the least squares
data fusion technique to propose an adaptive least squares and least median squares
TECHNIQUES FOR A SECURE LOCALIZATION
529
position estimator. The idea is to use the least squares in the absence of attacks and the
least median squares in the presence of attacks, since the latter alternative tolerates
up to 50% outliers and still provides correct estimates. The authors show that the
use of traditional Euclidean distance is not robust to intentional attacks against base
stations, and they introduce robustness to fingerprinting localization by means of a
median-based distance metric.
Liu et al. [31] propose a method that uses the MMSE (Minimum Mean Squared
Estimation), which is a data fusion technique for obtaining improved estimations, to
identify and remove malicious location information. In this method, sensor locations
are estimated using the MMSE-based method. Then, the method verifies whether the
estimated location can be estimated from a set of consistent location references. If
not, the most inconsistent reference is identified and removed, and the node location is
estimated again. It repeats this process until all inconsistent references are discarded.
The mean square error of the distance measurements is used as an inconsistency
indicator.
A second method proposed by Liu et al. [31] is a voting-based location estimation
technique. In this method, the sensor field is quantized into a grid of cells, and each
reference node votes on the cells to which an unknown node may belong. Then, the
method selects the most voted cell(s) and uses the “center” of these cell(s) as the
estimated location. Voting results can be refined interactively to improve accuracy.
This method requires few resources and is suitable for current resource constrained
sensor nodes.
18.4.4 Location Verification
If all previous techniques fail to provide the required security, it is still possible to
check the computed positions using redundant information available on the nodes and
in the network. This is the core of some proposed works that focus on the reliability
of the final position computations rather than on avoiding or detecting compromised
nodes and attacks.
For instance, LAD [32] uses deployment knowledge, with a group-based deployment model, to check whether the computed positions of the nodes are consistent
with the known model and observations.
In reference 33 an algorithm is proposed for in-region verification, in which a
certain node can check whether another node really is inside the particular region that
it claims to be. The proposed protocol, called Echo, uses known physical properties
of both radio-frequency and ultrasound to compute distances and check whether a
node really can be inside the claimed region. These techniques can be used to provide
a layer of security for all three localization system components, since they only verify
the result of the overall localization system.
18.4.5 Secure and Simple Algorithms
Localization systems are vulnerable mostly due to the number of components available to be attacked. Another way to secure a localization system is to use simple, less
530
SECURE LOCALIZATION SYSTEMS: PROTOCOLS AND TECHNIQUES
dependable localization algorithms, such as GPS-free, range-free, and/or one hop
algorithms.
One example is SeRLoc [25], in which beacon nodes are equipped with a set of
higher-power-directional antennas. These nodes send a packet using an asymmetric
transmission that contains their position and the sector of the antenna in which the
packets are sent. Because it is a range-free single-hop localization algorithm, it is
protected against attacks aiming at altering range measurements and against regular compromised nodes. However, it does not protect against wormholes, which are
avoided by checking network properties such as sector uniqueness and communication range.
A technique similar to SeRLoc is used in HiRLoc [24], which has a greater accuracy
but an increased computational and communication complexity. Techniques like these
can be used to protect the Localization Algorithm Component.
Another simple and secure algorithm is the already-mentioned Localization with a
Mobile Beacon (MBL) [23]. Although MBL is not designed to be a secure algorithm
and the authors do not mention anything about security in their work, the algorithm
used in the MBL is quite simple, where a mobile beacon walks the network and
broadcasts its position information to near nodes. In this case, regular nodes only need
to listen to the beacon node and they never exchange messages among themselves.
By being a simple single-hop algorithm, this localization system is secured against a
number of distributed attacks such as the wormhole attack.
18.4.6 Comparison of Current Solutions
A widely known fact in network security is that there is no system that is totally safe
and secure. There will always be weak points and the question is simply whether they
are acceptable. In WSNs, this issue becomes a little more complicated due to resource
limitations. In this case, we need to decide on the required level of security, which is
application-dependent, and how many resources can be spent in providing these levels
of security. Depending on this cost–benefit analysis, we can decide which solution or
what security techniques will be used to secure the WSN. In Table 18.1, we compare
each of the studied proposals, showing which type of security they use as well as
some observations about them and their potential weaknesses. As we can see, most
security proposals rely on some kind of lightweight cryptography as a second line
of defense combined with other security techniques such as misbehavior detection,
robust position computation, location verification, and simple algorithms combined
with extra hardware.
18.5 CONCLUSIONS
In this chapter, localization systems were studied under the viewpoint of security. We
showed how an insecure localization system can be attacked in a number of ways to
compromise the whole functioning of a WSN and thus lead to incorrect military plans
and decision-making. First, the localization systems were divided into three different
TABLE 18.1. Secure Localization Systems Comparison
Algorithm
HiRLoc [24]
SeRLoc [25]
ROPE [26]
SPINE [27]
DRBTS [29]
LAD [32]
Cryptography
Encryption and authentication of
beacons’ communication.
Global preloaded keys.
Encryption and authentication of
beacons’ communication.
Global preloaded keys.
Misbehavior
Detection
Robust Pos.
Computation
Location
Verification
Simple
Algorithms
—
—
—
Yes
—
—
—
Yes
—
—
Encryption and authentication of
beacons’ communication.
Global preloaded keys.
Symmetric or public-key
—
Verifiable multicryptography for authenticated
lateration.
distance estimations.
