1 collaborative processing in sensor networks lecture 4 - distributed in-network processing hairong...
TRANSCRIPT
1
Collaborative Processing in Sensor Networks
Lecture 4 - Distributed In-network Processing
Hairong Qi, Associate ProfessorElectrical Engineering and Computer ScienceUniversity of Tennessee, Knoxvillehttp://www.eecs.utk.edu/faculty/qiEmail: [email protected]
Lecture Series at ZheJiang University, Summer 2008
2
Research Focus - Recap
• Develop energy-efficient collaborative processing algorithms with fault tolerance in sensor networks
– Where to perform collaboration?– Computing paradigms
– Who should participate in the collaboration?– Reactive clustering protocols– Sensor selection protocols
– How to conduct collaboration?– In-network processing– Self deployment
3
A Signal Processing and Fusion Hierarchy for Target Classification
modality 1)( 1
11 tx
modality 1)(1
1 rtxmodality M
)(1 kM tx
Temporalfusion
Multi-modalityfusion
Multi-sensor fusion
……
Temporalfusion
…
… …
……
… …
node 1 node i node N
Localprocessing
Localprocessing
Localprocessing
Localprocessing
Localprocessing
Localprocessing
Mobile-agent Middleware
5
Local Signal Processing
Am
plitude stat.
Time series signal
Power Spectral Density (PSD) Wavelet Analysis
Shape stat.
Peak selection
Coefficients
Feature vectors (26 elements)
Feature normalization, Principal Component Analysis (PCA)
Target Classification (kNN)
Feature extraction
Classifier design
6
A Signal Processing and Fusion Hierarchy for Target Classification
modality 1)( 1
11 tx
modality 1)(1
1 rtxmodality M
)(1 kM tx
Temporalfusion
Multi-modalityfusion
Multi-sensorfusion
……
Temporalfusion
…
… …
……
… …
node 1 node i node N
Localprocessing
Localprocessing
Localprocessing
7
Temporal Fusion
• Fuse all the 1-sec sub-interval local processing results corresponding to the same event (usually lasts about 10-sec)
• Majority voting
,maxarg cj
i ωϕ = ],1[ Cc∈
number of possible localprocessing results
number of localoutput c occurrence
8
A Signal Processing and Fusion Hierarchy for Target Classification
modality 1)( 1
11 tx
modality 1)(1
1 rtxmodality M
)(1 kM tx
Temporalfusion
Multi-modalityfusion
Multi-sensorfusion
……
Temporalfusion
…
… …
……
… …
node 1 node i node N
Localprocessing
Localprocessing
Localprocessing
9
Multi-modality Fusion
• Difference sensing modalities can compensate each other’s sensing capability and provide a comprehensive view of the event
• Treat results from different modalities as independent classifiers – classifier fusion
• Majority voting won’t work• Behavior-Knowledge Space
algorithm (Huang&Suen)
Assumption: - 2 modalities - 3 kinds of targets - 100 samples in the training SetThen: - 9 possible classification combinations
c1, c2 samples from each class fused result
1,1 10/3/3 11,2 3/0/6 31,3 5/4/5 1,3
…3,3 0/0/6 3
10
A Signal Processing and Fusion Hierarchy for Target Classification
modality 1)( 1
11 tx
modality 1)(1
1 rtxmodality M
)(1 kM tx
Temporalfusion
Multi-modalityfusion
Multi-sensorfusion
……
Temporalfusion
…
… …
……
… …
node 1 node i node N
Localprocessing
Localprocessing
Localprocessing
11
Value-based vs. Interval-based Fusion
• Interval-based fusion can provide fault tolerance• Interval integration – overlap function
– Assume each sensor in a cluster measures the same parameters, the integration algorithm is to construct a simple function (overlap function) from the outputs of the sensors in a cluster and can resolve it at different resolutions as required
w5s5w4s4
w3s3w2s2w1s1
w6s6 w7s7
3
2
1
0
O(x)
Crest: the highest, widest peak of the overlap function
12
Fault Tolerance Comparison of Different Overlap Functions
