i

Data Processing Algorithms in

Wireless Sensor Networks for
Structural Health Monitoring

DANNA NIGATU MITIKU

MEKONNEN ESAYAS GETACHEW

Master of Science Thesis

Stockholm, Sweden 2011
 Data Processing Algorithms in
Wireless Sensor Networks for
Structural Health Monitoring

DANNA NIGATU MITIKU

MEKONNEN ESAYAS GETACHEW

December 2011
TRITA-BKN. Master Thesis 335, 2011
ISSN 1103-4297
ISRN KTH/BKN/EX-335-SE
©Nigatu Mitiku and Esayas Getachew, 2011
Royal Institute of Technology (KTH)
Department of Civil and Architectural Engineering
Division of Structural Engineering and Bridges
Stockholm, Sweden, 2011
 Preface

This master thesis was carried out by students from Blekinge Institute of Technol-
ogy (BTH) at the Swedish Institute of Computer Science (SICS), the division of
Networked Embedded Systems (NES), and the Department of Civil and Architec-
tural Engineering, the division of Structural Engineering and Bridges, at the Royal
Institute of Technology (KTH), in Stockholm on the VINNOVA STRUCT project.

We would like to express our gratutude to Thiemo Voigt for giving us the opportu-
nity to be part of the Network Embedded Systems (NES) group of SICS at Kista,
Stockholm. His fast response and guidance each time with a positive feedback to
get sensible results were so tremendous. We also would like to acknowledge our
supervisor Luca Mottola for his valuable comments on every way and for guiding us
to concentrate on our work. Our supervisor and examiner, Prof. Raid Karoumi is a
well dedicated person giving us time, patience, understanding and helping even with
administration issues. Once again we would like to thank our supervisor, Ignacio
Gonzales Silva (PhD student) for his encouragement that made the interdisciplinary
work more interesting by sharing his experiences.

Thanks to Claes Kullberg, the lab assistant at KTH who facilitated the instrumen-
tation and tests. Many thanks to SICS staffs, specially Simon Duquennoy, Marcus
Lundén, Nicklas Wirström and Joakim Eriksson, for their technical support.

The most special thanks to our families who gave unlimited support, care and
love.

The last, but not the least, thanks to all BTH staffs and friends.

Stockholm, January 23, 2012

Nigatu Mitiku and Esayas Getachew

iii
 Abstract

The gradual deterioration and failure of old buildings, bridges and other
civil engineering structures invoked the need for Structural Health Monitoring
(SHM) systems to develop a means to monitor the health of structures. Dozens
of sensing, processing and monitoring mechanisms have been implemented and
widely deployed with wired sensors.

Wireless sensor networks (WSNs), on the other hand, are networks of large
numbers of low cost wireless sensor nodes that communicate through a wireless
media. The complexity nature and high cost demand of the highly used wired
traditional SHM systems have posed the need for replacement with WSNs.
However, the major fact that wireless sensor nodes have memory and power
supply limitations has been an issue and many efficient options have been pro-
posed to solve this problem and preserve the long life of the network. This
is the reason why data processing algorithms in WSNs focus mainly on the
accomplishment of efficient utilization of these scarce resources.

In this thesis, we design a low-power and memory efficient data processing

algorithm using in-place radix-2 integer Fast Fourier Transform (FFT). This
algorithm requires inputs with integer values; hence, increases the memory
efficiency by more than 40% and highly saves processor power consumption
over the traditional floating-point implementation. A standard-deviation-based
peak picking algorithm is next applied to measure the natural frequency of the
structure.

The algorithms together with Contiki, a lightweight open source operating sys-
tem for networked embedded systems, are loaded on Z1 Zolertia sensor node.
Analogue Device’s ADXL345 digital accelerometer on board is used to collect
vibration data. The bridge model used to test the target algorithm is a simply
supported beam in the lab.

Keywords: Bridge, FFT, Modal Analysis, Monitoring, Natural Frequency,

Peak Picking, Structure, Wireless Sensor.

v
List of Abbreviations

ADC Analogue to Digital Converter

ANPSD Averaged Normalized Power Spectral Densities

AR Auto Regressive

ASICS Application Specific Integrated Circuits

BSL BootStrap Loader

CFDD Curve-fit Frequency Domain Decomposition

CPU Central Processing Unit

DAQ Data Acquisition System

DFT Discrete Fourier Transform

DSP Digital Signal Processors

EFDD Enhanced Frequency Domain Decomposition

ERA Eigensystem Realization Algorithm

FDD Frequency Domain Decomposition

FFT Fast Fourier Transform

FPGA Field Programmable Gate Arrays

FRF Frequency Response Functions

HPBW Half-Power BandWidth

I2 C Inter-Integrated Circuit

IFFT Inverse Fast Fourier Transform

IoT Internet of Things

IP Internet Protocol

vii
IPv6 Internet Protocol Version 6

ISM Industrial Scientific and Medical

ITD Ibrahim Time Domain

KTH Kungliga Tekniska Högskola

LSB Least Significant Bits

MEMS Micro Electro-Mechanical Systems

MIPS Million Instructions per Second

MSP Mixed Signal Processor

ODS Operation Deflection Shapes

OS Operating System

PP Peak Peaking

PSD Power Spectral Density

RDT Random Decrement Technique

RF Radio Frequency

RISC Reduced Instruction Set Computer

ROM Read-Only Memory

SDOF Single Degree-of-Freedom

SEK Swedish Kronor

SHM Structureal Health Monitoring

SPI Serial Peripheral Interface

SRAM Static Random Access Memory

SSI Stochastic Subspace Identification

SVD Singular Value Decomposition

TCP/IP Transport and Communication Protocol

WSN Wireless Sensor Network

viii
Table of Contents

Preface iii

Abstract v

List of Abbreviations vii

List of Figures xiii

List of Tables xv

1 Introduction 1
1.1 Problem Statement and Research Contribution . . . . . . . . . . . . 2
1.2 Aim and Scope of the Study . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Survey of Selected Related Works . . . . . . . . . . . . . . . . . . . . 3
1.4 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Structural Health Monitoring 5

2.1 Trends in SHM Systems . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Advaced Approaches in SHM . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Signal Analysis in Structural Health Monitoring . . . . . . . . . . . . 6
2.3.1 Sampling Theorem . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3.2 Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.3 Discrete Fourier Transform (DFT) . . . . . . . . . . . . . . . 7
2.3.4 Fast Fourier Transform (FFT) . . . . . . . . . . . . . . . . . 8
2.4 Categories of Structural Health Monitoring . . . . . . . . . . . . . . 8
2.4.1 Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . 9
2.4.2 Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.5 Input-Output Modal Analysis . . . . . . . . . . . . . . . . . . . . . . 13
2.6 Output-Only Modal Identification . . . . . . . . . . . . . . . . . . . 13
2.6.1 Peak Picking Method (PP) . . . . . . . . . . . . . . . . . . . 15
2.6.2 Frequency Domain Decomposition (FDD) . . . . . . . . . . . 15
2.6.3 Enhanced Frequency Domain Decomposition . . . . . . . . . 16
2.6.4 Curve Fitting Frequency Domain Decomposition . . . . . . . 16

ix
TABLE OF CONTENTS

3 Wireless Sensor Networks for SHM 19

3.1 Applications of WSNs . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 Wireless Sensor Nodes . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Hardware Components of a Wireless Sensor Node . . . . . . . . . . . 21
3.3.1 Microcontroller . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.2 Transceiver . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3.3 Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3.4 Power source . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3.5 Sensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.3.6 Energy harvester (Optional) . . . . . . . . . . . . . . . . . . . 25
3.4 Software Components of a Wireless Sensor Node . . . . . . . . . . . 26
3.4.1 Operating systems . . . . . . . . . . . . . . . . . . . . . . . . 26
3.4.2 Application programs . . . . . . . . . . . . . . . . . . . . . . 27
3.5 Challenges in Using WSNs . . . . . . . . . . . . . . . . . . . . . . . . 27
3.5.1 Memory constraints . . . . . . . . . . . . . . . . . . . . . . . 28
3.5.2 Power consumption constraints . . . . . . . . . . . . . . . . . 28

4 Design and Methodology 29

4.1 Hardware Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1.1 Wireless system . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1.2 Wired system . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2 Software Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
4.2.1 Realtime software tools . . . . . . . . . . . . . . . . . . . . . 33
4.2.2 Offline software tools . . . . . . . . . . . . . . . . . . . . . . . 38

5 Accelerometers 39
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.2 What Accelerometers Are? . . . . . . . . . . . . . . . . . . . . . . . 39
5.3 Selection Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3.1 Analogue or digital . . . . . . . . . . . . . . . . . . . . . . . . 40
5.3.2 Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.3.3 Full-Scale range . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.3.4 Number of axes measured . . . . . . . . . . . . . . . . . . . . 40
5.3.5 Bandwidth response . . . . . . . . . . . . . . . . . . . . . . . 40
5.3.6 Noise level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.3.7 Power consumption . . . . . . . . . . . . . . . . . . . . . . . . 41
5.3.8 Extra features . . . . . . . . . . . . . . . . . . . . . . . . . . 41
5.4 Which Ones Did We Use? . . . . . . . . . . . . . . . . . . . . . . . . 41
5.4.1 Analogue devices ADXL345 . . . . . . . . . . . . . . . . . . . 41
5.4.2 Si-Flex™ accelerometer . . . . . . . . . . . . . . . . . . . . . 43
5.4.3 Other accelerometers . . . . . . . . . . . . . . . . . . . . . . . 43
5.5 Validating the Accelerometer . . . . . . . . . . . . . . . . . . . . . . 43
5.6 Cares to Be Taken . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.6.1 Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

x
TABLE OF CONTENTS

5.6.2 Attachment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.6.3 Axis alignment . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.6.4 Sampling rate . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

6 Experimental Setup 47
6.1 Instrumentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
6.2 Experimental Procedures . . . . . . . . . . . . . . . . . . . . . . . . 50
6.3 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

7 Results and Discussion 53

7.1 Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
7.1.1 Validating the accelerometer . . . . . . . . . . . . . . . . . . 53
7.1.2 Validating the FFT . . . . . . . . . . . . . . . . . . . . . . . 54
7.2 Time Domain Data Analysis . . . . . . . . . . . . . . . . . . . . . . 56
7.3 Frequency Domain Analysis . . . . . . . . . . . . . . . . . . . . . . . 57
7.3.1 Experiment 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
7.3.2 Experiment 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
7.3.3 Experiment 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
7.4 The Cause of the Frequency Differences . . . . . . . . . . . . . . . . 59

8 Conclusion, Recommendations and Future Works 63

8.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
8.2 Recommendation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
8.3 Future Works . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

Bibliography 65

xi
List of Figures

1.1 The catastrophic failure of I-35W and Point Pleasant Bridges . . . . . . 1

2.1 Frequency response function. . . . . . . . . . . . . . . . . . . . . . . . . 10

2.2 Single degree of freedom model. . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Half power bandwidth method. . . . . . . . . . . . . . . . . . . . . . . . 12
2.4 An Impact hammer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3.1 Example application of SHM using WSN. From [23]. . . . . . . . . . . . 19

3.2 The Z1 Zolertia wireless sensor unit. . . . . . . . . . . . . . . . . . . . . 21
3.3 Hardware architecture of a wireless sensor node. . . . . . . . . . . . . . 21
3.4 Internal components of a microcontroller. . . . . . . . . . . . . . . . . . 22
3.5 The 3-axis ADXL MEM accelerometer mounted on a circuit board. . . . 24
3.6 An example sensor board with accelerometer sensor on board. . . . . . . 24
3.7 Solar panel and rechargeable battery for Jindo Bridge SHM [41] . . . . 25
3.8 Position of the operating system. . . . . . . . . . . . . . . . . . . . . . . 26

4.1 Functional block diagram of Z1 WSN module. . . . . . . . . . . . . . . . 30

4.2 The service architecture of Contiki. (From [2]) . . . . . . . . . . . . . . 34
4.3 A frequency spectrum plot divided into non-overlapping windows. . . . 37

6.1 Part of the 3.5 m long test beam . . . . . . . . . . . . . . . . . . . . . . 47

6.2 A simple beam model of experiment 1 showing the parts 1 to 4 for each
attachment (all measurements are in cm). . . . . . . . . . . . . . . . . . 48
6.3 The wooden support at the ends of the experimental beam. . . . . . . . 48
6.4 The front end of the instrumentation on the experimental beam . . . . . 49
6.5 Attachment of the sensor units on the experimental beam. . . . . . . . . 50
6.6 Exciting the experimental beam with an impulse hammer. . . . . . . . . 50

7.1 Time and frequency domain data from iPhone . . . . . . . . . . . . . . 54

7.2 Time and frequency domain data from the wireless sensor . . . . . . . . 54
7.3 Comparison between Floating point and Integer point FFT . . . . . . . 55
7.4 Time domain data from wired and wireless sensors . . . . . . . . . . . . 56
7.5 Measurement from the wired sensor . . . . . . . . . . . . . . . . . . . . 57
7.6 Node 1 experiment 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

xiii
List of Figures

7.7 Node 2 experiment 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58

7.8 Node 3 experiment 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.9 Node 4 experiment 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
7.10 Experiment 2 wireless sensor data . . . . . . . . . . . . . . . . . . . . . 60
7.11 Experiment 3 wireless sensor data . . . . . . . . . . . . . . . . . . . . . 60
7.12 Experiment 3 wired sensor data . . . . . . . . . . . . . . . . . . . . . . . 60

xiv
List of Tables

4.1 Summary of parametric features of MSP430F2617. . . . . . . . . . . . . 31

4.2 Summary of CC2420 features. . . . . . . . . . . . . . . . . . . . . . . . . 31
4.3 Approximate current consumption of Z1 components. . . . . . . . . . . 32

5.1 Summary of the ADXL345 accelerometer features. . . . . . . . . . . . . 41

5.2 Summary of the Si-Flex™ accelerometer features. . . . . . . . . . . . . 43
5.3 Summary of the most common accelerometers in SHM. . . . . . . . . . 44

7.1 Comparison of floating point and fixed point FFT on the node . . . . . 55

xv
Chapter 1

Introduction

The process of implementing a damage characterization and detection method for

engineering structures is referred to as Structural Health Monitoring (SHM). Al-
though it had been quite a while since the science of SHM was introduced, its use
was confined to mechanical structures like airplanes, ships, machineries, etc. It had
never been applied to civil engineering structures until its significance was noticed
in the frequent deterioration and collapse of large and prestigious structures.