—
Encryption using a network-wide Reputationgroup key.
and
trust-based
—
—
—
Echo [33]
—
—
Li et al. [30]
—
—
Liu et al. [28]
Liu et al. [31]
531
Beacon packets are authenticated Distances
using shared pairwise keys.
comp. and
RTT
Authentication with pairwise key
—
establishment.
Beacons verify
distances.
Yes
—
—
—
—
Deployment
knowledge
—
Physical
propagation of
sound/RF.
Robust statistical
—
methods.
—
—
—
Voting-based
—
—
Observation
Requires extra hardware
(directional antennas) in
beacon nodes.
Requires extra hardware
(sectored antennas) in beacon
nodes. Doesn’t consider
hostile beacons.
Requires extra hardware
(directional antennas) in
beacon nodes.
Nanosecond clocks. Uses
ultrasound. High number of
beacons.
Benign observations must be the
majority. Dense network.
—
Requires deployment knowledge.
—
Majority benign. No hostile
beacons. Uses Ultrasound.
—
Benign observations must be the
majority.
Requires redundant beacon
nodes.
Benign beacons must be the
majority.
532
SECURE LOCALIZATION SYSTEMS: PROTOCOLS AND TECHNIQUES
components: distance/angle estimation, position computation, and localization algorithm. We then went over each of these components, showing several techniques that
could be used to compromise them and, consequently, the whole localization system
when in the presence of compromised nodes and/or a compromised environment.
After that, we showed some examples of current insecure localization systems and
how they could be easily confused by some of the studied attacks. Finally, some of
the techniques used by current secure localization systems were studied in order to
show how to perform localization in the presence of hostile nodes and compromised
environments.
18.6 EXERCISES
1. Cite the main differences in providing security in (a) wired networks, (b) wireless
ad hoc networks, and (c) wireless sensor networks.
2. Why are localization systems for WSNs so hard to secure? Cite the main challenges in providing security in each of the three components of the localization
systems: (a) distance estimation, (b) position computation, and (c) localization
algorithm.
3. Cryptography is not so easy to be implemented in a WSN. Cite and explain the
main challenges in providing a WSN with a cryptographic system.
4. Cite the main techniques to provide a security layer in a WSN. Show the pros
and cons of each technique.
5. Choose a different nonsecure localization system (not studied in this chapter),
explain how it works, show how it is possible to be attacked, and finally, show
possible solutions to secure this chosen localization system.
BIBLIOGRAPHY
1. A. Boukerche. Handbook of Algorithms for Wireless Networking and Mobile Computing (Chapman & Hall/Crc Computer & Information Science), Chapman & Hall/CRC,
Boca Raton, FL, 2005.
2. I. F. Akyildiz, W. Su, Y. Sankarasubramaniam, and E. Cyirci. Wireless sensor networks: A
survey. Computer Networks, 38(4):393–422, 2002.
3. D. Estrin, L. Girod, G. Pottie, and M. Srivastava. Instrumenting the world with wireless
sensor networks. In International Conference on Acoustics, Speech, and Signal Processing
(ICASSP 2001), Salt Lake City, UT, June 2001, pp. 2033–2036.
4. M. Ilyas and I. Mahgoub. Handbook of Sensor Networks: Compact Wireless and Wired
Sensing Systems, CRC Press LLC, Boca Raton, FL, 2004, Chapter 20.
5. G. J. Pottie and W. J. Kaiser. Wireless integrated network sensors. Communications of the
ACM, 43(5):51–58, 2000.
6. H. A. B. F. Oliveira, E. F. Nakamura, A. A. F. Loureiro, and A. Boukerche. Directed position
estimation: A recursive localization approach for wireless sensor networks. In 14th IEEE
BIBLIOGRAPHY
533
International Conference on Computer Communications and Networks, San Diego, October
2005, pp. 557–562.
7. T. He, C. Huang, B. M. Blum, J. A. Stankovic, and T. Abdelzaher. Range-free localization
schemes for large scale sensor networks. In MobiCom ’03: Proceedings of the 9th Annual
International Conference on Mobile Computing and Networking, ACM Press, New York,
2003, pp. 81–95.
8. J. Albowicz, A. Chen, and L. Zhang. Recursive position estimation in sensor networks. In
The 9th International Conference on Network Protocols, November 2001, pp. 35–41.
9. L. Doherty, K. S. J. Pister, and L. E. Ghaoui. Convex position estimation in wireless sensor
networks. In IEEE Conference on Computer Communications 2001, Vol. 3, Anchorage,
AK, April 2001, pp. 1655–1663.
10. A. Savvides, C.-C. Han, and M. B. Strivastava. Dynamic fine-grained localization in ad-hoc
networks of sensors. In 7th ACM/IEEE International Conference on Mobile Computing and
Networking, Rome, Italy, 2001, pp. 166–179.
11. Y. Shang, W. Ruml, Y. Zhang, and M. P. J. Fromherz. Localization from mere connectivity.
In Proceedings of the 4th ACM International Symposium on Mobile Ad Hoc Networking
and Computing, Annapolis, MD, 2003, pp. 201–212.
12. D. Niculescu and B. Nath. Ad hoc positioning system (aps). In IEEE Global Communications Conference (GlobeCom ’01), San Antonio, TX, November 2001, pp. 2926–
2931.
13. N. Bulusu, J. Heidemann, D. Estrin, and T. Tran. Self-configuring localization systems: Design and experimental evaluation. Transactions on Embedded Computing Systems, 3(1):24–
60, 2004.
14. H. A. B. F. Oliveira, E. F. Nakamura, A. A. F. Loureiro, and A. Boukerche. Error analysis
of localization systems in sensor networks. In 13th ACM International Symposium on
Geographic Information Systems, Bremen, Germany, November 2005, pp. 71–78.