n: number of sensorsf: number of faulty sensorsM: to find the smallest interval that contains all the intersections of n-f intervalsS: return interval [a,b] where a is the (f+1)th left end point and b is the (n-f)th right end pointΩ: overlap functionN: interval with the overlap function ranges [n-f, n]
13
Multi-sensor Fusion for Target Classification
• Generation of local confidence ranges (For example, at each node i, use kNN for each k5,…,15)
• Apply the integration algorithm on the confidence ranges generated from each node to construct an overlapping function
confidencerange
confidence level
smallest largest in this column
Class 1 Class 2 … Class nk=5 3/5 2/5 … 0k=6 2/6 3/6 … 1/6 … … … … …k=15 10/15 4/15 … 1/15
2/6, 10/15 4/15, 3/6 … 0, 1/6
14
Distributed Integration Scheme
• Integration techniques must be robust and fault-tolerant to handle uncertainty and faulty sensor readouts
• Provide progressive accuracy• A 1-D array serves to represent the partially-
integrated overlap function• Mobile-agent-based framework is employed
15
Mobile-agent-based Multi-sensor Fusion for Target Classification
Node 1
Node 2
Node 3
[2/6, 10/15] Class 1
[3/15, 3/6] Class 2…
Confidence range CR1
[3/6, 4/5] Class 1
[0, 5/15] Class 2…
Confidence range CR2
[3/5, 10/12] Class 1
[1/15, 2/8] Class 2…
Confidence range CR3
At node 2, fusion code generates partially integrated confidence range CR12
[3/6, 10/15] Class 1
[3/15, 5/15] Class 2…
Fusion result at node 2: class 1 (58%)
Mobile agentcarries CR1
Mobile agentcarries CR12
Fusion result at node 3: class 1 (63%)
At node 3, fusion code generates partially integrated confidence range CR123
[3/5, 10/15] Class 1
[3/15, 2/8] Class 2…
16
An Example of Multi-sensor Fusion
[0.10 0.29][0.46 0.65][0.10 0.21]
[0.05 0.14][0.05 0.41][0.22 0.58]
[0.05 0.15][0.05 0.15][0.49 0.59]
[0.08 0.16][0.08 0.16][0.51 0.60]
node 1
node 2node 3
node 4
accwhc ××=h: height of the highest peakw: width of the peakacc: confidence at the center of the peak
17
Example of Multi-sensor Integration Result at Each Stop
stop 1 stop 2 stop 3 stop 4
c acc c acc c acc c acc
class 1 1 0.2 0.5 0.125 0.75 0.125 1 0.125
class 2 2.3 0.575 4.55 0.35 0.6 0.1 0.75 0.125
class 3 0.7 0.175 0.5 0.25 3.3 0.55 3.45 0.575
19
SITEX02 Scenario Setup
• Acoustic sampling rate: 1024HzSeismic sampling rate: 512 Hz
• Target types: AAV, DW, and HMMWV• Collaborated work with two other universities (Penn State, Wisconsin)• Data set is divided into 3 partitions: q1, q2, q3• Cross validation
– M1: Training (q2+q3); Testing (q1)– M2: Training (q1+q3); Testing (q2)– M3: Training (q1+q2); Testing (q3)
20
Confusion Matrices of Classification on SITEX02
Acoustic (75.47%, 81.78%)
Seismic (85.37%, 89.44%)
Multi-modalityfusion
(84.34%)
Multi-sensorfusion
(96.44%)
AAV DW HMV
AAV 29 2 1
DW 0 18 8
HMV 0 2 23
21
Result from BAE-Austin Demo
Compact Cluster Laydown
0.0
20.0
40.0
60.0
80.0
100.0
120.0
140.0
0.0 20.0 40.0 60.0 80.0 100.0 120.0 140.0
Ft
Ft Series1
Suzuki VitaraFord 350Ford 250 Harley Motocycle
24
Demo
• Mobile-agent-based multi-modal multi-sensor classification
– Mobile agent migrates with a fixed itinerary (Tennessee)
– Mobile agent migrates with a dynamic itinerary realized by distributed services (Auburn)
26
Target Detection in Sensor Networks
• Single target detection– Energy threshold– Energy decay model:
• Multiple target detection– Why is multiple target detection necessary?– Requirements: energy-efficient & bandwidth efficient– Multiple target detection vs. source number estimation
αd
EE source
obs =
source
WXS =
:S
:X
:W
source matrix
observation matrixunmixing matrix
Assumption:
)()( XsizeSsize =
Target separation Speaker separation
27
Source Number Estimation (SNE)
• Task:• Algorithms for source number estimation
– Heuristic techniques– Principled approaches: Bayesian estimation, Markov chain Monte
Carlo method, etc.
