You are on page 1of 5

NEW RESULTS ON ROBUST ADAPTIVE BEAMSPACE PREPROCESSING

Aboulnasr Hassanien and Sergiy A. Vorobyov


Dept. of Electrical and Computer Engineering
University of Alberta
Edmonton, Alberta, CANADA
vorobyov@ece.ualberta.ca
hassanie@ece.ualberta.ca
ABSTRACT
In this paper, we develop an algorithm for data-adaptive
beamspace preprocessing with robustness against out-of-sector sources. Our algorithm yields an orthogonal beamspace
matrix and, hence, it preserves the white noise property at
the output of the beamspace preprocessor. The beamspace
matrix is designed as a matrix filter that maintains an almost
distortionless response towards sources within the beamspace sector while maximally rejects all out-of-sector sources.
The columns of the beamspace matrix are designed sequentially, one column at a time. This sequential implementation is curried out by imposing orthogonality constraints between beamspace matrix columns. The proposed algorithm
is computationally less expensive as compared to the existing data-adaptive beamspace design techniques. Simulation
results are provided to validate the robustness of the developed algorithm, and show its effectiveness.
1. INTRODUCTION
Direction-of-Arrival (DOA) estimation of weak sources that
are observed in the presence of severe interfernces/jammers
has received much interest in the literature [1][5]. It is
known that the performance of conventional array processing techniques degrades in the case when signals with low
signal-to-noise ratio (SNR) are mixed with signals with high
SNR [1], [3]. Processing by means of using beamspace has
emerged as an efficient approach commonly used in array
processing to reduce the computational burden and improve
the performance of high-resolution direction finding techniques [4][11]. However, conventional beamspace preprocessing techniques lack robustness in the presence of strong
sources that are located outside of the beamspace angular
sectors-of-interest. This motivated the authors of [4] to de This work is supported in parts by Natural Sciences and Engineering Council (NSERC) of Canada and Alberta Ingenuity Foundation, Alberta, Canada. S.A. Vorobyov is on leave from the Joint Research Institute,
Heriot-Watt and Edinburgh Universities, UK.

978-1-4244-2241-8/08/$25.00 2008 IEEE

velop several formulations to the design of beamspace preprocessing with robustness against out-of-sector sources. In
particular, the data-adaptive beamspace preprocessing algorithm has been developed for the case of weak signal
sources observed in the background of strong interference.
The essence of the adaptive beamspace approach to the
design of beamspace preprocessor is to include additional
adaptive out-of-sector interference cancelation features. Therefore, it potentially provides better performance than the
off-line beamspace techniques. In [4], the M L (L is the
beamspace dimension which is much smaller than the elementspace dimension M ) beamspace matrix B is obtained
by solving an optimization problem that involves a constraint that upper-bounds the Frobenius norm of the difference between the sought matrix B and a quiescent response
beamspace matrix Q. Meanwhile, multi-constraints (that
lower-bound out-of-sector attenuation) are used for sidelobe
control. However, the solution to the aforementioned optimization problem is not always feasible. To guarantee solution feasibility, an ad hoc based pre-selection of the upper
and lower bound parameters must be done. Another disadvantage of the adaptive beamspace technique of [4] is the
distortion that occurs to in-sector sources. Moreover, the
technique of [4] suffers from the problem that the resulting beamspace matrix B is not column orthogonal and ,
therefore, it is required to perform a prewhitening step by
replacing B with (BH B)1/2 B. Unfortunately, this prewhitening step results in rising the beamspace attenuation
in the out-of-sector areas. As a result, the performance of
direction finding techniques deteriorates especially at low
SNR region.
In the present paper, we develop a new data-adaptive
technique to the design of beamspace preprocessing that
produces an orthogonal beamspace matrix. Unlike the adaptive beamspace algorithm of [6], our technique does not
involve the use of a quiescent response matrix in the design process. The proposed technique produces a column
orthogonal beamspace matrix and requires lower computational complexity than the algorithm of [4]. The proposed

