acoustic and seismic array processing in sensor network
Post on 05-Oct-2021
4 Views
Preview:
TRANSCRIPT
Acoustic and Seismic Array Processing in Sensor Network
Kung YaoUCLA
CTW, May 23, 2006
Outline0. Introduction1. Source localization using randomly distrib.
arrays based on TDOA-LS methods2. Source localization of near/far field
sources based on AML method3. Near field seismic source localization using
accelerometric data4. Distributed sensor node localization5. Conclusions6. Challenging Problems
0. Introduction• The purpose of a sensor network is to sense
one or more specific physical phenomena • Often it is necessary to detect, track, localize,
classify these physical sources• We consider three different approaches for
acoustic and seismic source localization• Often it is also necessary to localize the
sensing nodes/arrays relative to the sources • We also consider a distributed Gauss-Newton
iterative node localization method
1. Randomly distributed array processing based on TDOA-LS methods
Sensor 1 delay= 0
Sensor 2 delay= t12
Sensor R delay= t1R
Source
x1(t)
sensor input1
Propagationdelays
R
x2 (t)
xR(t)
+coherentlycombinedoutput y(t)
t1R
t122
sensor output
Sensor 1 delay= 0
Sensor 2 delay= t12
Sensor R delay= t1R
Source
x1(t)
sensor input1
Propagationdelays
R
x2 (t)
xR(t)
+coherentlycombinedoutput y(t)
t1R
t122
sensor output
Coherent Processing for Narrowband Beamforming
Space-Time-Frequency Wideband Beamforming
D
+
+
D
D
+y(t) = filtered array output
w10*
w11*
w20*
w1(L-1)*
D
+
+
D
D
w21*
w2(L-1)*
D
+
+
D
D
w30*
w31*
w3(L-1)*
x1(n) x2(n) x3(n)
Data from D number of sources collected by Nrandomly distributed sensors
Output of the beamformer
td,n - time-delayL - number of FIR tapsw - array weight
1*
1 0( ) ( )
N L
nl nn l
y t w x t l−
= =
= −∑∑
,1
( ) ( ) ( )D
n d d n nd
x t s t t m t=
= − +∑
Sample Auto- and Cross-Correlation Matrices
Array Weight Obtained by Dominant Eigenvector of Cross-Correlation Matrix
(For simplicity, consider N = 3)
Maximizing Beamformer Output
where w3L = [w10, w11,…,w1(L-1),…,w20,…,w2(L-1),…,w30,…,w3(L-1)]T
Time delays td, n est. from the w3L (Yao et al, IEEE JSAC, Oct. 1998)
the eigenvector of largest eigenvalue of R3Lw3L=λ3Lw3L
is
Intuitive & Formal Interpretation of the Beamformer
Intuitive interpretation
Max array output ~ coherently sum the “strongest
part” of the source
Szego asymptotic distribution of eigenvalues (1915)
S S dL
LL
max.
.limmax{ ( )} ( )= ∫ =
− →∞ω ω λ
0 5
0 5
≈ ==
λ LL
wLLH
L Lw R w( ) max { }1
S dL
LLL
L( ).
.lim
( ) ... ( )ω ω λ λ
−∫ =
→∞
+ +
0 5
0 51
for sufficiently large L
Simple Example
• Second order AR source with spectral peak at f = 0.2
• t12 = 3 ; t23 = 2 ; t13 = 5
• Array wL = φ3L (rL : -1 : (r-1)L+1), r = 1, 2, 3.