• Limitations– Centralized structure
– Large amount of raw data transmission– Computation burden on the central processing unit
Environment
Sensor 1 Sensor i
central processing unit
data data data
…… ……
decision
Sensor n
)|(maxarg Xmm
HPm =∧
28
Distributed Source Number Estimation Scheme
SensorSensor
Sensor
… SensorSensor
SensorSensor
clusterhead
clusterhead
Fusion Center
decision decision
cluster 1 cluster L
System architecture
Algorithm structure
)(1 mL
m targets present in sensor filed
Local estimation at cluster 1
progressive
Local estimation at cluster L
progressive
Posterior probability fusion
)(mLL
…centralized centralizedcentralized centralized
29
Centralized Bayesian Estimation Approach
∑=
allH
mmm HPHp
HPHpHP
)()|(
)()|()|(
X
XX
∫
∏=
=
aaaxx
xX
dRAtpRAtp
RAtpRAp
nn
tnn
)(),,,|)((),,|)((
),,|)((),,|(
πϕϕ
ϕϕ
: sensor observation matrixX
: hypothesis of the number of sourcesmH
: source matrixS
: mixing matrix,A SX A=: unmixing matrix,W and TT AAAW 1)( −=XS W=
: latent variable,a Xa W= and )(aS φ=: non-linear transformation functionφ: noise, with variancenR
: marginal distribution of)(⋅π aβ/1
)|(log)( )(m
t HpmL x=
2))()((2
log2
1)
2log()(
2
1))((log tAtAAmnt
T ∧∧∧
∧∧∧
∧
−−−−+= axaβ
πβπ
∑=
∧∧
++−m
j
j mnanmn
1
2
]log)log(2
)2
log(2
[ γπ
β
Source: S. J. Roberts, “Independent component analysis: source assessment & separation, a Bayesian approach,” IEE Proc. on Vision, Image, and Signal Processing, 145(3), pp.149-154, 1998.
30
Progressive Bayesian Estimation Approach
• Motivation– Reduce data transmission– Conserve energy
)(1 mL
m targets present in sensor filed
Local estimation at cluster 1
progressive
Local estimation at cluster L
progressive
Posterior probability fusion
)(mLL
…centralized centralized
31
Progressive Bayesian Estimation Approach
1I
Environment
Sensor 11x
Sensor 22x
2I … 1−iI Sensor iix
iI … 1−nI Sensor nnx
)(mLn)(mLi)(2 mL
Estimation accuracy is progressively improved
includesiI
i][A
it ][ )(a
iε
iTerm ]1[
iTerm ]2[
…
iTerm ]7[
Transmitted information
)|(log)( )(m
t HpmL x=
2))()((2
log2
1)
2log()(
2
1))((log tAtAAmnt
T ∧∧∧
∧∧∧
∧
−−−−+= axaβ
πβπ
∑=
∧∧
++−m
j
j mnanmn
1
2
]log)log(2
)2
log(2
[ γπ
β
32
Progressive Update at Local Sensors
)()()( )(1
tiii xfmLmL += −
)(mL 2))()((2
log2
1)
2log()(
2
1))((log tAtAAmnt
T ∧∧∧
∧∧∧
∧
−−−−+= axaβ
πβπ
∑=
∧∧
++−m
j
j mnanmn
1
2
]log)log(2
)2
log(2
[ γπ
β
)(,
)(,1
)(
1 ]1)2)[exp(][2exp(1
]1[]1[ tiik
tiiki
t
kii xwxwaTermTerm −−−−−= −
∧
− ααα
1
1
2)(
,)(
1
)(
2)
1log(
2]2[
1]2[
−
=
∧
−
∑−⋅
−+
−−−−
+−−
−=
i
m
k
t
kkit
i
ii
aAxmi
mimimi
Termmi
miTerm
ε
4
1
)(
,)(
1
)(
,)(
11 )()(2]4[1
]4[ ∑∑=
∧
=
∧
−− −+−+−−−
=m
k
t
kkit
i
m
k
t
kkit
iiii aAxaAxTermmimi
Term ε
only depends on the updating rule of mixing matrixiTerm ]3[ A
Progressive update of log-likelihood function:
33
Progressive Update at Local Sensors
)(mL 2))()((2
log2
1)
2log()(
2
1))((log tAtAAmnt
T ∧∧∧
∧∧∧
∧
−−−−+= axaβ
πβπ
∑=
∧∧
++−m
j
j mnanmn
1
2
]log)log(2
)2
log(2
[ γπ
β
∑=
−
∧
−
∧
−
++
−=
m
ki
t
k
i
t
kt
iikt
iikii
a
axwxwiTerm
i
iTerm
1 21
)(1
)()(
,2)(
,1
)]([
][2)(
2]6[
1]6[
1
1
2)(
,)(
1
)(
2)
1log(
2]5[
1]5[
−
=
∧
−
∑−⋅+
−−−
+−
=i
m
k
t
kkit
i
ii
aAxim
mimiim
Termii
Termε
only depends on the updating rule of mixing matrixiTerm ]7[ A
Progressive update of mixing matrix : BFGS method (Quasi-Newton method)A
2
1
)(
,)(
1 )( ∑=
∧
− −+=m
k
t
kkit
iii aAxεεProgressive update of error:
36
