Detection and Estimation in Wireless Sensor Networks ˙ Israfil Bahçeci Department of Electrical Engineering TOBB ETÜ June 28, 2012 1 of 38
Detection and Estimation in Wireless SensorNetworks
Israfil Bahçeci
Department of Electrical EngineeringTOBB ETÜ
June 28, 2012
1 of 38
Outline
Introduction
Problem Setup
Estimation
Detection
Conclusions
References
2 of 38
Outline
Introduction
Problem Setup
Estimation
Detection
Conclusions
References
3 of 38
Wireless Sensor Networks
I Many nodes, preferablycheap
I Power/energy/bandwidth limitedI Wireless medium
4 of 38
Functionality and Utility
I DetectionI False alarm and
detection probabilityI Estimation
I Estimation error
5 of 38
Typical Problems
I Deployment optimizationI Node densityI Node location
I Wireless networking and communicationsI Achievable rate/distortion regionsI Source/channel coding problemsI Quantization/coding/analog transmissionI Power control and interference management, energy efficiencyI Centralized vs. distributedI Multiple access vs. Orthogonal accessI Single vs. multiple fusion centerI Path selection and shortest path algorithms
I Self-organizationI Node failure & self-healing
I Information securityI Access to informationI Node intrusion, e.g. Byzantine attack
6 of 38
Typical Network Configurations
Parallel network
Serial network
Hierarchical network
7 of 38
Outline
Introduction
Problem Setup
Estimation
Detection
Conclusions
References
8 of 38
System Model
I Let n0, n1, . . . , nN−1 denote the sensor nodesI Let ui is the observed samples at node ni
I Let hi,j is the channel gain from nj to niI hi,j include the effect of antenna gains and long term channel losses
I For transmission from nm to nk, received signal:
rk[t] = hk,msm[t] +N−1∑
i=0,i 6=m
Ii[t]hk,isi[t] + wk[t]
I Transmitted signal at nm: sm[t] = g(um, rm)
I Bandwidth and energy constraints
9 of 38
Metrics
I DetectionI Detection probability (correct decision)I False alarm/miss probability (erroneous decision)
I EstimationI Mean-square error, E(|θ − θ|2)
10 of 38
Observation Statistics
I Independent observationsI Correlated observations
I Dense deployment
11 of 38
Outline
Introduction
Problem Setup
Estimation
Detection
Conclusions
References
12 of 38
Estimation Under Bandwidth Constraints I
I Universal estimation [1]I Each sensor has 1 sample from a noisy observation and can send
1 bit (0 or 1 per local estimateI ui = θ + ni, θ ∼ [−V,V] and ni ∼ fU(u), V = UI if fU(u) ≥ µ is known, N ≥ 1
4ε2µ2
I if fU(u) is unknown, N ≥ U2
4ε2 , e.g., binary messaging requires only atmost 4 times more sensor nodes
I Sample mean estimation [2, 3]I ui = θ + ni, ni ∼ N(0, σ2)I Maximum likelihood estimator available for both identical
thresholds, non-identical thresholdsI Fixed step size difference, τk+1 − τk > σ equal to noise variance is
close to optimalityI Parameter with a small dynamic range: 1 bit quantization is
sufficientI Relaxing 1 bit constraint, a step size equal to noise variance is
good for practical cases
13 of 38
Estimation Under Bandwidth Constraints II
I Inhomogeneous environment [4]I Local information compressed to a number of bits proportional to
logarithm of its local observation SNRI Fusion center only needs the received quantized messages and
use the length of the message in final estimationI No need for noise pdf at the FC, each sensor needs its local SNRI The MSE of this estimator achieves 25/8 times the MSE of BLUE
14 of 38
Compression and Estimation
I The above bandwidth constrained schemes compress the signalsto a few bits
I An overview of several cases of distributed estimation [5]I Same order of MSE performance achieved by a centralized
estimation is doable under various bandwidth constrainedschemes under different knowledge levels for the observationnoise statistics
15 of 38
Power Control for Distributed Estimation
I Estimation with digital modulation [6]I Joint design of universal estimator and uncoded QAM modulationI Optimal quantization and transmit power levels to minimize MSEI Bah channel or bad observation⇒ lower quantization level, or
inactiveI Estimation with analog modulation [7]
I Correlated data observation, e.g., a random fieldI Non-linear measurement issues also consideredI Linear MMSE + numerical power control optimization
16 of 38
Source-Channel Coding for Distributed Estimation
I Wyner-Ziv source coding based strategies for a general treenetwork [8]
I Achievable region for a generic one-step communication withside-information