15. A. Boukerche, H. A. B. F. Oliveira, E. F. Nakamura, and A. A. F. Loureiro. A Voronoi approach for scalable and robust dv-hop localization system for sensor networks. In Computer
Communications and Networks, 2007. ICCCN 2007. Proceedings of 16th International
Conference on, 2007, pp. 497–502.
16. A. Boukerche, H. A. B. F. Oliveira, E. F. Nakamura, and A. A. F. Loureiro. Localization
systems for wireless sensor networks. IEEE Wireless Communications—Special Issue on
Wireless Sensor Networks, 2007, pp. 25–30.
17. H. A. B. F. Oliveira, E. F. Nakamura, A. A. F. Loureiro, and A. Boukerche. Localization
in time and space for sensor networks. In AINA’07: 21st IEEE International Conference
on Advanced Information Networking and Applications, Niagara Falls, Canada, May 2007,
pp. 539–546.
18. A. Boukerche, H. A. B. F. Oliveira, E. F. Nakamura, and A. A. F. Loureiro. Vehicular ad-hoc
networks—A new challenge for localization. Elsevier Computer Communications, 2008,
to appear.
19. A. Boukerche, H. A. B. F. Oliveira, E. F. Nakamura, and A. A. F. Loureiro. A novel locationfree greedy forward algorithm for wireless sensor networks. In Proceedings of the 2008
IEEE International Conference on Communications (ICC 2008), Beijing, China, May 2008,
IEEE CD-ROM.
20. A. Boukerche, H. A. B. F. Oliveira, E. F. Nakamura, and A. A. F. Loureiro. Towards an
integrated solution for node localization and data routing in sensor networks. In ISCC ’07:
534
SECURE LOCALIZATION SYSTEMS: PROTOCOLS AND TECHNIQUES
12th IEEE Symposium on Computers and Communications, Aveiro, Portugal, July 2007,
pp. 449–454.
21. B. Karp and H. T. Kung. Gpsr: Greedy perimeter stateless routing for wireless networks.
In Proceedings of the 6th Annual International Conference on Mobile Computing and
Networking, Boston, MA, 2000, pp. 243–254.
22. Y. Yu, R. Govindan, and D. Estrin. Geographical and energy aware routing: A recursive data
dissemination protocol for wireless sensor networks. Technical Report CSD-TR-01-0023,
Computer Science Department, UCLA, 2001.
23. M. L. Sichitiu and V. Ramadurai. Localization of wireless sensor networks with a mobile
beacon. In Proceedings of the 1st IEEE International Conference on Mobile Ad hoc and
Sensor Systems (MASS 2004), Fort Lauderdale, FL, October 2004, pp. 174–183.
24. L. Lazos and R. Poovendran. Hirloc: High-resolution robust localization for wireless sensor
networks. IEEE Journal on Selected Areas in Communications, 24:233–246, 2006.
25. L. Lazos and R. Poovendran. Serloc: Secure range-independent localization for wireless
sensor networks. In Proceedings of WiSe’04, ACM Press, New York, 2004, pp. 21–30.
26. L. Lazos, R. Poovendran, and S. Capkun. Rope: Robust position estimation in wireless
sensor networks. In Proceedings of IPSN, April 2005, pp. 324–331.
27. S. Capkun and J.-P. Hubaux. Secure positioning of wireless devices with application to
sensor networks. In Infocom’05, Miami, March 2005, pp. 1917–1928.
28. D. Liu, P. Ning, and W. Du. Detecting malicious beacon nodes for secure location discovery in wireless sensor networks. In 25th ICDCS, Washington, DC, 2005, IEEE Computer
Society, pp. 609–619.
29. A. Srinivasan, J. Teitelbaum, and J. Wu. Drbts: Distributed reputation-based beacon trust
system. In 2nd IEEE DASC, 2006, pp. 277–283.
30. Z. Li, W. Trappe, Y. Zhang, and B. Nath. Robust statistical methods for securing wireless localization in sensor networks. In IPSN ’05: Proceedings of the 4th International
Symposium on Information Processing in Sensor Networks, Los Angeles, CA, IEEE Press,
New York, 2005, p. 12.
31. D. Liu, P. Ning, and W. Du. Attack-resistant location estimation in sensor networks. In
IPSN ’05: Proceedings of the 4th International Symposium on Information Processing in
Sensor Networks, Los Angeles, CA, IEEE Press, New York, 2005, p. 13.
32. W. Du, L. Fang, and P. Ning. Lad: Localization anomaly detection forwireless sensor
networks. In 19th IPDPS, Washington, DC, 2005, IEEE Computer Society, p. 41.1.
33. N. Sastry, U. Shankar, and D. Wagner. Secure verification of location claims. In WiSe’03,
New York, 2003, ACM Press, New York, pp. 1–10.
