Академический Документы
Профессиональный Документы
Культура Документы
(1)
where ( ) F is a the radiation pattern at different observation
angles, Q is the number of identical antenna elements in the
array, R
i
is the excitation of the ith element located at d
i
, k is
the wavenumber in the freespace. The objective of the
problem is to synthesize the desired radiation pattern ( ) F
with the minimum number of elements under a small tolerance
error [3].
A. Linear array
Suppose the array is composed of N identical antenna
elements. For linear antenna arrays, the array factor is given
by
( )
cos
1
i
N
jkd
i
i
F Re
=
=
(2)
where R
i
is the excitation coefficient of the ith element located
at x = d
i
, k is the wave number in the freespace. Suppose that
all the antenna elements are symmetrically distributed within a
range of x
s
to x
s
along the x direction as shown in Figure 1.
Fig1. Geometry of nonuniformly spaced linear symmetric array
Considering the symmetric geometry of the array, the
pattern of the array can be written as
( )
[ ] / 2
2 cos( cos )
N
i i
i
F R kd =
(3)
where[ ] / 2 N is the minimum integer no less than / 2 N . In
order to solve the above equation we can assume that all the
antenna elements are equally spaced from x
s
to x
s
with a
small interelement spacing d. Although it is supposed that
there is one element at each position, not each antenna element
is necessarily radiating waves or excited with current. All the
antenna elements can be in two states: on state (when the
element is in the supposed position or has an excitation) or
off state (when there is no element in the supposed position
or without an excitation). Through discretization, (3) can be
written in a matrix form
[ ] [ ] [ ]
1 1 m m n n
F M r
= (4)
where m is the number of sampled antenna radiation pattern, n
is the minimum integer no less than 2 /
s
x d , F is a vector
containing the sampled radiation pattern at different angles,
( ) ( ) ( )
1 2
, ,
T
m
F F F =
F , M is an m n
matrix with the (i, j)-th element 2cos( cos )
ij i j
M kd = .
Choosing m n < , M forms an overcomplete dictionary. r is
the excitation vector, R
i
= 0 is equivalent to the off state
which means the antenna in the ith position is not excited or is
absent from the supposed position.
All of the existing L
2
minimization algorithms for antenna
array synthesis do not exploit a prior sparsity information
about the antenna location space. Although many equally
spaced antennas are supposed to be placed in the array, not
each antenna is required to be excited. The antenna element
without excitation is equivalent to being absent in the
supposed position. Thus we can exploit very useful a priori
information that the antenna location space is sparse. The
spasity of the antenna location space is not well exploited in
current antenna array synthesis methods. In this paper the
problem of antenna array synthesis with minimum number of
antenna elements is cast as sparseness constrained
optimization problem
1
min . . s t < r F Mr (5)
where is a small tolerance error. In the above equation we
seek to find the smallest number of non-zero elements in the
excitation vector r. Comparing with traditional antenna
synthesis problem solved with L
2
minimization, this paper
casts the array synthesis problem as a sparseness constrained
optimization problem and changes the L
2
minimization to L
1
minimization. The L
2
minimization is a quadratic optimization
problem which requires a large computation resource and does
not ensure the minimum number of elements in the array. In
contrast, the sparseness constrained optimization problem
solves a convex, nonquadratic optimization problem and seeks
the sparsest solution to the linear equation, which is also much
more computationally efficient than the global minimization
methods for solving L
2
minimization problems. In this paper
the Bayesian compressive sensing inversion algorithm for
solving the L
1
minimization problem is employed to solve the
antenna array synthesis problem [14]. A brief summarization
of Bayesian compressive sensing algorithm is given in
Appendix A. Empirical studies demonstrate that the Bayesian
learning algorithm has a significantly accelerated rate of
convergence compared to other existing L
1
minimization
algorithms.
B. Planar Array
Fig. 2 Geometry of 2Nx 2Ny elements symmetrical array [4]
3
Planar antenna arrays are currently under study for
spacecraft communication and phased array radar systems due
to the low side lobe level (SLL) and ease of deployment [4, 5].