I Application of one-step solution to a tree network: A sensor usesits own observations, all messages it received + statisticalinformation for the observation made by decoder and messagesreceived by the decoder
I Rate-distortion bounds for the Quadratic Gaussian case isdetermined
17 of 38
CEO problem and distributed estimation
I CEO problem: Estimation with a parallel configurationI Admissible sum-rate distortion regions [9]I Local observations separately encoded and transmitted to a CEOI Closed form solution to rate allocation for the Quadratic Gaussian
CEO problemI Rate-constrained estimation for CEO problem [10]
18 of 38
Distributed Quantization and Estimation
I Adaptive quantization [11]I Bandwidth constraint, so only one bit quantizationI Dynamic adjustment of quantization threshold based on feedback
from other sensor nodesI Distributed Delta modulation
I Quantizer precision for large networks [12]I xi = θ + ni for all nodesI Identical, noncooperative uniform scalar quantization at each node
achieves same asymptotics as optimal schemeI If observation SNR is high, few nodes with fine quantization is betterI There exists an optimal number of sensors for this quantization, not
all sensors needed
19 of 38
Cooperative Communications
I Cooperative diversity for distributed estimation [13]I Several cooperative relaying schemes exists that achieve spatial
diversityI Multiple access channel, r[n] =
∑Ni=1 xi[n] + w[n]
I Amplify-forward or decode-forward based distributed estimationachieve same asymptotic performance
I Collection of correlated data: spatial sampling (one sensor out of agroup of correlated sensor nodes)
I Selective transmission is good for loose distortion, but needsimproved cooperation for strict distortion constraint
20 of 38
Linear Distributed Estimation
I Parallel configuration with all linear processing for a coherentGaussian network with MAC [14]
I Linear observation modelI Linear encoding at the transmitter: MAC allows for a closed-form
expression for encodingI Linear MMSE at the fusion centerI Optimal power allocation allows distributed implementation
21 of 38
Estimation Diversity and Energy Efficiency
I Analog transmission of xi = θi + ni, to a fusion center [15, 16]I Fixed data vs. correlated dataI BLUE vs. MMSEI Estimation outage and estimation diversity (slope of outage
probability)I Full diversity can be achieved on the number of sensor nodesI Power control for fixed data vs. correlated data
22 of 38
Distributed Kalman Filtering
I Distributed estimation of a dynamically varying signal with a linearobservation model [17]
I Need to exchange messages between neighbor nodesI 2-step estimation
I Step 1: Kalman-like estimation based only on local observationsI Step 2: Information fusion via a consensus matrix after receiving
messages from neighborsI Design problems: Optimal Kalman gain, and consensus matrix,
based on the amount of message exchange
23 of 38
Distributed Data Gathering with a Dense SensorNetwork
I Estimation of a observable random field at a collector node [18]I Transport capacity of many-to-one channel ∼ O(logN) can be
achieved by an amplify-forward scheme, even under subject tototal power constraint
I Unbounded transport capacity for many-to-one channel with onlyfinite total average power
I Gaussian spatially bandlimited processes are observable (e.g., itcan be estimated at a collector node with a finite MSE for a certainbandwidth and total average power level)
I This is true even for lossy source encoder composed of asingle-dimensional quantization followed by a Slepian-Wolf encoder
24 of 38
Outline
Introduction
Problem Setup
Estimation
Detection
Conclusions
References
25 of 38
Distributed Detection with Multiple Sensors
I An overview on various distributed detection strategies [19]I Error-free transmission of local decisions to a fusion centerI Independent local observationsI Likelihood ratio tests, for both Neyman-Pearson and Bayes’
formulation, are optimal at both local sensor nodes and fusioncenters
26 of 38
Bandwidth/Power Constrained Distributed Detection I
I Binary detection over a parallel network with MAC [20]I Specifying the power, bandwidth, error tolerance fixed the
information rates of sensors for this MACI Minimization of Chernoff exponent for the decision at the FCI Asymptotically, for Gaussian and exponential observation, having R
identical binary sensors, e.g., 1 bit/sensor for a rate-R MACchannel, is optimal
I Not true for some other statistical distributionsI Having more sensors is better than having detailed information