INDEX
Abstraction-based sensor programming, 46
ACE algorithm, 173
pseudocode for, 173
Active sensor network (ASN) project, 228
Ad hoc networks, 51, 63, 267
epidemic models, 63
paradigm of, 267
techniques, 438
Ad hoc nodes, 130
Ad hoc positioning system (APS), 325,
326–328, 523, 526
Ad hoc routing protocols, 66
Adaptive threshold-sensitive energy-efficient
sensor network protocol (APTEEN),
144
Advertisement messages (ADV), 55
Aggregation-and-forwarding (AFN), 236
Algorithm design, 90
Algorithm for cluster establishment (ACE),
171
Algorithm for robust routing in volatile
environments (ARRIVE), 155–156
Algorithmic models, 96
Alternating-current power adaptor, 22
Ambient conditions, 225
humidity, 225
light intensity, 225
pressure, 225
temperature, 225
Analog to digital converter (ADC), 356
Angle estimation method, 344
Angle of arrival (AOA) estimation, 348–349
Angle of arrival (AOA) measurements, 359
Angle/direction of arrival (AoA/DoA),
314
APIT algorithm, 323, 526
Arbitrary nodes, 82
Area coverage, 232
ARRIVE algorithm, 155
ARRIVE protocol, 155
flow chart, 155
Art-gallery model, 230
Attribute-based routing protocols, 133–135
Autonomous underwater vehicles (AUV),
268
Base station, 492
placement, 238–241
position, 242
protection, 256
relocation, 254
Beacon identification, 355
Beacon nodes, 334
Beacon vector routing (BVR), 214
BFS, see Breadth first search
Binary interference models, 89
Blom’s key pre-distribution method, 491
Bounded independence graph (BIG), 80
model, 80
Braided multipaths, 149
design, 150
Breadth first search (BFS) 164, 166, 204
BFS algorithm, 171
Broadcast-based dissemination, 54
Broadcast transmission protocol, 55
Broadcasting technique, 187
BVR algorithm, 215
overview of, 215
BVR protocol, 214
Carrier sensing multiple access (CSMA),
506
Algorithms and Protocols for Wireless Sensor Networks, Edited by Azzedine Boukerche
Copyright © 2009 by John Wiley & Sons Inc.
535
536
INDEX
Cartesian coordinate system, 213
Cascaded sensors movement, 252
Center for embedded networked sensing
(CENS), 44
Centralized analytical model, 272
Chessboard clustering protocol, 33
Cipher block chaining (CBC), 499
Cipher in counter mode (CTR), 496
Civilized graphs, 97
Cluster-based distributed localization
scheme, 359
Cluster-based graph network, 187–188
Clustered network, 250
SMART, 250
Clustering techniques, 163
Communication range method, 315
advantage of, 315
Compromised node, 514
Computational geometry, 78
based algorithm, 240
Connection-tree (C-tree), 152
Constant transmission power, 85
Constrained shortest-path algorithm, 135
Controlled sink mobility, 281–283
Correlation model, 107
Cricket compass project, 348
Cricket location support system, 334
Cross-link detection protocol (CLDP),
208
Data aggregation, 65
Gossip algorithms, 65
Data collectors, 268
AUVs, 268
Robots, 268
Data dissemination process, 54–56
SPIN, 54
Data dissemination protocol, see Scalable
energy-efficient asynchronous
dissemination
Data fidelity, 237
Data forwarding phase, 498
Data MULEs, 293
Data packets, 278
Data propagation protocols, 476
Data redundancy, 28
Delay-constrained traffic, 255
Deluge protocol, 61–63
Deluge protocol, 62
MAINTAIN states, 62
RX states, 62
TX states, 62
Deluge state machine, 61
Deployment objectives, 231
Deployment schemes, 228
Depth first search (DFS), 164, 166
Design optimization strategies, 227
Differential-equation-based approach, 70
Differential-rate-equation-based modeling
methods, 71
Dijkstra’s algorithm, 233
Dijkstra’s least-cost path algorithm, 252
Directed diffusion algorithm, 134
pseudocode, 134
Directed position estimation (DPE),
329–331
Direct transmission phase, 427
Directed position estimation (DPE), 523
Distance/Angle estimation, 523
Distributed reputation based beacon trust
system (DRBTS), 528
Disjoint multipaths, 149
Disk graph, 78
Disk graph model (QUDG), 79
Distance-based coordinates, 210
Distance estimates, 349
Distance/angle estimation, 310, 311
Distances estimation methods, 343–344
mechanisms for, 343
Distributed algorithms, 91–92, 98, 271
Distributed dominating set-based
algorithms, 175–176
Distributed sensor networks with collective
computation (DSN-CC), 27
Divide-and-conquer method, 178
DM model, 89
Dominating set problem (DS), 90
DPE algorithm, 331
phases, 331
Drift error, 504
DSN-CC project, 28, 167, 176
DSN-CC system, 28
Dust-sized smart sensors, 342
DV-distance, 326
DV-Hop, 326
DV-Hop algorithm, 326
DV-Hop solution, 359
Dynamic packet state (DPS), 151
INDEX
EAD algorithm, 476
Echno protocol, 529
E-health systems, 268
telemedicine equipment, 268
wearable sensors, 268
Efficient aggregation, 65
geographic Gossip, 65
smart Gossip, 65–66
Embedded networked sensing, 44–45
End-to-end transmission rate drops, 30
performance of, 30
Energy-aware data-centric routing (EAD),
135
Energy balanced data propagation problem,
453
Energy-constrained nodes, 32
nodes E1, 32
nodes E2, 32
Energy dissipation, 452, 463
Energy-efficient algorithms, 423
Energy-efficient communication protocols,
106
