IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO.
5, MAY 1998
selected as B (q ) = I222 and
^(q ) = I222 +
00 6 : 0 01 + 0:1 0 02 :
A
0 01 4 :
q
0 0:5
q (5.5)
Notice that the initial guess is not too far from the true values. It is
seen from Tables I and II that the proposed approaches are successful
in removing the bias due to measurement noise that results when
time-domain least-squares method is used. The output error method
yields consistent estimates in noise, but it is known to have false local
minima. This is reflected in the tables where the performance of the
output error method does not improve with increasing record length;
however, the proposed approaches do show this improvement.
VI. CONCLUSIONS
Two new frequency-domain solutions to the linear multivariable
system identification problem given time-domain IO data were con-
sidered. The results of [9] pertaining to SISO models were extended to
MIMO models. The main focus is on a pseudo-maximum likelihood
approach where explicit knowledge or explicit modeling of the noise
model was not required. The PML approach requires a good initial
guess to converge, which, in this correspondence, was provided by an
equation-error approach that yields a closed-form, consistent solution.
Some further properties of the proposed approaches may be found in
[14].
REFERENCES
[1] P. A. Regalia, “An unbiased equation error identifier and reduced-
order approximations,” IEEE Trans. Signal Processing, vol. 42, pp.
1397–1412, June 1994.
[2] R. Pintelon et al., “Parametric identification of transfer functions in the
frequency domain—A survey,” IEEE Trans. Automat. Contr., vol. 39,
pp. 2245–2260, Nov. 1994.
[3] J. K. Tugnait and Y. Ye, “Stochastic system identification with noisy
input-output measurements using polyspectra,” IEEE Trans. Automat.
Contr., vol. 40, pp. 670–683, Apr. 1995.
[4] T. Söderström and P. Stoica, System Identification. London, U.K.:
Prentice-Hall Int., 1989.
[5] L. Ljung, System Identification: Theory for the User. Englewood Cliffs,
NJ: Prentice-Hall, 1987.
[6] T. Söderström and P. Stoica, “On the stability of dynamic models
obtained by least-squares identification,” IEEE Trans. Automat. Contr.,
vol. AC-26, pp. 575–577, Apr. 1981.
[7] D. R. Brillinger, Time Series Data Analysis and Theory. New York:
Holt, Rhinehart & Winston, 1975.
[8] P. A. Regalia, Adaptive IIR Filtering in Signal Processing and Control.
New York: Marcel Dekker, 1995.
[9] J. K. Tugnait, “Identification of stochastic linear systems via spectral
analysis: Reduced-arder approximation, performance analysis and trans-
fer function bias,” in Proc. 35th IEEE Conf. Decision Contr., Kobe,
Japan, Dec. 11–13, 1996, pp. 2042–2047.
[10] J. K. Tugnait, “Stochastic system identification with noisy input using
cumulant statistics,” IEEE Trans. Automat. Contr., vol. 37, pp. 476–485,
Apr. 1992.
[11] P. van Overschee and B. de Moor, Subspace Identification of Linear
Systems. Boston, MA: Kluwer, 1996.
[12] R. J. Serfling, Approximation Theorems of Mathematical Statistics.
New York: Wiley, 1980.
[13] J. K. Tugnait, “Frequency-domain identification of multivariable linear
systems with noisy input,” in Proc. 34th IEEE CDC, New Orleans, LA,
Dec. 13–15, 1995, pp. 1246–1251.
[14] , “Identification of multivariable stochastic linear systems via spec-
tral analysis given time-domain data,” in Proc. 1997 ACC, Albuquerque,
NM, June 4–6, 1997, pp. 622–626.
1053–587X/98$10.00  1998 IEEE
Authorized licensed use limited to: IEEE Xplore. Downloaded on March 30, 2009 at 04:28 from IEEE Xplore. Restrictions apply.
1463
Broadband DOA Estimation Using
Frequency Invariant Beamforming
Darren B. Ward, Zhi Ding, and Rodney A. Kennedy
Abstract— A new method of direction-of-arrival (DOA) estimation
for multiple broadband farfield signals is presented. The technique
uses a beamspace preprocessing structure based on frequency-invariant
beamforming. Specifically, a set of beam-shaping filters focus the received
array data in the time domain, thereby avoiding the need for frequency
decomposition. Hence, the proposed method is conceptually different
from most other broadband DOA estimators, which require frequency
decomposition. Numerical results are presented to demonstrate the use
of the new method and compare it with conventional coherent signal
subspace methods.
I. INTRODUCTION
Several methods of broadband direction-of-arrival (DOA) esti-
mation have been proposed based on the coherent signal subspace
(CSS) approach introduced by Wang and Kaveh [1] (for example,
[2]–[4]). The CSS method is based on decomposing the wideband
data into several nonoverlapping narrowband frequency bins and
finding focusing matrices that transform the data in each bin to
a reference frequency bin. The focused data is then combined to
form a composite covariance matrix. Conventional narrowband DOA
estimators such as MUSIC [5] may then be directly applied by
eigendecomposition of the composite covariance matrix. Incoherent
methods such as [6] find DOA estimates for each frequency bin and
then statistically average these to form a broadband DOA estimate.
CSS methods have been found to have lower SNR resolution and
estimate variance than incoherent methods.
Recently, Lee [4] proposed a CSS method based on beamspace
processing (as opposed to elementspace processing, which is per-
formed directly on the received sensor data). For each frequency
bin, Lee constructs a beamforming matrix chosen such that the
resulting beampatterns are essentially identical for all frequencies.
The beamformers are designed by solving a multiparameter least
squares optimization problem. We will refer to this method as
beamspace CSS (BS-CSS).
In this correspondence, we present a new DOA estimator that
performs broadband focusing using time-domain processing, rather
than frequency decomposition. Specifically, we perform beamspace
processing using frequency-invariant beamformers [7], i.e., beam-
formers whose beampatterns are constant over a wide frequency
band. A set of appropriately designed beam-shaping filters ensure
that the same array manifold is produced for all frequencies within
the design band. Unlike CSS-based methods (including BS-CSS), the
proposed estimator removes the requirement to perform frequency
Manuscript received October 22, 1996; revised September 19, 1997. This
work was supported in part by the Australian Research Council and the
U.S. Army Research Office. Parts of this work have previously appeared
in ICASSP’96. The associate editor coordinating the review of this paper and
approving it for publication was Dr. Yingbo Hua.
D. B. Ward was with the Department of Engineering, Australian National
University, Canberra, Australia. He is now with the Acoustics and Speech
Research Department, Bell Laboratories, Lucent Technologies, Murray Hill,
NJ 07974 USA.
Z. Ding is with the Department of Electrical Engineering, Auburn Univer-
sity, Auburn, AL 36849-5201 USA.
R. A. Kennedy is with the Telecommunications Engineering Group, Aus-
tralian National University, Canberra, Australia.
Publisher Item Identifier S 1053-587X(98)02012-1.
1464 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1998