315

method designs the beamspace matrix by solving L successive optimization problems. The solution of each optimization problem produces one of the columns of the sought
beamspace matrix.
2. ARRAY SIGNAL MODEL
Consider a linear array of M sensors that receives the signals from D narrowband sources. The M 1 array snapshot
vector can be modeled as [1]
x(t) = A()s(t) + n(t)

(1)

where A() is the MD direction matrix, = [1 , . . . , D ]T


is the D 1 vector of the source DOAs {d }D
d=1 , s(t) is the
D 1 vector of signal waveforms, n(t) is the M 1 vector
of sensor noise, and ()T denotes the transpose. Assuming
that the noise is zero-mean white process with the variance
2 , the M M array covariance matrix can be written as
Rx

, E{x(t)xH (t)}
= A()SAH () + 2 I

(2)

where E{} denotes the statistical expectation operator, ()H


stands for the Hermitian transpose, I is the identity matrix
and
S , E{s(t)sH (t)}
(3)
is the D D source covariance matrix.
The beamspace snapshot vector z(t) of a reduced dimension L1 (L < M ) is defined as a linear transformation
of the original (elementspace) data snapshot x(t) as
z(t) = BH x(t)

(4)

where B is the M L beamspace matrix. It is required that


BH B = I so that the beamspace sensor noise remains spatially white. It is important to preserve this spatially white
property of sensor noise because it facilitates subsequent
processing stages, e.g. DOA estimation. Then, the L L
beamspace array covariance matrix is given by
Rz

where N is the number of snapshots available.


The eigendecomposition of (7) can be written as
z = Es s EH + En n EH
R
s
n

D
diagonal matrix s contains the largest
where the D

(signal-subspace) eigenvalues, the columns of the L D


D
matrix Es are the corresponding eigenvectors, and D
is the number of sources within the sector-of-interest. Simi
diagonal matrix n contains the
larly, the (L D)(L
D)

smallest (noise-subspace) eigenvalues while the L(L D)


matrix En is built from the corresponding eigenvectors. The
subspaces Es and En in (8) are used for DOA estimation in
beamspace using any subspace-based technique, e.g. rootMUSIC.
3. ADAPTIVE BEAMSPACE PREPROCESSING: A
REVIEW
The adaptive beamspace preprocessing approach is introduced in [4] to include additional adaptive out-of-sector interference cancelation features and to provide better performance than off-line beamspace design techniques. The
data-adaptive formulation for the beamspace matrix design
problem aims at minimizing the output power of the transformed data [4]. This power can be written as
E{zH (t)z(t)}

x =
R

N
X

x(t)xH (t).

(10)

t=1

The data-adaptive beamspace matrix design problem is


formulated as [4]
x B}
min tr{BH R
s. t. kB QkF ,
k = 1, . . . , K (11)
kBH a(k )k , k ,

(5)

(6)

() is the new steering vector after the beamspace


where a
transformation. In practice, the exact covariance matrix Rz
is not available and, therefore, its sample estimate is used
N
X
z = 1
z(t)zH (t)
R
N t=1

(9)

From (5), it can be seen that the beamspace transformation


changes the array manifold as
() = BH a()
a

= tr{E{z(t)zH (t)}}
= tr{BH Rx B}

where tr{} denotes the trace of a square matrix. Note


that in practical applications, the true covariance matrix Rx
should be replaced in (11) by the sample matrix

, E{z(t)z (t)}
= BH A()SAH ()B + n2 I.

