Distance Based DecisionFusion in a Distributed

Wireless Sensor Network

Marco Duarte and Yu-Hen HuElectrical and Computer Engineering

Introduction Increasingly feasible miniature ad-hoc sensor

network integration. Conventional centralized information and

data fusion are unsuited because of amountof data.

Distance-based fusion algorithm will selectsensors that give reliable results toparticipate.

Wireless Sensor Nodesand Network Sensor node: computer, battery, sensors,

transceivers. Deployment of sensor nodes outdoors. Regions: geographical clusters, each with a

manager node. Objective: detect vehicles as they pass

region, identify vehicle type, estimate location(EBL).

Sensor Network SignalProcessing Tasks

CFAR TargetDetection

Target Classification Target Localization Target Tracking

-200.0

-100.0

0.0

100.0

200.0

300.0

400.0

-200.0 -150.0 -100.0 -50.0 0.0 50.0 100.0

road

Node1-6

Node15

Node41-42

Node46-56

Node58-61

Sensor Network SignalProcessing Tasks Maximum Likelihood

50-dim spectral feature: Sampling frequency 4960 Hz 512-point FFT, resolution ~9

Hz Average first 100 points by

pairs, describes ~900 Hz Signal + noise classification

rate depends on SNR. SNR proportional to vehicle-

node distance.

11( | ) ~ exp ( ) ( )

2

T

k k kP x k x x x x

!" #! ! $ !% &

' (

!

Classification Success

•Correct•Incorrect

All Vehicles AAV

DW

Classification Rate Distribution

0

1

ClassificationProbability

All Vehicles AAV

DW

Distance Based DecisionFusion

Current system architecture allows localizationprior to classification.

Accurate localization allows for estimation ofsensor-vehicle distance, estimation of probabilityof correct classification based on distance.

Data fusion: function of marginal results fromeach node.

Some events may be rejected by fusionalgorithm.

Measurements: classification and acceptancerates.

Decision Fusion x(i) is the feature vector for ith sensor, Ck is the kth

vehicle class, we must establish a function

g(zk) is the maximum function:

This is called Decision Fusion

( | (1), , ( )) ( | )

( ( ( | ( ))),1 )

k k

k

P x C x x N P x C x

f g P x C x i i N

! !

" ! # #

… !

1 ,( )

0 otherwise

k j

k

z z k jg z

> !"= #$

Distance Based DecisionFusion Multiplicative form: for statistically

independent feature vectors x(i), x(j).

Not realistic for sensor network. Additive form: weighted sum of marginal

posterior probabilities.

If wi=1 for all i, simple voting.

1

ˆ( | ) ( | ( )) i

N

w

k k

i

P x C x P x C x i

=

! = !"

1

ˆ( | ) ( ( | ( )))N

k i i k

i

P x C x w g P x C x i=

! = !"

Maximum A PosteriorDecision Fusion Weighting factor as function of distance and

SNR, determined using CFAR and EBLinformation.

We formulate a Maximum A Posterior (MAP)Probability Gating Network, using Bayesianestimation:

Probabilities from experiment data.ˆ( ) ( | , , ) ( | , ) ( , )

k k i i i i i iP x C P x C x d s P x d s P d s! = ! " "

Maximum A PosteriorDecision Fusion This amounts to assigning

weights:

Other methods: Distance gating:

Nearest Neighbor:

Baseline: simple voting(wi=1 for all i)

( | , ) ( , )i i i i iw P x d s P d s= !

1

0 otherwise

i max

i

d dw

!"= #$1 ,

0 otherwise

i j

i

d d j iw

! " #$= %&

Experiment Results

Results

Closest node gives highest acceptance,classification rates for accurate localizationestimates

MAP Fusion has smaller dependence onlocalization error than other methods

Both of these methods can reducecommunication needed for decision

Further Work MAP Classifier allows for exclusion of those

samples with low classification rates (i.e. onlysamples with wi > 0.5 are allowed).

This will allow for reduction of communicationbandwidth used for classification fusion.

This method can be applied to other signalprocessing tasks.