where rFI () is the frequency-invariant response. See [7], [14] for
details of how to design the FIR filters bn [m]:
Note that throughout this correspondence, we will use bold low-
ercase and uppercase characters to denote vectors and matrices,
respectively. Time series will be indicated by square brackets, e.g.,
b[m], and frequency responses will be denoted by round brackets,
e.g., b(f ):

III. FREQUENCY-INVARIANT BEAMSPACE

PROCESSING FOR DOA ESTIMATION
Having described the structure of the FIB, we will now consider
its application to the problem of broadband DOA estimation.

A. Problem Statement
Consider a linear array of N sensors that are not necessarily
uniformly spaced. Assume thatD < N farfield broadband signals
arrive from directions 2 = [1 ; 1 1 1 ; D ], where d is the direction
to the dth source measured relative to the array axis. The time series
received at the nth sensor is
Fig. 1. Block diagram of general frequency invariant beamformer. D
yn [k] = sd [k 0 n()] + vn [k] (6)
d=1

where sd [k] is the dth source signal, and vn [k] is additive white noise.
decomposition; instead, we exploit an underlying frequency-invariant
The frequency response of yn [k] is given by
beamforming structure (which is based on FIR filtering) to implicitly
perform focusing over a wide frequency band. Furthermore, use of
this underlying structure eliminates the need to perform computation- D
ally intensive numerical optimization procedures in order to design yn (f ) = ej2f ( )
sd (f ) + vn (f ): (7)
the beamformers (as is required by BS-CSS). d=1

