Variable step-size multichannel frequency-domain LMS algorithm for blind identication of nite impulse response systems

M.A. Haque and M.K. Hasan Abstract: The choice of step size is a critical factor in blind identication of single-input multioutput systems using adaptive multichannel frequency-domain least-mean-squares (MCFLMS) algorithm. The proper step size is dependent on the signal power and it inuences the speed of convergence and the nal misalignment error. Applying normalised MCFLMS (NMCFLMS) algorithm, the dependency of the step-size parameter on signal power can be eliminated; however, it is shown that even for a moderate signal-to-noise ratio, the algorithm fails to converge to the eigenvector corresponding to the minimum eigenvalue of the data correlation matrix and hence misconverges to a ctitious solution. A self-adaptive variable step-size MCFLMS (VSS-MCFLMS) algorithm that optimises the performance of the algorithm in each iteration in order to achieve minimum misalignment with the true channel impulse response is proposed. The convergence analysis of the proposed algorithm is performed and it is shown that the VSS-MCFLMS algorithm converges, both in noisefree and noisy conditions, to the eigenvector corresponding to the minimum eigenvalue and is therefore more noise robust when compared with the NMCFLMS. The algorithm also shows the optimal speed of convergence in a noise-free case and the proposed variable step size guarantees the stability of the algorithm. Simulation results are presented to demonstrate that the proposed VSS-MCFLMS algorithm gives lower nal misalignment when compared with the NMCFLMS algorithm without sacricing the speed of convergence and computational efciency.

Introduction

Traditionally, channel identication is done by using a training sequence that is known to both the transmitter and the receiver. Blind channel identication aims at identifying the channel impulse response without using a training signal; instead, it uses only the channel output along with certain a priori statistical information on the input to identify the channel. Both single and multichannel identication schemes are reported in the literature by many researchers. Multichannel identication schemes, however, are increasingly becoming popular due to their suitability in removing the unknown channel effects more effectively than their single channel counterparts. Various techniques reported so far can be categorised into two big groups, adaptive and non-adaptive techniques. Some examples of using adaptive techniques are least-squares (LS) approach [1], recursive LS algorithms and, the least-mean-squares (LMS) algorithm [2]. Among the adaptive ltering algorithms, the LMS algorithm is considered as a benchmark [3]. The main shortcoming of the LMS algorithm, however, is related to the selection of appropriate step size, which greatly inuences the speed, nal misalignment and stability of the algorithm. The idea of using an adaptive step size for non-blind single-channel LMS algorithm has been reported by many researchers for both noise-free [4] and noisy conditions
# The Institution of Engineering and Technology 2007 doi:10.1049/iet-spr:20070022 Paper rst received 20th February and in revised form 25th September 2007 The authors are with the Department of Electrical and Electronic Engineering, Bangladesh University of Engineering and Technology, Dhaka, Bangladesh E-mail: arifulhoque@eee.buet.ac.bd

[5]. Most of the algorithms use a heuristic approach and hence their performance is highly dependent on some inherent parameters. A variable step size for time-domain multichannel LMS (VSS-TDMCLMS) has been reported [6, 7]. The convergence analysis of the variable step-size VSS-TDMCLMS algorithm, however, is not presented. Moreover, the algorithm fails to converge to the desired solution even in moderate signal-to-noise environments [8]. In recent era, adaptive ltering in the frequency domain has attracted a great deal of research interest with a view to reducing the computational complexities of the convolution and correlation operations needed in the time-domain multichannel LMS algorithm. The frequency-domain implementation of the multichannel Newton algorithm known as the normalised multichannel frequency-domain LMS (NMCFLMS) has been suggested as a more efcient and effective method for BCI [9]. The method is effective in a noise-free environment. Its performance, however, deteriorates with noise and shows misconvergence [10]. The convergence analysis of the NMCFLMS algorithm with noise is yet to be reported in the literature. In this paper, we give the convergence analysis of the NMCFLMS algorithm and show that it is very likely that the algorithm fails to converge, in the presence of noise, to the eigenvector corresponding to the minimum eigenvalue of the data correlation matrix. We propose a selfadaptive variable step-size MCFLMS (VSS-MCFLMS) algorithm which optimises the performance of the algorithm in each iteration in order to achieve minimum misalignment with the true channel impulse response. The convergence analysis and the computational complexity of the the proposed algorithm are also discussed. We show that the proposed algorithm ensures the convergence of the algorithm
IET Signal Process., 2007, 1, (4), pp. 182 189

182

Authorized licensed use limited to: Bangladesh Univ of Engineering and Tech. Downloaded on March 11,2010 at 02:15:41 EST from IEEE Xplore. Restrictions apply.