Figure 1.1: (Left) The catastrophic failure of I-35W Bridge in Minneapolis, Min-
nesota after collapse on August 1, 2007; (right) The Point Pleasant Bridge collapse.

For example, the catastrophic failure of the I-35W Bridge in Minneapolis, Min-
nesota (Figure 1.1 left) and the Point Pleasant Bridge (Figure 1.1 right) were among
the failures that alerted the need to devise some means to tell the status of struc-
tures before anything worse happens. Consequently, a continuous health monitoring
of structures is important and a mechanism should be developed by which efficient
and accurate information could be obtained.

1
1.1. PROBLEM STATEMENT AND RESEARCH CONTRIBUTION

Researchers, hence, gave special attention to this discipline and proposed their own
customized solutions in the last couple of decades which eventually gave birth to
the science of SHM. SHM is thus one of the multidisciplinary fields that integrates
the contribution of researchers from mechanical, electrical, civil and architecture
engineering. Due to the easy access to, the wide availability and reliability of wired
systems, many solutions have been implemented using wired sensor networks [37].
High installation cost, need for specially trained professionals for setting up and
maintain and their bulky nature made the research community to divert its atten-
tion towards WSNs. In addition, the advance of low cost wireless sensors has given
a new dimension to the field.

Wireless Sensor Networks (WSN) is an upcoming technology which has a wide

range of applications including infrastructure protection, industrial sensing and di-
agnostics, environment monitoring, context-aware computing (for example intelli-
gent home and responsive environment) and so on. This kind of network usually
consists of a large number of nodes that communicate together to form a wireless
network. It is however essential to improve the energy efficiency for WSNs as the
energy designated for sensor nodes is usually extremely limited.

The important thing one should have in mind, while considering the design of appli-
cations on wireless sensor networks, is that WSNs have the constraints of memory
and energy consumption. Thus, designing customized systems involves trade-offs.

1.1 Problem Statement and Research Contribution

The use of wired sensors and data acquisition systems (DAQs) in SHM incur a high
cost to purchase and setup the instruments on structures. The one which is in use
at the Division of Structural Engineering in KTH , for example, is in the order of
more than 300,000 SEK per unit. Densely instrumenting structures with these sys-
tems is thus economically infeasible. Hence, low-cost options are worthy to consider.

However, the use of WSNs for SHM systems is a significat task as there are some
limitations that need to be addressed when it comes to utilizing reliable and robust
monitoring systems. The fact that the process of SHM systems involves collecting
huge amounts of data, applying complex data processing algorithms and extract-
ing significant information, requires the use of high memory and processing power
nodes. This poses an issue to the trend of currently available wireless sensor nodes
with small memory size and processor power. Fully autonomous, fast and reliable
systems to-date use wired systems. To completely shift the realm of SHM to a fully
WSN-based solution, we need to tackle the challenges.

In this thesis, we explore the use of WSNs for SHM and implement a low-power and
memory efficient data processing algorithm thereby paving the way for promising

2
1.2. AIM AND SCOPE OF THE STUDY

future deployment and applications. The algorithms are developed using C pro-
gramming language and the Contiki operating system on the Z1 Zolertia sensor
nodes.

1.2 Aim and Scope of the Study

The thesis work aims at understanding to what extent existing algorithm run on
WSN devices in face of computing and memory limitations, studying the trade-off
in terms of quality of the output versus resource consumption, and implementing a
customized algorithm to better fit the characteristic constraints of WSN devices.

1.3 Survey of Selected Related Works

In this section, a review of some selected important works in the field of SHM is
given. The review follows chronological order from the earliest to the latest with
some jumps.

In 1998, Straser and Kiremidjian [44] were the first to develop a method for de-
termining the health of a civil structure using wireless sensors on Alamosa Canyon
Bridge, NM, USA. The method detected the general state of a structural immedi-
ately following a seismic event. They used a program at the core of their embedded
analysis to indirectly measure the kinetic energy of a structure to detect when en-
ergy is dissipated during damage. It is better conceived as a decentralized damage
detection method than a monitoring system as it measures the after-effects of a
structural damage.

The Cooley-Tukey [20] implementation of FFT was successfully embedded in the

computational core of a wireless sensing unit developed by Lynch and his research
team [36]. The FFT embedded in the wireless sensing unit was utilized during field
deployments of the wireless monitoring system to provide the Frequency Response
Functions (FRFs) of instrumented structures. The accuracy of the complex-valued
Fourier amplitude spectra computed by the wireless sensor node was shown to pro-
vide identical results to those generated by Matlab using the same time-history data.

While many researchers proposed the use of modal frequencies as a primary damage
indicator, the method lacks the sensitivity to capture ambient vibrations in struc-
tures where environmental factors also contribute to modal frequency shifts [22].
To fully account for the environmental and operational variability of structures, a
damage detection methodology based upon a pattern recognition framework was
proposed by Sohn and Farrar [42]. An autoregressive (AR) time series model is fit
to the stationary response time history of the structure.

Hackmann et al. [24] presented a way of structural monitoring using a holistic

3
1.4. THESIS OUTLINE

distributed damage localization technique where each sensor node calculates the
partial curve fitting factor and then sends its data to the central station for final
integration and correlation based damage localization.

The effect of time synchronization on the performance of WSNs for SHM that af-
fects the quality and accuracy of mode shapes was addressed by V. Krishnamurthy
et al [33] using the popular frequency domain decomposition (FDD) technique.

Musiani et al.[40] described the design of an active sensor platform - Shimmer-

that uses super capacitor to store solar energy harvesting that serves for twenty
years once charged.

Recently, Gluaco Feltrin et al. [23] used accelerometers and strain gauge sensors to
measure vibration data. They built a separate sensor board supplied with separate
power source from the host node to avoid errors that maybe introduced due to the
initial sensor warm up phase while powered on during duty cycling.

1.4 Thesis Outline

As the work is multidisciplinary and readers from both the electrical engineering
side and structural engineering side should fully be able to understand the concept,
the thesis is organized in such a way that detailed analysis of the subjects is given
to make the report as self-complete as possible.

- Chapter 2 presents a detailed background on SHM.

- Chapter 3 introduces WSNs and their use in SHM.

- Chapter 4 addresses the design and methodology that we used while doing
experiments.

- Chapter 5 gives a detailed look into accelerometer selection.

- Chapter 6 explains the experimental evaluation to validate the designed SHM

system.

- Chapter 7 Results and discussion are presents and analyzed as well.

- In Chapter 8 conclusions are drawn along with recommendations and insights

on future works.

4
Chapter 2

Structural Health Monitoring

It has been mentioned in the previous chapter that the process of implementing
a damage detection and characterization strategy for engineering structures
is referred to as SHM. Damage may refer to all changes to the properties of a
structure. It is an interdisciplinary study that involves identifying structural
changes and effects on the overall integrity of the structure. In this chapter,
we discuss the trends and practices of SHM system by analyzing frequently
used approaches.

2.1 Trends in SHM Systems

There are several methods that are practiced among structural engineers in
the study of changes in the health of civil engineering structures. Visual
inspection is one of the predominant methods used quite often in studying
changes and effects on structures. It is labor intensive and is only capable of
observing changes on the surface of the structures.
The modern practice towards SHM is designing a continuous methodology for
the study of structural changes. This has a wide benefit over traditional visual
inspection methods by providing real-time, continuous analysis and detection
of damage on civil structures. Thus, in order to avoid the failure of bridges
and buildings, implementing SHM system to monitor the safety of structures
is important.

2.2 Advaced Approaches in SHM

An advanced and alternative approach to visual structural inspections, SHM

steps have been modified by Sohn [42]. These steps are:

5
2.3. SIGNAL ANALYSIS IN STRUCTURAL HEALTH MONITORING

1. Operational evaluation of damage identification process to features which

are unique to the system being monitored.
2. Data acquisition, fusion, and cleansing which are used to collect sensor
measurements data and to store in a centralized location.
3. Feature extraction and information condensation in identification of data
features that allow one to distinguish between healthy and damaged part.
4. Statistical model development, concerned with the implementation of
algorithms that operate on the extracted features to quantify the damage
state of the structure.

2.3 Signal Analysis in Structural Health Monitoring

In bridge monitoring, accelerations, strains, deflections, temperature and ap-

plied forces, etc., are the important signals interesting to measure for analysis.
These measurements are recorded for a given period which shows the time flow
of a parameter under the study. It is difficult to see directly from the time
domain data representation with what frequencies the bridge vibrates. The
characteristics of a signal can be better understood in a frequency domain rep-
resentation rather than the intuitive time domain representation. Therefore,
it’s better to use a Fourier transform, a mathematical operation that decom-
poses a signal into its constituent frequencies. Currently, most SHM systems
depend on measuring structural dynamics characteristics and analyzing these
data in the frequency domain by performing modal analysis which directly
involves Fast Fourier Transform (FFT).

2.3.1 Sampling Theorem

The Nyquist theorem states that any continuous baseband signal (signal ex-
tending down to zero frequency) may be similarly reconstructed if the signal
is bandwidth limited and the sampling frequency is at least twice the band
width of the signal. If a time-domain signal is sampled uniformly, then the
frequency corresponding to one-half that rate is called the Nyquist frequency,
N1
1
N1 = (2.1)
2∆t
where ∆tis the time between successive samples. Appropriate choice of sam-
pling interval is necessary for obtaining correct representation of the original
signal.

6
2.3. SIGNAL ANALYSIS IN STRUCTURAL HEALTH MONITORING

2.3.2 Fourier Transform

The Fourier transform is a mathematical transform that converts the time

domain waveform into the frequency domain. Hence, Fourier Transform pro-
duces a complex number valued sinusoid which can be displayed with two
parts, either with the real and imaginary part or with magnitude and phase.
When this sinusoid of different frequency components is summed, it would
give the original waveform. The Fourier transforms F (w) of a function can be
mathematically expressed from the time domain data f (t) as:
Z +∞
F (w) = f (t)e−iwt dt. (2.2)
−∞
and the inverse Fourier transform as:
1 +∞
Z
f (t) = F (w)eitw dw. (2.3)
2π −∞
√
where i is the imaginary value and in both cases, i = −1. The complex
valued output can be expressed as the sum of two terms
F (k) = R(k) + iI(k) (2.4)
where R(k) is the real component and I(k) is the imaginary component. The
magnitude is: q
|F (k)| = R(k) ∗ R(k) + I(k) ∗ I(k) (2.5)
and the phase can be calculated as:
I(k)
6 f (x) = tan−1 (2.6)
R(k)

2.3.3 Discrete Fourier Transform (DFT)

DFT is essential because it takes a discrete signal in the time domain and
transforms that signal into its discrete frequency domain representation while
the inverse DFT performs the reverse operation from frequency domain to an
equivalent time domain sequence. It transforms a discrete signal to a sum of
sinusoidal-shaped signals, all with their different amplitudes, frequencies and
phases. The DFT of N uniformly sampled data points xj (where j = 0,. . . ,
N-1) and its inverse are defined in mathematical representation as:

N −1
xj e−i2πjk/N
X
Xk = (2.7)
j=0

and

−1
1 NX
xj = Xk ei2πjk/N (2.8)
N k=0

7
2.4. CATEGORIES OF STRUCTURAL HEALTH MONITORING

2.3.4 Fast Fourier Transform (FFT)

Fast Fourier Transform (FFT) is an efficient mechanism for transforming time

domain data into frequency domain. This fast implementation of discrete
Fourier transform reduces the number of computations needed for N-point
DFT from O(N 2 ) to O(N log2 N ) in (Radix-2 Cooley-Tukey) [22]. The only
requirement of the most popular implementation of this algorithm is that the
number of points in the series has to be a power of 2. The computing time
for the radix-2 FFT is proportional to N. The FFT considerably reduces the
computational requirements of the DFT. For example, if we take a transform
on 1024 points using the FFT it is about 100 times faster than using the DFT,
which is a significant speed increase.
The frequency resolution is given as:

fN yquist 2fN yquist

fr = = (2.9)
N/2 N

Where:

fr = frequency resolution
fN yquist = Nyquist frequency
fmax = maximum frequency

Hence, to determine the corresponding frequency,fN , at particular sample

point. It can be written as:
fN = fr × i (2.10)

Where i = instantaneous sample number point

2.4 Categories of Structural Health Monitoring

SHM process involves the study of a system over time by collecting measure-
ments via data acquisition systems, extraction of damage-sensitive features
from these measurements, and making statistical analysis of these features to
determine the current state of system health. In order to make all these detail
analysis we need to implement data processing algorithms in efficient man-
ner. Until recently qualitative and non-continuous methods have been used to
evaluate structures for their structural health to provide their intended use.
In the last few years SHM technologies have developed well due to the ad-
vancement in the other related science and technology fields. SHM system
can be wired or wireless. Due to the technology development and the benefits
that wireless systems have over the wired system, wireless SHM became an
interesting research area. The wired system is robust and reliable, but it is

8
2.4. CATEGORIES OF STRUCTURAL HEALTH MONITORING

costly and less flexible. On the other hand, the wireless systems are cheap,
flexible and easy for deployment. The drawbacks with the wireless systems
are problems with lifetime, memory, reliability and also having energy effi-
cient data processing algorithms [36]. Taking the advantage of those wired
system and customizing to wireless system is a major development to SHM
in the wireless network environment. For the appropriate choice of efficient
data processing algorithm having the background of these data processing al-
gorithms is important. The goal of this chapter is to give a brief background
of data processing algorithms in relation to wireless sensor networks ability
to process huge volume of data, reliability and flexibility. We will give brief
description of existing methodologies for structural health monitoring.