Here we consider a uniformly spaced 2N
x
2N
y
elements
rectangular planar array with uniform phase excitation shown
in Figure 2. The excitation of the (n, m)-th antenna element,
located at r
nm
= (x
nm
, y
nm
), is denoted as I
nm
. d
x
and d
y
are the
inter-element spacing between the antennas in x and y
directions. Assume that the excitation and inter-element
spacing are symmetric with respect to both x and y axes, the
array factor can be written as [8, 12]
( ) ( ) ( )
1 1
, 4 cos cos
y x
N N
nm nm nm
n m
F u v I kx u ky v
= =
=
(6)
where
sin cos u = , sin sin v = (7)
Through discretization, (6) can be written in following matrix
form
= F MI (8)
Assuming that N
u
and N
v
samples are taken in the u and v
planes and by vectoring
( )
,
i j
F u v , which is sampled at
different angles, into a column vector the kth element of F can
be written as
[ ] ( )
, , ( 1)
i j v
k
F F u v k j i N = = + (9)
I is the excitation vector whose lth element is
[ ] , ( 1)
nm y
l
I I l m n N = = + (10)
The (k, l)-th element of the matrix M can be derived as
[ ] ( ) ( )
4cos cos
( 1) , ( 1)
nm i nm j
kl
y v
M kx u ky v
l m n N k j i N = + = +
=
(11)
F is an N
u
N
v
dimension column vector and M is an (N
u
N
v
)
(N
x
N
y
) matrix. Similar to the linear antenna array
synthesis problem, the sparse planar array synthesis can be
solved as the following sparseness constrained optimization
problem
1
min . . s t < I F MI (12)
Although all elements in the planar array are assumed to be
equally spaced with very small inter-element spacing, not each
element is necessarily radiating EM waves or excited with
current. If the excitation of the lth element is 0 (off state), it
is equivalent to there being no element in the corresponding
location. The sparseness constrained optimization problem in
(12) seeks as few non-zero element as possible in the
excitation vector I, which coincides with the objective of the
antenna array synthesis with minimum number of elements.
III NUMERICAL RESULTS
In this section some numerical results are presented to show
the effectiveness of sparseness constrained optimization for
both linear and planar sparse antenna array synthesis.
A. Synthesis of Chebyshev Pattern with linear array
The first example synthesizes a twenty elements uniformly
spaced Chebyshev array. The side lobe level (SLL) of the
Chebyshev array is SLL = -30dB. If not specified, in all the
following simulations x
s
= 10 and the inter-element spacing
is chosen to be d = /10. Table1 gives the antenna location
and excitation amplitude in [3] and the result obtained by the
sparseness constrained optimization in this paper. Figure 3 is
the reconstructed radiation pattern with sparseness constrained
optimization and compared with the matrix pencil method
result in [3]. It is observed that for cos() between 0.85 and
1.0 the side lobe level is a bit higher than that of the desired
radiation pattern. This is due to the reason that in the
computation we discarded one antenna element with excitation
amplitude smaller than 1/20 of the maximum excitation
amplitude. In order to characterize the difference between the
desired pattern and the synthesized pattern the mean squared
error (MSE) is used in this paper. The MSE is defined as norm
of the error between the desired pattern and the synthesized
pattern, normalized to the norm of the desired pattern. The
MSE in this example is 1.57e-4. From the result we can find
that through sparseness constrained optimization only 14
antenna elements are required to achieve the desired radiation
pattern. The number of antenna elements is reduced by 30% as
compared to the uniformly spaced array and the synthesized
pattern is quite satisfactory for practical use.