to the eigenvector corresponding to the minimum eigenvalue both in noise-free and noisy conditions. Based on these analytically obtained insights, we argue that the VSS-MCFLMS algorithm is more noise robust when compared with the NMCFLMS algorithm. Using theoretical analysis and numerical example, it is shown that this step size guarantees the stability of the algorithm. The rest of the paper is organised as follows. The problem scenario is presented in Section 2. The MCFLMS algorithm is reviewed in Section 3. Section 4 presents the proposed variable step size for the MCFLMS algorithm. The convergence analysis and stability of the VSS-MCFLMS algorithm are discussed in Section 5. A comparison of the VSS-MCFLMS algorithm with the NMCFLMS algorithm is presented in Section 6. The simulation results are given in Section 7. The paper concludes with some remarks in Section 8. 2 Problem formulation

It can be expressed in matrix notation as yij (m) W 01 C xi (m)W 10 bj (m) L2L 2LL h (4)

where ^ denotes an estimate. The matrix C xi (m) is a circulant matrix with its rst column xi (m), and yij (m) [ yij (mL) yij (mL 1) yij (mL L 1)]T W 01 [0LL I LL ] L2L xi (m) [xi (mL L) xi (mL) xi (mL L 1)]T W 10 [I LL 0LL ]T 2LL bj (m) [bj,0 (m) bj,1 (m) bj,L1 (m)]T h h h h where I denotes an identity matrix and 0 is a matrix of zeros. The frequency-domain block error based on the cross-relation between the ith and jth channel is determined as eij (m) yij (m) yji (m) W 01 [C xi (m)W 10 bj (m) C xj (m)W 10 bi (m)] L2L 2LL h 2LL h (5) Let F LL be the discrete Fourier transform (DFT) matrix of size L L. Then the block error sequence in the frequency domain can be expressed as eij (m) FLL eij (m) W 01 [Dxi (m)W 10 bj (m) Dxj (m)W 10 bi (m)] L2L 2LL h 2LL h (6) where underline denotes the frequency domain and the circulant matrix C xi (m) is decomposed as C xi (m) F 1 Dxi (m)F 2L2L 2L2L (7)

The input output relationship of a single-input multiple-output (SIMO) nite impulse response (FIR) channel is given by ui (n) s(n) hi (n) xi (n) ui (n) vi (n),
L1 X l0

hi,l (n)s(n l)

(1) (2)

i 1, 2, . . . , M

where M is the number of sensors, L the length of the impulse response and s(n), ui (n), xi (n), vi (n) and hi (n) denote respectively, the common source signal, ith channel output, ith channel output corrupted by background noise, observation noise and impulse response of the source to ith sensor. It is assumed that the additive noise on M channels is uncorrelated white random sequence, that is, E{vi (n)vj (n)} 0 for i = j and E{vi (n)vi (n n0 )} 0 for n0 = 0. It is also assumed that vi (n) are uncorrelated with s(n). Using vector notation, (1) can be written as ui (n) hT (n)s(n) i (3)

where hi [hi,0 hi,1 hi,L1 ]T and s(n) [s(n) s(n 1) s(n L 1)]T . A BCI algorithm estimates hi , i 1, 2, . . . , M solely from the observations xi (n), n 1, 2, . . . , N . The identiability conditions [1] commonly stated are: 1. the channel transfer functions do not contain any common zeros; 2. the autocorrelation matrix of the source signal is of full rank. 3 MCFLMS algorithm for BCI

where Dxi (m) is a diagonal matrix whose elements are obtained from the DFT coefcients of the rst column of C xi (m) and W 01 F LL W 01 F 1 L2L L2L 2L2L W 10 F 2L2L W 10 F 1 2LL 2LL LL bj (m) F LLbj (m) h h The frequency-domain cost function Jf (m) using the frequency-domain block error signal eij (m) is dened as Jf (m)
M 1 X X M i1 ji1

For SIMO systems, the outputs are correlated since they are driven by the same input. This correlation can be exploited for the derivation of blind identication algorithms [1]. The derivation of MCFLMS is based on the cross-relation criteria in the frequency domain using the overlap-save method. In this section, we review the MCFLMS algorithm reported in [9]. The linear ltering of a signal can be obtained in block mode using circular convolution. We dene yij (m) as the ltered signal block which is produced by ltering the mth signal block of the ith channel by the estimate of the impulse response of the jth channel.
IET Signal Process., Vol. 1, No. 4, December 2007

eH (m)eij (m) ij

(8)

where H denotes the Hermitian transpose. The MCFLMS algorithm approaches the desired solution by going along the opposite direction of the gradient at each iteration. The update equation of the MCFLMS algorithm is given by bk (m 1) bk (m) mf rJk (m), k 1, 2, . .. , M h h (9)

where mf is the step size and the gradient vector rJk (m) can
183

Authorized licensed use limited to: Bangladesh Univ of Engineering and Tech. Downloaded on March 11,2010 at 02:15:41 EST from IEEE Xplore. Restrictions apply.