A large number of sensors are used in most SHM system to study the structural
elements of which include accelerometers, strain gauges, displacement trans-
ducers, temperature sensors, etc,. After collection of these measurements,
data has to be extracted and appropriate analysis should be done to investi-
gate the general condition of the system. There are broadly two categories of
SHM: Feature extraction and modal analysis.

2.4.1 Feature Extraction

Feature extraction algorithm depends on features or signatures gathered from

the recorded structural response signals such as acceleration, strain or other
data that change with the occurrence of damage. Then, the measured vibra-
tion responses are used to identify damage-sensitive properties which allow
one to distinguish between the damaged and the healthy structure. Such
algorithms mainly involve the following steps:

i. The evaluation of a structure’s operational environment which include

loading conditions, temperature and humidity.

ii. The acquisition of structural response measurements and data prepro-

cessing.

iii. The extraction of features that are sensitive to damage and

iv. The development of statistical models for feature discrimination

The downfall of this method is that it needs previous record of damage features
and development of statistical models in order to monitor structural changes.
Therefore, this process is computation intensive and requires more application
of artificial intelligence at the sensor nodes which lead to the design of power-
hungry algorithms [32].

9
2.4. CATEGORIES OF STRUCTURAL HEALTH MONITORING

2.4.2 Modal Analysis

Modal analysis has been widely applied and commonly used approach in vibra-
tion based structural health monitoring in aerospace, mechanical and civil en-
gineering to investigate the integrity of structures based on natural frequency,
modal damping, and a mode shape. Modal analysis describes a structure in
terms of its natural characteristics which are frequency, damping ratio and
mode shapes. The objective is the evaluation of modal parameters of a struc-
ture under ambient vibrations and dynamic loads. Modal parameters provide
information that help to detect changes in structure.

In modal analysis, we use Frequency response function (FRF) as a mathe-

matical representation to describe the relationship between the input and the
output of a system. FRF shows how much displacement, velocity, or accelera-
tion response a structure has at the output point with respect to input force;
if we take measurement of the time data and transform it to the frequency
domain using FFT. Then, looking at the FRF we will see some very interest-
ing characteristics of the structure as we can see, e.g., in Figure 2.1. We will
be able to identify that there are peaks in this function which occur at the
resonant frequencies of the system. Modal analysis can further be classified as
output only and input/output based on the excitation [6]. This section gives
a brief introduction to Modal parameters.

Figure 2.1: Frequency response function.

10
2.4. CATEGORIES OF STRUCTURAL HEALTH MONITORING

Single degree-of-freedom

The simplest vibratory system that can be described by a single mass con-
nected to a spring. The mass is allowed to travel only along the spring elon-
gation direction. Such systems are called Single Degree-of-Freedom (SDOF)
systems and are shown in Figure 2.2. where m is the mass, cv is the viscous

Figure 2.2: Single degree of freedom model.

damping coefficient, k is the stiffness, x is the absolute displacement of the

mass, lo is the base input displacement.

Natural frequency

When an object is hit and the external force is removed it will vibrate at its
natural frequency. The natural frequency is defined as the number of times a
system will oscillate (move back and forth) between its original position and
its displaced position assuming there is no outside interference. The natural
frequency can be calculated by the formula
s
1 k
f= (2.11)
2π m

where k is the stiffness and m is the mass of the structure.

Damping Ratio

Damping ratio, a dimensionless measure that tells us how oscillations in a

system behave after an excitation or due to environmental loadings. For de-
termining damping ratio from frequency domain data, half-power bandwidth
(HPD) method can be used. In this method, for each natural frequency there

11
2.4. CATEGORIES OF STRUCTURAL HEALTH MONITORING

is a peak in FRF amplitude and 3 dB down from the peak there are two points
corresponding to half power points which are depicted in the Figure 2.3. HBD
is defined as the ratio of the frequency range between the two half power points
to the natural frequency at this mode where fa and fb are the frequencies as-
sociated with the half power points on either side of the peak as shown in
Figure 2.3. A1 is the amplitude at the peak A2 can calculated as

Figure 2.3: Half power bandwidth method.

√
A2 = A1 / 2 (2.12)

Thus damping ratio associated with each natural frequency can be mathe-
matically calculated as obtained using the formula

fb − fa
ζ= (2.13)
fb + fa

Mode Shape

Modes are defined as inherent properties of a structure, and are determined

by the material properties (mass, damping, and stiffness), and boundary con-
ditions of the structure. Different mode shapes will be related with different
frequencies and each mode has a natural frequency associated with it. Basi-
cally, there are characteristics that depend on the weight and stiffness of the
structure that determine where these natural frequencies and mode shapes
will exist. Mode shapes are helpful because they represent the shape that
the structure will vibrate and simplifies the vibration response of a complex
structure into a set of modal parameters that can be easily analyzed. Exper-
imentally to determine the mode of a structure, modal testing is done, e.g,
by measuring operational deflection shapes (ODS) which are deflections of a

12
2.5. INPUT-OUTPUT MODAL ANALYSIS

Figure 2.4: An Impact hammer

structure at a specific particular frequency then post processing them in a

specific manner to define mode shapes [13].
Mode shapes do not have unique value but they are unique in shape. These
mode shapes can be arbitrarily scaled to any set of values and the relationship
of one shape component to the other is distinct. One of the common ways to
scale mode shapes is to scale them so that the modal masses are one (unity).
Modal Testing is based on the estimation of a set of FRFs relating the ap-
plied force and the corresponding response at several pairs of points along the
structure, with enough high spatial and frequency resolution [6].

2.5 Input-Output Modal Analysis

Input-Output modal analysis is based on the estimation of a set of FRFs re-

lating an applied force to the corresponding response at several points along
the structure. In small and medium-sized structures, the excitation can be
induced by an impact hammer (Figure 2.4) or special impulse devices specifi-
cally designed to excite bridge and large buildings. The main problem in input
output modal analysis associated to the performance of forced vibration tests
in large engineering structures like bridges, dams and high-rise building, is
the difficulty to excite with sufficient energy and in controlled manner, which
needs a sophisticated testing instrument. Thus, it would not be a good option
for using input-output modal analysis in WSN for structural health monitor-
ing.

2.6 Output-Only Modal Identification

Output only modal identification is the study of a response under ambient

excitations which usually includes temperature, humidity, wind or under traf-
fic loadings. Random forces such as vehicle trafficking, wind excitation, etc.

13
2.6. OUTPUT-ONLY MODAL IDENTIFICATION

may act on bridges and excite the structure together so it’s almost impossible
to measure all these forces simultaneously. If the forces are not measured
correctly, then input-output modal analysis cannot give accurate estimates
of the modal parameters. Natural conditions under which structure operates
are very difficult to be produced in the laboratory. So, output-only modal
analysis is the best option available in such conditions as it depends on the
responses only and the responses could be measured with high accuracy both
on wired and wireless environments. One assumption made in output only
modal analysis is the excitation input is taken as zero mean Gaussian white
noise. The advantage to measure the natural (or ambient) response and then
estimate the modal parameters by performing an output-only modal identifi-
cation includes.

– The test is inexpensive and fast; since special equipment for excitation
is not needed. Thus, there is no need for vibration shake or impacts
hammer
– The test can be performed with out creating obstacle for normal opera-
tion of the structure.
– The real operating conditions of the structure are represented by the
output response of FRF.

The great advancement in sensing technologies such as MEMS sensors, wire-

less sensors, very high computational efficiency and communication technolo-
gies allows to record ambient response of a structure accurately. This devel-
opment makes output modal identification method preferable for monitoring
structural changes. With output-only modal identification it has been realized
the instrumentation of different bridges and buildings with wireless monitoring
systems[27]. Mathematical models on the output only identification technique
can roughly be classified as:

I. Parametric methods: These are methods that involve time domain anal-
ysis.
– Ibrahim Time Domain (ITD)
– Eigensystem Realization Algorithm (ERA)
– Random Decrement Technique (RDT)
– Stochastic Subspace Identification (SSI)
II. Non-parametric methods: These are methods that involve frequency do-
main analysis.
– Peak Picking (PP)
– Frequency Domain Decomposition (FDD)
– Enhanced Frequency Domain Decomposition (EFDD)
– Curve-Fitting Frequency Domain Decomposition (CFDD)

14
2.6. OUTPUT-ONLY MODAL IDENTIFICATION

These methods are advantageous on natural frequency and mode shape ex-
traction. There is uncertainty in damping ratio estimation but the EFDD and
CFDD perform better than FDD in this regard. In the next subsections we
will see in depth non parametric methods for structural health monitoring for
their implementation in wireless sensor networks.

2.6.1 Peak Picking Method (PP)

Peak picking (PP) method is an output-only modal analysis technique for

estimating the modal properties of a structure. The method is based on the
values of PSD spectrum that show changes in extreme values around the
natural frequencies. PSD shows the strength of the variations (energy) as
a function of frequency. Thus, it shows at which frequencies variations are
strong and at which frequencies variations are weak. The frequency at which
this extreme value occurs considered a good estimate for natural frequency of
the system. In this method the natural frequencies are determined from the
observation of the changes in peaks on the graphs of the averaged normalized
power spectral densities(ANPSDs)[10]. The ANPSDs are basically obtained
by converting time data to the frequency domain by performing DFT. The
main drawback of this method is it is difficulty in identifying closely spaced
modes and results are highly biased in modal frequencies values and mode
shapes in case of closely spaced modes. Even though PP has this weak point,
the method is strong and useful since the identification is very fast.
In structural health monitoring, using WSN the implementation of peak pick-
ing algorithm starts with a user setting the maximum number of peaks which
will be determined by the algorithm. Next, using data acquisition system,
a set of acceleration time history data will be collected at each sensor node
and converted to a PSD using an embedded FFT algorithm. Each node picks
largest peaks from its PSD function by searching for frequencies at which the
value of the PSD found to be greater than some threshold level set by the user.
If less than a preset number of peaks are found, zeros will be returned in place
of the missing peaks [50]. This implementation as a method is strong, fast
and useful for indication of changes in structures [34]. As we have constraints
with memory and power in the wireless sensor network this algorithm greatly
reduces the amount of data to be transmitted by the wireless sensing network
which will result in energy efficient algorithm.

2.6.2 Frequency Domain Decomposition (FDD)

FDD is a popular technique based on decomposing the power spectral density

functions matrix at each discrete frequency using the Singular Value Decompo-
sition (SVD) algorithm, for more understanding of this method refer a paper
by Brincker [15]. It decomposes the spectral matrix response into a set of

15
2.6. OUTPUT-ONLY MODAL IDENTIFICATION

single degree-of-freedom systems, each corresponding to one individual mode.

In FDD method, the mode shapes are estimated as the singular vectors at
the peak of each auto power spectral density function corresponding to each
mode. The results can be accurate if the structure assumed is lightly damped.
In wireless sensor network using FDD the method is to have each sensor send
its vibration data to a central sensor node which computes the SVD of the
PSD power spectral density matrix and then distributes the mode shapes back
to each sensor [29]. This would require significant computational power and
memory both at the sensor node and at the central node.
FDD method can only be applied in the frequency domain if the dominant
frequencies are determined at a priori. This method has played a major role in
WSN SHM and has been implemented for output-only system identification
through the ambient vibration measurements due to its reliability, straight-
forwardness and effectiveness [15]. The drawback of this method is with un-
certainty in damping ratio estimation. Other methods like EFDD and CFDD
which are the extension of the classical FFD would perform better with de-
termining damping ration than FDD.

2.6.3 Enhanced Frequency Domain Decomposition

EFDD technique is an extension to the FDD technique. In this method, the

auto spectral density functions are transformed into the time domain using
IFFT, resulting in autocorrelation functions for each mode of a system [17].
Extraction of modal parameters are done from the inspection of the decay of
auto correlation functions.

2.6.4 Curve Fitting Frequency Domain Decomposition

This method is an extended new version of the Enhanced Frequency Domain

Decomposition (EFDD). The auto-spectral function is curve-fitted directly
in to the frequency domain using SDOF curve-fitter to produce high-quality
estimates of the natural frequency and damping ratio. There is no need of
converting spectral density functions in to time domain using Inverse Fourier
transform as needed in EFDD method. [26].
In summary, for effective determination and study of overall integrity of a
structure accurate data processing algorithms in SHM are required. The de-
sign of continuous real time monitoring of changes in structural elements is
needed to avoid failures in bridges, dams and civil structures. Thus, identify-
ing changes by using output-only modal analysis without caring about input
excitation and determining modal parameters is important for clearly obtain-
ing changes in structures. There exist two categories within the output-only
identification method, namely as time domain and frequency domain analysis.

