ISSN 1818-4952
© IDOSI Publications, 2010
Department of Electronic Engineering, Muhammad Ali Jinnah University, Islamabad, 44000, Pakistan
Abstract: Smart Antenna systems have become a practical reality after the advent of powerful, low-cost and
digital signal processing components. Smart Antenna applications range from RADAR (Radio Detection and
Ranging), SONAR (Sound Navigation and Ranging) and wireless communications to geophysical and
astrophysical explorations and medical ultrasound imaging. Since early 1950's, many adaptive beam forming
algorithms have been proposed. These algorithms use different criterion to adapt the system for better
performance and steer the beam towards signal of interest. An algorithm with less complexity, low computation
costs, good convergence rate, robust to signal and steering vector errors is usually preferred. This research
work investigates the performance of adaptive beamforming algorithms for different channel models which
include Additive White Gaussian Noise (AWGN), Rician and Rayleigh Channels. The comparison is made on
the performance of beam steering ability and nullifying capability of the algorithms. Sidelobe levels,
computational complexity and cost of online implementation of adaptive beamforming algorithms have also
been considered for the comparison. Robust Kalman beamformer has better performance and steers deep nulls
towards interferers than other adaptive algorithms like Recursive Least Square (RLS), constrained Kalman and
Matrix Inversion Normalized Least Mean Square (MI-NLMS).
Key words: Adaptive beamforming % MI-NLMS % RLS % CMA % Least Square CMA % Kalman Beamformer
Corresponding Author: Dr. Noor Muhammad Khan, Department of Electronic Engineering, Muhammad Ali Jinnah University,
Islamabad Campus, 44000, Pakistan E-mail: noor@ieee.org.
775
World Appl. Sci. J., 11 (7): 775-785, 2010
776
World Appl. Sci. J., 11 (7): 775-785, 2010
E[|e|2] = E[|d|2] – 2wHr + wHRxxw (5) using these estimates, we can calculate optimum weights
for SMI beamformer according to MMSE criteria.
In general, we can find the minimum value by taking
the gradient of the MSE with respect to the weight w smi = R −xx1r (12)
vectors and equating it to zero. Thus the Wiener-Hopf
equation is given as Recursive Least Squares (RLS): The basic approach
taken by RLS is to recursively perform the matrix inversion
Lw(E[|e|2] = 2Rxx w – 2r = 0 (6) as required by the direct calculation approach so that at
no time direct matrix inversion computation is required
w MSE = R −xx1r (7) [11, 4]. We can find µ −1 recursively using standard RLS
R xx
algorithm and array weights using equation
Where
w(k) = w(k – 1) + g(k)[d*(k) – xH(k)w(k – 1)] (13)
Rxx = E[x xH]
r = E[d*. x] Where g(k) is gain vector of standard RLS algorithm.
Adaptive Beamforming Algorithms: This section Constant Modulus Algorithm (CMA) & Least Square
presents the adaptive algorithms and their CMA: The constant modulus algorithm (CMA) is a blind
mathematical analysis for weight update procedure of adaptive algorithm proposed by Goddard et al. [12]. The
beam steering. weight update vector for CMA is
Least Mean Square (LMS): The least mean squares w(k + 1) = w(k) – µe*(k)x(k) (14)
algorithm is a gradient based approach. Monzingo [3]
and Godara [5] gives an excellent fundamental treatment Where e(k) is defined according to CMA algorithm
of this approach. The steepest descent iterative
1
approximation is given as e(k ) = [1 − 2
]y (k )
y (k ) (15)
1
w ( k + 1) = w ( k ) − µ∇ w ( J (w)) (8)
2 When the CMA algorithm converges, it converges to
Where, µ is the step-size parameter, J(w) is the cost the optimal solution, but convergence of this algorithm is
function and L( w ) is the gradient of the performance not guaranteed. The slow convergence of CMA limits
surface. Upon simplifying, we have the final LMS solution the usefulness of the algorithm in dynamic environments
for weight update vector. where the signal must be captured quickly. The least-
squares CMA (LSCMA) [13] is a variation of CMA that
W(k + 1) = w(k) – µe*(k) x (k) (9) uses a direct matrix inversion to overcome slow
convergence. The weights are calculated using equation.
Where e(k) is defined in (3).
w LSCMA = R −xx1rxd (16)
Sample Matrix Inversion (SMI): The Sample Matrix
Inversion Algorithm estimates the array weights by With only difference that here desired signal is
replacing the array correlation matrix (Rxx) with its estimate calculated as d = y .