be obtained as rJ k (m) @Jf (m) hk @b (m)

M X i1

Setting, @ Jm (m)=@mf (m) 0 gives

mopt (m) f
Di (m)W 01 eik (m) x 2LL (10)

bH (m) (1=a) hH h rJf (m) jjrJf (m)jj2

(16)

W 10 L2L

where denotes complex conjugate and W 10 F LL W 10 F 1 L2L L2L 2L2L W 01 F 2L2L W 01 F 1 2LL 2LL LL W 10 [I LL 0LL ] L2L W 01 2LL [0LL I LL ]
T

Equation (16) is not implementable in practical applications because of the presence of the true channel vector h which is unknown in our case. It has been reported in [6] that for noise-free observations, the gradient vector is approximately orthogonal to the true impulse response vector in the time domain. Since two orthogonal vectors in the time domain remains orthogonal in the frequency domain, we approximate hH rJf (m) 0 (17)

Using (10), the MCFLMS update equation can be expressed as bk (m 1) bk (m) mf W 10 h h L2L W 01 eik (m), 2LL
M X i1

Therefore the optimal step size for MCFLMS algorithm in noise-free case can be obtained from (16) as

Di (m) x

mopt (m) f

bH (m) h rJ (m) jjrJf (m)jj2 f

(18)

k 1, 2, . .. , M

Concatenating the M impulse response vectors into a longer one, we can write the update equation for the MCFLMS algorithm as b 1) b mf rJ f (m) h(m h(m) where b [b1 (m) b2 (m) bM (m)]T h(m) h T hT hT rJf (m) [rJ T (m) rJ T (m) rJ T (m)]T 1 2 M 4 Optimum variable step-size for MCFLMS (11)

In this work, we investigate the effectiveness of the MCFLMS algorithm in (11) but replacing the xed step size mf by mopt (m) both in noise-free and noisy conf ditions. We also give a performance comparison of the proposed VSS-MCFLMS with the NMCFLMS [9] in different noisy environments and show the superiority of our method. Before we do this, the stability and convergence analysis of the proposed VSS-MCFLMS and the algorithmic difference between the two approaches in the presence of noise would be interesting. 5 Convergence analysis of the VSS-MCFLMS algorithm In this section, we give a theoretical analysis of the mean convergence of the VSS-MCFLMS algorithm. In particular, we focus on the mechanism how the adaptive algorithm for BCI converges to the eigenvector corresponding to the minimum eigenvalue both in noise-free and noisy cases. ~ Using the ML ML autocorrelation matrix R(m), the gradient vector rJf (m) can be represented as ^ ~ rJf (m) R(m)h(m) (19)

In this section, we derive an expression for the self-adaptive step size mf (m) of the MCFLMS algorithm in the LS sense. The step size is adapted so that the distance between b 1) and h is minimum at each iteration. We dene h(m the cost function as h(m h(m Jm (m) [h ab 1)]H [h ab 1)] (12)

where a is a constant used to resolve the scaling ambiguity between h and b 1). It is known that the blind LMSh(m type algorithms can only estimate the channel impulse response up to a scaling factor [2]. The MCFLMS update equation for adaptive step-size mf (m) can be written as b 1) b h(m h(m) mf (m)rJf (m) Substituting (13) into (12), we obtain h(m) amf (m)rJf (m)] [h ab h(m) Jm (m) [h ab
H

(13)

~ where R(m) is dened as [9] 2P 3 ~ ~ ~ R21 (m) RM1 (m) i=1 Rii (m) P 6 7 ~ ~ ~ RM2 (m) 7 6 R12 (m) i=2 Rii (m) ~ 7 R(m) 6 6 7 . .. . . . . . 4 5 . P . . . ~ 1M (m) ~ 2M (m) ~ ii (m) R R i=M R Substituting (19) into (13), the update equation of the VSS-MCFLMS algorithm can be written as

amf (m)rJf (m)] Rearranging and simplifying (14), we obtain h(m) 2amf (m)hH rJf (m) Jm (m) hH h 2ahHb a2b Hb h(m) h(m) 2a2 mf (m)b H rJf (m) h(m) a2 m2 (m)jjrJf (m)jj2 f

(14)

b 1) b ~ h(m h(m) mf (m)R(m)b h(m)

(20)

