UWCSP University of Wisconsin Collaborative Signal Processing

1
UWCSPUniversity of Wisconsin Collaborative
Signal Processing
Presented at DARPA SensIT PI Meeting

• University of Wisconsin Madison
• Dept. Electrical Comp. Engr.
• parmesh,hu, saluja ,akbar_at_engr.wisc.edu

2
Outline

• Overview
• Region-based Sensor Network Signal Processing
• UWCSP
• Node signal processing
• Energy based detection
• Acoustic feature classification
• Region-based collaborative signal processing
• Region detection and classification
• Energy based localization
• Least square, robust tracking
• Achievements
• Off-line simulation using Matlab and Sitex02 data
• On-line simulation using Sitex02 and synthetic
  data

3
Design Objectives

• Less communication!
• Conserve bandwidth, energy
• Distributed processing ?
• Centralized processing ?
• Fewer computation
• Conserve power
• Simple implementation
• Deep theory Complicate algorithm
  formulation
• Easier to debug

• Adaptive
• Energy aware
• Performance energy tradeoff
• Task oriented
• Accuracy on-demand
• Robust and Resilient
• Less sensitive to measurement error and noise
• Less sensitive to timing error

4
Sensor Network Collaborative Signal Processing

• A network of sensors working cooperatively to
  engage signal processing tasks
• Known but unstructured sensor locations
• Limited communication bandwidth
• Sensor communication via ad hoc wireless channels
• Low power operation
• Sensor nodes are powered by local batteries
• Collaborative signal processing needed to save
  power, conserve bandwidth

Sitex 02 experiment sensir field
5
Region-Based Sensor Network Signal Processing

• A region is a cluster of sensor nodes that
  locate within a contiguous, convex geographic
  region.
• Region assignment is based on tasks, sensing and
  comm. capabilities.
• One node can belong to multiple regions,
  executing multiple threads of process
• Adjacent regions collaborate to accomplish a
  query such as detection, classification,
  localization and tracking.

Region III
Region I
Region II
Region partition in Sitex02 sensor field
6
UWCSP
Node Detection

• Node signal processing
• Energy Detection
• Target classification
• Region signal processing
• Region detection and classification -- voting
• Energy based localization
• Least square tracking
• Hand-off policy

Node Classi- fication
7
NSP Node Signal Processing

• CFAR (constant false alarm rate) energy detection
• Computes one energy per 0.75 sec. (changeable)
• Can down-sample to reduce computation
• Detection rate can be improved by classification

• Acoustic signature classification
• Features
• FFT magnitude ?
• FFT real and imaginary
• Feature dimension 50
• Classifiers
• K-nearest neighbor ?
• Maximum likelihood classifier ?
• Support vector machine
• Learning vector quantization
• Decision trees C4.5, etc
• MLP neural network
• Baysian belief network
• etc

signal
noise
Signal energy pdf ?
Noise p.d.f.
energy
threshold
False alarm rate
8
Sample CFAR Detection Result

• Sitex02 node 4 channel 1, recorded on Mon Nov 13,
  2001 151724 528 msec to 154502 84 msec.
  Length 8 minutes 32 seconds
• Ground truth? Hearing is believing

Green line energy _at_ 0.75s interval Upper dash
line 3s Lower dash line s
Detection results
9
CFAR Detection of Acoustic Signals
energy detection
Node number
time
10
CFAR Detection of PIR Signals
energy detection
Node number
time
11
Additive Region Detection Fusion
12
Sample Feature Vectors

• Currently, 3 classes AAV, DW, and LAV
• Trained with Sitex00 broadband data from BAE
  broadband (08030800, 08031035, 08020800,
  08040820) and tested with Sitex02 AAV, DW data.
• Simple feature extraction algorithm
• 512-pt FFT on time series. Get 256-pt magnitude.
  Take the first 100 of them. Averaging 2 adjacent
  FFT magnitude to make a 50 x 1 feature.
• The feature covers 968.8 Hz, each dim. (frequency
  bin) covers 19.38 Hz.

0Hz 194 Hz 388Hz 582Hz 776Hz 970Hz
13
Classification Results

• Confusion matrices
• Training 78
• Testing 61.86

• Use Maximum Likelihood Classifier
• Discriminant function
• ?I, Ci mean, Covariant matrix of class i
  training samples
• i arg max gi

14
Energy Based Localization (EBL)

• Acoust
• Ic
• energy

• Energy decay Model
• If the energy decay exponent is a constant and
  can be measured, then distance between source and
  sensor can be estimated based on energy reading
  at individual sensors.

Distance from source
15
Implementations

• Ratio of acoustic energy between two sensors.
  Solve
• yield
• n(n-1)/2 energy ratios of n sensors gives
  n(n-1)/2 circles that should intersect at source
  location.
• Factors affecting location estimate accuracy
• Energy estimate y(t)
• Sensor locations ri
• Energy decay exponents a
• Sensor gain variation gi
• Nonlinear cost function that may contain multiple
  local minimum

16
Relative Energy at each Sensor Node
17
Corresponding Localization Results
Dot sensor nodes o sensor with energy
detection ground truth location
estimate Green line contour of 1 of global
minimum of cost function
18
Simulations Synthetic data

• Square sensors
• circle EBL estimated position
• dots ground truth
• Triangle predicted position _at_ 6 seconds later.
• Use Sitex02 time series and GPS ground-truth to
  generate synthetic data
• Each point is 0.75 sec.
• Two regions nodes 2-9, 10-15.
• Simple hand-off.

19
Simulations Sitex02 data

• Demonstrated on Tuesdays demo session
• Data Sitex02,
• run-2001-11-14-2149-aav,
• run-2001-11-14-2300-dw.
• Edited to include only east west section
• One region nodes 51, 52, 54, 55, 59

• Method
• Feed data into BAE repository to individual nodes
• Perform node detection in all 5 nodes
• Perform localization and tracking on manager node
• Dump result into a log file
• Use perl script to extract relevant information
  into an ASCII file and plot using Matlab