INSTRUMENTING THE WORLD WITH WIRELESS SENSOR NETWORKS

D.Estrin, L. Giro@ G. Pottie, M. Srivastavat

UCLA UCLA
Department of Computer Science Department of Elecrical Engineering
email: (destrin,girod)@cs. ucla.edu email: (mbs,pottie)@ee.ucla.edu

ABSTRACT many envisioned applications, the environment being mon-

itored does not have installed infrastructure for either com-
Pervasive micro-sensing and actuation may revolutionize the way munications or energy, and therefore untethered nodes must
in which we understand and manage complex physical systems: rely on local, finite, and relatively small energy sources, as
from airplane wings to complex ecosystems. The capabilities for well as wireless communication channels.
detailed physical monitoring and manipulation offer enormous op-
portunities for almost every scientific discipline, and it will alter 0 Why distributed processing? Finally, although sensors
the feasible granularity of engineering. are distributed to be close to the phenomena, one might still
We identify opportunities and challenges for distributed sig- consider an architecture in which sensor outputs could be
nal processing in networks of these sensing elements and investi- communicated back to a central processing unit. However,
gate some of the architectural challenges posed by systems that are in the context of untethered nodes, the finite energy budget
massively distributed, physically-coupled, wirelessly networked, is a primary design constraint. Communications is a key en-
and energy limited. ergy consumer as the radio signal power in sensor networks
drops off with r4[4] due to ground reflections from short
antenna heights. Therefore, one wants to process data as
1. INTRODUCTION much as possible inside the network to reduce the number
of bits transmitted, particularly over longer distances.
The availability of low-power micro-sensors, actuators, embedded
processors, and radios is enabling the application of distributed
wireless sensing to a wide range of applications, including en- 2. MOTIVATING APPLICATION
vironmental monitoring, smart spaces, medical applications, and
The potential applications of wireless sensor networks are highly
precision agriculture [ 1][2]. Most deployed sensor networks in-
varied: e.g., Physiological monitoring; Environmental monitoring
volve relatively small numbers of sensors, wired to a central pro-
(air, water, soil chemistry); Condition based maintenance; Smart
cessing unit where all of the signal processing is performed [3]. In
spaces; Military; Precision agriculture; Transportation; Factory in-
contrast, this paper focuses on distributed, wireless, sensor net-
works in which the signal processing is distributed along with the strumentation and inventory tracking
Habitat monitoring [Cerpa-eta10I , Hamilton, Steere-eta1001pro-
sensing.
vides a rich collection of sensing modalities and environmental
Why distributed sensing? When the precise location of conditions and we use it to motivate our technical discussion. Con-
a signal of interest is unknown in a monitored region, dis- sider the goal of supporting data collection and model development
tributed sensing allows one to place the sensors closer to of complex ecosystems. Scientists and environmental impact mon-
the phenomena being monitored than if only a single sen- itoring authorities would like to monitor soil and air chemistry, as
sor were used. This yields higher SNR, and improved op- well as plant and animal species populations and behavior. For
portunities for line of sight. While S N R can be addressed the latter, the primary modalities are imaging and acoustics to lo-
in many cases by deploying one very large sensitive sen- calize, identify and track species or phenomena based on implicit
sor, line of sight, and more generally obstructions, cannot signals (acoustic and seismic), or explicit signals (RF tags). These
be addressed by deploying one sensor regardless of its sen- facilities must be deployable in remote locations that lack installed
sitivity. Thus, distributed sensing provides robustness to energy and communication infrastructures, motivating the need for
environmental obstacles. low-power wireless communication.
The strategy for node cooperation strategy has significant con-
0 Why wireless? When wired networking of distributed sen- sequences in terms of communication bandwidth and energy con-
sors can be easily achieved, it is often the more advanta- sumption. For example, consider the task of identifying bird species
geous approach. Moreover,when nodes can be wired to re- in view of several cameras. If it is to be accomplished through
newable (relatively infinite) energy sources, this too greatly image analysis, we could stream all the video back to a human
simplifies the system design and operation. However, in operator-a very costly approach. Alternatively, we could stream
Supported by the DARPA SensIT program and NSF Special Projects audio to a central location, which then performs signal processing
under ANIR. to identify and stream back only those streams that are most likely
t Supported by the DARPA SenslT program. to contain a target species. While this reduces communications