[4, 5]. we can estimate Rxx and r as y
K
Matrix Inversion-Least Mean Square (MI-NLMS):
µ xx = 1
R
K ∑x(k )x H
(k ) (10) In Matrix Inversion-Normalized Least Mean Square
algorithm [14], the SMI algorithm is utilized to
k =1
K
determine the optimum weight vectors assigned to each
∑d∗ (k )x(k )
1 (11)
r=
K k =1
777
World Appl. Sci. J., 11 (7): 775-785, 2010
µ ( k/ k ) = w
w µ ( k/k − 1) + G ( k )[d ( k ) − x H ( k ) w ( k/k − 1)]
(20)
K ( k/k − 1) x (k )
Fig. 2: Flow Chart for MI-NLMS Beamforming G(k ) = (21)
x H ( k )K ( k/k − 1) x ( k ) + σ 2 ( k )
of the antenna elements of the array instead of
K (k/k − 1)x(k )x H (k )K (k/k − 1)
arbitrary value before calculating the final weight vector. K (k/k ) = K (k/k − 1) −
The weight is calculated only for the first few samples σ 2 (k ) + x H (k )K (k/k − 1) x(k ) (22)
or for a small block of incoming data. The weight
coefficients derived by SMI algorithm are set as initial Constrained Kalman Beamforming: Y.H. Chen and C.T.
coefficients and are updated by introducing NLMS Chiang [16] provided an improved version of kalman type
algorithm. The flowchart of the MI-NLMS algorithm is beamforming. They presented a constraint on the array
shown in Figure 2. response along the look direction to the Kalman filter.
The weight vector of the constrained Kalman beamformer
Kalman Filtering: The adaptive control problems was derived and shown to converge to that of the
presented by small communications and data collection minimum-variance distortionless-response (MVDR)
arrays are relatively simple and are often adequately beam-former Their proposed algorithm had better
handled by means of adaptive control algorithms based convergence rate and nulling capability than Baird
on gradient methods. For more complex problems such as kalman beamforming. They constrained the gain of the
command and control of remote vehicles or rapid angular system as unity on the desired signal from look direction
tracking in radar systems, however, more sophisticated 2 and minimize the system output power subject to
processing is often required. For such demanding constraint,
applications adaptive processing methods based on
Kalman filtering have been proposed which is a state Chw(k) = 1 (23)
space approach [3]. Here we have discussed three
variants of kalman filtering in our comparison. Where the steering vector is defined as
778
World Appl. Sci. J., 11 (7): 775-785, 2010
µ (k ) = w
w µ ( k − 1) − G ( k )[ Y − B H ( k ) w
µ ( k − 1)] (29) µ (k ) = w
w µ ( k − 1) + G ( k )[ z − z$ ( k | k − 1)] (37)
Where Where
779
World Appl. Sci. J., 11 (7): 775-785, 2010
Where R is noise covariance matrix represented as spaced half a wavelength apart stationary, Rayleigh and
Rician. The desired signal is assumed to impinge from
σ 2 0
R= 1 (42) 2 = 3° with SNR = 0 dB and two interferers are assumed to
0 σ 22 impinge from 30° and 50°. The kalman algorithms are
initialized as in [16, 17]. We have considered four cases
and finally the updated weight vector covariance can be for interferers. In case-1, we have assumed that both
expressed as interferers have INR = -20 dB, in case-2 both interferers
have INR = 30 dB and in 3rd and 4th case rayleigh and
P(k k) = P(k k – 1) – G(k)S(k)GH(k) (43) rician fading has been introduced respectively with
INR = -20 dB. For interferers having INR = -20 dB, beam
Hw(k,w(k)) and H (2) are Jacobian and Hessian matrices pattern and ampliude pattern of all beamforming
ww
of h(w(k)) are given as algorithms are shown in Figure 3 and Figure 4
respectively. It can be seen that all algorithms steer the
main beam towards desired signal accurately but a
x H (k )
H w ( k , w (k )) = (44) difference can be seen in steering the nulls toward the
ε 2 w H ( k ) − (aa H w ( k )) H + a H interferers. LMS has one null at 30° and one sidelobe at
50° towards the interferers with powers -22.5 dB and -
(45) 19.88 dB respectively. The closest null towards the
ww = ε I − aa
H(2) 2 H
second interferer is at 58°. SMI has the null depth of -
Simualtions: In this section, we present the comparison 48.75 dB towards first interferer and -17 dB towards the
of the adaptive beamforming algorithms discussed in second interferer. The closest null towards the second
this paper. The performance of beamforming algorithms interferer is at 56°. RLS steers -22.50 dB towards the 1st
depends upon ability to steer beam in desired direction, interferer and -22.00 dB towards the other interferer.
level of sidelobes and ability to steer minimum power The closest nulls towards the both interferers are at
towards interferers. Three type of environments are 33° and 54°. CMA has the closest nulls at 22° and 40° and
assumed with a uniform linear array (ULA) of ten sensors side lobes of -8 dB and -7 dB towards the interferers.