Chebyshev
(N=20)
Position/excitation
in [3]
Position/excitation
in this paper
i d
i
/ R
i
d
i
/ R
i
d
i
/ R
i
1 0.25 1.0 0 1 0.4 1
2 0.75 0.97010 0.8206 0.95818 1.2 0.9295
3 1.25 0.91243 1.6381 0.84113 2 0.7951
4 1.75 0.83012 2.4481 0.67176 2.8 0.6143
5 2.25 0.73147 3.2432 0.48115 3.6 0.4172
6 2.75 0.62034 4.0071 0.30046 4.5 0.2409
7 3.25 0.50461 4.7145 0.23345 5.0 0.0804
8 3.75 0.39104
9 4.25 0.28558
10 4.75 0.32561
Table 1 Positions and amplitudes of the reconstructed nonuniform array
elements for the Chebyshev radiation pattern
Fig. 3 20 elements Chebyshev pattern synthesis with 14 non-uniform antenna
elements with sparseness constrained optimization
B. Synthesis of a Taylor-Kaiser Pattern with linear array
The side lobe level (SLL) of the Taylor-Kaiser array to be
synthesized in this example is SLL= -25dB. Using the matrix
pencil method in [3], only 17 elements are required to produce
the desired pattern. Table 2 gives the positions and amplitudes
of the reconstructed nonuniform array element with sparseness
4
constrained optimization. For the convenience of comparison
the result in [3] is also given in the table. Figure 4 is the
reconstructed radiation pattern with sparseness constrained
optimization and the results obtained in [3]. The MSE between
the desired radiation pattern and the synthesized pattern is
2.09e-4 in this example. It is observed that for cos() between
0.7 and 1.0 the side lobe level is a bit higher than that of the
desired radiation pattern. This is due to the reason that in the
computation process we discard one antenna element whose
excitation amplitude is 1/20 smaller than the maximum
excitation amplitude. From the reconstruction results we can
find that only 18 antenna elements are required to achieve the
desired radiation pattern using sparseness constrained
optimization. It requires about 38% fewer antenna elements
than the uniformly space antenna array. Although one more
element is used than the matrix pencil method in [3], the result
is still very satisfactory.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
-60
-50
-40
-30
-20
-10
0
cos( )
P
a
t
t
e
r
n
(
d
B
)
Fig.2 in [3]
This paper
Fig. 4 Reconstruction of the 29 element Taylor-Kaise arrays with 18 antenna
elements with sparseness constrained optimization
Position/
excitation in [1]
Position/excitation
in [3]
Position/excitation
in this paper
i d
i
/ R
i
d
i
/ R
i
d
i
/ R
i
1 0 1 0 1 0.4 1
2 0.5 0.99328 0.8831 0.97859 1.2 0.96431
3 1 0.97329 1.7652 0.91634 2.0 0.89689
4 1.5 0.94063 2.6451 0.81903 2.8 0.8045
5 2 0.89622 3.5211 0.69547 3.6 0.69429
6 2.5 0.84132 4.3905 0.55651 4.4 0.57146
7 3 0.77748 5.2485 0.4137 5.2 0.43998
8 3.5 0.70645 6.0842 0.27782 6.0 0.30694
9 4 0.63017 6.8 0.18562
10 4.5 0.55065
11 5 0.46994
12 5.5 0.39004
13 6 0.31282
14 6.5 0.24001
15 7 0.17309
Table 2 Positions and amplitudes of the reconstructed nonuniform array for
the Taylor-Kaise radiation pattern
C. Synthesis of Electrically Large Linear Array
In order to show the efficiency of sparseness constrained
optimization for solving electrically large scale array synthesis
problem which involves a large number of unknowns in the
optimization, a 300 element Chebyshev array is tested in this
example. In the simulation x
s
= 150 and d = /10, the SLL
of the Chebyshev pattern is -35dB. The synthesized pattern is
presented in Figure 5 and the total number of element used is
224. The MSE between the desired radiation pattern and the
synthesized pattern is 1.85e-4. The computation time of this
example is 7.5s which is much shorter than the global
optimization methods and is similar to the matrix pencil
method in [3]. The reducing of the number of elements in
large antenna arrays and short computation time show that
sparseness constrained optimization is very efficient for sparse
antenna array synthesis.