• Propagation delay information contained in the array weights wL
(r)
• Each array FIR acts as a narrowband filter centered at frequency of Smax
(3L)(r)
0 50 1000
0.01
0.02
0.03
0.04
0.05
0.06
0.07
0.08
0.09
W eight index
Mag
. of t
hree
arr
ay w
eigh
ts
0 0 .5 10
1
2
3
4
5
6
f
Mag
. of t
rans
fer
func
tion
of a
n ar
ray
chan
nel
n |wL(1)| |wL
(2)| |wL(3)|
1 0.0062 0.0011 0.00022 0.0088 0.0020 0.00053 0.0114 0.0036 0.00114 0.0140 0.0062 0.00205 0.0165 0.0088 0.00366 0.0191 0.0114 0.00627 0.0216 0.0140 0.00888 0.0241 0.0165 0.01149 0.0266 0.0191 0.014010 0.0291 0.0216 0.016511 0.0315 0.0241 0.019112 0.0339 0.0266 0.021613 0.0363 0.0291 0.024114 0.0386 0.0315 0.026615 0.0409 0.0339 0.0291
Mag. of three arrayweights
Mag. of transfer function of any array channel vs. freq
Values of |wL(1)|, |wL
(2)| and |wL(3)|
Continuing
Source Localization from EstimatedTime Difference of Arrival (TDOA)
-20 -15 -10 -5 0 5 10 15 20-20
-15
-10
-5
0
5
10
15
20
+1 +2
+3
x-coordinate
y-co
ordi
nate
sensor #1 at (0,0)
sensor #2 at (10,0)
sensor #3 at (9,7)
constant contour of time delay t13=5
constant contour of time delay t12=3
constant contour of time delay t23=2
• Least-squares solution is then given as follows after algebraic manipulation
Aw b=
TDOA - Least-Squares
=
=
−
−−
=
2
23
22
212
1
213
212
1
13333
12222
21,,
2/
2/2/
n
t
t
t
nnnnn r
rr
b
v
vszyx
w
t
tt
tzyx
tzyxtzyx
A
We write , where
w A b= + A A A AT T+ −= ( ) 1, where the pseudoinverse
An overdetermined solution of the source location and speed of propagation can be given from the sensor data as follows
-20 -10 0 10 20 30 40 50-10
-5
0
5
10
15
X (feet)
Y (fe
et)
Fig.4(a)
0 5 10 15 200
2000
4000
6000
8000
10000
Time slot
velo
city
(ft/s
ec)
Fig.4(b)
Source Localization and Speed of PropagationResults Using Seismic Array Sensors
2. Approx. ML Estimation Method
• ML method is a well-known statistical est. tool
• We formulated an approx. ML (AML) method for wideband signal for DOA, source localization, and optimal sensor placement in the freq. domain (Chen-Hudson-Yao, IEEE Trans. SP, Aug. 2002)
• AML method generally outperforms many suboptimal techniques such as closed-form least squares and wideband MUSIC solutions
• Has relative high complexity
AML Metric Plot
Near-field case Far-field case
• Peak at source location in near-field case• Broad “lobe” along source direction in far-field
case• Sampling frequency fs = 1KHz, SNR = 20dB
− sensor locations
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Y-a
xis
(met
er)
X-axis (meter)-5 0 5 10 15
-5
0
5
10
15
Semi-Anechoic Room at Xerox Parc
-2 -1 0 1 2 3 4
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
X-axis (meter)
Y-a
xis
(met
er)
Sensor locations Actual source location Source location estimates
-2 -1 0 1 2 3 4
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
X-axis (meter)
Y-a
xis
(met
er)
Sensor locations Actual source location Source location estimates
Indoor Convex Hull Exp. Results
AML LS
• Semi-anechoic room, SNR = 12dB• Direct localization of an omni-directional loud speaker
playing the LAV (light wheeled vehicle) sound• AML RMS error of 73 cm, TOA-LS RMS error of 127cm
Outdoor Testing at Xerox -Parc
-4 -2 0 2 4 6 8 10 12 140
5
10
15
A B
C
X-axis (meter)
Y-a
xis
(met
er)
Sensor locations Source location estimatesActual traveled path
-4 -2 0 2 4 6 8 10 12 140
5
10
15
A B
C
X-axis (meter)
Y-a
xis
(met
er)
Sensor locations Source location estimatesActual traveled path
Outdoor Moving Source Exp. Results
AML LS
• Omni-directional loud speaker playing the LAV sound while moving from north to south
• Far-field situation: cross-bearing of DOAs from three subarrays
Optimum Sensor Placement via CRB Approach
−1
0
1
2
3
4
CRB of source localization in log scale
−20 −10 0 10 20 30 40
−20
−10
0
10
20
30
40
Source should be inside the convex hull of the sensors
Low CRB value
High CRB value
Sensor Network at 29 Palms
-150 -125 -100 -75 -50 -25 0 25 50 75 100 125 150-100
-75
-50
-25
0
25
50
75
100
Node 1
Node 61
X-axis (meter)
Y-a
xis
(met
er)
Sensor locations Source location estimates
29 Palms Field Measurement• Single Armored Amphibious Vehicle (AAV) traveling at
15mph• Far-field situation: cross-bearing of DOAs from two
subarrays (square array of four microphones, 1ft spacing)
Outdoor Experimentusing iPAQ testbed
One and Two Source AML Localization
Space-Frequency Classification of Targets
• Previous discussions show that AML is able to estimate the DOAs of multiple targets
• Then the targets are spatially separated, detected, and located
• Since the AML algorithm not only estimates the DOA spatially, but also can yield the dominant spectral contents of each target
• The spectral signature of the target at a given DOA can provide information for classification
Spectra of Two Targets Used in Simulation for AML DOA Estimation
Tracked vehicle data Simple sound dataOne FFT index no. = 14 Hz
Excellent estimation of spectrum of source 1 at subarray 1
Excellent estimation of spectrum of source 2 at subarray 1
Other Use of AML Algorithm
• We are using the AML algorithm to detect, localize, and classify woodpeckers in collaboration with bio-complexity researchers
• We are using the AML metric as the LR function in the recursive formulation of particle filtering approach for real-time tracking of a human speaker in a reverberant room
Tracking Experiment
6 m
9m
2.53 mMic
speaker
Tracking experiments in a reverberant room
Experimental Results
0 2 4 6 8 10 12 141
1.5
2
2.5
3
3.5
4
4.5
5
Time (sec)
Loca
tion
(m)
Tracking results in the x−axis
PFAMLground truth
0 2 4 6 8 10 12 141
2
3
4
5
6
7
8
9
Time (sec)
Loca
tion
(m)
Tracking results in the y−axis
PFAMLGround truth
Tracking Results in x-axis Tracking Results in y-axis
Update rate: 128 ms, # of particles=200
3. Garner Valley, CA Acoustic Array Experiment
• 8 low cost Behringer XM200S microphones
• Each array consists of 4 microphones 1 meter apart in a square
• 2 acoustic arrays• The output of 8 microphones
are sent to the PresonusFirepod 8 channel 94bit/96kb firewire-based recording sys.