Energy-efficient protocols, 425
energy-balanced protocol (EBP), 425
local target protocol (LTP), 425
probabilistic forwarding protocol (PFR),
425
variable transmission range protocol
(VTRP), 425
Energy-efficient multipath routing, 149
mechanisms, 149
Energy-efficient routing protocol, 269
use, 269
Epidemic algorithms, 68
Epidemic algorithms classification, 58
Epidemic broadcast-based dissemination,
54
Epidemic models, 68
Epidemic parameter, 53
Epidemic routing protocol, 62, 66–67
Epidemic theory, 52–54
definition, 52
overview of, 52–54
Epidemic theory, 52, 53
epidemic parameter, 53
overview of, 52
Epidemiological studies, 53
definition, 53
ERUP protocol, 273
537
Estimating angles method, 344
mechanisms for, 344
Euclidean graph, 78
Euclidean method, 327
Euclidean plane, 82
Eulerian broadcasting procedure, see
Broadcasting technique
Event-driven heterogeneous WSNs, 43
Event-to-sink reliability notion, 122
Exponential autocorrelation function, 113
Exposure-based coverage assessment, 232
Fault-tolerant data propagation protocols,
461 447
Fault tolerance, 131
Field sources, 118
spatiotemporal characteristics of, 118
Firecracker dissemination componets, 58
broadcast protocol, 58
mechanism, 57–59
routing protocol, 58
seed selection, 58
Firecracker protocol, 57–59
Flat homogeneous WSN, 36
Flat network topology, 241
Flooding time synchronization protocol
(FTSP)
Gabriel graph construction, 205
Galileo device, 98
Galileo device, see GPS
Gallager, Humblet, and Spira (GHS)
algorithm, 177
Gaussian distribution, 350
Gaussian random variables (JGRVs), 109
GDSTR, 208
General graph (GG), 79
General metric spaces, 81
General weighted graph (GWG), 89
Generalized network of miniature
environmental sensors (GNOMES), 44
Geographic routing, 138, 140
energy-efficient forwarding strategies, 140
Geographic routing algorithm (GRA), 139,
201
Geographical positioning system (GPS),
141, 325, 335, 341
device, 345
GPS receiver, 334, 341, 346
538
INDEX
Geosensor network, 68
definition, 68
epidemic approach, 68
flooding approach, 68
location-constrained approach, 68
Global algorithms, 91, 92, 93
Global positioning system (GPS), 22, 98,
196, 334
advantages, 334
disadvantages, 334
Gossip-based approach (63–66) 63
GOSSIP protocol, 64
parameters, 64
Gossip algorithms, 65
Gradient broadcast (GRAB), 137
Graph-theoretic modeling technique, 70
Graph theory, 77
Greedy algorithms, 91, 92
Greedy distributed spanning tree protocol
(GDSTR) algorithm, 208
Greedy maximum residual energy (GMRE),
288, 293, 294
protocol, 288
Grid-based sensor network, 37
Grid points, 38
graphical presentations, 38
Ground-based VOR stations, 344
Hard-wired MAC addres, 130
HELLO packet, 140
Hello flood attacks, 484
Heterogeneous camera sensor network,
25
Heterogeneous transmission, 86
Heterogeneous wireless sensor networks, 23,
33, 36, 37, 38–42, 47
applications, 41
architectures for, 23
coverage in, 37
differentiated coverage, 38
goal of, 40
onadequate theory of, 42
stochastic coverage, 39
systems infrastructure, 42
Heterogeneous wireless sensor networks
projects, 42
resource-oriented protocol, 33
Hierarchical architecture, 23
Hierarchical protocols, 142
High-end nodes, 35, 45
intel XScale-based nodes, 45
High-end sensor nodes, 34, 35, 38
High-resolution data, 132
Higher-fidelity image, 27
Higher-priority neighbors, 95
HiRLoc techniques, 530
Homogeneous ad hoc networks, 30
Homogeneous WSNs, 30
Homogeneous mixing model, 53
Homogeneous WSNs, 37
Hop-based coordinates, 214
Hop-b-hop data propagation protoco,
440–445
Hop-by-hop transmissions, 451
Hop interference (UHI), 87
ID distributions, 98
ILP model, 274
In-home sensor nodes, 32
In-network processing, 28
Integer linear programming (ILP), 145, 242
Inter-base-station network, 260
Interference issues, 84
Interference models, 84, 90
overview of, 90
Interfering transmissions, 86
Internet routing techniques, 195
Intersection graph, 78
Initialization vector, 496
Intrusion-tolerant routing in wireless sensor
networks, 497
IP-like routing techniques, 196
background, 196
overview, 196
Iterative multilateration algorithms, 351
Joint source-channel coding, 116
K-hop neighborhood, 88
K-local algorithm, 95, 96
Kephart-White (KW) model, 69, 70
Key management schemes, 487–492
basic random key pre-distribution scheme,
488–489
extended random key pre-distribution
scheme, 489–490
master-key-based key predistribution
scheme, 487–488
INDEX
multiple space key pre-distribution
scheme, 491–492
Kruskal’s algorithm, 91
Large phased-array antennas, 344
Lamport’s method, 493
Layering-based security approach, 484–485
LEACH, 144
LEACH protocol, 143
pseudo-code, 143
LID values, 152
Limited destination information, 141
Linear chain of causality, 93
Lithium-ion battery, 22
Local algorithms, 94
Local dominating set algorithm, 95
Local randomized greedy algorithm (LRG),
176
Local target protocol (LTP), 462
Low-cost sensor devices, 437
Localization algorithm, 324, 325, 334, 523
categories, 324
Localization schemes, 349
Localization system, 307–310
components of, 310, 523