Taking the statistical expectation of (20), we obtain [11]   ^ ^ h(m 1) h(m) m f Rh(m) (21)

(15)

 ^ ^ m f E{mf (m)} and h(m) E{h(m)}, where, ~ E{R(m)} R and we assume statistical independence ~ among mf (m), R(m) and b h(m). The autocorrelation matrix, R, is Hermitian and hence it can be represented as R UL UH (22)

where jj jj is the l2 norm. To obtain the optimal step-size mf (m), the partial derivative of Jm (m) is equated to zero.
184

IET Signal Process., Vol. 1, No. 4, December 2007

Authorized licensed use limited to: Bangladesh Univ of Engineering and Tech. Downloaded on March 11,2010 at 02:15:41 EST from IEEE Xplore. Restrictions apply.

where U is the unitary matrix whose columns are the eigenvectors of R and L is a diagonal matrix with diagonal elements lk , 1 k ML, equal to the eigenvalues of R. Substituting (22) into (21), we obtain o o ^ h (m 1) (I m f L) h (m) where o  ^ ^ h (m) U H h(m) (24) (23)

We see that, in the noise-free case, the VSS-MCFLMS algorithm converges to a scaled version of the eigenvector of the autocorrelation matrix R corresponding to the minimum eigenvalue. From (26) and (28), we nd that the speed of convergence of the MCFLMS algorithm depends on how quickly we can o ^ diminish the hk (m) corresponding to non-zero eigenvalues. With this point of view, we dene a cost function, Jo (m) as Jo (m)
ML 2 X  o ^  hk (m 1) k1 ML i2 X  o h ^  hk (m) 1 m f lk  k1

The set of ML rst-order difference equations in (21) are now decoupled. Therefore the solution of the kth equation can be obtained as o ^ hk (m) Dk (1 m f lk )m u(m), k 1, 2, . . . , ML (25) o ^ where hk (m), k 1, 2, . . . , ML are the components of o ^ h (m), Dk an arbitrary constant, that depends on the initial ^ value of h(m) and u(m) the unit step function. Now using  ^ (24), the channel estimate h(m) can be expressed as  o ^ ^ h(m) U h (m) 2 o 3 ^ h1 (m) 6 7 . 6 7 . 6 7 . 7 6  o ulML 6 h (m) 7 ^ 6 k 7 6 7 . 6 7 . 4 5 . o ^ hML (m)

(29)

Differentiating (29) with respect to m f , we get

ML h i X o @Jo (m) ^ 2 jhk (m)j2 1 m f lk ( lk ) @m f k1

(30)

Equating (30) to zero gives PML o jh k (m)j2 lk m f Pk1 o ML 2 2 k1 jh k (m)j lk (31)

ul1

Now, we verify the similarity between mf (m) obtained in (18) and that of (31). Substituting (19) into (18), we get

ulk

mf (m)

o o ^ ^ ul1 h1 (m) ulk h k (m) o ^ ulML hML (m) (26)

bH (m)R(m)b ~ h h(m) H H b (m)R (m)R(m)b ~ ~ h h(m) h iH bo (m) L(m)bo (m) h h h iH bo (m) L2 (m)bo (m) h h PML bo jh (m)j2 lk (m) Pk1 k ML bo 2 2 k1 jhk (m)j lk (m) (32)

where, ulk is the eigenvector corresponding to the eigenvalue lk . Therefore at any iteration, adaptive estimate of the channel is the weighted combination of all the eigeno ^ vectors with hk (m) as the weighting function for the eigenvector, ulk . o ^ hk (m) in (25) diminishes exponentially to zero except one which corresponds to the minimum eigenvalue lmin 0 in o ^ the noise-free case. Now the nal value of hk (m) in the noise-free case can be obtained as o ^ hk (1) 0, when lk = 0 when lk lmin (27)

Dlmin ,

It shows that although step size is computed from complex vectors, it is a real number. Equations (31) and (32) are similar in form. It reveals that the minimum norm solution of (29) is identical to the LS solution of (12). Therefore the variable step size represented by (18) gives the fastest convergence speed for the MCFLMS algorithm in the noise-free case. We now discuss the stability of the MCFLMS algorithm when the proposed variable step-size is adopted. From (25), we see that the VSS-MCFLMS algorithm will be unstable if for any value of k, j1 m f lk j becomes greater than 1 and hence bo m will rise exponentially. In that case, according hk to (32), mf m becomes (neglecting all other components compared with the rising one)