from each nodeI Asymptotic detection for power constrained network [21]
I Joint power constraint + AWGN at sensor-to-fusion center channelI Having identical sensor nodes, e.g., each node using the same
scheme, is asymptotically optimalI Optimal transmission power levels for binary nodes observing
Gaussian source
27 of 38
Energy Efficient Distributed Detection
I Energy and bandwidth constraints taken into account [22]I Detection performance subject to system cost due to transmission
power and measurement errorsI Randomization over the choice of measurements and when to
send/no sendI Joint optimization over sensor nodes allows the optimization per
node
28 of 38
Parallel and Serial Detection over Fading Channel
I The communications are all Rayleigh fading [23, 24]I Binary detection and binary antipodal modulation for decision
transmissionI Channel state information need to be obtained at the receiver
nodeI Suitable likelihood ratios at all nodes are optimal with known CSI
29 of 38
Optimal Distributed Detection over Noisy Channel
I Non-ideal channels to fusion center [25, 26]I Detection at fusion center needs to consider the CSI in case of
fadingI LRTs are shown to be optimal in the sense that they minimize
error probability at the fusion center
30 of 38
Type-Based Distributed Detection
I Each local sensor generates a histogram, or type of itsobservation over time and forwards the type to fusion center[27, 28]
I MAC where fusion center receives a superposition of transmittedlocal signals attain a better detection performance relative toorthogonal MAC
I Histogram fusion at the fusion center is asymptotically optimal andobservation statistics need to be known only at the fusion cener
31 of 38
Outline
Introduction
Problem Setup
Estimation
Detection
Conclusions
References
32 of 38
Remarks
I Diverse applications and many research topicsI Many open problems
I Joint design of local/global processing and network operation,routing
I Cooperative sensing/routingI Network life time maximization via data aggregation, joint
source/channel coding, and power controlI New paradigms for detection estimation under constraints of WSNsI Detection/estimation at multiple fusion center, distributed
congestion control
33 of 38
Outline
Introduction
Problem Setup
Estimation
Detection
Conclusions
References
34 of 38
References I
[1] Zhi-Quan Luo,“Universal decentralized estimation in a bandwidth constrained sensor network,”Information Theory, IEEE Transactions on, vol. 51, no. 6, pp. 2210 –2219, june 2005.
[2] A. Ribeiro and G.B. Giannakis,“Bandwidth-constrained distributed estimation for wireless sensor networks-part i: Gaussian case,”Signal Processing, IEEE Transactions on, vol. 54, no. 3, pp. 1131 – 1143, march 2006.
[3] A. Ribeiro and G.B. Giannakis,“Bandwidth-constrained distributed estimation for wireless sensor networks-part ii: unknown probability density function,”Signal Processing, IEEE Transactions on, vol. 54, no. 7, pp. 2784 –2796, july 2006.
[4] J.-J. Xiao and Z.-Q. Luo,“Decentralized estimation in an inhomogeneous sensing environment,”Information Theory, IEEE Transactions on, vol. 51, no. 10, pp. 3564 –3575, oct. 2005.
[5] Jin-Jun Xiao, A. Ribeiro, Zhi-Quan Luo, and G.B. Giannakis,“Distributed compression-estimation using wireless sensor networks,”Signal Processing Magazine, IEEE, vol. 23, no. 4, pp. 27 – 41, july 2006.
[6] Jin-Jun Xiao, Shuguang Cui, Zhi-Quan Luo, and A.J. Goldsmith,“Power scheduling of universal decentralized estimation in sensor networks,”Signal Processing, IEEE Transactions on, vol. 54, no. 2, pp. 413 – 422, feb. 2006.
[7] Jun Fang and Hongbin Li,“Power constrained distributed estimation with correlated sensor data,”Signal Processing, IEEE Transactions on, vol. 57, no. 8, pp. 3292 –3297, aug. 2009.
[8] S.C. Draper and G.W. Wornell,“Side information aware coding strategies for sensor networks,”Selected Areas in Communications, IEEE Journal on, vol. 22, no. 6, pp. 966 – 976, aug. 2004.
35 of 38
References II
[9] Jun Chen, Xin Zhang, T. Berger, and S.B. Wicker,“An upper bound on the sum-rate distortion function and its corresponding rate allocation schemes for the ceo problem,”Selected Areas in Communications, IEEE Journal on, vol. 22, no. 6, pp. 977 – 987, aug. 2004.
[10] P. Ishwar, R. Puri, K. Ramchandran, and S.S. Pradhan,“On rate-constrained distributed estimation in unreliable sensor networks,”Selected Areas in Communications, IEEE Journal on, vol. 23, no. 4, pp. 765 – 775, april 2005.