Define the N -dimensional vector of stacked array data as

II. FREQUENCY–INVARIANT BEAMFORMING
In this section, we consider the design of a beamformer in which y [k] = [y1 [k]; 1 1 1 ; yN [k]]T (8)
the resulting beampattern is constant within a wide frequency band.
with a frequency response given by
We will refer to this as a frequency-invariant beamformer (FIB).
Several methods of designing a FIB have been proposed [7]–[14]. y (f ) = A(2; f )s(f ) + v (f ) (9)
In general, these methods consist of a filter and sum structure as
shown in Fig. 1, where Hn (f ) is the FIR filter on the nth sensor, where s(f ) = [s1 (f ); 1 1 1 ; sD (f )]T is the D 21 source signal vector,
and Hs (f ) is an optional normalization FIR filter. The sensor filters v (f ) = [v1(f ); 1 1 1 ; vN (f )]T is the N 2 1 additive noise vector, and
perform the role of beam shaping and ensure that the beam shape is A(2 ; f ) = [a(1 ; f ); 1 1 1 ;aa(D ; f )] is the N 2 D source direction
constant as a function of frequency. matrix.
The response of this beamformer to plane waves arriving from an Without loss of generality, we assume that the source signals
angle  (measured relative to the array axis) is and noise have finite (but arbitrarily wide) bandwidth [fL ; fU ]:
N The problem we consider is determining the source directions 2
r(; f ) = Hs (f ) Hn (f )ej2f ( )
(1) from the observed array data vector y [k] over a finite time period
n=1 k = 1; 1 1 1 ; K:
where n () is the propagation delay to the nth sensor. This may
be rewritten as B. Beamspace Formulation
r(; f ) = b (f )a(; f )
H
(2) Assume we apply an FIB to the received array data. The beam-
former output is
where
N M 1 0
b(f ) = Hs (f )[H1 (f ); 1 1 1 ; HN (f )]H (3) z[k] = bn [m]yn[k 0 m] (10)
is the N 2 1 filtering vector, and n=1 m=0

a(; f ) = [ej2f () ; 1 1 1 ; ej2f () ]T (4)
is the N 2 1 array manifold vector.
Assume that each row of b(f ) is an FIR filter with filter coefficients
bn [m]; m = 0; 1 1 1 ; M 0 1: Through proper design of the filters
bn [m], the response of this beamformer may be made approximately
constant with respect to frequency over the design bandwidth, i.e.
r(; f ) = bH (f )a(; f )
 rFI (); 8; 8f 2 [fL ; fU ]
where bn [m] is the set of FIR filter coefficients on the nth sensor; note
that these filters are designed as outlined in the previous section such
that the resulting beamformer spatial response satisfies (5), i.e., the
beamformer spatial response is frequency invariant. The frequency
response of the beamformer output is
z(f ) = bH (f )y(f ): (11)
Assume we now form J (D < J  N ) such beamformers using
(5) J different sets of filtering vectors. Denote the stacked vector of

Authorized licensed use limited to: IEEE Xplore. Downloaded on March 30, 2009 at 04:28 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1998 1465

Fig. 2. Superposition of the three beampatterns calculated at f0 = 0:3:

beamformer responses as A broadband FIBS data covariance matrix can now be formed as
bH1 (f ) f
Rz = Rz (f ) df = Ac (2)RsAc (2)H + Rv
z (f ) = [z1 (f ); 1 1 1 ; zJ (f )]T =
(16)
..
. y (f ) (12) f
bJH (f ) where
()
where bi f is the ith set of beam shaping filter responses. These
f
Rs = Rs (f ) df
beamformers are designed to cover a spatial sector  in which 1 f
(17)