^ As a result, the nal estimate of h(m) in the noise-free case using (26) and (27) can be expressed as  o ^ ^ h(1) U h (1) 0 6 . 7 . 6 . 7 7 6 6D 7 ulML 6 lmin 7 6 7 6 . 7 4 . 5 . 0 Dlmin ulmin
IET Signal Process., Vol. 1, No. 4, December 2007

mf (m)

jbo (m)j2 lk hk bo (m)j2 l2 jh

k k

1 lk

(33)

ul1

ulmin

(28)

Therefore we nd that the proposed mf (m) is autoregulated, which forces the blowing component to decay as quickly as possible and thus ensures the stability of the MCFLMS algorithm. To verify the above statement, a xed step size, mf , was arbitrarily selected for a random multichannel system with signal-to-noise ratio (SNR) 20 dB that causes the estimated ^ channel coefcients, h(m), to blow and thus results in the divergence of the MCFLMS algorithm. After certain
185

Authorized licensed use limited to: Bangladesh Univ of Engineering and Tech. Downloaded on March 11,2010 at 02:15:41 EST from IEEE Xplore. Restrictions apply.

6.1

Performance analysis of the NMCFLMS

The update equation of the NMCFLMS [9] is expressed as bk (m 1) bk (m) W 10 Pk (m)1 h h L2L where Pk (m)
M X i1,i=k M X i1

Di (m) x (35)

W 01 eik (m), 2LL

k 1, 2, . . . , M

Di (m)Dxi (m) x

In order to differentiate VSS-MCFLMS from NMCFLMS algorithm, we rearrange the update equation of the NMCFLMS algorithm in (35) as follows
Fig. 1 Initially a large mf was arbitrarily selected for a random multichannel system with SNR 20 dB that makes hnorm kh(m)k blow After 500 iterations, our proposed adaptive mopt(m) comes into action f
and forces the MCFLMS algorithm to converge

bk (m 1) bk (m) 2W 10 P k (m)1 W 10 W 10 h h L2L 2LL L2L

M X i1

Di (m)W 01 eij (m) x 2LL

bk (m) 2P k (m)W 10 h L2L iterations, the large xed mf was replaced by the proposed variable step-size mf (m). The results are shown in Fig. 1. ^ The norm of h(m) blows as long as xed mf is present and the corresponding normalised projection misalignment (NPM) shows the divergence of the algorithm. But as soon as it is replaced by the optimal mf (m), the algorithm returns back from divergence and thereafter converges to the desired solution. However, in the presence of noise, none of the eigenvalues (lk ) of the autocorrelation matrix R is zero. o ^ Therefore for a stable system all the hk (m) decay with itero ^ ations. Although, it is clear that all components of h (m) decay exponentially to zero, each component decays at a different rate depending on the value of lk . Using (25), the ratio of the kth component to minimum eigenvalue component can be expressed as  ^ 1 m f lk D hk (m) k : o ^ hlmin (m) Dlmin 1 m f lmin
o

M X i1

Di (m)W 01 eik (m) x 2LL

(36)

where we have used the relation [9]

10 10 W 10 2L2L 0:5I 2L2L W 2LL W L2L

and W 10 P k (m)1 W 10 P k (m) L2L 2LL Therefore the NMCFLMS algorithm can be expressed as b 1) b ~ h(m h(m) 2P(m)R(m)b h(m) where P 1 (m) 6 0 6 P(m) 6 . 4 . . 0 2 3 0 0 P 2 (m) 0 7 7 . 7 .. . . 5 . . . . 0 P M (m) (37)

(38)

!m u(m) (34)

Taking the statistical expectation of (37), we obtain    b 1) b h(m h(m) 2PRb h(m) (39)

Now for any k except one; which corresponds to the minimum eigenvalue, we can write j1 m f lk j , j1 m f lmin j. As a result, (34) diminishes exponentially to zero as m tends to innity. It means after a large number of iterations all the orthogonal components become negligible when compared with the minimum eigenvalue component. Therefore similar to the noise-free case, we can conclude that when the received data are corrupted by noise, the MCFLMS algorithm converges to the eigenvector corresponding to the minimum eigenvalue.

where P E{P(m)} and we assume the statistical indepen h(m). Substituting (22) into (39) dence between P, R and b and premultiplying by U H we obtain o o o ^ ^ ^ h (m 1) h (m) 2U H PU Lh (m) o ^ (I 2Lp L)h (m) where Lp Diag U H PU (40)

(41)

6 VSS-MCFLMS against NMCFLMS: algorithmic difference In this section, we highlight the difference between the proposed VSS-MCFLMS and the NMCFLMS algorithms in terms of (i) nal solution in noisy environments and (ii) computational complexity.
186

and Diag[.] refers to a diagonal matrix with diagonal elements of U H PU. We have found that U H PU is very close to a diagonal matrix. Therefore our approximation in (40) introduces insignicant error. Observing (23) and (40), we can clearly visualise the algorithmic difference between the proposed VSS-MCFLMS and
IET Signal Process., Vol. 1, No. 4, December 2007

Authorized licensed use limited to: Bangladesh Univ of Engineering and Tech. Downloaded on March 11,2010 at 02:15:41 EST from IEEE Xplore. Restrictions apply.