(8)

(7)

where Q is an optimal beamspace transformation matrix


(which does not take into account any out-of-sector sources),
> 0 is the parameter which bounds the passband distortion of the designed beamspace matrix B with respect to Q,
> 0 is the parameter of the user choice that character comizes the worst acceptable stopband attenuation, and
bines a continuum of all out-of-sector directions (i.e., direc k=
tions lying outside the sectors-of-interest ), k ,
1, . . . , K is the angular grid chosen (uniform or nonuni
form) which properly approximates the stopband region

316

by a finite number K of directions. It is worth noting that


there is a tradeoff between the in-sector (passband) and outof-sector requirements. This tradeoff can be satisfied by
properly choosing the values of and . Choosing to
be small leads to low sidelobe levels which comes at the
price of high signal distortion within the sector-of-interest.
On the other hand if is chosen small to minimize distortion within sector-of-interest, then has to be large which
means high sidelobe levels.
In [4], the matrix Q is computed using the spheroidal
sequences based technique [6] as
B = [u1 , u2 , . . . , uL ]

(12)

where {ul }L
l=1 are L principal eigenvectors of the positive
definite matrix
Z
C,
a()aH () d.
(13)

The optimization problem (11) is convex and can be


easily solved using second-order cone programming. However, performance improvement offered by this formulation
comes at the price of relatively high computational complexity that is stated in [4] to be of O(L3.5 M 3.5 ). Moreover, the solution of (11) produces a transformation matrix
B that is not orthogonal. Hence, a prewhitening step is required where the matrix B(BH B)1/2 is used instead of
the matrix B. This prewhitening step does not guarantee
that k(BH B)1/2 BH a(k )k is less than kBH a(k )k and,
therefore, it may easily result in less out-of sector attenuation.
Another shortcoming of the formulation (11) is that it
distorts the in-sector signals due to the use of the constraint
kB QkF . In fact, larger values of lead to larger
distortion. This distortion can be quite harmful in the case of
weak desired sources observed in the background of strong
interferences.

Additionally, the orthogonality constraint BH B = I is imposed to avoid the need for prewhitening. Hence, the adaptive beamspace design problem can be formulated as the
following optimization problem
x B}
min tr{BH R
B

s. t. BH B = I,
bH
l a(l ) = 1,

Unfortunately, the optimization problem (15) is not convex due to the orthogonality constraint and, therefore, it is
hard to solve in a computationally efficient manner. However, it can be solved sequentially where the columns of the
beamspace matrix B are obtained one at a time. Then, the
lth column of B can be computed by solving the following
optimization problem

min bH
l Rx bl
bl

s. t. bH
l a(l ) = 1
bH
1<i<l
l bi = 0,
H
k = 1, . . . , K (16)
|bl a(k )| , k ,
where > 0 is the parameter of the user choice that characterizes the worst acceptable stopband attenuation. The
optimization problem (16) is convex and can be formulated
as a second-order cone programming problem [12] which
can be efficiently solved using interior point methods [13].
Although the parameter is of the user choice, it should be
chosen carefully to ensure that there is a feasible solution
to the optimization problem (16). The minimum value of
that ensures this feasibility can be obtained by solving the
following optimization problem
min

bl ,

4. PROPOSED ADAPTIVE BEAMSPACE


PREPROCESSING ALGORITHM
In this section we develop a new formulation to the adaptive
beamspace preprocessing problem. Our formulation leads
to an algorithm that avoids the shortcomings of the adaptive
beamspace algorithm of [4]. Specifically, we propose to replace the constraint on the upper-bound of the Frobenius
norm of the difference between B and Q in (11) by another
set of constraints. In other words, we propose to replace the
quadratic constraint kB QkF with the set of linear
distortionless constrains
bH
l a(l ) = 1,

l = 1, . . . , L

(14)

where bl is the lth column of B and {l }L


l=1 are L different angles which properly represent the sectors-of-interest.

l = 1, . . . , L,

k = 1, . . . , K. (15)
kBH a(k )k , k ,

s. t. bH
l a(l ) = 1
1<i<l
bH
l bi = 0,
H
k = 1, . . . , K. (17)
|bl a(k )| , k ,
Note that the optimization problem (17) does not depend on
the array data and, therefore, can be solved off-line whenever the feasibility issue arises. If min denotes the solution
of (17), then using min in (16) would always result in
a feasible solution to that problem.
Note that the computational complexity of solving the
optimization problem (16) is of O(M 3.5 ) [14]. Since (16)
has to be solved L times, the algorithm requires a total computational complexity of O(LM 3.5 ). This computational
load is less than the computational complexity of the algorithm of [4] which is of O(L3.5 M 3.5 ).