the sources are assumed to lie, and they should have uniformly
is the broadband source covariance matrix, and
low sidelobes to attenuate unwanted out-of-sector sources [4], e.g.,
()
Chebyshev beamformers. We will refer to z f as the frequency
Rv =
f
Rv (f ) df
invariant beamspace (FIBS) data observation vector. Let C f ()= f
(18)
[ () ( )]
b1 f ; 1 1 1 ; bJ f be the N 2 J beamforming matrix. This ma-
()
trix transforms the received elementspace data, y f ; to a reduced is the broadband FIBS noise covariance matrix.
dimension beamspace z f : () The broadband FIBS data covariance matrix (16) is now in a form
Using (9), the FIBS data vector can be rewritten as in which conventional eigen-based DOA estimators may be applied.
Denote the eigendecomposition of Rz ; R v as R z E ( 3Rv E , ) =
z (f ) = C H (f )A(2; f )s(f ) + C H (f )v (f ) (13) where
= Ac (2; f )s(f ) + v c (f ) (14) 3 diagonal matrix of sorted eigenvalues;
=[ ]
E E s jE v corresponding eigenvectors;
(2 ) =
where Ac ; f C H f A ; f is the J 2 D FIBS source
( ) (2 ) Es eigenvectors corresponding to the largest D
direction matrix, and v c f C H f v f is the J 2 FIBS noise
( )= ( ) ( ) 1 eigenvalues;
vector. Ev eigenvectors corresponding to the smallest J 0 D
Because the beamformers are designed to satisfy the frequency- eigenvalues.
invariant property (5), the FIBS source direction matrix is approx- The ranges of E s and E v are referred to as the signal and noise
imately constant for all frequencies within the design band, i.e., subspaces, respectively. Specifically, for the MUSIC algorithm [5],
(2 ) (2) [ ]
Ac ; f  Ac ; 8f 2 fL ; fU : Hence, the broadband source the source directions are given by the D peak positions of the
directions are completely characterized by a single beamspace source following FIBS-MUSIC spatial spectrum:
direction matrix Ac(2) : H
Assuming the source signals and the noise are uncorrelated, the 8() = a (a)cH(E) EacH(a) () (19)
FIBS data covariance matrix is c v v c
Rz (f ) = E fz (f )z H (f )g where ac () are the FIBS source location vectors.
= Ac (2)Rs (f )Ac (2)H + Rv (f ) (15)
C. Approximation of Broadband Covariance Matrices
where R s f E fs f s f H g is the D
()= ()() 2 D source covariance Consider approximation of the broadband FIBS noise covariance
()= ()
matrix, and R v f C H f E fv f v f ( ) ( )H gC (f ) is the J 2 J matrix (18). Assuming the element space noise covariance matrix, i.e.,
FIBS noise covariance matrix. E fv f v f H g, is known, then the broadband FIBS noise covariance
()()

Authorized licensed use limited to: IEEE Xplore. Downloaded on March 30, 2009 at 04:28 from IEEE Xplore. Restrictions apply.
1466 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1998

(a)

(b)
Fig. 3. Beampatterns for the center beamformer calculated at 50 frequencies within the normalized band [0.2, 0.4]. (a) Superposition of all 50 beampatterns.
(b) Variation of beampattern with frequency.

matrix (18) is fixed (for a given set of beamformers). Hence, it may filters in the time domain, it is unnecessary to perform frequency
be calculated off line. In the specific case in which the noise is decomposition in order to form the FIBS data covariance matrix.
spectrally white and uncorrelated from sensor to sensor, the FIBS Assuming the received array data contains no frequency components
noise covariance matrix is outside the design band [fL ; fU ],1 we can write the broadband FIBS
Rv = 2V (20)
where V = f H
sf C (f )C (f ) df:
1 Note that because of the structure of the FIB (see Fig. 1), it is straight-
forward to include a bandpass filter on the normalization filter of each
Consider approximation of the broadband FIBS data covariance beamformer, thus ensuring that z (f ) contains no frequency components
matrix (16). Note that since focusing is performed by a set of FIR outside of the design band.

Authorized licensed use limited to: IEEE Xplore. Downloaded on March 30, 2009 at 04:28 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1998 1467

Fig. 4. Comparison of the resolution performance of CSS, BS-CSS and FIBS for several SNR values (100 independent trials per plot point).

Fig. 5. Comparison of the average RMSE’s for CSS, BS-CSS and FIBS for several SNR values (100 independent trials per plot point).

data covariance matrix as where z [k] is the J -dimensional stacked time series of data observed
f =2 at the J beamformer outputs, i.e., it is the inverse Fourier transform
Rz = Rz (f ) df (21)
of (12). Note that each row of z [k] is given by (10). Hence, the
0f =2
f =2 broadband FIBS covariance matrix is formed directly from the ob-
= E fz (f )z H (f )g df (22) served beamspace array data. This represents a major departure from
0f =2
where fs is the sampling frequency. Assuming the data is wide-sense
conventional broadband DOA estimators. In CSS, the broadband data
covariance matrix is formed by performing frequency decomposition
stationary, and using Parseval’s relation, we have
and summing the focused covariance matrix in each frequency bin.
K
R^ z = K1 z [k]z H [k] (23) The proposed method differs in that the broadband data matrix is
k=1 formed by time-domain processing only.