NMCFLMS. We nd that an additional multiplying factor, Lp appears in the NMCFLMS algorithm, which modulates the eigenvalues of the data correlation matrix. From (41), we can derive the analytic expression of the diagonal components of Lp which can be expressed in vector form as 2 6 6 6 6 6 6 6 6 6 6 6 4 u2 p1 u2 p2 u2 pML 11 12 1(ML) u2 p1 u2 p2 u2 pML 21 22 2(ML) . . . 2 2 uk1 p1 uk2 p2 u2 pML k(ML) . . . u2 p1 u2 p2 u2 (ML)1 (ML)2 (ML)(ML) pML 3 7 7 7 7 7 7 7 7 7 7 7 5

where uk1 uk2 uk(ML) are the components of eigenvector ulk with k 1, 2, . . . , ML and p1 , p2 , . . . , pML are the diagonal components of P. Since ulk is a unit norm vector, if p1 , p2 , . . . , pML were equal, the diagonal elements of Lp would be equal. But p1 , p2 , . . . , pML are computed from the received data of different channels and hence they are in general unequal. As a result, the diagonal elements of Lp are also unequal. Therefore the scaling factor of the minimum eigenvalue is usually not the minimum one. In the noise-free case, the minimum eigenvalue of the data correlation matrix is zero. As a result, the minimum value in the resultant eigenvalue prole of the NMCFLMS algorithm remains zero, even after a larger scaling factor is attached to it. But in noisy conditions, the eigenvalues of the data correlation matrix come closer to each other. Therefore it is very likely that the minimum eigenvalue remains no longer minimum in the resultant eigenvalue prole. In that case, the NMCFLMS algorithm will fail to converge to the eigenvector corresponding to the minimum eigenvalue of the data correlation matrix. But for the proposed VSS-MCFLMS algorithm, (23) shows that no such scaling factor appears to modulate the eigenvalue prole and hence it converges to the minimum eigenvalue solution both in noise-free and noisy conditions. To verify the above statement, we present in Fig. 2 the eigenvalue proles of original data correlation matrix and those obtained from the NMCFLMS algorithm for a 5-channel acoustic system with 128 coefcients at SNR 20 dB. The scaling factor Lp is also depicted in the gure. It is seen that though the eigenvalue in position 29 is the minimum in the original correlation matrix, the situation is no longer maintained in the NMCFLMS case. The eigenvalue in position 31 takes the minimum position. As a result, the NMCFLMS algorithm will misconverge completely which can be veried from the adaptive solution in the simulation section.

Fig. 2 Eigenvalue proles of original data correlation matrix and those obtained from the NMCFLMS algorithm for a 5-channel acoustic system
a Eigenvalue prole of the data correlation matrix Minimum eigenvalue is located at position 29 which can be seen from the zoomed result b Scaling factor Lp It is seen that the values are completely arbitrary, which cannot ensure minimum scaling factor for the minimum eigenvalue c Resultant eigenvalue prole of the NMCFLMS algorithm Minimum eigenvalue is now located at position 31 which leads the algorithm to complete misconvergence

where b10 (m) W 10 b h 2LL h(m) rJ 10 (m) f W 10 rJ f (m) 2LL W 10 , L2L we obtain (45) (46) (43) (44)

Premultiplying (43) and (44) by

b h(m) W 10 b10 (m) L2L h rJ f (m) W 10 rJ 10 (m) L2L f W 10 W 10 L2L 2LL

I LL . where we have used the relation Substituting (45) into (18), the expression of the optimal step size can be expressed as

mopt (m) f
We can write

[W 10 b10 (m)]H L2L h W 10 rJf10 (m) L2L 10 jjW L2L rJf2L (m)jj2

(47)

[W 10 ]H W 10 0:5 W 10 W 10 L2L L2L 2LL L2L 0:25 I 2L2L Using (48),

(48)

mopt (m) f

in (47) can be written as [b10 (m)]H h rJf10 (m) jjrJf10 (m)jj2 (49)

mopt (m) f
6.2 Comparison of computational complexity

The proposed VSS-MCFLMS algorithm is implemented in L length for the purpose of convergence analysis and derivation of the optimal step size. However, we can form an equivalent representation of the VSS-MCFLMS in 2L length in order to compare its computational complexity with the NMCFLMS which was originally implemented in 2L length. Premultiplying (11) by W 10 , we obtain 2LL b10 (m 1) b10 (m) mf rJ 10 (m) h h f
IET Signal Process., Vol. 1, No. 4, December 2007