• The recording system synchronizes the acoustic signals and sends to a PC for AML-based DOA/loc. algorithm
Garner Valley, CA Seismic Array Experiment
• Episensor tri-axial/bi-axial accelerometer sensors
• Accelerometers have widefrequency/amplitude ranges,and wide dynamic range
• Outputs of accelerometersare fed to the low power, highresolution Quanterra Q330srecording systems
• 9 sensor (6 tri-axial and 3 bi-axial), 8 of them on theperimeter of a 100 feet square, and the last one in the center
Garner Valley Acoustic and Seismic Field Measurements/Results
• Metal hammer struck a heavy metal plate at the two separate locations
Acoustic and Seismic Sensor Map Acoustic Array Map
−50 0 50 100 150−50
0
50
100
150Hammer Strike Experiment A
X (feet)
Y (
feet
)
Seismic SensorsAcoustic ArraysMetal Plate
1 2 3
4 5 6
7 8 9
1
2
−50 0 50 100 150−50
0
50
100
150Hammer Strike Experiment B
X (feet)
Y (
feet
)
Seismic SensorsAcoustic ArraysMetal Plate
2
4
1 3
5
7 9
6
8
1
2
Acoustic Array Localization
• AML-based acoustic DOA/localization using whitening pre-processing yields metal plate at (78.4, 23.8), close to the true location of (75, 25)
• AML-based acoustic DOA/localization using whitening pre-processing yields metal plate at (23.4, 80.1), close to the true location of (25, 75)
Case A Case B
Seismic Event Detection Results• The simple eigen-decomposition procedure can be
performed on sliding time windows through the data record to find signification events of interest
• The data record with 6 significant hammer strikes was selected for analysis
• The eigenvalue plot clearly indicates the 6 peaks above 10-4 is where the 6 significant hammer strikes happened
0 5 10 15 20 25 30 35 40−0.015
−0.01
−0.005
0
0.005
0.01
0.015Original Hammer Strike Signal, Sensor 1 Z Direction
Time (sec)
Sign
al S
treng
th
0 50 100 150 200 250 300 350−11
−10
−9
−8
−7
−6
−5
−4
−3
Window Number
Log
of E
igenv
alues
Eigen Value Progression As a Function of Window Number
Eigenvalue #1Eigenvalue #2Eigenvalue #3
Single-Station Tri-Axial Seismic DOA Est. via Covariance Matrix Analysis
• Use polarization analysis developed for long-range seismic data analysis
• The largest eigenvalue of the covariance matrix corresponds to the average energy of the strongest seismic mode polarized in the direction of the corresponding eigenvector
• Same can be said for the second and the smallest eigenvalues and their corresponding eigenvectors
Single-Station Tri-Axial Seismic DOA Est. via Surface Wave Analysis
• Rayleigh wave is elliptically polarized in two mutually orthogonal directions
• Love wave is rectilinearly polarized in the direction orthogonal to the two Rayleigh directions
• The three-dimensional space can be rotated such that two dimensions only pick up the Rayleigh wave, and the one other, orthogonal dimension picks up only the Love wave
Seismic Source LocalizationCovariance Analysis Source Localization
84.8481.2381.4379.7275Hmr. B y-coordinate
13.76 0.39
31.1038.3930.8152.4325Hmr. B x-coordinate
28.5154.4931.8264.1125Hmr. A y-coordinate
74.35 1.71
83.0761.4885.4359.1575Hmr. A x-coordinate
S.D. Range
Wt. L1Unw. L1Wt. TLSUnw. TLS
True SrcLocLocation in Feet
Surface Wave Analysis Source Localization
79.6665.7277.3669.6275Hmr. B y-coordinate
3.170.37
26.9731.1427.5629.8925Hmr. B x-coordinate
26.6734.1527.5932.1925Hmr. A y-coordinate
5.991.66
80.6485.2081.2780.3175Hmr. A x-coordinate
S.D. Range
Wt. L1Unw. L1Wt. TLSUnw. TLS
True SrcLocLocation in Feet
Combined Acoustic/Seismic Localization Case A