importance of, 309
requirements of, 310
Localization system division, 523
Localization with a mobile beacon (LMB),
331–334, 526
advantage of, 333
algorithm of, 333
Localized MDS algorithm, 93, 94, 171
Location-aware anchor nodes, 359
Location discovery schemes, 342–343
Location errors (imprecise GPS), 141
Location estimation, 341
aircraft navigation, 341
maritime, 341
robotics, 341
tactical missions, 341
transportation, 341
Location fingerprinting method, 323
Logical Coordinate Routing (LCR), 216
Los Alamos National Laboratory, 27, 29
Lossy wireless sensor networks, 140
geographic routing in, 140
Low-complexity techniques, 342
Low-cost techniques, 342
539
Low-cost sensor devices, 437
Low-end sensor node, 34, 35, 38
Low-energy directional broadcast, 427
Low-fidelity cameras, 25, 27
Low-power sensor devices, 437
Malicious code propagation, 69–70
epidemic models, 69
Malory, 514
Manhattan norm, 81
Markov chain, 65
Matroid theory, 92
Mediator-wrapper, 41
Medium access control (MAC), 96, 119
MAC algorithms, 496, 499
MAC protocol, 86–87, 120
Medium access mechanism, 97
Meshed multipath routing (M-MPR), 145
steps, 145
Meshed multipath routing algorithm, 148
pseudocode description, 148
Message complexity, 93
Message’s destination, 98
Message-passing model, 96
Metric space, 81
Microelectromechanical (MEMS) systems,
423
MFR, see Most forward within radius
scheme
MILP formulation, 283–288
Minimum cost forwarding algorithm
(MCFA), 138
Minimum dominating set (MDS), 91, 167
Minimum mean square estimate (MMSE),
350
Minimum sanning tee (MST), 176
Min-two uniform targets protocol (M2TP),
444
Mixed integer linear programming (MILP)
model, 275
Mobile ad hoc networks (MANET), 67–69,
480
Mobile beacons, 331
Mobile element scheduling (MES), 272
Mobile nodes, see MULEs
Mobile relays, 271
Mobile sensors, 249
Mobile sensor network, 67
Mobile sensor nodes, 32
540
INDEX
Mobile sinks, 273, 274
use of, 274
Mobile ubiquitous LAN extension
(MULEs), 271, 276
approach, 276, 277, 279
architecture, 293
beacons, 278
mobility, 273
nodes, 280
nonshareable (NS) nodes, 280
shareable (SH) nodes, 280
system, 277
Mobile wireless sensor networks, 183
Moore’s law, 479
Most forward within radius scheme, 198,
199
Motion control algorithm, 279
Multi-base-station clustered sensor network
architecture, 259
Multi-base-station positioning, 260
Multidimensional scale (MDS), 323
Multihop (ad hoc) communications, 267
Multihop network topology, 510
Multihop routing, 281, 290
protocol, 269
Multilateration algorithm, 352
illustration of, 352
Multinode relocation, 258
Multipath routing, 145–148
Multiple base-stations, 241
Multiple hops, 351
Multiple sensor indoor surveillance (MSIS)
project, 228
Multitier multimodal camera sensor
network, 45
Network algorithm, 278
Network connectivity, 234
Network-controlled sink mobility, 274
Network lifetime, 35, 99, 235
Network operation model, 253
Nearest forward progress, (NFP), 198,
199
challenges, 206
drawbacks, 206
Nearest forward progress, NFP
Node-move-out algorithm, 188
Node identifiers, 98
Node supervision, 482
Nodes repositioning, 245
NoGeo algorithm, 211
NoGeo method, 210
Non-power-constrained nodes, 34
Object recognition, 27
Omnidirectional radio antennas, 78
Online algorithms, 96
Optical (laser) transmission, 423
Optimal node placement, 227
Optimal offline algorithm, 96
Optimal transmission ranges (OTR)
approach, 201
Optimized sensor placement, 232
Packet delay components, 506
Pan-tilt-zoom (PTZ) cameras, 25
Passive information gathering, 481
Path-loss exponent, 85, 98
Peer-to-peer computing, 427
energy efficiency of, 436
Peer-to-peer generalized clustering model,
185–186
PEGASIS, see LEACH
Perimeter mobility (PM), 291
Perimeter node, 140
Periodic data collection model, 238
PFR, 448–450
correctness of, 435
energy efficiency of, 436
properties, 448–449
protocol, 445
robustness of, 451
Placement algorithm, 235
illustration of, 235
optimized positioning, 238
Planar methods, 207
Planar subgraph methods, 204–206
Point-in-triangulation (PIT) test, 359
Point-to-point routing solutions, 196
Poisson models, 97
Poisson process, 71
Polynomial-time approximation algorithm,
177
Polynomial time approximation scheme
(PTAS), 91
Position-based routing, 196
algorithms, 197
protocols, 195
INDEX
Position component, 523
Position computation methods, 315, 528
Post-deployment relocation, 250
Post-deployment sensor relocation, 248–254
Post facto synchronization approach, 509
Power-aware chessboard-based adaptive
routing (PCAR), 31
Power-constrained nodes, 34
Power-efficient gathering in sensor
information systems (PEGASIS),
143–144
Power exponential model, 120
Pre-deployed sensor network, 431
Probabilistic approaches, 319
Probabilistic forwarding protocol (PFR),
445–448
Protocol model (PM) 88
Public key cryptography, 472
Quality-of-service (QoS), 27, 201
parameters, 151
Quality-of-service support, 47
Quasi unit disk graph (QUDG), 79, 80, 82,
83
model, 79, 81
Query-driven heterogeneous WSNs, 43
RADAR indoor location system, 348