Authorized licensed use limited to: IEEE Xplore. Downloaded on March 30, 2009 at 04:28 from IEEE Xplore. Restrictions apply.
1468
D. FIBS Source Location Vectors
In calculating the FIBS-MUSIC spatial spectrum (19), focused
()
FIBS source location vectors ac  are required. These location
vectors are defined as ac   C f H a ; f ; 8f 2 fL ; fU : In
() () ( ) [ ]
practice, these location vectors can be computed at any frequency in
the design band. Hence, we define the FIBS source location vectors
as ac  C f0 H a ; f0 for some f0 2 fL ; fU : It is important
( )= ( ) ( ) [ ] Fig. 2 were used as the desired beampatterns in BS-CSS.
to note that this is not the same as the focusing frequency used in
CSS methods. In CSS, all the source location vectors at different
frequencies are transformed to the focusing frequency and the data
covariance matrix calculated from the focused data. In our method, all
FIBS source location vectors are virtually identical for all frequencies,
no explicit focusing calculations are performed, and f0 plays a far
lesser role. Thus, any frequency in the design band can be used as the
frequency at which the FIBS-MUSIC spatial spectrum is calculated.
E. Comment on Fully Correlated Source Signals
Consider two broadband source signals, one of which is a delayed
and attenuated version of the other. In this case, the source covariance
()
matrix Rs f will be singular at every frequency. However, as shown
in [1], as long as the time delay between the two signals is nonzero,
the spectral averaging implicit in the calculation of the broadband
source covariance matrix (17) removes this singularity. Hence, just as
CSS is able to resolve fully correlated sources, our proposed method
is also able to resolve fully correlated sources.
F. Summary of Proposed Algorithm
An outline of the proposed broadband DOA estimator is given as
follows:
1) Design J FIB’s that cover the selected spatial region : 1 REFERENCES
2) Calculate the broadband FIBS noise covariance matrix Rv (18).
This can be done once off line (assuming the noise covariance
does not change over the observation time).
3) Collect data from each of the J beamformers over the obser-
vation time period, k =1
; 1 1 1 ; K , and estimate the broadband
FIBS data covariance matrix Rz (23). ^ [2] J. Krolik and D. Swingler, “Focused wide-band array processing by
^
4) Find the source locations from R z and R v using a conventional
narrowband DOA estimator such as MUSIC.
IV. SIMULATIONS
To demonstrate the use of the FIBS-MUSIC algorithm, we consider
an array of 27 uniformly spaced elements, with a spacing of = ,
where  is the wavelength corresponding to the maximum frequency
component of the received signals.
Three FIB’s were designed (using FIR filters with 201 taps)
according to [14] to be frequency invariant over the normalized
frequency band [0.2, 0.4]2 and to cover the spatial sector 
; 
: [80 100 ]
The corresponding beampatterns are shown in Fig. 2 and calculated
at the center frequency f0 = 03
: : The variation of the center
beampattern with frequency is shown in Fig. 3, from which it is seen
that the resulting beampattern is approximately frequency invariant
over the entire band. Similar results were obtained for the other two
beamformers.
A set of simulations was performed to compare the performance
of FIBS with that of conventional element space CSS [1], as well
as BS-CSS [4]. Signals from two wideband uncorrelated sources
arrived at the array from directions 89 and 92 . Spatially white
Gaussian noise was present at each array element. All signals had
flat spectral densities over the design band and zero spectral density
outside the design band. For CSS and BS-CSS, the received data
2 The normalized sampling frequency was 1.
Authorized licensed use limited to: IEEE Xplore. Downloaded on March 30, 2009 at 04:28 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1998
was decomposed into 26 frequency bins using an unwindowed FFT
of length 128, and ten snapshots were used to calculate each DOA
estimate. Thus, a total of 1280 data samples were used for each DOA
estimate, and the same amount of data was also processed by FIBS for
each DOA estimate. Diagonal focusing matrices [1] were calculated
for CSS using a focusing angle of 90.5 . The beampatterns shown in
The probability of resolving the two sources is shown in Fig. 4 for
various SNR levels. Results are based on 100 independent trials for
each SNR, using the same array data for each method.3 The resulting
sample root-mean-squared error (RMSE) of the DOA estimates are
shown in Fig. 5 for various SNR levels. Note that each RMSE value
represents the average of the sample RMSE’s for the two sources.
These results show that the performance of FIBS is comparable with
that of BS-CSS, and that, as expected, the beamspace methods exhibit
better resolution performance than conventional element space CSS.
V. SUMMARY
In this correspondence, we have presented a new method of
broadband DOA estimation. Unlike DOA estimators based on CSS,
no frequency decomposition is required. Instead, a set of beam
shaping filters focus the array data in the time domain. This avoids the
problem of designing a separate focusing matrix for each frequency
bin. In addition, the proposed technique does not require preliminary
DOA estimates and is able to resolve correlated source signals.
Numerical results indicate that the new estimator has comparable per-
formance with the recently proposed BS-CSS method [4]. However,
by eliminating the need for frequency decomposition, the proposed
estimator is computationally less expensive.
[1] H. Wang and M. Kaveh, “Coherent signal-subspace processing for the
detection and estimation of angles of arrival of multiple wide-band
sources,” IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-33,
pp. 823–831, Aug. 1985.
spatial resampling,” IEEE Trans. Acoust., Speech, Signal Processing,
vol. 38, pp. 356–360, Feb. 1990.
[3] B. Friedlander and A. J. Weiss, “Direction finding for wide-band signals
using an interpolated array,” IEEE Trans. Signal Processing, vol. 41,
pp. 1618–1634, Apr. 1993.
[4] T. S. Lee, “Efficient wideband source localization using beamforming
invariance technique,” IEEE Trans. Signal Processing, vol. 42, pp.
2 1376–1387, June 1994.
[5] R. O. Schmidt, “Multiple emitter location and signal parameter estima-
tion,” IEEE Trans. Antennas Propagat., vol. AP-34, pp. 276–280, Mar.
1986.
[6] M. Wax, T. J. Shan, and T. Kailath, “Spatio-temporal spectral anal-
ysis by eigenstructure methods,” IEEE Trans. Acoust., Speech, Signal
Processing, vol. ASSP-32, pp. 817–827, Aug. 1984.
[7] D. B. Ward, R. A. Kennedy, and R. C. Williamson, “Theory and design
of broadband sensor arrays with frequency invariant far-field beam
patterns,” J. Acoust. Soc. Amer., vol. 97, no. 2, pp. 1023–1034, Feb.
1995.
[8] R. Smith, “Constant beamwidth receiving arrays for broad band sonar
systems,” Acoustica, vol. 23, pp. 21–26, 1970.
[9] J. Lardies and J. P. Guilhot, “Realization of a broadband constant
beamwidth end-fire line array,” Acoust. Lett., vol. 10, no. 8, pp. 122–127,
1987.
[10] J. H. Doles, III and F. D. Benedict, “Broad-band array design using the
asymptotic theory of unequally spaced arrays,” IEEE Trans. Antennas
Propagat., vol. 36, pp. 27–33, Jan. 1988.
[11] M. M. Goodwin and G. W. Elko, “Constant beamwidth beamforming,”
in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP),
1993, vol. 1, pp. 169–172.
3 The two sources were considered resolved if the values of the MUSIC
spectrum at 89 and 92 were both above the value at 90.5 .
IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 46, NO. 5, MAY 1998 1469