• A simple average of the acoustic and seismic coordinates gives the fusion result
−20 0 20 40 60 80 100 120 140
−20
0
20
40
60
80
100
120
140
160
180
X (feet)
Y (
feet
)
Acoustic and Seismic Source Localization Results, Hammer Strike Case A
Seis. Snsr Loc.Acou. Snsr Loc.Metal Plate Loc.Seis. SrcLocAcou. SrcLocFusion SrcLoc
1 2 3
4 5 6
7 8 9
1
2
25.279.5Fusion
26.780.6Seismic
23.878.4Acoustic
2575True
YXLocation in Feet
Combined Acoustic/Seismic Localization Case B
• The fusion result in this case is closer to the true source than either the acoustic or the seismic result
78.725.5Fusion
77.427.6Seismic
80.123.4Acoustic
7525True
YXLocation in Feet
−20 0 20 40 60 80 100 120 140
−20
0
20
40
60
80
100
120
140
160
180
X (feet)
Y (
feet
)
Acoustic and Seismic Source Localization Results, Hammer Strike Case B
Seis. Snsr Loc.Acou. Snsr Loc.Metal Plate Loc.Seis. SrcLocAcou. SrcLocFusion SrcLoc
1 2 3
4 5 6
7 8 9
1
2
4. Sensor Node LocalizationSystem Description
• m anchor nodes (locations known) and n sensor nodes (locations to be estimated)
• Each sensor node has distance measures to all of the sensor and anchor nodes in its neighborhood
Example• 4 anchor nodes at each
corner and 15 sensor nodes in the middle
• Radio range = 0.35• Every link represents a
communication link• Distance measurement dij
available -0.5 0 0.5-0.5
0
0.5sensor nodesanchor nodes
Metric for Minimization2
,
222
,
22)( ∑∑ −−+−−=
kiikki
jiijji daxdxxF x
Centralized Method : • Gauss-Newton (non-linear least squares)• Multi-Dimensional Scaling• SDP
Require communications to a central computer
[ ]TTn
T xx1=x is the i-th sensor location estimate2Rxi ∈is the k-th anchor node location2Rak ∈
1 2ˆ argmin ( , ,..., ), and are known (estimated) range values
n
ij ik
F x x xd d
= xx
Distributed Gauss-Newton Sequential/Parallel Node Localization Methods
Estimation Results
• 10 anchor nodes at each coroner and 100 sensor nodes in the middle
• Radio range = 0.35
-0.8 -0.4 0 0.4 0.8-0.8
-0.4
0
0.4
0.8true locationestimation anchor
Some Average Iteration/ Processing Time Behaviors
• Keep the density of the sensor, the ratio of sensor/anchor and the radio range to be constant
• Each sensor stops the algorithm if
0
5
10
15
20
aver
age
num
ber
of it
erat
ion
40 80 120
160
200
netwrok size (n)
Sequential Parallel
1
10
100
1000
aver
age
proc
essi
ng ti
me
40 80 120
160
200
netwrok size (n)
Sequential Parallel
01.0≤ip
Comparison to Centralized Algorithm
0
50
100
150
200
com
m. e
nerg
y co
nsum
ptio
n
40 80 120
160
200
network size (n)
Sequential Parallel Centralized
5. Conclusions• Considered three source localization methods:
1. TDOA-LS algorithms(centralized processing)2. AML algorithm (distributed processing)3. Seismic localization: covariance method
(centralized processing) and surface wavemethod (distributed processing)
• Distributed Gauss-Newton Node LocalizationMethod (Sequential and Parallel algorithms)
6. Challenging Processing Problems in SN
• Perform distributed computation of eigenvalue/ eigenvector/singular value/singular vector without sending raw data from the sensors to a central proc.
• Theoretical/practical (i.e., low complexity) use of data fusion from different types of sensors (video/acoustic/ seismic) of different quality and sampling rate
• Theoretical (e.g., CRB)/practical issues of placement of sensors of different types
• Using random set theory method for determining number of sources and nearest neighboring nodes
• Real-time/practical (i.e., low complexity) use of particlefiltering methodology for source/sensor node tracking
top related