Mobile-agent-based Implementation of Progressive Estimation
Sensor 1)(
1tx
Sensor 2)(
2tx
Sensor 3)(
3tx
Initialization
Update
Update terms in log-likelihood function
Compute estimated latent variable
Update estimation error
A
)(mL)(t∧
a
ε
Update
Update terms in log-likelihood function
Compute estimated latent variable
Update estimation error
A
)(mL)(t∧
a
ε
Output m
37
Distributed Fusion Scheme
• Bayes theorem
• Dempster’s rule of combination– Utilize probability intervals and
uncertainty intervals to determine the likelihood of hypotheses based on multiple evidence
)(1 mL
m targets present in sensor filed
Local estimation at cluster 1
progressive
Local estimation at cluster L
progressive
Posterior probability fusion
)(mLL
…centralized centralized
∏
∏
=
=== L
l
tl
m
L
lm
tl
tmm
tt
m
p
HPHp
p
HPHpHP
1
)(
1
)(
)(
)()(
)(
)()|(
)(
)()|()|(
x
x
x
xx
∑=
=L
lm
tll
tm HpcHP
1
)()( )|(log)|(log xx
where
∑∑==
==ll K
k kl
K
kk
ll dK
EK
c1
21
111
∑∑
≠=∩
≠=∩
∗−
∗=
jiHallH jgif
jiHHallH jgif
m
gf
mgf
HPHP
HPHPHP
,
,
)|()|(1
)|()|()|(
φXX
XXX
38
Performance Evaluation
Target types
Sensor nodesclustering
Signals: 1-second acoustic, 500 samples each
)(1 mL
m targets present in sensor filed
Local estimation
progressive
Local estimation
progressive
Posterior probability fusion
)(mLL
…centralized centralized
Bayesian Dempster’s rule
m targets present in sensor filed
centralized progressive
Source number estimation
centralized
Experiment 1
progressive
Experiment 2
Bayesian
Experiment 3
Dempster’s rule
Experiment 4
centralized centralizedprogressive progressive
Experiment 5Experiment 6
39
Average Log-likelihood Comparison
Experiment 1(Centralized)
Experiment 2(Centralized + Bayesian)
Experiment 3Centralized + Dempster’s rule
Experiment 4(Progressive)
Experiment 5 & 6(Progressive (cluster 1) vs. fusion)
Experiment 5 & 6(Progressive (cluster 2) vs. fusion)
40
Output Histogram ComparisonExperiment 1(Centralized)
Experiment 2(Centralized + Bayesian fusion)
Experiment 3(Centralized + Dempster’s rule)
Experiment 4(Progressive)
Experiment 5(Progressive + Bayesian fusion)
Experiment 6(Progressive + Dempster’s rule)
42
Discussion
• Develop a novel distributed multiple target detection scheme with– Progressive source number estimation approach (first in literature)– Mobile agent implementation– Cluster-based probability fusion
• The approach (progressive intra-cluster estimation + Bayesian inter-cluster fusion) has the best performance
Kurtosis Detection probability(%)
Data transmission(bits)
Energy comsumption
(uW*sec)
Centralized -1.2497 0.25 144,000 486,095
Centralized + Bayesian fusion
5.6905 0.65 128,160 433,424
Progressive + Bayesian fusion
4.8866 0.55 24,160 (16.78%) 89,242(18.36%)
43
Reference
• X. Wang, H. Qi, S. Beck, H. Du, “Progressive approach to distributed multiple-target detection in sensor networks,” pp. 486-503, Sensor Network Operations. Editor: Shashi Phoha, Thomas F. La Porta, Christopher Griffin. Wiley-IEEE Press, March 2006.
• H. Qi, Y. Xu, X. Wang, “Mobile-agent-based collaborative signal and information processing in sensor networks,” Proceedings of the IEEE, 91(8):1172-1183, August 2003.
• H. Qi, X. Wang, S. S. Iyengar, K. Chakrabarty, “High performance sensor integration in distributed sensor networks using mobile agents,” International Journal of High Performance Computing Applications, 16(3):325-335, Fall, 2002.