16
2.6. OUTPUT-ONLY MODAL IDENTIFICATION

Due to their computation intensiveness, the need of prior knowledge of the

mathematical model and the need of database in feature extraction we pre-
fer frequency domain analysis. These methods are mainly on PSD response,
which are advantageous on natural frequency and mode shape determination.
SHM using WSN has the advantage of flexibility, low cost and easy deploy-
ment as compared with wired system, but it suffers from memory and energy
constraints. Therefore, appropriate choice of and implementation of data pro-
cessing algorithms on WSN is a good step moving to an advanced WSN SHM
system. Parametric methods like ITD, ERA or RDT are advantageous for
estimating modal damping ratio, but have difficulty in determining natural
frequencies and mode shapes extraction. On the other hand, nonparametric
methods such as PP, FDD or EFDD are better on determining natural fre-
quencies. Among the non-parametric methods, PP is selected for our work in
this thesis because of its fast and less energy demanding requirement [25].

17
Chapter 3

Wireless Sensor Networks for SHM

In the science of instrumentation, there is a trend of adopting more recent

achievements of sensing technologies into the traditional ones. In the latter,
wireless sensors are one of the emerging technologies that lessen the tedious
works imposed while using wired sensors. They take the sensing technology
into a new realm giving new ways of sensing paradigm. Mass production

Figure 3.1: Example application of SHM using WSN. From [23].

and wide utilization of wireless sensors is made possible due to the extreme
progress of micro electro-mechanical systems (MEMS) technology in the pro-
duction of very small counterparts of huge mechanical tools. Thanks to this
technology, today almost all sensors are being replaced for special applications
that demand the integration of very small components into a system yielding
a great reduction of cost, size and also ease of deployment [8].
As it is the case in other fields, it has been quite a while since the adoption of
wired sensor applications in structural health monitoring to WSNs paved its

19
3.1. APPLICATIONS OF WSNS

way into the field of structural engineering. A reliable SHM requires a dense
instrumentation of structures which is infeasible deploying such structures
with tethered sensors [18]. Fortunately, it is possible to densely deploy wire-
less sensors with comparatively reduced costs and this gives wireless sensors
advantage over the traditional wired sensors.

Beyond just as a simple replacement of wired sensor applications due to their

high savings on budget, WSNs have gained also a great attention due to their
great distributed signal processing capability. They come up with analogue
to digital converter (ADC) and microcontrollers integrated in one unit. The
presence of a microcontroller helps the sensor node to apply some simple signal
processing algorithms. This capability gives the whole system an additional
feature to locally process collected data than sending the whole bulk of data
to the central base station.

In this chapter, the overall features of WSNs for structural health monitoring
is presented. Hardware and software architectures are reviewed.

3.1 Applications of WSNs

Inspired by the developments of battle field surveillance applications in mili-

tary [37], their application has shown much more progress that such networks
are used in many industrial and consumer applications, such as industrial
process monitoring and control, machine health monitoring, environment and
habitat monitoring, health care applications, home automation, and traffic
control [8]. Figure 3.1 shows an application of WSNs for bridge monitoring.

3.2 Wireless Sensor Nodes

A sensor node, also known as a mote is the basic building block of a wire-
less sensor node [40]. Every node in a WSN is capable of performing some
processing, gathering sensory information and communicating with other con-
nected nodes in the network. The size of a sensor node might vary depending
upon the number of components integrated on the unit [31]. According to
Moore’s law and the fast progress of nanotechnology, the size is even expected
to diminish extensively.

Today, a wide variety of commercial and academic wireless sensor node pro-
totypes have been developed and validated. A detailed summary of academic
wireless sensing unit prototypes is found in [37].

20
3.3. HARDWARE COMPONENTS OF A WIRELESS SENSOR NODE

Figure 3.2: The Z1 Zolertia wireless sensor unit.

3.3 Hardware Components of a Wireless Sensor Node

Each sensor network node has typically several parts: a radio transceiver with
an internal antenna or connection to an external antenna, a microcontroller, an
electronic circuit for interfacing with the sensors and an energy source, usually
a battery and/or an embedded form of energy harvesting [21]. Many off-the-
shelf sensor nodes on the market come up with all of the above components
integrated with one or more sensors. Nevertheless, there comes a situation in

Figure 3.3: Hardware architecture of a wireless sensor node.

21
3.3. HARDWARE COMPONENTS OF A WIRELESS SENSOR NODE

which the hardware components that are pre-built on the node do not suffice
for a specific project at hand while designing a system that works with a
wireless sensor node. In such cases, the engineers are required to assemble
hardware components together to form the customized wireless sensor node.
Attaching a separate sensor board is quite a common practice [37] and seldom
independent power supply is provided for such boards while considering duty
cycling to avoid some measurement errors due to the time and power needed
to warm up those external devices.
A detailed explanation on each hardware component follows.

3.3.1 Microcontroller

The microcontroller is the core component of a sensor node which performs

several tasks, processes data and controls the functionality of other compo-
nents in the sensor node. It is the most common controller type in embedded
systems. It has a set of hardware components built around a central processing
unit (CPU), Figure 3.4.

Figure 3.4: Internal components of a microcontroller.

The CPU controls a range of peripherals with both digital and analogue func-
tions such as timers and analogue to digital converters. Small microcontrollers
usually incorporate volatile and/or nonvolatile memory. They can easily be
programmed with C or C++. They are cheaper, flexible to connect to other
devices, easier to program and consume a very low power when compared to
digital signal processors (DSPs), Field-Programmable Gate Arrays (FPGAs)
and Application-Specific Integrated Circuits (ASICs) [30].

3.3.2 Transceiver

Transceivers designate the communication interface of the sensor node with

the functionality of both transmitter and receiver combined into a single de-
vice. Communication is the most energy demanding process in wireless sensor

22
3.3. HARDWARE COMPONENTS OF A WIRELESS SENSOR NODE

nodes and care must be taken while selecting a communication means. Sen-
sor nodes often make use of the industrial, scientific and medical (ISM) radio
band which gives free radio, spectrum access and global availability. The pos-
sible choices of wireless transmission media are radio frequency (RF), optical
communication, Bluetooth and infrared. Infrared and optical communication
media need no antenna and require less energy, but need line-of-sight for com-
munication and are sensitive to atmospheric conditions. Bluetooth has a small
range. RF based communication is the most common means that fits most
of the WSN applications. WSNs tend to use license-free communication fre-
quencies: 173 Hz, 433 Hz, 868 Hz, and 915M Hz; and 2.4G Hz. The operational
states are transmit, receive, idle, and sleep.

3.3.3 Memory

When it comes to processing large amounts of data for quite long time, we need
at least a temporary storage to store a transitional bulk of data before, during
and after processing. In such cases, the built in memory on the microcontroller
hardly suffices. However, from an energy efficiency point of view and speed
of operation, the on-chip memory of a microcontroller and flash memory are
most relevant [8]. Flash memories are typically used due to their low cost
and better storage capacity. Two categories of memory are used based on
the purpose of storage: RAM, random access user memory, which is used for
storing application related or personal data, and ROM, read only program
memory, that is dedicatedly used for storing data used for programming the
device. The MSP430F2617 has 92 KB of Flash memory for program and
data and 8 KB SRAM, a static random access memory, which is common in
microcontrollers [39], that retains its data even when the clock is stopped
provided that power is maintained.

3.3.4 Power source

The sensor node consumes power for sensing, communicating and data pro-
cessing. More energy is required for data communication than any other
process. Power is stored either in batteries or capacitors. Batteries, both
rechargeable and non-rechargeable, are the main source of power supply for
sensor nodes. As wireless sensor nodes are typically very small electronic de-
vices, they can only be equipped with a limited power source of less than
0.5-2 AH and 1.2-3.7 V [8].

3.3.5 Sensors

Sensors are hardware devices that produce a measurable response to a change

in a physical condition like temperature or pressure. They can be analogue

23
3.3. HARDWARE COMPONENTS OF A WIRELESS SENSOR NODE

or digital sensors based on the nature of the output they give to the external
world. In case of Analogue sensors, an external ADC is required to convert
the analogue signals to digital. The ADCs can be integrated in the micro-
controllers or they can be integrated into the same circuit board with the
microcontroller and/or the sensor board. Digital sensors have built in config-
urable ADCs hence we can bypass the external ADCs of the microcontroller.

Figure 3.5: The 3-axis ADXL MEM accelerometer mounted on a circuit board.

Sensors are commonly built on a separate sensor board as can be seen in

Figure 3.6 and are interfaced with the main sensor node externally [36]. In
such cases, the host node should contain an interface to which sensing boards
can be plugged. The sensing interface is largely responsible for converting
analog output of sensors into a digital representation that can be understood
and processed by digital systems.

Figure 3.6: An example sensor board with accelerometer sensor on board.

24
3.3. HARDWARE COMPONENTS OF A WIRELESS SENSOR NODE

The quality of the sensor interface is a function of the conversion resolu-

tion, sampling rate, and number of channels available on its ADC. For most
structural monitoring applications, an ADC resolution of 16-bits or higher is
preferred. Ordinarily, low sampling rates of less than 500 Hz are adequate for
structural monitoring [23].

3.3.6 Energy harvester (Optional)

Energy harvesting (also known as power harvesting or energy scavenging) is

the process by which energy is derived from external sources (e.g., solar power,
thermal energy, wind energy, salinity gradients, and kinetic energy), captured,
and stored. This is an optional component which gives a supplementary source
of power to the battery.

Figure 3.7: Solar panel and rechargeable battery for Jindo Bridge SHM [41]

Energy harvesters provide very small amount of power for energy constrained
systems such as wireless networks to increase their life time by extracting
energy from their environment [7]. Energy harvesting mechanisms based on
solar cells are the ones that are widely used and there are some wireless sensor
units that are incorporated in a box with solar cells on it cover, Figure 3.7.
But, for places like Sweden, where the sun light is not such a reliable source of
energy, other energy harvesting options such as vibration based piezoelectric
energy harvesters would be more feasible. To the best of our knowledge,
although few successful products are available with such features, they are
not yet applicable for small devices like wireless sensor units. But, we hope
that they would be available soon and will help in the long time deployment
of WSNs.

25
3.4. SOFTWARE COMPONENTS OF A WIRELESS SENSOR NODE

3.4 Software Components of a Wireless Sensor Node

The CPU spends much of its processing power performing some calculations
on the data collected by sensors and interacting with peripherals. Both the
algorithms to do the computations on sensor nodes and the policy to interact
and coordinate all the operation functions should be defined and loaded to
the node. These are software components of a wireless sensor node loaded
into the microcontroller for debugging through bootstrap loader (BSL) and
are used to drive the hardware components and the CPU. These are broadly
classified into operating system and application software.

3.4.1 Operating systems

An operating system (OS) is software that manages hardware resources and

provides common services for execution of various application software [43].
It acts as an intermediary between application programs and the hardware for
input output and memory allocation. Operating systems are found on almost
any device that contains a microprocessor or a microcontroller, from tiny
to much complex systems. Unlike other large systems with larger processors,

Figure 3.8: Position of the operating system.

operating systems for WSN nodes are typically less complex and take kilobytes
of memory rather than megabytes [3]. They are written with the memory and
power consumption constraints of the node in mind. Operating systems on
general purpose computers like the desktop are complex with multitasking
features in which several applications and tasks run concurrently on a device.
This requires the availability of large amounts of memory and power on board

26
3.5. CHALLENGES IN USING WSNS

to run all those processes. In contrast, OS in wireless sensor nodes are basically
of event driven systems to load and unload applications on the encounter
of interrupts from hardware or software. Moreover, the operating systems
for wireless sensor networks should be portable across a range of hardware
platforms.
Based on the above requirements, several operating systems for WSNs have
been proposed, written and deployed so far with a focus of one or more func-
tionality. To name some: TinyOS, Contiki, a virtual Java machine based
MagnetOS [9], SensorWare with an abstract scripting language for program-
ming sensors [12], the Mantis system that used a traditional preemptive multi-
threading model of operation [1], etc.
TinyOS is perhaps the earliest operating system that addresses the specific
constraints and limitations of motes [35]. It is a tiny (the core is less than
400 B), flexible, low power, application specific operating system for sensor
networks. It is written in NesC, a dialect of C, programming language and
developers also use this language to write applications.
As this thesis is done with Conitiki, we give a detailed explanation of the
Contiki operating system latter in section 4.2.1.

3.4.2 Application programs

Application programs are any software in the mote other than the operating
system that determine the way the node accomplishes specifically tailored
applications. These programs should be developed with algorithms which are
simple and efficient that ensure effective utilization of memory and power
consumption of the node. The number of instruction cycles that each piece of
program runs should also be limited so that the flow of the control would not
end up in a long loop that eventually triggers the watchdog timers to reset the
hardware. In our case, the FFT algorithm and the peak picking algorithms
are carefully selected in accordance to the requirements.

3.5 Challenges in Using WSNs

The small physical size and simplicity nature of the components that make
up the motes results in corresponding constraints on resources such as mem-
ory, computational speed, communications bandwidth and power supply [5].
Hence, selection of an appropriate wireless sensor node is the first important
step towards the design of applications, operating systems, and even other
plug-in components. Because, the performance of the entire system is depen-
dent upon the individual node capacity and applications should be developed
with those constraints in mind. We discuss the major constraints as follows.

