Você está na página 1de 39

Robust Adaptive Beamforming

with application to
Matched Field Processing

Yong Han Goh & Y. Jin


Outline
„ Background
„ 4 Robust Adaptive Beamforming Methods
„ LCMP
„ White Noise Gain Constraint
„ Gershman et al.
„ Stoica et al.
‰ Simulations
„ Intro to Matched Field Processing
‰ MFP simulations
‰ MFP results on the actual data
„ Conclusion
To Start with…
10
MPDR

0
Beampattern B(θ) in dB

-10

-20

Shortcoming: -30
„ Does not provide sufficient robustness against mismatch
-40
between presumed
0 and
50 actual 100
signal steering
150 vector
θ -space
„ Tends to suppress the SOI by adaptive nulling
Effect of mismatch in beamscan

15-elt ULA
NL = 50 dB
No
SNRin = 63 dB
mismatch
case

Source
SL = 140 dB
R = 7 km
Mismatch
case
Steering Vector Mismatch Due to…

„ look direction mismatch


„ array perturbation
„ array manifold mismodeling
„ wavefront distortions
„ source local scattering
„ …
Approaches to
Robust Adaptive Beamforming
„ Linear Constraint Minimum Power Beamformer
C w=g
H

‰ Directional Constraints
10 10 ⎡1⎤
⎡ ⎛ 0.891 ⎞ ⎤
g = ⎢⎢1⎥⎥
MPDR LCMP
⎛ 0.891 ⎞
0
C = ⎢1# v u ⎜ ⎟# v u ⎜ − ⎟⎥
0
Beampattern B(θ) in dB

Beampattern B(θ) in dB
⎣ ⎝ N ⎠ ⎝ N ⎠⎦
⎢⎣1⎥⎦
-10 -10

-20 -20

-30 -30

-40 -40
70 80 90 100 110 70 80 90 100 110
θ -space θ -space
White Noise Gain Constraint R = σ s 2 vs vs H + σ n 2 Sn

H 2
w v 1
„ Array Gain: G= H
⎯⎯⎯⎯→
distortionless
white noise 2
w Sn w w

„ Signal perturbed by uncorrelated random errors:


R = σ s 2 ( vs vs H + ξ I ) + σ n 2 S n
„ Sensitivity:
⎛ dG / d ξ ⎞ distortionless 1
S =⎜ ⎟ ⎯⎯⎯⎯⎯ → w =
2

⎝ G ⎠ G
„ Sensitivity increases as white noise gain decreases
„ UWLA: low sensitivity; MVDR: high sensitivity
White Noise Gain Constraint
„ Goal: reduce sensitivity
„ Impose quadratic constraint on white noise gain to
increase robustness
1
−2
min w Rw subject to w d = 1, and Sw≤
H H
≥. δ 2
.
w δ 2

„ Solution is No simple relation

(R + ε I ) −1d
wWNGC = H
d (R + ε I ) −1d
„ Adjust ε until white noise gain constraint is satisfied
Shortcoming of WNGC

„ Relationship between ε and δ2 is not simple


„ Need approach to compute the ε based on the
uncertainty of the steering vector
„ Iterative procedure is required to adjust the
diagonal loading factor
„ In practice, can use ad hoc method to determine it
Looking for methods that…
„ Have sufficient robustness against arbitrary steering
vector mismatch
„ Have sound mathematical framework
„ Are computationally easy to implement

One approach:
2003: Sergiy A. Vorobyov, Alex B. Gershman, and Zhi-Quan Luo
Gershman et al. :
Worst-Case Performance Optimization
„ Steering vector distortion: ∆
Δ ≤ε

„ The actual steering vector belongs to


A ( ε )  {c | c = v + e, e ≤ ε }

Any vector satisfies constraint


Set of vectors presumed s.v.
A ( ε )  {c | c = v + e, e ≤ ε }

• Uncertainty Set
• Search space

v
c1
c2

„ Worst-Case optimization
For all vector in A(ε), the array response should NOT be
smaller than 1
w c ≥1
H
for all c ∈ A(ε )
Formulation of Robust BF
min wH Rw subject to wH c ≥ 1 for all c ∈ A(ε )
w

„ Looks good, but


‰ Non-linear
‰ Non-convex
‰ Infinite # of constraints

λ
minw0w= RwH subject2 to−1 −w1 ( Rv+−λεε Iw) ≥v 1
H H 2 −1

w λ v ( R + λε I ) v

min w H Rw subject to w H v ≥ ε w + 1 & Im{w H v} = 0


w
Another Method: Stoica et al. 2003:
Directly Estimate Signal Power σs2
R = σ s 2 v s v s H + ∑ σ i 2 vi vi H + Sn

„ Using w0HRw0 as σs2 estimate, MPDR gives


i 2 1
σs =
v s H Rv s
„ It can be shown that it’s the solution to

max σ 2
subject to R − σ v s v s ≥ 0
2 H
2
σ

Covariance fitting problem


Adding in uncertainty
max σ 2
subject to R − σ 2
v v
s s
H
≥0
σ
2

„ Given an uncertainty “ellipsoid”

( p ) (v − vp ) ≤1
H −1
v − v C

„ Estimate the power of SOI by


max σ 2
subject to R − σ 2
vv H
≥0
σ ,v
2

& ( v − v p ) C −1 ( v − v p ) ≤ 1
H
By derivation, we get
m2ax σ 2
subject to R − σ 2 vv H ≥ 0
σ ,v

(v − v ) C −1 ( v − v p ) ≤ 1
H
& p

subject to ( v − v p ) C −1 ( v − v p ) ≤ 1
−1 H
min vR v H
v

„ Assume C = εI, we have


2
min vR v −1 H
subject to v − vp ≤ε
v

−1
⎛R −1
⎞ R −1 v 0
Solution: v0 = ⎜ + I ⎟ vp w0 = H −1
⎝ λ ⎠ v0 R v0
Direct estimation of the actual steering vector
Relationship
Between Stoica’s and Gershman’s method
„ It can be shown that,