317

10

SPHEROIDALSEQUENCES BEAMSPACE ALGORITHM [6]


ADAPTIVE BEAMSPACE ALGORITH [4]
PROPSED BEAMSPACE ALGORITHM

10

SPHEROIDALSEQUENCES BEAMSPACE ALGORITHM [6]


ADAPTIVE BEAMSPACE ALGORITH [4]
PROPSED BEAMSPACE ALGORITHM

10

RMSE (DEGREES)

BEAMSPACE ATTENUATION

20
30

10

40
1

10
50
60
80

60

40

20
0
20
ANGLE (DEGREES)

40

60

80

15

10

10

SNR (dB)

Fig. 1. Beamspace beampattern.

Fig. 2. RMSE versus SNR.

5. SIMULATION RESULTS

poor sidelobe attenuation. The algorithm of [4] has good


sildelobe attenuation and excellent interference rejection capabilities at the price of high signal distortion within sectorof-interest. It is also noted from Fig. 1 that our proposed
algorithm enjoys the advantages of the algorithm of [4] in
stopband regions and the advantages of the algorithm of [6]
in passband regions and, therefore, is superior to both.
The root-MUSIC algorithm is used to compute the DOAs
of the desired signals in beamspace for all methods. The
root-mean-square error (RMSE) versus SNR is shown in
Fig. 2 for all methods tested. RMSEs are averaged over the
signal sources and over 200 independent simulation runs.
Fig. 3 shows the probability of source resolutions where the
signal sources are said to be resolved in the lth run if [9]

A uniform linear array of M = 30 omnidirectional sensors


spaced half a wavelength apart is assumed. The beamspace
dimension L = 3 and the number of snapshots N = 500 are
taken. We assume two signals-of-interest with the DOAs
1 = 0.5 and 2 = 0.5 and two out-of-sector (interfering) sources with DOAs 3 = 30 and 4 = 10 . The
interference-to-noise ratio (INR) is fixed at 30 dB. The desired signals are assumed to impinge on the array within the
sector [5 5 ]. The SeDuMi convex optimization MATLAB toolbox [13] is used to solve the optimization problems (11) and (16). The beamspace attenuation pattern
g() ,

aH ()BBH a()
k
a()k2
=
ka()k2
aH ()a()

(18)

is used to compare between different beamspace methods.


The value min = 0.018, obtained by solving (17), is the
minimum value of that guarantees a feasible solution to
(16). In our simulations, the value = 0.02 is used which
results in 30 dB out-of-sector attenuation. The value =
0.15 is used to solve (11) which also results in 30 dB outof-sector attenuation. By using trial and error method, the
value = 1.0 is found to be the smallest value that makes
the solution to (11) feasible and, therefore, it is used in our
simulations.
The beamspace attenuation pattern (18) is shown Fig. 1
for the spheroidal sequences-based beamspace algorithm [6],
the adaptive beamspace algorithm [4], and the proposed robust adaptive beamspace algorithm at SNR = 0 dB. As
shown in Fig. 1, the spheroidal sequences-based beamspace algorithm has good in-sector response at the price of

2
X

|i (l) i | < |1 2 |

i=1

where i (l) is the DOA estimate of the ith source in the lth
run.
It can be noted from Figs. 2 and 3 that the non-adaptive
spheroidal-sequences based algorithm [6] has totally failed
to resolve the two sources of interest because it is unable to
eliminate strong interference observed in the background. It
can also be noted that there is a substantial improvement in
performance achieved by the proposed method as compared
to the adaptive beamspace technique of [6]. This performance improvement offered by the proposed algorithm is a
result of robust adaptive cancelation of interference while
preserving the signals of interest with minimal distortion
which is achieved by using the proposed orthogonal set of
constraints given in (14).