Therefore the optimal step size of the proposed VSS-MCFLMS can be evaluated using the 2L representation of the algorithm given in (42). With this optimal step size, (42) can be written as b10 (m 1) b10 (m) mopt (m)rJ 10 (m) h h f f (50)

(42)

In Fig. 3, we give a comparison of the computational complexities of the proposed VSS-MCFLMS and NMCFLMS algorithms for a 3-channel system with random impulse response. We have used the notations
187

Authorized licensed use limited to: Bangladesh Univ of Engineering and Tech. Downloaded on March 11,2010 at 02:15:41 EST from IEEE Xplore. Restrictions apply.

Fig. 5 NPM of the VSS-MCFLMS and NMCFLMS algorithms for M 5 acoustic channel L 128 coefcients SIMO FIR system at SNR 25 dB Fig. 3 Comparison of the computational complexities of the proposed VSS-MCFLMS and NMCFLMS algorithms for a M 3 channel system with random impulse response at different channel length, L

7.1

Acoustic multichannel system

VSS-MCFLMS(L), VSS-MCFLMS (2L) and NMCFLMS (2L) to indicate the L and 2L length implementation of the VSS-MCFLMS and NMCFLMS algorithms. As shown in the gure, there is a linear increase in oating point operations (FLOPS) with the increase channel length for all three implementations. The VSS-MCFLMS(2L) and the NMCFLMS(2L) show almost identical computational complexities with channel length. However, the VSS-MCFLMS(L) has more computational complexity than the other two implementations.

Simulation results

In this section, we investigate the performance of the proposed VSS-MCFLMS algorithm and present comparative results with the NMCFLMS algorithm [9] for both acoustic and random multichannel systems. The performance index used for measurement of improvement and comparison is the NPM [12] dened as NPM(m) 20 log10 jjY(m)jj dB jjhjj

Y(m) h

^ hT h(m) ^ h(m) ^ jjh(m)jj2

The dimension of the room was taken to be (5 4 3) m. A linear array consisting of M 5 microphones with uniform separation of d 0.2 m was used in the experiment. The rst microphone and source were positioned at (1.0, 1.5, 1.6) m and (2.0, 1.2, 1.6) m, respectively. The positions of the other microphones can be obtained by successively adding d 0.2 m to the y-coordinate of the rst microphone. The impulse responses were generated using the image model reported in [13] for reverberation time T60 0.1 s and then truncated so as to make the length 128. The sampling frequency was 8 kHz. The source signal was Gaussian white noise. Fig. 4 shows NPM against SNR for both the algorithms. It is seen that up to 25 dB, the NMCFLMS algorithm completely fails to give an estimate of the channel impulse response. This happens because the scaling matrix Lp misconverges the algorithm to a spurious solution. At any SNR, VSS-MCFLMS algorithm gives a better estimate when compared with the NMCFLMS algorithm as revealed from this gure. We now provide the NPM prole of the estimated channel for both the algorithms at SNR 25 dB in Fig. 5. In the case of the NMCFLMS algorithm, we see a rapid convergence at initial iterations, but with increased iterations the NPM deteriorates until complete misconvergence. But the proposed VSS-MCFLMS algorithm shows more robustness to noise as it maintains a reasonable NPM level in the steady-state solution.

where jj jj is the l2 norm.

Fig. 4 Noise robustness of the NMCFLMS and the proposed VSS-MCFLMS algorithms for a M 5 channel L 128 coefcients acoustic system at low SNR
188

Fig. 6 NPM of the VSS-MCFLMS and NMCFLMS algorithms for M 5 acoustic channels L 128 coefcients SIMO FIR system at SNR 40 dB
IET Signal Process., Vol. 1, No. 4, December 2007

Authorized licensed use limited to: Bangladesh Univ of Engineering and Tech. Downloaded on March 11,2010 at 02:15:41 EST from IEEE Xplore. Restrictions apply.