[12] F. Khalil, J. P. Jullien, and A. Gilloire, “Microphone array for sound
pickup in teleconference systems,” J. Audio Eng. Soc., vol. 42, no. 9,
pp. 691–700, Sept. 1994.
[13] T. Chou, “Frequency-independent beamformer with low response error,”
in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP),
Detroit, MI, May 1995, pp. 2995–2998.
[14] D. B. Ward, R. A. Kennedy, and R. C. Williamson, “FIR filter design for
frequency-invariant beamformers,” IEEE Signal Processing Lett., vol. 3,
pp. 69–71, Mar. 1996.
Maximum Likelihood Joint Channel and
Data Estimation Using Genetic Algorithms
S. Chen and Y. Wu
Abstract— A batch blind equalization scheme is developed based on
maximum likelihood joint channel and data estimation. In this scheme,
the joint maximum likelihood optimization is decomposed into a two-
level optimization loop. A micro genetic algorithm is employed at the
upper level to identify the unknown channel model, and the Viterbi
algorithm is used at the lower level to provide the maximum likelihood
sequence estimation of the transmitted data sequence. As is demonstrated
in simulation, the proposed method is much more accurate compared with
existing algorithms for joint channel and data estimation.
Index Terms—Blind equalization, genetic algorithms, maximum likeli-
hood estimation.
I. INTRODUCTION
This algorithm may be viewed as an “enhanced” Viterbi algorithm
(VA) that retains several surviving sequences and associated channel
estimates for each state of the trellis. The quantized channel algorithm
[11] is a batch procedure that maintains a family of candidate channels
with discrete parameters. Each channel model is used by the VA
to decode data, and the algorithm selects the most likely quantized
channel. In this correspondence, we propose a novel scheme for
joint channel and data estimation using genetic algorithms (GA’s)
[13]–[16].
We develop a two-layer strategy for joint optimization over channel
and data by combining the GA with the VA. At the top layer, an
efficient version of GA known as the micro-GA (GA) [15] searches
the channel parameter space to optimize the ML criterion. The bottom
layer consists of a number of VA units: one for each member of the
channel population. Each VA unit decodes data based on the given
channel model and feeds back the corresponding likelihood metric
value to the GA. The performance of this GA scheme is investigated
in a simulation study. The results obtained clearly demonstrate that
the GA-based scheme has superior performance over other existing
methods for joint channel and data estimation.
II. MAXIMUM LIKELIHOOD BLIND EQUALIZATION
Throughout this study, the channel is modeled as a finite impulse
response filter with an additive noise source [17]. Specifically, the
received signal at sample k is given by
n 01
r(k) = ai s(k 0 i) + e(k) (1)
i=0