Radiation detection nodes, 29
Radiation detectors, 28, 29
acoustic sensors, 28
atmospheric sensors, 28
magnetometers, 28
seismic sensors, 28
video cameras, 28
Radioactive source detection, 28
staged architecture, 28
Radio frequency (RF), 423
channel, 334
Radiological dispersal devices (RDDs), 27
Radio propagation models, 312
Random bit string, see Initialization vector
Random deployment schemes, 231
Random distribution model, 37
Random graph model, 71
Random key distribution technique, 71
Random mobility (RM), 290
model, 67
Random node distribution, 97
541
Random transmission errors, 99
Randomized sensor placement, 230
Range-free techniques, 359
Rayleigh distributed random variable, 80
RBS protocol, 509
Real-time data, 27
Received signal strength indicator (RSSI),
311–313, 343
RECRUIT message, 173
Recursive position estimation, (RPE), 325,
526
algorithm, 328–329
Reference broadcast synchronization, 509
Relative neighborhood graphs (RNGs),
204
Relative removal rate, 53
Reliable event communication, 122
spatiotemporal correlation, 122
Reliable information forwarding using
multiple paths (ReIn-ForM) protocol,
149
Reprogramming algorithm, 60
properties, 60
Reprogramming protocol, 59
Request message, 55
Resource-oriented protocol (ROP), 32, 33
analysis of, 36
performance of, 33
RFID tags, 98
Robust distributed algorithms, 424
Robust distributed protocols, 424
Robust position computation, 528
Round trip time (RTT), 279
Route discovery phase, 498
Route request (RREQ) packet, 33
Routing algorithms, 130
Routing protocols, 130
applications, 131
design issues, 131, 132
Routing protocols taxonomy, 129
Replay attack, 525
SAR algorithm, 138
Scalable energy-efficient asynchronous
dissemination (SEAD), 273
Scale-free topology, 53
Search Phase, 441 427
Secure node-to-node communication, 472
Secure localization technique, 527
542
INDEX
Secure sensor networks, 70
compromise propagation, 70
Secure time synchronization approaches,
515–518
Security protocols for sensor networks,
495–496, 481
Semidefinite program (SDP), 323
SensEye, 25, 26
staged architecture, 26
SensEye heterogeneous camera sensor
network, 27
SensEye system, 27, 28
Sensing model, 232
Sensor field broadcasting messages, 331
Sensor network, 37, 51, 77, 90, 97, 98, 132,
171, 195, 342, 346–347, 353, 424, 438,
452, 479
approaches and obstacles, 195
angle estimation, 353
distributed algorithms for, 98
objective of, 97
Sensor networks algorithm, 77
Sensor network artitechture, 226
Sensor networks localization, 345–347
Sensor network models, 77
Sensor networks routing protocols, 132–135
cryptography, 486–487
distributed protocols, 424
role, 479
security attacks, 482–484
security classes, 481–484
symetric cryptography, 486
Sensor networks limitation, 480
network, 480
node, 480
physical, 480
Sensor networks time synchronization, 506
challenges, 506
design issues, 507
Sensor network topology, 123
Sensor node, 22, 138
components, 22
connectivity, 78
modeling the, 78
placement, 228
Sensor nodes mobility, 270–271
Sensor repositioning schemes, 247–248
Sensor routing protocols, 133
categories of, 133
Sensor-to-MULE communication, 277
Sensor-to-sink transmissions, 271
Sequential assignment routing (SAR),
138–139
SeRLoc techniques, 530
SER protocol, 151
SER protocol parameters, 153
Shorted path first (SPF) algorithm, 164
Shortest-path energy-aware routing, 135
Short-range communications, 274
Signal signature database, 323
Signal-to-interference-plus-noise ratio,
85, 89
Signal to interference ratio, 85
Signal to noise ratio (SINR), 85, 463
Signal to noise ratio model, 84, 85, 86
Single base station, 238
Single node, 54
base station, 54
multiple sensor nodes, 54
Sink mobility rates, 286
S-I-S model, 53, 67, 69, 70
Sleep-awake probabilistic forwarding
protocol (SW-PFR), 462
Sleeping time, 99
Small-scale robot squads, 268
Smart dust cloud, 426
Smart dust propagation protocol, 442
Smart dust protocols, 443
Smart gossip argument, 66
Space complexity, 93
Spatiotemporal correlation theory, 105, 106
spatial correlation, 106
temporal correlation, 106
SPEED architecture, 139
SPEED protocol, 139
SPIN-BC protocol architecture, 55, 56
SPIN family, 54
SPIN protocol, 54
SPIN-BC, 54
SPIN-EC, 54
SPIN-PP, 54
SPIN-RL, 54
Stargate family processors, 22
Stateless geographic nondeterministic
forwarding (SNFG), 139
Static base-station positioning, 244
approaches for, 244
Static sinks, 269
INDEX
Stochastic coverage, 39
analytical expressions, 39
Stream-enabled routing (SER), 151
Susceptible infected recovered (S-I-R)
model, 52
Susceptible infected susceptible (S-I-S)
model, 52
Sybil attacks, 484, 525
Synchronization protocols, 509
Synchronization schemes components
access time, 506
propagation time, 506
receive time, 507
send time, 506
Synchronous dynamic random access
memory, 22
System lifetime, 30
Table-driven multipath approach, 138
Tag-based data dissemination technique, 58
TASC algorithm, 181
pseudocode for, 181
TASC cluster’s nodes distribution, 180
Task-tree (T-tree), 152
TESLA protocol, 496