27
3.5. CHALLENGES IN USING WSNS

3.5.1 Memory constraints

Most sensor nodes make use of only the memory available in the microcon-
troller which is not large enough to give room for memory demanding oper-
ating systems, processes and data. The MSP430F2617 microcontroller, for
example, has 92 KB of flash memory for program and data and 4 KB of RAM
for data. All software and data should fit into this small compartment. Thus,
proper selection of small size operating system, less bit representation of data
like data type integer shorts instead of floating points, developing applications
with algorithms that efficiently utilize the memory would help to tackle this
problem.

3.5.2 Power consumption constraints

When it comes to power consumption, sensor nodes operate with limited-

power-supplied batteries. This is a decisive thing that determines the life
span of the sensor network. Hence, power should be redeemed from being
depleted for long term deployments. Changing the batteries frequently would
augment the maintenance costs severely putting one of the main assets of the
WSN monitoring systems into jeopardy; not forgetting the fact that the WSN
monitoring systems could be deployed in remote and/or risky places like high
buildings and bridges.
Possible ways of power consumption reduction are:

a) Low-power components: building the mote from low power compo-

nents (sensors, microcontrollers, radio transceivers) that need less input
power and expend less amount of power [11]
b) Duty cycling: keeping the node switched-on for a fraction of time within
a given period of time. Care should be taken in such cases that some de-
vices, e.g. the radio unit, draw quite a significant amount of power when
they are powered on; and others to warm-up introducing error in data
collection. This corruption, however, can be eliminated by supplying a
separate power source to those components continuously supplying low
power [23].
c) Node-level data processing: since the energy cost of transmitting
1 Kb data a distance of 100 m, e.g., is approximately the same as that used
for the execution of 3 million instructions by a 100 million instructions per
second (MIPS) processor [21], sending information instead of raw data
is beneficial when monitoring vibration based processes which produce
large samples of raw data. This strategy is the most powerful energy
saving method for a longer system lifetime.

28
Chapter 4

Design and Methodology

The design and methodology that we followed is based on the resources that
we have at hand. As the aim of the thesis is to implement a data processing
algorithm on WSNs, the environment on which the designed target algorithm
would be implemented should be considered carefully. In this chapter, we
address the detailed explanation of the methodology and tools that we used
throughout our experiments. We look into the hardware and the software
sections for both the wired and wireless measurement systems used from the
beginning of the thesis work.

4.1 Hardware Used

We used two sets of hardware for measurement. Wireless system and wired
system.

4.1.1 Wireless system

We used different wireless sensor nodes during our thesis work. A t-mote
sky, a Sentilla Jcreate node and the Z1 Zolertia node. The first two were
not suitable for our thesis as a t-mote sky does not have accelerometer while
the acceleromter in Sentilla node was of low quality and unreliable. Thus,
the primary hardware that we used in the wireless system is the Z1 Zolertia
sensor node. Its functional block diagram is given in Figure 4.1. It is one of
the recently developed sensor nodes to test applications and prototypes with
the best trade off between time of development and hardware flexibility. It
is a general purpose low-power WSN module that can be used as a develop-
ment platform for WSN applications. Its core architecture is based upon the
MSP430+CC2420 family of microcontrollers and radio transceivers respec-
tively by Texas Instruments, which makes it compatible with other motes,

29
4.1. HARDWARE USED

like Crossbow’s TelosB, Moteiv’s Tmote sky, and alike. Immersing devices
in the Internet of Things (IoT), Personal health care monitoring, environ-
mental monitoring, emergency detectors, safe and rescue devices, long-term
unattended monitoring, energy metering, agricultural monitoring are some
of the applications that the mote is entitled for. It has a temperature and
accelerometer sensors already included on board.

Figure 4.1: Functional block diagram of Z1 WSN module.

Based on the hardware components of a sensor node from Section 3.3, a de-
tailed explanation on hardware component which are relevant to this thesis
follows.

Microcontroller

The Z1 Zolertia mote, see Figure 3.2, is based upon the MSP430 family of
microcontrollers from Texas instruments, specifically a 64 pin MSP430F2617
[51]. MSP stands for mixed signal processor (analogue or digital) [39] and
the processor has ADC12, a 12-bit 8 channel analogue to digital converter for
mixed signal use. The F indicates that the program memory in the microcon-
troller is of flash memory, the most common type of memory that can be both
programmed and erased electrically. It is a 16-bit processor, both the address
bus, the data bus and all the 16 registers in the CPU are 16-bits, with a Von
Neumann architecture, designed for low-power applications. Its CPU is often
described as a reduced instruction set computer (RISC) in which simplified
sets of instructions are provided for faster execution.
In addition, other feature of the MSP430 that makes it preferable for low power
applications is that the device can easily be put into a low-power mode. The

30
4.1. HARDWARE USED

Table 4.1: Summary of parametric features of MSP430F2617.

Parameter Value
Frequency support for a 32768 Hz
Flash 92 KB
Static RAM 8192 B
Timers - 16-bit 2
Multiplier 16x16
ADC 12 bit 8 channels

mode can be controlled by bits in the status registers without the need for
special instructions. Many of the peripherals can run autonomously without
the CPU for most of the time and the processor can be awakened from a
standby mode rapidly by an interrupt and returns automatically to its low-
power mode after handling the interrupt. This feature is extremely important
in structural health monitoring as the state of the device is controlled based
on the traffic load on the structure. Table 4.1 summarizes some parametric
features of MSP430F2617.

Transceiver

The communication module in Z1 Zolertia is a Texas Instruments CC2420

[51], a 48 pins, low cost 2.4 GHz IEEE 802.15.4 compliant RF transceiver,
Table 4.2.

Table 4.2: Summary of CC2420 features.

Parameter Value
RX Current Consumption 19.7 mA
TX Current Consumption 17.4 mA
Data Rate(Max) 250 Kbps
Frequency Range 2400 Hz - 2483.5 Hz
Operating Temperature Range -40℃to 85℃
Operating voltage range 2.1 V - 3.6 V
Output Power Range -25 dBm to 0 dBm
RSSI Output Digital

It is connected with an integrated Yageo/Phycomp ceramic antenna [51]. Ex-

ternal antennas can be used when large range communication is needed.

31
4.1. HARDWARE USED

Memory

The MSP430F2617 has 92 KB of Flash memory for program and data and
8 KB RAM for storing temporary data. In addition to these on chip memories,
it is integrated with an external flash memory. This can be used as extension
to store data while programming or collecting data; however, care should be
taken while doing so as writing and reading operations from this memory take
extra time and CPU cycles which add unnecessary cost to the entire system.

Power supply

The Z1 Zolertia mote should nominally be powered at 3 V although it may

work partially or totally from 1.8 V, without radio, to 2.1 V, with radio. The
absolute maximum rating for the whole component is, Table 4.3, in the range
-0.3 V to +3.6 V. No battery is needed if it is connected to the mini-USB port
for programming or communication.

Table 4.3: Approximate current consumption of Z1 components.

Component Operating Range Current Consumption Case
1.8 V to 3.6 V 0.1 µA OFF Mode
0.5 µ A Standby Mode
MSP430F2617
0.5 mA Active Mode @ 1 MHz
<10 mA Active Mode @ 16 MHz
2.1 V to 3.6 V <0.1 µA OFF Mode
20 µA Power Down
CC2420 426 µA IDLE Mode
18.8 mA Rx Mode
17.4 mA Tx Mode @ 0 dBm
1.8 V to 3.6 V 0.1 µA Standby
ADXL345
40 µA to 145 µA Active Mode
2.7 V to 3.6 V 1 µA Deep Power Down
M25P16
40 mA to 15 mA Active Mode
1.4 V to 3.6 V 1 µA Shutdown Mode
TMP102
15 µA Active Mode

Sensors

The Z1 Zolertia mote comes with TMP102 Thermometer from Texas Instru-
ments and ADXL345 digital accelerometer from Analogue Devices. We used
the accelerometer to measure vibration data from our lab beam model.

32
4.2. SOFTWARE USED

4.1.2 Wired system

The wired system was used to validate the instruments and the algorithms
used in the wireless system. The system available in the division of struc-
tural engineering and bridges, KTH consists of a high resolution 24-bit data
acquisition system HBM MGC Plus1 , and a very sensitive, extremely low-
noise level measuring Si-Flex™ MEMS accelerometer and a laptop computer
loaded with Windows XP operating system. Although high quality data is
collected using this system, it incorporates a bulky set of cables and huge
high cost hardware equipments depicting the wireless system replacement of
the overall system quite a big relief.

4.2 Software Used

In this section, we discuss the software tools that we used. They are sub-
categorized as realtime and offline software tools.

4.2.1 Realtime software tools

By realtime software tools, we mean software components embedded into the

mote to run while processing the entire monitoring. According to section 3.4,
these are the operating system and the application software. The operating
system that we loaded onto the the sensor nodes is the Contiki operating
system and the application program is fixed-point Real-valued FFT, RFFT.

Contiki

Contiki is a lightweight, highly portable, small open source operating system

primarily developed for networked embedded systems. Contiki is written with
the C programming language and applications are also written using the same
language. It is portable on a number of memory-constrained systems with
processors ranging from 8-bit to high level microcontrollers [2]. A typical full
functioning system on a device with a graphical user interface, for instance,
requires just about 30 KB of memory. It is highly flexible whereby memory
efficiency can be increased by reconfiguring it to as low as tens of bytes with
a code footprint on the order of kilobytes. It gives a built-in IPV6 compatible
TCP/IP connected into the Internet to reap the benefit of the Internet of
things.
A Contiki system is partitioned into the core and the loaded programs, see
Figure 4.2, at compile time depending on the specific deployment.
1
http://www.hbm.com

33
4.2. SOFTWARE USED

Figure 4.2: The service architecture of Contiki. (From [2])

Embedded Fast Fourier algorithm

Some embedded microprocessors may have an external unit capable of per-

forming floating point arithmetic, but most low-end embedded systems like
MSP430 ultra low power microcontrollers have no floating point arithmetic
operation. Thus, MSP430s do not apply floating point rule on their program;
hence, fixed point arithmetic operation is required to obtain proper results.
The main reasons why we use fixed point arithmetic operations in our imple-
mentation for emulating floating point calculations are:

– Limited memory of The microcontroller which has 8 KB of RAM and

92 KB of flash memory. The algorithm requires 2N 16-bit variables for
FFT data, and our microcontroller can’t perform FFTs for values of
N greater than 1024. Even this is difficult because other parts of the
firmware will also require a few bytes of RAM. For our implementation,
we therefore limit N to 512 due to overflow problem. Having small num-
ber of samples would have effect on frequency resolution; thus having
large number of sample for FFT is recommended.
– Limited speed: low power microcontrollers have low speed for performing
FFT which is significantly slower.

We used 16-bits variables in our implementation of fixed point FFT to rep-

resent the real and imaginary parts of every value. In doing so, for a 16 bit
digit to represent a number:

– Bits 0 to 6 represent the fractional part

34
4.2. SOFTWARE USED

– Bits 7 to 14 represent the integer part

– Bit 15 holds the sign bit l

This format has no effect on adding and subtracting but we must be careful
when performing multiplication that would cause overflow. For the forward
FFT (time to frequency domain), fixed scaling is performed to prevent arith-
metic overflow while multiplying two 16 bit numbers. It’s important to store
the intermediate product in 32 bit representation and perform shifting by 15
bit to get the 16 bit result. If we want the format of the output to be the
same as the format of the input, we must restrict the range of the inputs to
prevent overflow.

Lookup Tables

The twiddle factors calculations in FFT are one of the processes that consume
too much processing time of the CPU. In each and every instance of their oc-
currence, the processor is required to compute the values of the trigonometric
sine and cosine functions. The number of this computation is directly pro-
portional to the size of the FFT. Therefore, by using a lookup table of these
trigonometric functions, we can extremely reduce the extra computation time
claimed. In doing so, we load pre-calculated values of those trigonometric
functions into an array of memory to be called with array indexing. This is
based on the fact that it is much economical and faster to load data from
memory than do computations on the run.
One decisive thing we should note is that the range of sine and cosine func-
tions lies in the interval [-1,1]. For any angle θ the following condition should
be fulfilled: | sin(θ)| ≤ 1 and similarly | cos(θ)| ≤ 1. Which means that the
magnitude of most of the values are less than 1 and floating point. Hence
proper scaling is important here as well. Keeping in mind the 16-bits mi-
croprocessor, we scaled the trigonometric table multiplying by 215 = 32768
and to compensate for the truncation error introduced while representing the
floating numbers with integers, we added 0.5 to each value. Of course, we
need to scale back this result latter dividing by the same number somewhere
in the computation.

Peak Picking

The peak picking is the next experiment that is done after the FFT. To do the
peak picking, the frequency spectrum is needed to be calculated before hand.
The standard way of doing the peak picking is by calculating the magnitude
of the FFT output samples. This is done with the Euclidean norm as:
q
|X| = |Re{X}|2 + |Im{X}|2 (4.1)

35
4.2. SOFTWARE USED

But, to implement the standard square root function we need to include the
math C library header in our code and this would add to the memory consump-
tion and processing cycle. to avoid this loss, we approximate the magnitude
simply as:

|X| = |Re{X}| + |Im{X}| (4.2)

Although this is not 100% accurate, it better approximates the magnitude

with so much save on the processing resources.