Fig.5 Reconstruction of the 300 element Chebyshev array with 224 antenna
elements with sparseness constrained optimization
D. Synthesis of square planar array
In the last example we synthesized a square planar antenna
array with sparseness constrained optimization. Planar array
synthesis has more antenna elements, resulting in more
unknowns in the optimization than the linear array. This is a
very challenging problem for the traditional global
optimization methods to synthesize the desired radiation
pattern in a short computation time while maintaining a
minimum number of elements in the array. In this paper, the
1024 element planar array in [4] is synthesized with
sparseness constrained optimization. A total of 2500 samples
of the radiation pattern are sampled at different azimuth and
elevation angles. The antenna locations in the x and y
directions have total array lengths of 10 with an inter-element
spacing d
x
= d
y
= /10, resulting in 101 101 unknowns. For
the large amount of unknowns the planar array synthesis
problem can still be efficiently solved within 10 seconds,
which is much more computationally efficient than other
existing planar array synthesis methods. Figure 7 gives the
synthesized 3D radiation pattern using the sparseness
constrained optimization. The MSE between the 3D desired
radiation pattern and the synthesized pattern is 2.41e-4. For
the convenience of comparison the radiation pattern in the
E-plane is also presented in Figure 8. In the synthesized planar
array 496 elements are required to achieve the desired
radiation pattern. Approximately 50% fewer antenna elements
are required by solving the planar array synthesis as
sparseness constrained optimization problem.
IV. CONCLUSION
In this paper the antenna array synthesis problem is studied
from the new perspective of sparseness constrained
optimization. Most of the existing antenna array synthesis
algorithms are based on L
2
minimization which does not
exploit any sparse information about the antenna location
5
space. By exploiting the a priori information that the antenna
location space is sparse, the antenna array synthesis problem is
cast as sparseness constrained optimization problem and
solved with Bayesian compressive sensing inversion algorithm.
Numerical examples are presented to show the high efficiency
of both planar and linear arrays synthesis with sparseness
constrained optimization.
Acknowledgement:
The authors would like to thank S. Ji, Y. Xue and L. Carin for
the permission to use the Bayesian Compressive Sensing
Codes (BCS_demo) for solving the problems in this paper.
Fig. 6 3D radiation pattern of the 1024 elements planar array in [4]
Fig. 7 3D synthesized pattern of the 496 elements array with sparseness
constraint optimization
Fig. 8 Reconstructed pattern on the E plan (=0) with sparseness constrained
optimization
APPENDIX A: BAYESIAN COMPRESSIVE SENSING
In the appendix, the optimization algorithm exploited by
Bayesian compressive sensing algorithm is briefly
summarized [14, 16, 17].
Assume that the equation relating the measurement
N
R b and the unknowns
M
R x is n b Ax + = ,where
( ) I n n
2
, 0 | ~ N is i.i.d Gaussian measurement noise,
2
is
the variance. From Bayes rule, one can get the posterior
distribution of x
( ) ( ) ( ) x x b b x Pr | Pr | Pr (A.1)
It is noted that Pr (x) describes the prior distribution of
unknown x. A sparseness-promoting prior used in the
Bayesian compressive sensing is the so-called Laplace density
function, in particular,
( ) ( )
1
exp
2
Pr x x
\
|
=
N
(A.2)
where is the reduced hyperparameters. Now, one can get the
estimation of unknown x via maximizing its posterior
distribution, in particular,
( )
1
2
2
min arg x b Ax x
x
+ =
(A. 3)
which is the standard sparsity-promoted linear inverse
problem via L
1
-norm.
It is noted that, since the distribution (A.2) is not conjugated
to the conditional distribution in (A.1), it does not allow a
tractable Bayesian analysis. Therefore, the hierarchical priors
or Gaussian mixtures are firstly employed to describe the
distribution of
( ) ( )
i i i i
x x , 0 | | Pr N = (A.4)
The specification of the hierarchical prior model can be
completed by defining hyperpriors over
i
x s. To make the
priors on the hyperparameters non-informative (i.e. flat), we
choose
i
to be very small. So in the second stage of the
hierarchy the Gamma distribution are assigned to
i
, in
particular,
( )
|
\
|
=
2
, 1 | | Pr
i i
Gamma (A.5)
Now, the original unknowns x and reduced hyperparameters
and can be iteratively estimated via the following
formulations, in particular, at the (n+1) iteration, one has
( ) ( ) |
\
|
+ =
=
+
N
i
i
n
i
n
n
1
2
1 2
2 1
min arg x b Ax x
x
(A.6)
and
( ) ( )
( )
( )
n n
n
n
x b
, | , , Pr max arg , ,
2
, ,
1
1
2 1
2
+
+
+
=
(A.7)
Additionally, this fast algorithm has been developed in [16]
and [17] by analyzing the properties of the marginal likelihood
function in [14]. This enables a principled and efficient
sequential addition and deletion of candidate basis function to
monotonically maximize the marginal likelihood. We omit the
6
detailed discussion of this fast algorithm and refer the reader
to [16] and [17] for more details.
