Transcript Distributed Signal Processing

Distributed Signal Processing
Woye Oyeyele
March 4, 2003
Outline
Distributed Signal Processing
Matched Filter Detection
Examples
2
AICIP Seminar Spring 2003
Definition

Distributed Signal Processing is
–
–
–
3
Distributed topology
Distributed task; sensing,localization,tracking
Distributed hypothesis
AICIP Seminar Spring 2003
Why Distributed concept?



Offers opportunities for improved line of sight
Finite energy budget at nodes
Communications consumes significant power
–
4
Need to transmit minimal data
AICIP Seminar Spring 2003
Two Dimensions

Hardware DSP
–
–

Software DSP
–
5
Focus on distributed computation
Efficient computation strategies
Distributed Sensing,Fusion, Detection
AICIP Seminar Spring 2003
Software Distributed Signal
Processing

Distributed Detection
–

Distributed Estimation
–
6
Identify presence of phenomenon
Determine particular target, locate and track
AICIP Seminar Spring 2003
Typical Topologies[1][3]



7
Parallel
Serial or tandem
Tree
AICIP Seminar Spring 2003
Parallel Topology
Phenomenon
y1
s1
y2
s2
y3
s3
...
Fusion Center
u0
8
yN
yN-1
Sn-1
sn
Fusion Center may be
omitted to form a
variant of this topology
AICIP Seminar Spring 2003
Serial or Tandem topology
Phenomenon
u2
u1
s1
9
s2
…
u3
s3
u0=un
un-1
Sn-1
sn
AICIP Seminar Spring 2003
Tree Topology
s3
s4
u3
s5
u5
u4
s6
u6
s2
u2
s1
u1
s0
10
u0
AICIP Seminar Spring 2003
Mobile Agent topology

11
Combined parallel, serial and short hierarchical
AICIP Seminar Spring 2003
Overall aims



12
Maximize probability of detection
Minimize probability of error
Maximize accuracy of classification
AICIP Seminar Spring 2003
Outline
Distributed Signal Processing
Matched Filter Detection
Examples
13
AICIP Seminar Spring 2003
Matched Filter Approach to
Distributed Detection

Two fold
–
–
14
Restricted matched filter(RMF) – restricted to choice
of sensors
Match observations to a template of expected
cues/features
AICIP Seminar Spring 2003
Aims and objectives


Model as a distributed antenna problem
Cast problem in proper domain
–

15
Frequency, time, time-frequency
Develop robust, scalable and highly reliable
algorithms
AICIP Seminar Spring 2003
Sequential vs. Fixed Sample size
detection
Sequential



16
Number of observations
combined to reach
decision is dependent on
nature of observations
Stopping rule is required
at each sensor
Number of observations
used is random
Fixed Sample size

Operate on observation
size predetermined
during design
AICIP Seminar Spring 2003
Matched Filter Algorithm
input
Perform FFT
Knowledge
Base
Construct matched
signal
-Filter response
Correlator
-To be based on
knowledge of
expected target
signatures
output
17
-Omitted in testing
phase
Normally, possible transmissions are built to be easily
differentiated, but target signatures are more complicated.
AICIP Seminar Spring 2003
General Challenges


Information is spatially distributed
Data is multimodal
–
–

Asynchronous nature of observations
–
18
Common statistics not sufficient to differentiate
Measurement variability exists for different temporal
measurements of same mode
Data captured in out-of-order fashion
AICIP Seminar Spring 2003
General Challenges – contd.

Energy Efficiency
–

Hierarchical data
–
19
Can algorithm recover/slowly degrade on power
failure
Local, global
AICIP Seminar Spring 2003
RMF detection aims


Obviate high communication costs occurring
due to use of all sensor observations
Choose adequate subset
–
20
Optimal subset balances signal energy and noise
correlation
AICIP Seminar Spring 2003
Mobile Agent RMF challenges

Size constraints
–

Communication costs
–

Need to minimize
Time
–
21
Limit for buffer growth vs. attainment of sufficient
threshold
Need to serve real time environments
AICIP Seminar Spring 2003
RMF algorithm
y
Select
Sensors
Criterion:
Balance signal
Energy and
noise correlation
l = skT-1yk
=  h[i] y[i]
22
yk
Filter
h[i]
l
j=0
Matched to
sufficient
statistic l
{
yk =
nk – target absent(noise)
sk + nk – target present
AICIP Seminar Spring 2003
RMF Characteristics



23
Optimal choice of sensors ensures maximum
SNR
Sufficient statistic (with tolerance) is a
threshold for evaluating likelihood of
hypothesis
Spatial processing independent of algorithm
topology
AICIP Seminar Spring 2003
Focus of current work


Focus is on filter design i.e. second stage of
algorithm
Future – technique for estimating noise
correlation between sensors
–
24
assumption that noise is independent between
sensors may not be right
AICIP Seminar Spring 2003
Distributed Signal Processing
Matched Filter Detection
Examples
25
AICIP Seminar Spring 2003
Example from SensIT data
Channel
26
Mode
1
Seismic
2
Acoustic
3
PIR
-SITEX00 Data used
-Event 08020830 (Dragon Wagon)
-Sensor Node A01
AICIP Seminar Spring 2003
Signal Plot –Channel 1
27
AICIP Seminar Spring 2003
Signal Plot – Channel 2
28
AICIP Seminar Spring 2003
Signal Plot – Channel 3
29
AICIP Seminar Spring 2003
Power Spectral Density – Channel 1
30
AICIP Seminar Spring 2003
Power Spectral Density – Channel 2
31
AICIP Seminar Spring 2003
Power Spectral Density – Channel 3
32
AICIP Seminar Spring 2003
Time domain plot - 3 seconds
33
AICIP Seminar Spring 2003
Freq. domain plot
34
AICIP Seminar Spring 2003
Filter Response
35
AICIP Seminar Spring 2003
Final Output
36
AICIP Seminar Spring 2003
Recap


37
Distributed Detection and Estimation offers
tremendous benefits
Matched Filter forms a conceptual basis for
sensor selection and local processing
AICIP Seminar Spring 2003
Research Directions

Develop spatial techniques for matched filter
detection
–

Algorithm to select optimal sensor subset
–
–
38
Combine with beamforming and direction of arrival
mechanisms to detect and localize target (target
tracking is continuous localization)
Estimate noise correlation between sensors
Achieve optimal ROC - quadratic
AICIP Seminar Spring 2003
References
1.
2.
3.
4.
5.
6.
39
R. Viswanathan and P. K. Varshney, "Distributed detection with multiple sensors:
Part I - fundamentals," Proceedings of the IEEE, vol. 85, no. 1, pp. 54-63, Jan.
1997
R. Viswanathan and P. K. Varshney, "Distributed detection with multiple sensors:
Part I - fundamentals," Proceedings of the IEEE, vol. 85, no. 1, pp. 54-63, Jan.
1997
Couch II, Leon W., Digital and Analog Communications Systems, Fourth Edition,
Macmillan Publishing Co, 1993, ISBN 0-02-325281-2.
P. K. Varshney, Distributed Detection and Data Fusion, Springer-Verlag 1996
D. Estrin, L. Girod, G. Pottie, M. Srivastava, Instrumenting the world with
Wireless Sensor Networks.
Charles Sestok, Alan Oppenheim, The Restricted Matched Filter for distributed
detection(presentation), DARPA SensIT PI Meeting, Jan. 16, 2002.
AICIP Seminar Spring 2003