Set Threshold

The crucial part in applying a peak picking algorithm is to set the threshold
so that a limit should be given to pick the peak. This is the lower threshold to
discard the data below the limit and concentrate on the ones above the limit.
we used a standard deviation and mean based limit. Setting the lower limit
just from the mean value would end up in losing lots of significant peak val-
ues at higher frequencies. Thus, we added the standard deviation into the
arithmetic so that the limit should be set fairly in an appropriate way. In
probability and statistics, the standard deviation, σ, is calculated as:

q
σ(X) = (X − X̃)2 (4.3)

Where X is sample and X̃ is mean of the samples. It is a widely used stochastic

index to show how much dispersion there is from the average (mean).
The equation that we used to set the lower limit, X, is:
¯

X = X̃ + 2σ(X) (4.4)
¯
This gives a peak picking limit with 95% confidence.

Sample Window Size

After setting the limit, care should be taken while setting the sample size to
compare the peaks within. Fortunately, it was not a problem in the exper-
iments we did, but there are cases where the mode frequencies are large or
closely spaced. In those situations, setting the windows we would lose those
closely spaced modes next to the dominant peak in the range. Observing the
nature of the frequency spectrum would help to set a good approximation.

36
4.2. SOFTWARE USED

Figure 4.3: A frequency spectrum plot divided into non-overlapping windows.

In our experiments, we used a non-overlapping moving window size of 50

samples sizes.

How the Algorithm Works

Having set the lower limit, the algorithms picks a sample from the beginning
of a window and compares it against the next two consecutive samples. If
the sample is found to be the peak value, it sets it as a local maximum value
and then moves to the next sample and does the same thing, but in this and
the next cases, it does one more comparison with the immediate previously
found local maximum value. If the new sample is found to have higher peak
than the previous local maximum value, it is set to be the new local maximum
value and the next comparison is done against this new value and it goes on
like this with in the window and the whole process is repeated in the following
windows up to the end.

Computing the Natural Frequency

After the peak values are found. The next step is to calculate the natural
frequencies. From the theory of vibration analysis, the resonance natural
frequencies are believed to occur at those peak amplitude values. Multiplying
the instant sample number with the frequency resolution gives the natural
frequency.

After all the above procedures are performed, the resonance frequencies are
ready to be transmitted to the sink station where final modal analysis are
done and the overall work is finalized.

37
4.2. SOFTWARE USED

4.2.2 Offline software tools

These software tools are signal processing tools that we used to analyze the
data collected by the wireless and wired sensor systems and hence validate
the algorithms implemented and the instruments used in the entire process.
Matlab, a high level language with advanced signal processing toolkits, is
used on Windows based computers. SciLab2 , which is an open source free
software, was used for Unix based machines. Catman™3 a powerful data
acquisition software for configuring, visualizing and analyzing measurements,
was integrated with the wired measurement system. It has interactive user
interface and provides a wealth of mathematical and graphical functions for
analyzing and evaluating measurement data and data can be exported into all
common standard formats (for example Excel or ASCII).
On the wireless side, most of the experiments were latter done by collecting
data with Matlab through serial communication. We wrote a code so that
four nodes are connected to our laptop through a USB hub. After the data
are collected, they are post processed and plotted in a time and frequency
domain subplots.

2
http://www.scilab.org
3
http://www.hbm.com

38
Chapter 5

Accelerometers

5.1 Introduction

Accelerometers are the major instruments used for vibration measurement and
data collection in this thesis. We used different accelerometers for wireless and
wired systems. While exploring which accelerometer to use, we went through
different sets of choices and in this section, we will present a detailed walk
through of the ways we chose the ones that we ended up working with. After
a generic explanation of accelerometers, a tabular summary of the features of
some of them that we encountered is given latter in the section.

5.2 What Accelerometers Are?

An accelerometer is an electromechanical device that is useful to measure

static or dynamic acceleration forces in units of meters per second squared
(m/s2 ) or G-force (g). By static force, it is meant that the accelerometer
measures constant force of gravity pulling. This helps, e.g., to find out the
angle by which the device attached to the accelerometer is tilted at with
respect to the earth. And by dynamic force, it means that it measures force
caused by moving or vibrating the accelerometer as in SHM.

5.3 Selection Criteria

The following characteristics should be considered while selecting accelerom-

eters [49] for SHM systems:

39
5.3. SELECTION CRITERIA

5.3.1 Analogue or digital

Analogue output accelerometers output a continuous voltage that is propor-

tional to the sensed acceleration. As most microcontrollers come up with a
built in analog to digital converter, these accelerometers are the easiest to in-
terface with. Where as in microcontrollers with purely digital inputs a digital
output accelerometer is preferable. Digital accelerometers usually incorporate
a serial interface, SPI or I2 C to communicate with the microcontroller.

5.3.2 Sensitivity

It is the rate at which the signal reading changes with the acceleration. The
higher the sensitivity of the accelerometer the better is the measurement as
ambient vibrations can accurately be measured.

5.3.3 Full-Scale range

This determines the range of readings one does. The full-scale range of an
accelerometer and its sensitivity are inversely proportional in that a smaller
full-scale range means a more sensitive output. A ±2 g range accelerometer
should suffice to measure vibration of a structure.

5.3.4 Number of axes measured

This is the number of possible axes (x, y, and z) out of which the accelerometer
senses. 1- and 2-axis accelerometers are the most common ones. Nonetheless,
if there is no other limitation that hinders from doing so, the 3-axis accelerom-
eters are more preferable as they give a surplus of information against the 2-
and 1- axis accelerometers. Besides, they come up with affordable prices with
negligible difference.

5.3.5 Bandwidth response

This is the frequency range of vibration data that the acceleromter can mea-
sure.

5.3.6 Noise level

This is the lower limit of the measurement below which is considered to be

unnecessary noise. It is the smallest amount of vibration an acceleromter can
detect and is an important factor to properly measure the ambient vibration
data. The lower is better as small vibrations are considered as noise if this
floor is higher.

40
5.4. WHICH ONES DID WE USE?

5.3.7 Power consumption

Accelerometers in wireless sensor nodes should not consume too much power.
It should be in the 100 µA range. Often, some accelerometers feature a sleep
functionality to conserve energy when not needed.

5.3.8 Extra features

Modern accelerometers may include features like sleep control, tap sensing,
free fall (0 g) detection and selectable measurement ranges. These can be
used to trigger certain functionalities.

5.4 Which Ones Did We Use?

In the next sections we are going to give a detailed explanation of the ac-
celerometers that we used.

5.4.1 Analogue devices ADXL345

The ADXL345 [4] is a digital accelerometer, Figure 3.5, found in Z1 Zoler-

tia node. It is a 14 pin, ultra-low power, accelerometer with 10- to 13-bit
resolution. Although it is a 3-axial accelerometer, we used a single axis mea-
surement in the vertical direction as most of the properties of a structure can
be identified in that dimension. With a 10000 g absolute shock rating, it is
well suitable for SHM.

Table 5.1: Summary of the ADXL345 accelerometer features.

Parameter Value Unit
Linear Output Range ±2 - ±16 g peak
Sensitivity 29 - 282 LSB/g
Bandwidth Response 0.1 to 3200 Hz
Dynamic Range 120 dB
Noise 0.75 to 1.1 LSB rms
Operating Temperature Range -40 to +85 °C
Input Voltage 2 to 3.6 Volts DC

Several extra features are also provided. Activity and inactivity sensing fea-
tures detect the presence or absence of motion by comparing user set thresh-
olds with the acceleration in any axis. These features help the ADC to con-
tinuously sample input data and check the amount to determine if the whole
system should be actively responding to the change in acceleration. Hence the

41
5.4. WHICH ONES DID WE USE?

system remains in sleep mode unless otherwise woke up by a rise of accelera-

tion above the threshold saving a tremendous amount of power. On the other
hand, tap sensing detects double taps and single taps in any direction while
free fall sensing detects whether the device is falling. Some more important
specifications are summarized in Table 5.1.

Data collection with ADXL345

ADXL345 provides two output pins for driving interrupts which can be en-
abled by setting the appropriate bit in the INT_ENABLE register. Differ-
ent set of actions are activated with these interrupts like Single_Tap, Dou-
ble_Tap, Free_Fall detection and advanced features such as Activity and
Inactivity detection when they pass some threshold values set by the user.
ADXL345 contains an integrated memory management system with a 32-level
first in, first out (FIFO) buffer. This buffer has four modes: bypass, FIFO,
stream, and trigger.
We first save memory space for buffer size of 512 samples to collect data. In
order to obtain a synchronized data as possible from the wireless sensor nodes
simple broadcasting is implemented on the nodes. The message sent is a simple
"hello" text to trigger other nearby sensors to simultaneously start sampling
after receiving the broadcast message. In this case, one of the wireless sensor
nodes act as an event triggering node. Then when the other sensor nodes
receive the broadcast message, the appropriate bit register will be enabled with
either of the pins mapped to the interrupts on the ADXL345 accelerometer.
In order to collect the X, Y, and Z -axes vibration data, set FIFO mode and
data are stored in FIFO. When the number of samples in FIFO equals the
level specified in the samples bits of the FIFO_CTL register,the WATER-
MARK interrupt is set. WATERMARK is an interrupt and a bit is set when
the number of samples in FIFO equals the value stored in the samples bits
(Register FIFO_CTL)[4]. The WATERMARK bit is cleared automatically
when FIFO is read, and the content returns to a value below the values stored
in the samples bits. The accelerometer does not stop working while FIFO con-
tinues collecting samples until it is full (32 samples from measurements of the
X-, Y-, and Z-axes). WATERMARK interrupt still continues to occur until
the number of samples in FIFO is less than the value stored in the samples
bits of the FIFO_CTL register.
when the memory space of the buffer for the set sample size is full it shows
warning by printing "WARNING: FIFO OVERRUN". This happens when
new data replaces unread data and it gives indication of data collection is
overrun. The actual overrun bit is set when FIFO is filled. The overrun bit is
automatically cleared when the contents of FIFO are read. Hence, the registers
will fill the buffer until it reaches the maximum buffer size and anymore data

42
5.5. VALIDATING THE ACCELEROMETER

will not be written to the accelerometer then finally it will print out the values
collected by the accelerometer.

5.4.2 Si-Flex™ accelerometer

This is a relatively low-noise 3-axial analogue accelerometer that is used in

the wired system. It is among a line of sensors produced by Si-Flex™ and the
sensors are characterized by features such as wide dynamic range, excellent
bandwidth, low distortion, high shock tolerance, and thermal stability.

Table 5.2: Summary of the Si-Flex™ accelerometer features.

Parameter Value Unit
Linear Output Range ±3 g peak
Sensitivity 1.2 V/g
Bandwidth Response DC to 1000 Hz
Dynamic Range 120 dB√
Noise 300 to 500 ngrms / Hz
Operating Temperature Range -40 to +85 °C
Input Voltage 6 to 15 Volts DC

Being the best accelerometer available, this accelerometer would have been
the right one to be used with wireless systems. However, the fact that it
needs much operating power input and its bulkiness are the factors that we
didn’t consider it in our selection.

5.4.3 Other accelerometers

We also encountered other dozens of accelerometers and in Table 5.3 we give

the overall summary of their characteristic features. Most of the accelerom-
eters in the list are also among the ones used by most other wireless SHM
system projects: SMB380 from Bosch Sensortec, LIS302DL from ST Micro-
electronics and MMA7361L. The LIS302DL accelerometer is the one similar
to the accelerometer in iPhones and MMA7361L is the accelerometer in Sen-
tilla node, the node that we were practicing with during the beginning of the
thesis.

5.5 Validating the Accelerometer

The validation of the accelerometer selected is made by comparing the mea-
surements against a well known accelerometer that measures accurate values.
In our case, we did this by comparing the measurement from our wireless
accelerometer against the wired one.

43
5.6. CARES TO BE TAKEN

Table 5.3: Summary of the most common accelerometers in SHM.

Parameter SMB380 LIS302DL MMA7361L
Linear Output Range ±2 - ±8 g ±2 - ±8 g ±1.5 - ±6 g
Sensitivity 64 - 256 LSB/g 16.2 - 79.2 mg/digit 206 - 800 mv/g
Bandwidth Response 25 - 1500
√ Hz 20 - 4000
√ Hz 400 Hz
√
Noise 500 µg/ Hz 50 µg/ Hz 350 µg/ Hz
Operating Temperature -40 to 85 °C -40 to 85 °C -40 to 85 °C
Input Voltage 2.4 - 3.6 V 2.16 - 3.6 V 2.2 - 3.6 V

5.6 Cares to Be Taken

Cares should be taken while using an accelerometer so that the measurement

should not go erroneous. Some of the most important factors that need be
considered before using accelerometers are:

5.6.1 Calibration

Accelerometers measure the acceleration of a structure by detecting the rate

of the relative change velocity. This relativity is taken by setting a reference
level. Hence, setting reference is the first thing to do while using accelerome-
ters. This is termed as calibration and in doing so accurate vibrations can be
acquired. This is generally done by taking random measurements by setting
a structure in its initial position. Those random measurements are averaged
and are considered as 0 g from which the next measurements are contrasted
within the range of the accelerometer. In case of measuring the vertical ac-
celeration, the random average value is rather considered as 1 g as it is in the
direction of the earths gravitation.
In our case, the calibration is done by measuring the ambient vibration of
a structure and is considered as 1 g and is taken as the reference level of all
coming measurements. Thus, this ambient measurement is averaged and the
average is subtracted from any further measurements.

5.6.2 Attachment

The way the accelerometer is attached determines the quality and accuracy
of the readings. The accelerometer should be attached as firmly as possible to
the structure so that the whole integrated structure should be stiff enough. If
the attachment is loose, this gives unnecessary extra acceleration due to the
free movement of the accelerometer from the main structure.