Since the pioneering work of Sato [1], three families of blind where
equalization techniques have emerged. The first family of blind na channel length;
adaptive algorithms, which is commonly known as Bussgang al- ai channel taps;
gorithms, constructs a transversal equalizer directly to unravel the e(k) Gaussian white noise with zero mean and variance e2 ;
effects of the channel impulse response [1]–[4]. This class of blind and the symbol sequence fs(k)g is independently identically dis-
equalizers has very low computational complexity but suffers from tributed with a variance s2 . We will assume that the multilevel pulse
the drawback of slow convergence. The second family of blind amplitude modulation (M -PAM) scheme is used. The signal-to-noise
equalization algorithms is based on higher order cumulants [5]–[8]. ratio (SNR) of the system is defined as
This second class of blind equalizers, although very general and n 01
powerful, requires a large amount of received data samples and SNR = s2 a2i =e2 : (2)
extensive computation to estimate higher order cumulants. The third i=0
family of blind adaptive algorithms uses some blind approximations
of the maximum likelihood sequence estimation (MLSE) to perform Joint channel and data estimation can be performed based on the
a joint channel and data estimation [9]–[12]. The resulting blind ML criterion. Let
equalizers are therefore computationally very expensive. A major r = [r(1)r(2) 1 1 1 r(N )]T
advantage of this third approach is that relatively few signal samples
s = [s(0na + 2) 1 1 1 s(0)s(1) 1 1 1 s(N )]T
are required to achieve the equalization objective.
When both the channel and transmitted data sequence are unknown, a = [a0 a1 1 1 1 an 01 ]T (3)
in theory, their optimal estimates can be obtained via the maximum be the vector of N received data samples, the transmitted data
likelihood (ML) optimization over channel and data jointly. The sequence, and the vector of channel taps, respectively. The probability
computational requirement of such a joint optimization procedure density function of the received data vector r conditioned on the
is, however, prohibitively large. In practice, approximations are channel impulse response a and the symbol vector s is
adopted. A straightforward way is to employ a batch iterative process
p(rja; s) =
1
between data decoding and channel estimation [9]. Seshadri [10]
N=2
presented a recursive algorithm for joint channel and data estimation. (2e2 )
N n 01 2
Manuscript received October 22, 1996; revised September 16, 1997. The
associate editor coordinating the review of this paper and approving it for 1 exp 0 21 2 r(k) 0 ai s(k 0 i) : (4)
e k=1 i=0
publication was Dr. Ali H. Sayed.
The joint ML estimate of a and s is obtained by maximizing p(rja; s)
The authors are with the Department of Electrical and Electronic Engineer-
ing, University of Portsmouth, Portsmouth, U.K.
Publisher Item Identifier S 1053-587X(98)03267-X. over a and s jointly. Equivalently, the ML solution is the minimum

1053–587X/98$10.00  1998 IEEE

Authorized licensed use limited to: IEEE Xplore. Downloaded on March 30, 2009 at 04:28 from IEEE Xplore. Restrictions apply.