Three-stage handshaking
(ADV-REQ-DATA), 54
Three-way handshaking mechanism, 62
Threshold-sensitive energy-efficient sensor
network protocol (TEEN), 144
Throughput-capacity networks, 141
Time difference of arrival (TDoA) method,
313, 334
Time division multiple access (TDMA), 505
data dissemination protocol, 56–57
data dissemination, 57
Time division multiple access-based
medium access layer, 56
Time division multiple access slot, 57
Time synchronization protocol sensor
networks (TPSN), 510–512
attacks, 512
TinySec security architecture, 498
Topologically aware worm propagation
model (TWPM), 70
Topology adaptive spatial clustering (TASC)
algorithm, 179, 179
Transmission energy, 98
Transmission power, 254
543
Tree-based clustering protocol, 171
Triangulation algorithm, 140
Triangulation method, 319
Trickle’s principles, 62
Trickle algorithm, 59–61
Trickle metadata, 61
Trilateration and multilateration method,
316
Triple-key management, 494
graphical presentation, 494
Two-dimensional euclidean plane, 82, 97
Two-dimensional sensor field, 35
Two-ray ground model, 85
Two-tier sensor network architecture, 236
UDI model, 86, 87
Uniform node distribution, 97
Unit ball graph (UBG), 81
definition, 81
Unit disk graph (UDG), 78, 79
model, 78, 80, 81, 85
Ultra low wireless sensor, 423
Variable transmission range protocol
(VTRP) 437, 451–461, 465
VOR, 249, 250
Voronoi-based (VOR) method, 248
Voronoi-based (VOR) stations, 354
Voronoi-based (VOR) systems, 344
Voronoi cells, 35
Virtual polar coordinate routing (VPCR),
210
Virtual polar coordinate space (VPCS), 210
Wake-up process, 26
Weight partitioning algorithm, 179
Well-known recognition algorithms, 27
Wide-sense stationary (WSS), 112
Wireless device, see Malory
Wireless sensor network (WSN), 1, 21, 30,
31, 36, 41, 42, 51, 52, 54, 84, 105, 107,
109, 112, 115, 119, 130, 161, 163, 164,
169, 170, 225, 237, 267, 307, 341
advantages, 105
applications, 31, 267
architecture, 107
code update protocols, 59
data dissemination, 54
definition, 521
544
INDEX
Wireless sensor network (WSN) (Continued)
epidemic models, 52
field of, 342
graph theory approaches, 169
heterogeneous, 21
homogeneous, 21
joint spatiotemporal correlation, 115
management of, 41
mobility, 267
properties of, 130
protocols/techniques, 521
self-configuring, 341
sound sensors, 42
spatial correlation, 109
spatiotemporal correlation, 119
temporal correlation, 112
Wireless sensor nodes, 130, 276
Wormhole attack, 525
Worst-case node distribution, 97–98
Zonal algorithm, 176
Zonal weakly connected clustering
algorithm, 183
Zone-based clustering, 183–185
WILEY SERIES ON PARALLEL AND DISTRIBUTED COMPUTING
Series Editor: Albert Y. Zomaya
Parallel and Distributed Simulation Systems / Richard Fujimoto
Mobile Processing in Distributed and Open Environments / Peter Sapaty
Introduction to Parallel Algorithms / C. Xavier and S. S. Iyengar
Solutions to Parallel and Distributed Computing Problems: Lessons from
Biological Sciences / Albert Y. Zomaya, Fikret Ercal, and Stephan Olariu (Editors)
Parallel and Distributed Computing: A Survey of Models, Paradigms, and
Approaches / Claudia Leopold
Fundamentals of Distributed Object Systems: A CORBA Perspective /
Zahir Tari and Omran Bukhres
Pipelined Processor Farms: Structured Design for Embedded Parallel
Systems / Martin Fleury and Andrew Downton
Handbook of Wireless Networks and Mobile Computing /
Ivan Stojmenović (Editor)
Internet-Based Workflow Management: Toward a Semantic Web /
Dan C. Marinescu
Parallel Computing on Heterogeneous Networks / Alexey L. Lastovetsky
Performance Evaluation and Characteization of Parallel and Distributed
Computing Tools / Salim Hariri and Manish Parashar
Distributed Computing: Fundamentals, Simulations and Advanced Topics,
Second Edition / Hagit Attiya and Jennifer Welch
Smart Environments: Technology, Protocols, and Applications / Diane Cook
and Sajal Das
Fundamentals of Computer Organization and Architecture /
Mostafa Abd-El-Barr and Hesham El-Rewini
Advanced Computer Architecture and Parallel Processing / Hesham El-Rewini
and Mostafa Abd-El-Barr
UPC: Distributed Shared Memory Programming / Tarek El-Ghazawi,
William Carlson, Thomas Sterling, and Katherine Yelick
Handbook of Sensor Networks: Algorithms and Architectures /
Ivan Stojmenović (Editor)
Parallel Metaheuristics: A New Class of Algorithms / Enrique Alba (Editor)
Design and Analysis of Distributed Algorithms / Nicola Santoro
Task Scheduling for Parallel Systems / Oliver Sinnen
Computing for Numerical Methods Using Visual C++ / Shaharuddin Salleh,
Albert Y. Zomaya, and Sakhinah A. Bakar
Architecture-Independent Programming for Wireless Sensor Networks /
Amol B. Bakshi and Viktor K. Prasanna
High-Performance Parallel Database Processing and Grid Databases /
David Taniar, Clement Leung, Wenny Rahayu, and Sushant Goel
Algorithms and Protocols for Wireless and Mobile Ad Hoc Networks /
Azzedine Boukerche (Editor)
Algorithms and Protocols for Wireless Sensor Networks / Azzedine Boukerche
(Editor)