0-7803-7041 -4/01/$10.00 02001 IEEE 2033

overhead greatly, it still suffers from communications latency and ena in the presence of communication noise and sensor di-
lacks scalability due to the need to stream audio thorough a central versity, the fidelity and timeliness ofthe signal processing at
processing point. Finally, we might distribute the problem further, individual sensor nodes can be adapted to energy resources
hosting the audio signal processing software on the nodes, and de- and latency requirements.
veloping algorithms that require only local cooperation to make a 0 A hierarchical, tiered architecture can greatly contribute
decision to capture images. This approach is scalable in that no to overall system lifetime and capability. Whenever pos-
long-range streaming of audio or video is necessary, resulting in sible, higher capacity system elements can be used to of-
more efficient use of communications bandwidth and limited en- fload drain on small form factor elements, while the latter
ergy resources. can be exploited to obtain the desired physical proximity to
In the remainder of this paper we identify some of the tech- stimuli. Moreover, even among elements with homogenous
nical challenges associated with the design of wireless sensor net- capabilities, creating clusters and assigning special com-
works and discuss several algorithmic approaches. bining functions to cluster heads can contribute to overall
system scalability. However, to avoid compromising ro-
3. TECHNICAL CHALLENGES bustness, such clustersihierarchy must be self-configuring
and reconfiguring in the face of environmental or network
Most envisioned sensor network applications encounter one or more changes.
of the following challenges:
0 Untethered for energy and communication requiring max- 4. TECHNICAL APPROACHES
imal focus on energy efficiency.
o Ad hoc deployment, requiring that the system identifies We now describe three generic techniques that would enable dis-
and copes with the resulting distribution and connectivity tributed signal processing tasks in wireless sensor networks.
of nodes.
0 Dynamic environmental conditions requiring the system to 4.1. Coherent processing algorithms
adapt over time to changing connectivity and system stim-
uli. Coherent signal processing algorithms are distinguished from non-
0 Unattended operation requiring configuration and recon- coherent methods in that information about the phase of the wave-
figuration be automatic (self-configuration) front impinging on the nodes must be conveyed. Beamforming
techniques allow localization of signals that originate within the
To address these challenging environments, several strategies convex hull of the participating nodes, higher SNR estimates of
are likely to be key building blockdtechniques for wireless sensor the signals compared to non-coherent methods, and determination
networks: of bearing angles for signals that originate outside the convex hull
0 Collaborative signal processing among nodes that have of the participating nodes. The price is a higher level of synchro-
nization (to within a small fraction of one oscillation), and com-
experienced a common stimulus will greatly enhance the
munication of relatively high bit rate data streams consisting of
efficiency (information per bit transmitted) of these sys-
tems. Develop both coherent signal processing on small sampled waveforms. Given its high resource cost, we should re-
sort to coherent processing only when we cannot attain adequate
clusters by a centralized entity within the cluster, and non-
coherent processing with much less stringent synchroniza- accuracy in the result with non-coherent methods such as combi-
tion requirements and applicable across larger numbers of nation of likelihood functions.
more loosely coupled elements. One way to organize the operations leading to coherent beam-
forming is as follows. Nodes go through a sequence of internal lev-
0 Exploiting redundancy of hardware elements to compen- els of signal processing before determining that neighbors should
sate for ad hoc deployment of systems. If elements cannot be involved in a detection/localization decision. An ad hoc net-
be carefully positioned relative to each other and the envi- work is constructed for non-coherent decision-making using for
ronment, then an altemate strategy to achieve “coverage” example the single winner election algorithm of [2]. The algo-
is to deploy a greater density of elements so that one can rithm is optimized to minimize the overhead in finding a fusion
make use of some subset that have the desired absolute and center, since relatively little data must actually be communicated.
relative position. In some contexts, even if elements can However, if the decision has insufficient certainty or resolution,
be uniformly placed in 3-space, environmental conditions the same set of nodes become involved in a new network set-up
might be such that coverage is not uniform due to obstacles that seeks to minimize the energy consumption in conveying sam-
and other sources of noise. Another application of redun- pled waveforms to a common central processing point. To this
dancy is when the incremental cost of a node during initial end significant overhead is acceptable since large amounts of data
deployment is much smaller than the incremental cost of will be conveyed in the local neighborhood. A multi-winner elec-
deploying new nodes or renewing node resources (e.g., en- tion algorithm to accomplish this is also described in [2]. Standard
ergy). In this case, one can exploit redundancy to extend beamforming techniques can now be applied using the data col-
system lifetime by adjusting duty cycle based on local den- lected from the cluster of nodes.
sity and local demand. There is no requirement for uniform lay-down of nodes to
0 Adaptive fidelity signal processing is another strategy that achieve beamforming [5]. To track distant sources, two or more
can be exploited in sensor networks to make trade-offs be- clusters of nodes can be used, and with the intersection of the
tween energy, accuracy, and rapidity of results. Recogniz- bearing lines used to establish location. Note that simply using
ing that one is trying to detect non-deterministic phenom- all nodes in the network to do one massive beamforming operation