Fig. 7 NPM of the VSS-MCFLMS and NMCFLMS algorithms for M 3 channel L 32 random coefcients SIMO system at SNR 20 dB

Fig. 8 NPM of the VSS-MCFLMS and NMCFLMS algorithms for M 3 channels L 32 random coefcients SIMO system at SNR 40 dB

Fig. 6 shows the convergence prole for both the algorithms at SNR 40 dB where the VSS-MCFLMS algorithm converges to NPM 240 dB with no sign of misconvergence. In contrast, the NMCFLMS algorithm rst converges towards NPM 235 dB and then diverges from this solution to a steady NPM of 226 dB. Even at this low-noise environment, VSS-MCFLMS algorithm shows an improvement of almost 215 dB in the NPM. 7.2 Random multichannel system

been demonstrated that the proposed variable step size guarantees the stability of the MCFLMS algorithm. The convergence analysis has revealed that the NMCFLMS algorithm [9], in the presence of noise, is very likely to misconverge to a solution other than the eigenvector corresponding to the minimum eigenvalue of the data correlation matrix. We have also shown that the VSS-MCFLMS algorithm is more noise-robust when compared with the NMCFLMS algorithm as it always converges to the minimum eigenvalue solution without sacricing the speed on convergence and computational efciency. 9 References

We now present blind identication results for a M 3 channel random coefcient impulse response system. The impulse responses were generated using the randn function of MATLAB. The length of each channel impulse response is L 32. The source signal was Gaussian white noise. First, we compare performances of the proposed VSS-MCFLMS and NMCFLMS algorithms for random channel estimation in a moderate signal-to-noise environment. Fig. 7 shows the results for both the algorithms at SNR 20 dB. The proposed VSS-MCFLMS algorithm shows lower nal misalignment compared with the NMCFLMS algorithm without sacricing the speed of convergence. The NPM of NMCFLMS algorithm reaches 220 dB in the initial stage of iterations which is indeed a good estimate of the channel coefcients at this noise level. But after this rapid convergence, it gradually diverges and nally settles between 2 10 and 212 dB. In contrast, the VSS-MCFLMS algorithm reaches 2 20 dB with the same speed of convergence but it shows no sign of divergence from initial convergence. At a relatively high SNR, both the algorithms perform almost identically as shown in Fig. 8. The gure shows the NPM prole at SNR 40 dB. The speed of convergence and nal misalignment show no signicant difference for these two algorithms. 8 Conclusion

In this paper, we have proposed a VSS-MCFLMS algorithm with an application to blind channel identication. An expression for the VSS has been derived in such a way that it minimises the misalignment of the estimated channel vector with the true one at each iteration. It has

1 Xu, G., Liu, H., Tong, L., and Kailath, T.: A least-squares approach to blind channel identication, IEEE Trans. Signal Process., 1995, 43, (12), pp. 2982 2993 2 Huang, Y., and Benesty, J.: Adaptive multi-channel least mean square and newton algorithms for blind channel identication, Signal Process., 2002, 82, (8), pp. 11271138 3 Haykin, S.: Adaptive lter theory (Prentice-Hall, Inc., 1996) 4 Kwong, R., and Johnston, E.: A variable step size lms algorithm, IEEE Trans. Signal Process., 1992, 40, (7), pp. 1633 1642 5 Aboulnasr, T., and Mayyas, K.: A robust variable step-size lms-type algorithm: analysis and simulations, IEEE Trans. Signal Process., 1997, 45, (3), pp. 631 639 6 Chen, J., Huang, Y., and Benesty, J.: Optimal step size of the adaptive multichannel lms algorithm for blind simo identication, IEEE Signal Process. Lett., 2005, 12, (3), pp. 173176 7 Gaubitch, N., Hasan, M., and Naylor, P.: Noise robust adaptive blind channel identication using spectral constraints. Proc. IEEE Int. Conf. Acoust., Speech, Signal Processing, 2006 8 Hasan, M.: Damped variable step size multichannel wiener lms algorithm for blind channel identication with noise. Proc. Communication Systems, Networks and Digital Signal Processing, 2006 9 Huang, Y., and Benesty, J.: A class of frequency-domain adaptive approaches to blind multichannel identication, IEEE Trans. Speech Audio Process., 2003, 51, (1), pp. 11 24 10 Hasan, M., Benesty, J., Naylor, P., and Ward, D.: Improving robustness of blind adaptive multichannel identication algorithms using constraints. Proc. European Signal Processing Conf., 2005 11 Proakis, J., Rader, C., Ling, F., Nikias, C., Moonen, M., and Proudler, I.: Algorithms for statistical signal processing (Pearson Education, Inc., 2002) 12 Morgan, D., Benesty, J., and Sondhi, M.: On the evaluation of estimated impulse responses, IEEE Signal Process. Lett., 1998, 5, (7), pp. 174176 13 Allen, J., and Berkley, D.: Image method for efciently simulating small-room acoustics, J. Acoust. Soc. Amer., 1979, 65, (4), pp. 943950

IET Signal Process., Vol. 1, No. 4, December 2007

189

Authorized licensed use limited to: Bangladesh Univ of Engineering and Tech. Downloaded on March 11,2010 at 02:15:41 EST from IEEE Xplore. Restrictions apply.