318

PROBABILITY OF SOURCE RESOLUTION

1.2

[4] A. Hassanien, S. Abd Elkader, A. B. Gershman, and


K. M. Wong, Convex optimization based beam-space
preprocessing with improved robustness against out-ofsector sources, IEEE Trans. Signal Processing, vol. 54,
pp. 15871595, May 2006.

SPHEROIDALSEQUENCES BEAMSPACE ALGORITHM [6]


ADAPTIVE BEAMSPACE ALGORITH [4]
PROPSED BEAMSPACE ALGORITHM

[5] M. Li and Y. LU, Dimension reduction for array processing with robust interference cancellation, IEEE
Trans. Aerospace Elect. Syst., vol. 42, pp. 103112, Jan.
2006.

0.8

0.6

0.4

[6] P. Forster and G. Vezzosi, Application of spheroidal sequences to array processing, in Proc. ICASSP87, Dallas, TX, May 1987, pp. 22682271.

0.2

0
15

10

10

SNR (dB)

Fig. 3. Probability of source resolution versus SNR.


6. CONCLUSIONS
An algorithm for data-adaptive beamspace preprocessing
with robustness against out-of-sector sources has been developed. Our algorithm produces an orthogonal beamspace
matrix and, hence, it preserves the white noise property at
the output of the beamspace preprocessor. The beamspace
matrix is designed as a matrix filter that maintains almost
distortionless response towards sources within the beamspace sector while maximally rejects all out-of-sector sources.
The columns of the beamspace matrix are designed sequentially, one column at a time. The proposed algorithm is
computationally less expensive as compared to previously
proposed data-adaptive beamspace design techniques and
offers better performance.
7. REFERENCES
[1] H. L. Van Trees, Optimum Array Processing.
York: Wiley, 2002.

New

[2] D. J. Rabideau,Hybrid mitigation of distortion in digital arrays, in Proc. International Radar Conference,
Washington , DC, May 2005, pp. 236241.
[3] A. Hassanien, Advanced array processing in the presense of complicated spatio-temporal sources, Ph.D.
dissertation, McMaster University, Hamilton, ON,
Canada, Nov. 2005.

[7] S. Anderson, Optimal dimension reduction for sensor array signal processing, Signal Processing, vol. 30,
pp. 245256, Jan. 1993.
[8] M. D. Zoltowski, G. M. Kautz, and S. D. Silverstein,
Beamspace root-MUSIC, IEEE Trans. Signal Processing, vol. 41, pp. 344364, Jan. 1993.
[9] A. B. Gershman, Direction finding using beamspace
root estimator banks, IEEE Trans. Signal Processing,
vol. 46, pp. 31313135, Nov. 1998.
[10] H. Kawk, E. Yang, and J. Chun, Vector sensor arrays
in DOA estimation for the low angle traching, in Proc.
International Waveform Diversity and Design Conference, Pisa, Italy, June 2007, pp. 183187.
[11] F. Li and H. Liu, Statistical analysis of beam-space
estimation for direction-of-arrivals, IEEE Trans. Signal Processing, vol. 42, pp. 604610, Mar. 1994.
[12] M. Lobo, L. Vandenberghe, S. Boyd, and H. Lebret, Applications of second-order cone programming, Linear Algebra and its Appl., 284, pp. 193228,
1998.
[13] J. F. Sturm, Using SeDuMi 1.02, a MATLAB toolbox
for optimization over symmetric cones, Optim. Meth.
Softw., vol. 1112, pp. 625653, Aug. 1999.
[14] Y. Nesterov and A. Nemirovsky, Interior-point polynomial methods in convex programming, volume 13 of
Studies in Applied Mathematics, SIAM, Philadelphia,
PA, 1994.

319

You might also like