2034
could accomplish this end, but excessive communications and sig- locally distributed iterative multilateration [8] whereby a node cal-
nal processing complexity would be required. Rather, for a scal- culates its position and is promoted to a beacon as soon as enough
able solution a signal processing step is required that recognizes of its 1-hop neighbors are beacons. Starting with a critical density
whether near or far objects are being tracked. A crude technique of beacons, a percolation-like phenomenon would result in grad-
is to consider the SNR variations among nodes in a cluster and to ually all the nodes discovering their position. With a sufficient
neighboring clusters. If the S N R is similar, then the signal source beacon density,a small number of successive multilateration steps
is likely to be distant. Having made this determination, clusters lead to rapid convergence of location estimates. The communi-
may decide to estimate lines of bearing or not, whether probabilis- cation overhead is much lower than in centralized approach as all
tically or according to a predetermined schedule. The information message exchange is strictly local and is easily piggybacked on
on the bearing lines is then conveyed to a central node designated routing messages.
to perform the (noncoherent) fusion. Thus, there is never a case in Another challenge in localization is estimation of distance be-
which sampled waveforms must be conveyed over a large number tween a pair of nodes. Using time-of-flight of radio signals (as in
of hops. GPS) is ruled out when the distances are too tiny and radio fre-
Achieving the required level of synchronism for coherent beam- quencies not very high. A readily available method would be to
forming is in principle relatively straightforward for systems in use the received signal strength indication (RSSI) provided be the
which every node possesses a radio. Since the propagation ve- radio. The RSSI data can be cheaply piggybacked on regular rout-
locity of seismic and acoustic signals is six orders of magnitude ing and data. The accuracy of this approach can be improved by
slower than that of radio waves, achieving data lock for RF com- using a parameterized channel, path loss model whose parameters
munications would seem to already be much better synchroniza- are also estimated together with position [8]. However, in practice,
tion than is required for beamforming. However, particular care the RSSI based approach works only in the absence of significant
must be paid to the node architecture to take advantage of this tim- multipath effects. In most environments other than open spaces
ing information. The typical interrupt cycles of general-purpose multipath is an issue. A promising alternative technology is to es-
proccssors can be tens of milliseconds, an eternity with respect to timate distance by time of flight of acoustic or ultrasound signals,
even acoustic signals. Thus, embedded real-time components are and using the much faster radio signal to establish time reference
required in the nodes to deal with time-stamping of the data. [91[101[111.