44
5.6. CARES TO BE TAKEN

5.6.3 Axis alignment

The right axis orientation should be done in the driver code at the time of
collecting data. For example, while the accelerometer is attached with the
vertical orientation bound towards the Z-axis, assigning X-axis in the driver
code to be measured will account for some unexpected results.

5.6.4 Sampling rate

In case of using a digital accelerometer, special care should be taken while se-
lecting the sampling rate. As described in section 5.3.1, digital accelerometers
are interfaced with the host device with serial interface. The quality of this
connection affects the rate at which data can be retrieved. In the ADXL345
accelerometer case, using higher sampling rates would lead to improper data
collection. Hence, using sampling rates of 100 Hz is what we did. data col-
lected by an accelerometer may vary depending on the sensitivity, resolution,
sampling rate and range of the accelerometer. Thus literal comparison on the
time domain is not the best option. However, no matter what the quality
of the time domain data from different accelerometers vary, they all should
measure the same frequency of the material.

45
Chapter 6

Experimental Setup

We used a simple steel beam, Figure 6.1, supported by wooden blocks from
two ends to experiment our laboratory measurement works. The length of
the beam was 3.5 m and the wooden supports, Figure 6.3, at the end points
take a total of 18 cm thus the whole length excluding the wooden support
becomes 3.32 m. This span was further divided into five subparts of each
66.4 cm long. By placing our sensor nodes on these subparts, we collected our
measurements for realtime and offline analysis. A simple illustration of the
steel beam is given in Figure 6.2 as a simple bar supported at both ends.

Figure 6.1: Part of the 3.5 m long test beam

The motivation behind the selection of this beam structure is that, it is very
simple in form, easily available to do repeated experiments and move around.
Since all we needed was a structure which can give a good output for a light
excitation to verify our algorithm, this was the simplest and best available in
the lab that can fulfill our requirements.

47
Figure 6.2: A simple beam model of experiment 1 showing the parts 1 to 4 for each
attachment (all measurements are in cm).

Figure 6.3: The wooden support at the ends of the experimental beam.

The steel beam is considered as simply supported beam structure thus we

can theoretically determine the frequency, ωn , by a simple equation from the
strength of materials [6].
s
2 EI
ωn = (nπ) (6.1)
ρAL4

Where:

– n, Mode number
– A, Cross-sectional area of the beam (m2 )
– ρ, Density(kg/m3 )E, Y oung 0 smodulus(N/m2 )

48
6.1. INSTRUMENTATION

– I, Area moment of inertia (m4 )

– L, Effective length of simply supported beam (m)

The second modal frequency is four times the first modal frequency.

6.1 Instrumentation

Both the wired and wireless sensor systems were used to instrument the beam.
For the wired system, HBM MGC Plus data aqcuisition system (DAQ) was
used.

Figure 6.4: The front end of the instrumentation on the experimental beam. a)
Catman® window b) Matlab window c) Rear view of the HBM MGC Plus

This system encompasses, a Si-Flex™ MEMS sensor to be firmly attached to

the beam with a heavy electromagnet attachment, a multichannel ADC and
a Catman® DAQ software (see section 4.2.2) installed on a laptop computer.
Our Z1 Zolertia wireless sensor nodes are placed at four points on the beam
while the Si-Flex™ , see Figure 6.5, accelerometer is placed at one fixed po-
sition on the beam and data is obtained on our laptop terminal through the
mini-USB cable while the data from the Si-Flex™ accelerometer are acquired
to the Catman® DAQ software.
We need to make sure not to excite above the ±1g range so that not to harm
the instruments and the setup.

49
6.2. EXPERIMENTAL PROCEDURES

Figure 6.5: Attachment of the sensor units on the experimental beam.

Figure 6.6: Exciting the experimental beam with an impulse hammer.

6.2 Experimental Procedures

In order to obtain as synchronized data as possible from the wireless sensor

nodes in the lab test, we used simple broadcasting message on the sensor
nodes. We used 4 wireless sensor nodes to collect vibration measurements from
the simple beam on the experimental lab test. Additional sensor node is used
to send a broadcast message to all the sensors to trigger the accelerometers
to collect the vibration at the same time. When the sensor nodes receive the
broadcast message, they automatically start sampling simultaneously for a
sample size of 512 samples.
There exists an average time drift of 0.22 s between measurements among the
wireless sensors due to lack of applying synchronization protocol. This time

50
6.3. EXPERIMENTS

drift is not constant and has effect on determining the correct mode shape of
the structure because it is impossible to determine the relative phase difference
between the senor nodes and the imaginary components of the FFT data.
When a time synchronization algorithm is applied, the time-delay occurrs is
less than a few milliseconds [38].

6.3 Experiments

We divided the experiments into three sets as:

– Experiment 1: Sensor nodes are placed at 66.4 cm between each other.

This is a setup where the sensors divide the beam into five equal subparts.
In doing so, it is possible to capture more vibrations using the sensor
array.
– Experiment 2: two sensor nodes are placed at the two wooden supports.
In this case, we can see if no or zero vibrations are measured at the
supports as they are defined to be fixed from our definition.
– Experiment 3: The sensors are placed 66.4 cm apart from each other as
in experiment 1 but here we measure ambient vibration. This setup helps
us see the sensitivity of the accelerometers to measure ambient vibrations
in the absence of external impact.

51
Chapter 7

Results and Discussion

In this chapter, the results of the measurement setups are analyzed and pre-
sented. The validation of the accelerometers and the FFT algorithm, time
domain and the frequency domains of both the wired and wireless systems are
addressed and relevant discussions are made.

7.1 Validation

We divided the validation into categories:

– Validating the accelerometer

– Validating the FFT

7.1.1 Validating the accelerometer

The accelerometer validation was done at our table before we continue the
final test in the lab. The validation was performed by comparing its measured
frequency components against an iPhone Vibration measurement.
Vibration is a spectrum analyzer in iPhone using the built in accelerometers.
It acquires and displays time series data, optionally removes DC bias, applies
a Hamming window and performs an FFT on each channel to produce fre-
quency spectra. The 3-channel accelerometer in the iPhone has a sensitivity
of approximately 0.02 g and a range of ±2 g. Other features of the application
includes:

– Variable start delay from 0 to 20 s

– Selectable data length from 128 to 1024 samples
– Adjustable sample rate from 10 Hz to 100 Hz

53
7.1. VALIDATION

The time domain and frequency plots obtained from the iPhone measurements
are given in Figure 7.1.

Figure 7.1: Time and frequency domain data from iPhone

Figure 7.2 shows the same measurement done with our wireless sensor node.
The frequency of the vibration from the iPhone, 19.34 Hz, is almost the same
as the one from the wireless node, 19.26 Hz.

Figure 7.2: Time and frequency domain data from the wireless sensor

7.1.2 Validating the FFT

The second step is validating our FFT. After collecting a time domain data
with the wireless sensors, we perform floating point FFT and compare the
result with floating Matlab FFT. Figure 7.2 shows the floating point FFT
result.
We used 512 samples to perform FFT. The total amount of samples that
can be collected with Zolerita node is 1024 samples with data type 16 bit
integer, but for performing FFT we have to first declare two variables with

54
7.1. VALIDATION

buffer size of 1024 for the real and imaginary components which would be
2048 samples and this, together with the program data, can not be supported
with the 8 KB Zolertia node flash memory. Thus, we used 512 data samples.
Figure 7.3 compares the frequency spectrum computed with a 16 bit integer
approximation FFT with the one obtained by the standard 32 bit floating
point offline FFT with Matlab/Scilab.

Figure 7.3: Comparison between Floating point and Integer point FFT

As we can see from Figure 7.3 the integer FFT gives a good approximation of
the floating FFT on top of saving memory space and reduce energy consump-
tion.

We also implemented floating point FFT on the nodes and we put the com-
parison in Table 7.1.

Table 7.1: Comparison of floating point and fixed point FFT on the node
Floating point FFT Integer point FFT
Number of samples 512 512
Size on memory 41 KB 24 KB
Running time 6s 0.6 s

One big issue while using floating point FFT on the nodes is that due to
the large memory size consumed during collecting the time domain data and
because of the large number of cycles in the FFT, the node resets prematurely
before printing out the FFT output values. Therefore the result obtained from
the floating point FFT is not reliable.

Given the low resolution data from the ADXL35 accelerometer, the most
important thing is noting the existence of the resonant peak frequencies. These
points are what we want to know to go further to extract the important
information to feature the behavior on the model bridge.

55
7.2. TIME DOMAIN DATA ANALYSIS

7.2 Time Domain Data Analysis

A time domain plot of the data from the wired and wireless systems is given
in Figure 7.4.
As can be seen from the Figure, the time domain data plot from the wired
system is relatively of better quality than the one from the wireless system.
The data obtained from the wired and wireless look similar to some extent
which indicates that both are measuring same vibration data from the bridge.

Figure 7.4: Time domain data from wired and wireless sensors

The close up look into those two measurements exposes some of the flaws in
the wireless sensing system. At low amplitudes, they introduce noise due to
the low resolution (10 bits per sample) of the ADC used in the wireless system
in contrast with the 24-bits high resolution one used in the wired system. Here,
the 14-bits difference in the fidelity of both data is considered as one of the
trade-offs in using wireless sensor system.
The difference for the signal quality can be attributed to:
a) Noise √level: The Si-Flex™ accelerometer has lower noise level (300
ngrms / Hz) than that of the ADXL345 accelerometer (<1.5 LSB rms).
Hence less amount of noise is introduced into the measurements of the
wired system.
b) ADC resolution: The MGC plus DAQ system has a 24-bit ADC which
is of much resolution than the ADC found in the ADXL345 accelerometer
(10 bits of resolution for a g-range of ±2g). The more the bit resolution of
the ADC, the more the quantization level of the converter and eventually,
this yields digital samples of more fidelity.
c) Attachment: The Si-Flex™ accelerometer is attached so firmly to the
structure with an electromagnetic attachment which gives it’s system a
great stiffness to freely vibrate with the whole structure giving a good
signal quality.

56
7.3. FREQUENCY DOMAIN ANALYSIS

Figure 7.5: Measurement from the wired sensor

The data from the wireless sensor also has higher amplitude fringes beyond
±1 g at the beginning of the vibration as a result of imperfect attachment
between the two sensors that make the wireless sensor resonate during the
initial conditions. Using stronger attachments like magnetic ones can improve
signal quality.

7.3 Frequency Domain Analysis

We made three experiments with different sensor placements and the results
are put accordingly.

7.3.1 Experiment 1

In experiment 1, Figures 7.6 - 7.9, peak frequencies can be seen at sensors 1

to 4 as 12.76 Hz 13.1 Hz 13.29 Hz and 13.1 Hz respectively. Where as in the
wired system in Figure 7.5 the peak is at around 13 Hz. These peaks are the
first modal frequencies of the model beam.
These peak frequencies are also the same with the ones that we found using
our pick peaking algorithm.

7.3.2 Experiment 2

Placing the sensors just at the two end points at the wooden support would
give measurements only at sensor positions 2 and 3 while sensors 1 and 4 give
no frequency readings as they are placed right on the firm supports where no
vibration is expected.

57
7.3. FREQUENCY DOMAIN ANALYSIS

Figure 7.6: Node 1 experiment 1

Figure 7.7: Node 2 experiment 1

Since the wired sensor placement is not changed, giving the same excitation,
the result obtained is similar to experiment 1 as shown in Figure 7.5. The
wireless sensors, Figure 7.10, provide the expected.

7.3.3 Experiment 3

In this experiment, since the ambient vibrations were so small, a vibration

can barely be detected on the wireless system.
The wired sensor and the wireless sensor positions are not changed as shown
in Figure 6.2; but the input excitation here is ambient vibration just walking
near the support of beam. The wireless sensors are not able to measure this
small vibration and provide zero natural frequency, Figure 7.11. In the wired
system; however, there exist very small vibration measurement, depicted in
Figure 7.12, and the natural frequency of the beam is obtained even if the
magnitude is 140 times smaller than the one obtained in experiment 1.

58
7.4. THE CAUSE OF THE FREQUENCY DIFFERENCES

Figure 7.8: Node 3 experiment 1

Figure 7.9: Node 4 experiment 1

In general, the number of peak frequencies observed are similar on the lab
experiments. However, the values of the frequencies at which these peaks are
found show a slight difference from the wired measurement and among the
experiments as well.

7.4 The Cause of the Frequency Differences

Before our final experiment, we had encountered high frequency measurement

differences between wired and wireless sensors upto 5 Hz.

59
7.4. THE CAUSE OF THE FREQUENCY DIFFERENCES

Figure 7.10: Experiment 2 wireless sensor data

Figure 7.11: Experiment 3 wireless sensor data

Figure 7.12: Experiment 3 wired sensor data

60
7.4. THE CAUSE OF THE FREQUENCY DIFFERENCES

Getting the same readings from both the wired and the wireless systems took
too much of our thesis time. We tried to figure out the possible sources of
the frequency differences between the readings from the two systems. As
the development of FFT or PP algorithms depend on the similarity of the
results of the measurements, we should make the readings similar enough to
proceed with the rest of the thesis. The readings from the wired system were
reliable as we verified them with the iPhone application called Vibration. The
two readings match well. Thus we deduced that the problem is with the
accelerometer on the WSN node.
We tried every possible way, from varying the placements and attachments
of the node on the beam to changing axes, to find the problem with the
accelerometer.
We divided the probably sources of the error into:

1. Programming Bugs - We used the ADXL345 driver code that came

up with the Contiki OS. The algorithm uses I 2 C connection to stream
two bytes of accelerometer readings from the registers.
2. Circuitry - The placement of the ADXL345 accelerometer in the PCB
of the WSN node and the electronic components around it also alter the
output of the readings of the accelerometer.
3. Mishandling - Extensive and high vibrations that result from inconve-
nient environment and mishandling of the equipment would as well alter
the calibrated device. Since WSN nodes are such tiny devices, they of-
ten fall on floors and this might be one of the reasons that affected the
readings.