[11] Hongbin Li and Jun Fang,“Distributed adaptive quantization and estimation for wireless sensor networks,”Signal Processing Letters, IEEE, vol. 14, no. 10, pp. 669 –672, oct. 2007.
[12] S. Marano, V. Matta, and P. Willett,“Quantizer precision for distributed estimation in a large sensor network,”Signal Processing, IEEE Transactions on, vol. 54, no. 10, pp. 4073 –4078, oct. 2006.
[13] Y-W. Hong, W.-J. Huang, F-H. Chiu, and C.-C.J. Kuo,“Cooperative communications in resource-constrained wireless networks,”Signal Processing Magazine, IEEE, vol. 24, no. 3, pp. 47 –57, may 2007.
[14] Jin-Jun Xiao, Shuguang Cui, Zhi-Quan Luo, and A.J. Goldsmith,“Linear coherent decentralized estimation,”Signal Processing, IEEE Transactions on, vol. 56, no. 2, pp. 757 –770, feb. 2008.
[15] Shuguang Cui, Jin-Jun Xiao, A.J. Goldsmith, Zhi-Quan Luo, and H.V. Poor,“Estimation diversity and energy efficiency in distributed sensing,”Signal Processing, IEEE Transactions on, vol. 55, no. 9, pp. 4683 –4695, sept. 2007.
[16] I. Bahceci and A. Khandani,“Linear estimation of correlated data in wireless sensor networks with optimum power allocation and analog modulation,”Communications, IEEE Transactions on, vol. 56, no. 7, pp. 1146 –1156, july 2008.
36 of 38
References III
[17] R. Carli, A. Chiuso, L. Schenato, and S. Zampieri,“Distributed kalman filtering based on consensus strategies,”Selected Areas in Communications, IEEE Journal on, vol. 26, no. 4, pp. 622 –633, may 2008.
[18] H. El Gamal,“On the scaling laws of dense wireless sensor networks: the data gathering channel,”Information Theory, IEEE Transactions on, vol. 51, no. 3, pp. 1229 –1234, march 2005.
[19] R. Viswanathan and P.K. Varshney,“Distributed detection with multiple sensors i. fundamentals,”Proceedings of the IEEE, vol. 85, no. 1, pp. 54 –63, jan 1997.
[20] J.-F. Chamberland and V.V. Veeravalli,“Decentralized detection in sensor networks,”Signal Processing, IEEE Transactions on, vol. 51, no. 2, pp. 407 – 416, feb 2003.
[21] J.-F. Chamberland and V.V. Veeravalli,“Asymptotic results for decentralized detection in power constrained wireless sensor networks,”Selected Areas in Communications, IEEE Journal on, vol. 22, no. 6, pp. 1007 – 1015, aug. 2004.
[22] S. Appadwedula, V.V. Veeravalli, and D.L. Jones,“Energy-efficient detection in sensor networks,”Selected Areas in Communications, IEEE Journal on, vol. 23, no. 4, pp. 693 – 702, april 2005.
[23] I. Bahceci, G. Al-Regib, and Y. Altunbasak,“Parallel distributed detection for wireless sensor networks: performance analysis and design,”in Global Telecommunications Conference, 2005. GLOBECOM ’05. IEEE, dec. 2005, vol. 4, pp. 5 pp. –2424.
[24] I. Bahceci, G. Al-Regib, and Y. Altunbasak,“Serial distributed detection for wireless sensor networks,”in Information Theory, 2005. ISIT 2005. Proceedings. International Symposium on, sept. 2005, pp. 830 –834.
37 of 38
References IV
[25] B. Chen and P.K. Willett,“On the optimality of the likelihood-ratio test for local sensor decision rules in the presence of nonideal channels,”Information Theory, IEEE Transactions on, vol. 51, no. 2, pp. 693 –699, feb. 2005.
[26] Biao Chen, Lang Tong, and P.K. Varshney,“Channel-aware distributed detection in wireless sensor networks,”Signal Processing Magazine, IEEE, vol. 23, no. 4, pp. 16 – 26, july 2006.
[27] Ke Liu and A.M. Sayeed,“Type-based decentralized detection in wireless sensor networks,”Signal Processing, IEEE Transactions on, vol. 55, no. 5, pp. 1899 –1910, may 2007.
[28] Gokhan Mergen, Vidyut Naware, and Lang Tong,“Asymptotic detection performance of type-based multiple access over multiaccess fading channels,”Signal Processing, IEEE Transactions on, vol. 55, no. 3, pp. 1081 –1092, march 2007.
38 of 38