REFERENCES
[1] B. Kumar, G. Branner, Design of Unequally Spaced Arrays for
Performance Improvement, IEEE Trans. on Antennas Propag., vol. 47, no.3,
pp. 511-523, 1999.
[2] . Teruel and E. Iglesias, Ant Colony Optimization in Thinned Array
Synthesis with Minimum Sidelobe Level, IEEE Antennas and Wireless
Propag. Lett., vol.5, pp. 349- 352 , 2006.
[3] Y. Liu, Z. Nie, Q. Liu, Reducing the number of elements in a linear
antenna array by the matrix pencil method, IEEE Trans. on Antennas
Propag. vol. 56, no. 9, pp. 2955-2962, 2008.
[4] J. Ye and W. Pang, Immune algorithm in array-pattern synthesis with
side lobe reduction, International Conference on Microwave and Millimeter
Wave Technology, vol. 3: 1127-1130, 2008.
[5] B. Kumar, G. Branner, Generalized Analytical Technique for the
Synthesis of Unequally Spaced Arrays with Linear, Planar, Cylindrical or
Spherical Geometry, IEEE Trans. on Antennas Propag., vol. 53, no.2, pp.
621-634, 2005.
[6] H. Unz, Linear arrays with arbitrarily distributed elements, IEEE Trans.
Antennas Propagat., vol.8, pp. 222223, 1960.
[7] K. Chen, X. Yun, Z. He and C. Han, Synthesis of Sparse Planar Arrays
Using Modified Real Genetic Algorithm, IEEE Trans. on Antennas Propag.,
vol. 55, no. 4, pp. 1067-1073, 2007
[8] Y. Liu, Q. Liu, and Z. Nie, Reducing the Number of Elements in the
Synthesis of Shaped-Beam Patterns by the Forward-Backward Matrix Pencil
Method, IEEE Trans. on Antennas Propag., vol. 58, no.2, pp. 604-608, 2010.
[9] Y. Chen, S. Yang, and Z. Nie, The Application of a Modified Differential
Evolution Strategy to Some Array Pattern Synthesis Problems, IEEE Trans.
on Antennas Propag., vol. 56, no.7, pp. 1919-1927, 2008.
[10]C. Balanis, Antenna Theory: Analysis and Design, 3rd ed. New York:
Wiley, 2005.
[11] D. Kurup, M. Himdi, and A. Rydberg, Synthesis of Uniform Amplitude
Unequally Spaced Antenna Arrays Using the Differential Evolution
Algorithm, IEEE Trans. Antennas and Propag., vol. 47, no. 3, pp. 511-513,
1999.
[12] K. Yan and Y. Lu, Sidelobe reduction in array pattern synthesis using
genetic algorithm, IEEE Trans. Antennas and Propag., vol. 45: 11171122,
1997.
[13] L. Li and B. Jafarpour, A sparse Bayesian framework for conditioning
uncertain geologic models to nonlinear flow measurements, Advances in
Water Resources, vol. 33, pp. 1024-1042, 2010.
[14] S. Ji, Y. Xue and L. Carin, Bayesian Compressive Sensing, IEEE
Trans. Signal Processing, vol. 56, no. 6, pp. 2346-2356, 2008.
[15] J. A. Tropp, Greed is good: algorithmic results for sparse
approximation, IEEE Trans. Inform. Theory, vol. 50, pp. 2231-2242, 2004.
[16]A. C. Faul and M. E. Tipping, Analysis of sparse Bayesian learning, in
Advances in Neural Information Processing Systems, pp. 383389, 2002.
[17] M. E. Tipping and A. C. Faul, Fast marginal likelihood maximization
for sparse Bayesian models, in Proc. 9th Int.Workshop Artificial Intelligence
and Statistics, 2003.