Stocia
2
min vR v−1 H
subject to v − vp ≤ε
v

R −1 v 0
Let v0 denote the optimal solution, and w0 = H −1
v0 R v0
Then w0 is the optimal solution to

Gershman
min wH Rw subject to wH v ≥ ε w + 1 & Im{wH v} = 0
w
10
Simulation 1: Beampatterns
MPDR of

Beampattern B(θ) in dB
LCMP 0
LCMP, WNGC, Stoica and -10
Gershman
-20

-30

-40
50 100 150 0 50 100
θ -space θ -space

10
MPDR

Beampattern B(θ) in dB
Stoica 0

-10

-20

„ -30
ε0=0.217, εSoitca=0.3, εGershman=√0.3
Compare the 4 methods using sample
covariance matrix
10 10
MPDR MPDR
Beampattern B(θ) in dB

Beampattern B(θ) in dB
0 LCMP 0 WNGC

-10 -10

-20 -20

-30 -30

-40 -40
0 50 100 150 0 50 100 150
θ -space θ -space

10 10
MPDR MPDR
Beampattern B(θ) in dB

Beampattern B(θ) in dB
0 Stoica 0 Gershman

-10 -10

-20 -20

-30 -30

-40 -40
0 50 100 150 0 50 100 150
θ -space θ -space
Comparison of algorithms in beamscan
space in presence of mismatch
MFP overview
Test source positions
Array for replica (a)


c(z) . . .
. . .
. . .
X
True source
position (atrue)

csed(z), ρsed(z)

B(a) = wH(a)S(atrue)w(a)
Peak of the output of the beamformer B(a) is at atrue
w depends on the beamformer used
SwellEx’ 96 experiment
Sound speed profile VLA
60 m
4 knots
source

21-elt
94.125 to
212.25 m
d = 5.63m
Processing of received array data
Vert. Received chunk’s FFTs snapshots
Array Time series
X00
1

.
. X01
2
.

49 Hz

64 Hz
.
chunk k .
.

L-1

(...)
Spectrogram of data

Nfft = 213
Lk = 213
Fs = 1500 Hz
Kaiser Window
(β=7.85)

Fi = 49, 64, 79, 94, 112, 130, 148, 166, 201, 235, 283, 338, 388 Hz
K
1
CSDM estimate, Sˆ i =
K
∑ ki ki
X
k =1
X H
Simulation results at F = 148 Hz
„ SNRin = 10 dB, 21 element array
„ Source at r = 3000 m, z = 60 m
Simulation results at F = 148 Hz

WNGC = 0.5N
Simulation results at F = 148 Hz
source
Experimental Results at F = 49 Hz
Source at r = 3000 m and z = 60 m
Experimental Results at F = 49 Hz
WNGC = 0.5N
Experimental Results at F = 49 Hz
Results averaged over first 5 frequencies
Results averaged over first 5 frequencies

WNGC = 0.5N
Results averaged over first 5 frequencies
Results averaged over all 13 frequencies
Results averaged over all 13 frequencies

WNGC = 0.5N
Results averaged over all 13 frequencies
Conclusions

„ Investigated various robust ABF algorithms


„ MFP results improved as we average over
frequencies (except for MPDR)
„ MUSIC best localized the source for this
particular set of MFP data
References
„ H. Cox, R. M. Zeskind, M. M. Owen, “Robust Adaptive Beamforming,” lEEE Trans.
Acoust., Speech, Signal Processing, vol. 35, pp. 1365-1376, Oct. 1987.
„ J. Li, P. Stoica, Z. S. Wang, “On Robust Capon Beamforming and Diagonal
Loading,” lEEE Trans. Signal Processing , vol. 51, pp. 1702-1715, Oct. 2003.
„ S. A. Vorobyov, A. B. Gershman, and Z. Q. Luo, “Robust adaptive beamforming
using worst case performance optimization,” lEEE Trans. Signal Processing , vol.
51, pp. 313-323, Feb. 2003.
„ A. B. Baggeroer, W. A. Kuperman, P. N. Mikhalevsky “An Overview of Matched
Field Methods in Ocean Acoustics,” lEEE Journal Oceanic Engineering , vol. 18,
pp. 401-424, Oct. 1993.
„ H. L. V. Trees, Optimum Array Processing, John Wiley & Sons, Inc, 2002.
„ F. B. Jensen, W. A. Kuperman, M. B. Porter, and H. Schmidt, Computational
Ocean Acoustics, Springer-Verlag New York, Inc, 2000.
„ SwellEx-96 Experiment data (http://www.mpl.ucsd.edu/swellex96/)..
„ H. Schmidt, A. B. Baggeroer, W. A. Kuperman, E. K. Scheer, “Environmentally
tolerant beamforming for high-resolution matched field processing: Deterministic
mismatch,” J. Acoust. Soc. Am. 88, pp. 1851-1862, 1990.
Normal mode representation of pressure
field
The normal mode representation of the field p(r,z) at a range r and depth z from the
source is given by
π
−i
4 U n ( z s )U n ( z ) ik n r
p(r , z ) =
ie
ρ (z s )

8π n kn
e

where ρ(zs) is the density at the source depth zs, kn is the mode propagation
constant for mode n, and Un are normalized eigenvectors of the following
eigenvalue problem,
d 2U n
dz 2
+ K 2
(
( z ) − k n U n (z ) = 0
2
)
.
The eigenvectors Un are zero at z = 0, and satisfy the local boundary conditions at
the ocean bottom.

Você também pode gostar