4.2. Localization 4.3. Distributed power management

Node location is employed by routing protocols that use spatial ad- Dynamic power management techniques such as shutdown and
dresses, and by signal processing algorithms (e.g. beamforming) dynamic voltage scaling have emerged as powerful methods for
that are used for tasks such as target tracking. The underlying al- power-aware computing. Power-aware operation is even more im-
gorithm problem is that of localization whereby the nodes in the portant for wireless sensor networks, and requires distributed ver-
network discover their spatial coordinates upon network boot-up. sions of power management techniques.
When the sensor nodes are deployed in an unplanned topology, As an example, consider shutdown, which is widely used in
there is no a priori knowledge of location. The use of GPS in portable computing systems such as laptops. In sensor nets one
sensor nodes is ruled out in many scenarios because of power con- could exploit redundant nodes by turning on only a time- varying
sumption, antenna size, and overhead obstructions such as dense subset of nodes, where the subset is selected for desired sensor and
foliage. The ad hoc nature of deployment rules out infrastructure radio coverage. The remaining nodes can be shutdown, only to be
for many scenarios of localization. It is critical that sensor net- woken up to provide additional sensor readings or communication
work nodes be able to estimate their relative positions without as- routes when something interesting happens [ 121.
sistance, using means that can be built-in. A key problem in such a distributed shutdown scheme is the
The localization problem in itself is a good example of a signal- strategy to select which node to shutdown and which to turn on
processing task that the sensor network needs to solve. The ba- at any given instant. A good way to model this problem is to op-
sic approach would be for sensor nodes to gather sufficient num- timally divide the sensor nodes into several subsets such that any
ber of pair-wise distance estimates via some suitable mechanism, given subset provides a baseline level of sensing and communica-
and then use multilateration algorithms to estimate positions of the tion coverage. The different subsets can then be tumed on and off
nodes. To begin with, a few nodes might know their position via according to a duty cycle determined by a repetitive schedule. As
other means (beacon nodes), but at the end of the localization pro- nodes die by depleting their batteries, the subsets are changed.
cess every node would hopefully know its position. Unfortunately, modeling the problem in this fashion requires
A key problem however is that in conventional formulations one to gather global information to find the subsets. Since com-
of multilateration [6][7] one needs to estimate the location of an munication is expensive in energy, the cost of the power manage-
entity given estimates of its distance to 3 or more beacons with ment algorithm would swamp the savings from power manage-
known positions. In sensor networks a very high density of bea- ment! This illustrates the dilemma that so often arises in prob-
cons nodes would be needed. To keep the required beacon density lems in sensor nets: the seemingly optimal way of solving a
and energies low, a preferred method would be to jointly estimate problem often results in algorithms whose communication en-
positions of all the non-beacon nodes via a collaborative multilat- ergy costs exceed their benefits. Therefore, a better strategy is
eration formulation based on criterion such as least-square error to use algorithms that only shoot for good though sub optimal re-
minimization. Besides being computationally hard for large num- sults but require only locally distributed processing with minimal
ber of nodes, doing this would require a centralized node where all communication costs.
the distance estimates would be collected at significant comniu- This suggests that the decision regarding when to shutdown
nication and associated energy cost. A more scalable solution is and wakeup a node should be made using information in the local