Fig. 3: Case-1: Beam pattern of beamforming algorithms with SNR 0 dB and INR -20 dB
780
World Appl. Sci. J., 11 (7): 775-785, 2010
Fig. 4: Case-1: Amplitude pattern of beamforming algorithms with SNR 0 dB and INR -20 dB
Fig. 5: Case-2: Beam pattern of beamforming algorithms with SNR 0 dB and INR 30 dB
LSCMA successfully steers one null towards one respectively. Constrained Kalman beamformer has
interferer having depth of -48.26dB and one side closest nulls at 28° and 56° and side lobes of -22 dB
lobe of -20.5dB towards other interferer. MI-NLMS and -20 dB towards interferers respectively. Finally
has null depth of -46.9 dB at 30° and of -20dB towards 50°. robust kalman beamformer steers both null successfully
Kalman beamformer has closest nulls at 27° and 58° towards the both interferers having null depth of -43 dB
and side lobes of -21 dB and -19 dB towards interferers, and -52 dB.
781
World Appl. Sci. J., 11 (7): 775-785, 2010
Fig. 6: Case-2: Amplitude pattern of beamforming algorithms with SNR 0 dB and INR 30 dB
For case-2 with INR = 30 dB for both interferers, we case-2, LMS has been performed for 1000 iterations as
have the beam pattern and amplitude pattern of all compared to 100 iterations in case-1. Similarly RLS has
algorithms as shown in Figure 5 and Figure 6. We have been simulated for 200 iterations as compared to the
increased the number of iterations in some of the 50 iterations in case-1, CMA has used 100 iterations in
algorithms to accommodate the strong interferers. In case-2 as compared to 50 iterations in case-1 and LS-CMA
782
World Appl. Sci. J., 11 (7): 775-785, 2010
uses 100 iterations in case-2 as compared to 3 iterations in to 100 times faster than the conventional CMA algorithm
case-1. Rest of algorithms are independent of number of as observed in both cases. RLS shows a better
iterations as they converge with in iterations used in case- performance using 200 samples and successfully steers
1. MI-NLMS is computed using 200 iterations, while nulls towards interferers. RLS shows a better performance
kalman, constrained kalman and robust kalman using 200 samples and successfully steers nulls towards
beamformer has computed using 100 iterations. Table 1 interferers. All variants of the kalman filtering show a
provides a quantitative comparison of the adaptive consistent performance in first two cases. Figure 7 and
beamforming algorithms discussed in this work. Figure 8 shows the amplitude pattern of algorithms over
Comparing the curves in Figure 5 and 6, it is clear that, Rayleigh and Rician fading channel. It is obvious from the
Constant Modulus Algorithm (CMA) is not only figures that in both cases, the performance of all
unstable algorithm but also unable to converge to its algorithms suffer degradation except robust kalman
optimum solution due to the presence of strong algorithm. The convergence rate and side lobe level of the
interferers and its dependence upon the theory of robust kalman beamformer are better than the other
constant modulus no matter what the number of beamformers discussed. Moreover, the computational
iterations is kept. Being unstable and diverged, it complexity of the Robust Kalman is lower as compared to
sometimes takes interferers as the desired signal and other robust algorithms available in the text such as
steer the beam towards them. The main advantage of a Second Order Cone Programming (SOCP) based robust
variant of CMA, the LS-CMA is that it can converge up beamformer [18].
783
World Appl. Sci. J., 11 (7): 775-785, 2010
784
World Appl. Sci. J., 11 (7): 775-785, 2010
16. Chen, Y.H. and C.T. Chiang, 1993. Adaptive 18. Vorobyov, S.A., A. Gershman and Z.Q. Luo, 2003.
beamforming using the constrained Kalman filter. Robust adaptive beamforming using worst-case
IEEE Trans. Antennas Propag., 41(11): 1576-1580. performance optimization: A solution to the signal
17. El-Keyi, A., T. Kirubarajan and A.B. Gershman, 2005. mismatch problem. IEEE Trans. Signal Process.,
Robust Adaptive Beamforming Based on the Kalman 51(2): 313-324.
Filter. IEEE Trans. Signal Process., 53(8): 3032-3041.
785