After careful investigation of the problem behind the errors, we found out that
it was a programming bug in the driver code that is included in the Contiki
OS. Together with the SICS developers, we debugged and rewrote the code.
The problem is described as follows:
The driver code for ADXL345 in the old Contiki OS works in such a way that
after the accelerometer function is initialized, it collects a preset amount of
samples of two bytes and when the buffer is full, it collects the next set of
samples without clearing the buffer. This causes some residual sample bits to
accumulate and results in the difference in the frequency content of the over
all sample.
Hence we debugged the codes so that the memory buffers are cleared before
the next set of samples are collected. This was possible to do by using the
advanced features of the ADXL345 accelerometer: FIFO and WATERMARK.

61
Chapter 8

Conclusion, Recommendations and

Future Works

8.1 Conclusion

In this paper, we used Zolertia node as a WSN platform for performing output
only modal analysis studies in laboratory test on a simple 3.5 m long beam
structure for analyzing how data processing algorithms can be implemented
with limitations of memory and power. This work would help us to know to
what extent we can trade off memory limitations and energy efficient data
processing algorithms on the implementation of less energy demanding SHM
system. We developed our programs on a platform with Contiki OS and cheap
wireless sensor node with built in ADXL345 accelerometer attached to it and
carried out the tests.
We did a series of measurements where time history data both from wired
and wireless system were collected and modal parameters were compared.
We then implemented fixed point integer FFT on 512 samples followed by
embedded peak picking algorithm. We evaluated this by off-line processing of
the time domain data collected from the wired system and got similar results.
Implementing fixed point FFT, we were able to manage the memory effectively
by reducing the size of the program from 41 KB to 24 KB.

8.2 Recommendation

After all the works we have done, we give the following recommendations for
whomsoever wants to continue the project and get a better result in less time
and efficiency:

– Building a sensor board with a better accelerometer.

63
8.3. FUTURE WORKS

– Using the advanced features of the accelerometers like activity/inactivity

detection so as to set a lower threshold value to trigger the data collection
hence reducing the wake up time of the nodes.
– Getting a sensor node with a better memory capacity and processing
power.
– Working in collaboration with hardware or sensor node producing com-
panies and get a good customized node.
– Building signal conditioning circuit could also improve the quality of the
signal collected.
– Using project management applications like Redmine™ so that a good
follow up of the project could be done between the students and super-
visors.

8.3 Future Works

This project mainly focused on data processing algorithms and SHM. The
project can be extended to a more and advanced system so that it could
encompass design and implementation of less energy consuming communi-
cation protocols. Some complex algorithms could also be implemented to
perform damage detection as well. The experimentation and the test could
also be taken to the real bridges in the field like the Södeström bridge. Imple-
mentation of a time synchronization would also increase the accuracy of the
determination of the mode shapes.

64
Bibliography

[1] Abrach H., Bhatti S., Carlson J., Dai H., Rose J., Sheth A., Shucker B.,
Deng J., and Han R.(2003). MANTIS: system support for MultimodAl
NeTworks of In-Situ sensors. In Proc. WSNA’03.

[2] Adam Dunkels, Björn Grönvall, and Thiemo Voigt, (2004). Contiki - a
Lightweight and Flexible Operating System for Tiny Networked Sensors,
IEEE Emnets.

[3] Adi M., Reddy V., Phani, K., Janakiram, D., andKumar, G. A. (2009).
Wireless Sensor Network Operating Systems: A Survey. Int. Jour. of
Sensor Net., Vol.5 Issue 4, Geneva, SWITZERLAND.

[4] Analogue Devices, ADXL345 datasheet.

[5] Akyildiz, Ian F., Su, W., Sankarasubramaniam, Y. and Cayirci, E.

(2002). "Wireless Sensor Networks: A Survey," Comp. Net. Elsevier Jour-
nal, Vol. 38, No. 4, pp. 393 - 422.

[6] Álvaro, C., Elsa, C., Filipe, M. and Carlos, M.(2006). From Input-Output
to Output-Only Modal Identification of Civil Engineering Structures.

[7] Aman K. and Mani B.S. (2003). An Environmental Energy Harvesting

Framework for Sensor Networks. Proc. of the int. symp. on Low power
elec. and des. NY, USA

[8] Arampatzis, T., et. al. (2005). A Survey of Applications of Wireless Sen-
sors and Wireless Sensor Networks. 13th Mediterranean Conf. on Control
and Automation. IEEE, Limassol, Cyprus.

[9] Barr R., Bicket J. C., Dantas D. S., Du B., Kim T. W. D., Zhou B.,
and Sirer E. (2002). On the need for system-level support for ad hoc and
sensor networks. SIGOPS Oper. Syst. Rev., 36(2).

[10] Bendat, J.S. and Piersol, A.G. (1993). Engineering Applications of Corre-
lation and Spectral Analysis. 2nd edition, John Wiley & Sons, New York,
NY.

65
BIBLIOGRAPHY

[11] Bischoff, R., Meyer, J. and Feltrin, G. (2009). Wireless sensor network
platforms. In Encyclopedia of Structural Health Monitoring. Chichester,
UK: John Wiley & Sons Ltd.
[12] Boulis A., Han C., and Srivastava M. B. (2003). Design and implementa-
tion of a framework for efficient and programmable sensor networks. In
Proc. MOBISYS’03, May 2003.
[13] Brian, J., Schwarz and Mark H. Richardson (1999). Experimental Modal
Analysis. Appl. note Vib. Tech., Inc.
[14] Ahlin, K., Brandt, A., Experimental modal Analysis Practice, saven
Edutech AB, Täby,Sweden,2001.
[15] Brincker, R., Zhang, L., Andersen, P. (2001), "Modal identification of
output-only systems using frequency domain decomposition", Smart Ma-
terials and Structures, 10, 441-445.
[16] CC2420 Datasheet, Texas Instruments.
[17] Brincker,P., Andersen,N., Jacobsen, J. (2007). Frequency Domain De-
composition for Operational Modal Analysis. Proc.of the 25th Int. Modal
Analys. Conf. (IMAC), Orlando, Florida,.
[18] Çelebi, M. (2002). Seismic Instrumentation of Buildings (with Emphasis
on Federal Buildings). Report No. 0-7460-68170, United States Geological
Survey, Menlo Park, CA.
[19] Chia, C., Lee, J. and Shin, H. (2009). Hot Target Inspection Using a
Welded Fibre Acoustic Wave Piezoelectric Sensor and a Laser-Ultrasonic
Mirror Scanner. Meas. Scien.and Techn. 20(12), vol. 20, 127003.
[20] Cooley, J.W. and Tukey, J.W. (1965). "An Algorithm for the Machine
Calculation of Complex Fourier Series", Math. Comp. 19, 297 - 30.
[21] Culler, D. E. and Wei, H. 2004. Wireless sensor networks. Communica-
tions of the ACM, 47(6): 30-33
[22] Doebling, S. W., et. al. (1996). "Damage Identification and Health Moni-
toring of Structural and Mechanical Systems from Changes in their Vibra-
tion Characteristics: A Literature Review. " Report No. LA-13070-MS,
Los Alamos National Laboratory, Los Alamos, NM.
[23] Feltrin, G., et al. (2011). "Structural Monitoring with Wireless Sen-
sor Networks: Experiences from Field Deployments , "First middle east
conf. on smart monitoring, assessment and rehabilitation of civil struct.,
Dubail, UAE.
[24] Hackmann, G., et. al. (2008). A Holistic Approach to Decentralized Struc-
tural Damage Localization Using Wireless Sensor Networks. In Proceed-
ings of IEEE Real-Time Systems Symposium. pp.35-46.

66
BIBLIOGRAPHY

[25] He, J. and Fu, Z.F.(2001). "Modal Analysis, Butterworth-Heinemann,"

Oxford.
[26] Jacobsen, N.J., Andersen,P., Brincker, R. (2008). Applications of Fre-
quency Domain Curve-fitting in the EFDD Technique. Proc. of the 26th
Inter. Modal Analys. Conf. (IMAC) Orlando, Florida USA.
[27] Dan-jiang Yu and W Ei-xin Ren.(2005), EMD-based stochastic sub-
space identification of structures from operational vibration measure-
ments,Engineering Structures,vol. 27,pages 1741-1751
[28] Jian, H. W., Chin, H. L., Lynch, J. P., Lu, K. C., Lin, P. Y., Wang, Y.
(2008). Output-only modal identification of a cable-stayed bridge using
wirelesse monitoring systems.
[29] Jindal A., and Liu M. (2010). Networked Computing in Wireless Sensor
Networks for Structural Health Monitoring, 2010.
[30] John Davies (2008). MSP430 microcontrollers basics, Elsevier.MA, USA.
[31] Kahn, J.M., Katz, R.H., and Pister, K.S.J. "Mobile Networking for Smart
Dust." ACM/IEEE Intl. Conf. on Mob. Comp. and Net.(MobiCom 99),
Seattle, WA, August 17-19, 1999.
[32] Kiremijdian, A.S. et al., H. (1997). Structural Damage Monitoring for
Civil Structures. International Workshop - Structural Health Monitoring:
pp. 371 - 382.
[33] Krishnamurthy, V., et. al. (2008 ). "The Effect of Time Synchronization
on the Performance of Wireless Sensor Networks for Structural Health
Monitoring, " Smarter Mater. Struct.
[34] Lei Y., Shen W. A., Song Y., and Wang Y..(2010). "Intelligent Wireless
Sensors with Application to the Identification of Structural Modal Pa-
rameters and Steel Cable Forces: From the Lab to the Field," Advances
in Civil Engineering, 2010.
[35] Levis, P., Madden, S., Polastre, J., Szewczyk, R., Whitehouse, K., Woo,
A., Gay, D., Hill, J., Welsh, M., Brewer, E. and Culler, D. (2005). Tinyos:
An operating system for wireless sensor networks. In Ambient intelligence.
New York: Springer-Verlag.
[36] Lynch, J. P. (2002), "Decentralization of Wireless Monitoring and Control
Technologies for Smart Civil Structures, " Ph.D. Thesis, Department of
Civil and Environmental Engineering, Stanford University, Stanford, CA.
[37] Lynch, J.P. and Loh, K.J. (2006). A Summary Review of Wireless Sensors
and Sensor Networks for Structural Health Monitoring. The shock and
vibration digest, 38(2): pp. 91-128.

67
BIBLIOGRAPHY

[38] Maroti, M., Kusy, B., Simon, G. and Ledeczi, A. (2004), "The flooding
time synchronization protocol", Proceedings of 2nd International Confer-
ence On Embedded Networked Sensor Systems, Baltimore, November.
[39] Texas Instruments, (2011)MSP430 datasheet
[40] Musiani, D., et. al. (2007). Active Sensing Platform for Wireless Struc-
tural Health Monitoring. In Proc. of the Sixth Inter. Conf. on Infor.
Processing in Sensor Networks , pages 390 - 399. ACM.
[41] Shinae, J., et. al. (2009). Structural Health Monitoring of a Cable-Stayed
Bridge Using Smart Sensor Technology: Deployment and Evaluation.
Smart Structures and Systems, Vol. 6, No. 5-6 (2010) 439-459.
[42] Sohn, H. and Farrar, C. (2001). "Damage Diagnosis Using Time-series
Analysis of Vibrating Signals," Smart Materials and Structures, Vol. 10,
No. 3, 446 - 451.
[43] Stallings, W. (2008), Operating Systems: Internals and Design Principles,
6ed, USA. ISBN:9780136006329.
[44] Straser, E. G. and Kiremidjian, A. S. (1998). "A Modular, Wireless Dam-
age Monitoring System for Structures, " Technical Report 128, John A.
Blume Earthquake Engineering Center, Stanford University, Stanford,
CA.
[45] Thai-Hoa, L., and Yukio, T. (2009). Modal Identification of Ambient Vi-
bration Structure Using Frequency Domain Decomposition and Wavelet
Transform. 7th Asia-Pacific Conf. on Wind Engin. Taipei, Taiwan.
[46] Wang, Y. (2007). Wireless Sensing and Decentralized Control for Civil
Structures: Theory and Implementation. Ph.D. thesis, Dep. of Civil and
Env. Eng., Stanford Univ., Stanford, Calif., USA.
[47] Wang, Y., Lynch, J. P., and Law, K. H. (2005). "Wireless Structural Sen-
sors Using Reliable Communication Protocols for Data Acquisition and
Interrogation", in Proceedings of the 23rd International Modal Analysis
Conference (IMAC XXIII), Orlando, FL, January 31-February 3.
[48] Zhang, L., Brincker R., Andersen, P. A. (2005). Overview of Operational
Modal Analysis: Major Development and Issues. Proc. of the 1st Int.
Operational Modal Analys. Conf. (IOMAC), Copenhagen, Denmark.
[49] http://www.visongtest.com/technical%20support1.htm. Visited on: 17
Dec. 2011.
[50] Zimmerman,A. T., M. Shiraishi, R. A. Swartz, and J. P. Lynch. Auto-
mated modal parameter estimation by parallel processing within wireless
monitoring systems. Journal of Infrastructure Systems,March 2008.
[51] Z1 Zolertia Data sheet, (2010).

68