2035
neighborhood. A simple scheme is to turn the nodes on an off with 6. REFERENCES
a certain duty cycle with random phase differences. When off, the
node power consumption can be reduced to microwatts. However, I ] D. Estrin, R. Govindan, and J. Heidemann, Eds., Special
the power savings come at the cost of reduction in detection and Issue on Embedding the Internet. Communications of the
classification accuracy as the sensor would miss a phenomenon ACM, vol. 43, no. 5 , May 2000.
when off. The cost of such missed events is very application spe- 21 B.R. Badrinath, M. Srivastava, K. Mills, J. Scholtz, and
cific: it may be okay to miss a frequent event, but very crucial to K. Sollins, Eds., Special Issue on Smart Spaces and Envi-
detect a rare event. In general, such a goal might be quantified as a ronments. IEEE Personal Communications, Oct. 2000.
quality of service requirement such as the following in the case of a
sensor net for surveillance: “a target with 20 mph speed following [ 3 ] Sensors: The Journal of Applied Sensing Technology.
this track will not pass undetected”. [4] G. Pottie and W. Kaiser, “Wireless sensor networks,” Coni-
More sophisticated locally distributed schemes for shutdown mtrnications o f t h e ,4CM, vol. 43, no. 5 , pp. SI-58, May
would coordinate the on and off periods of neighboring nodes to 2000.
improve energy efficiency for the same level of detection perfor- [5] K. Yao, R. Hudson, C. Reed, D. Chen, and F. Lorenzelli,
mance, and perhaps adapt the duty cycle parameters based on event “Blind beamforming on a randomly distributed sensor array
activity. A crucial problem here is that of waking up the node. system,” IEEEJSAC, vol. 16, no. 8, Oct. 1998.
In shutdown in PCs and laptops external events such as keyboard
[6] G.L. Stuber and J.J. Caffery jr., The Mobile Conirnirnications
presses or arrival of network packet result in the rest of the sys-
Handbook, J.D. Gibson and E.M. Gibson, ed., chapter 24,
tem waking up. However, in sensor nodes it is highly desirable to
Radiolocation Techniques, CRC Press, 1999.
turn off the radio, which is usually more power-hungry than the
processor and the sensors. Turning off the radio, unfortunately, [7] G. Turin, W. Jewell, and T. Johnston, “Simulation of urban
means that a neighboring node that detected an interesting event vehicle-monitoring systems,” IEEE Transactions on Vehictr-
cannot wake a node up. This can lead to missed events and packets. lar Technology, vol. 21, no. I , Feb. 1972.
Therefore, one technological challenge for effective power man- [8] A. Savvides, F. Koushanfar, A. Boulis, V. Karavas, and
agement of sensor networks is to have an ultra low-power paging M.B. Srivastava M. Potkonjak, “Location discovery in ad-
communication channel to wake up neighboring nodes on demand. hoc wireless networks,” Memorandum, Networked and Em-
An alternative is for low-power sensors to be constantly vigilant, bedded Systems Laboratory,UCL A, June 2000.
with radio wake-up according to signal processing results.
[9] A. Ward, A. Jones, and A. Hopper, “A new location tech-
The proposed form of adaptive duty cycle applies at multiple nique for the active office,” IEEE Personal Comniunications,
levels in the system. At short ranges, radios will consume nearly vol. 4, no. 5 , Oct. 1997.
as much energy whether transmitting or receiving. Consequently,
it makes sense to arrange for a time-division structure for commu- [IO] N. Priyantha, A. Chakraborty, and H. Balakrishnan, “The
nications. In [2] an algorithm is described for enabling distributed cricket location support system,” in Mobiconi 2000. ACM,
boot-up of the network, establishing both channel assignments and Aug. 2000.
synchronism in an energy-efficient fashion. It assumes that suffi- [ I I ] L. Girod and D. Estrin, “Robust range estimation for lo-
cient signal processing is done in each node to result on average calization in adhoc sensor networks,” Tech. Rep. CS-TR-
with traffic occupying a very small fraction of the available band- 2000XX, UCLA, Nov.20 00.
width. Nodes regularly communicate with their neighbors to keep [ 121 D. Estrin, R. Govindan, J. Heidemann, and S. Kumar, “Next
synchronism and to indicate whether larger slots are to be reserved century challenges: Scalable coordination in sensor net-
for bulk data transfers, in effect alerting the neighbor to turn on its works,” in Mobicom 99.A CM, Aug. 1999.
receiver for some specified time period.
[ 131 J. EIson and D.Estrin, “Time synchronization for wireless
Maintaining a constant level of synchronism is helpful in re-
sensor networks,” in IPDPS Workshop on Parallel and Dis-
ducing network latency and for health-keeping, but it is not the tributed Computing Issues in Wireless Networks and Mobile
most energy efficient strategy in all traffic scenarios. When the
Computing, Apr. 200 1.
signal sources of interest have a very low duty cycle, it may be
better to periodically exchange packets to re-synch only to within
some very coarse accuracy, and then spend significant energy in
achieving tighter synchronism when there is something of interest
to send [ 131.

5. CONCLUSIONS

In conclusion, wireless sensor networks present fascinating chal-

lenges for the application of distributed signal processing and dis-
tributed control. These systems will challenge us to apply appro-
priate techniques and metrics in light of the technology opportuni-
ties (cheap processing and sensing nodes) and challenges (energy
constraints).

2036