IEEE SIGNAL PROCESSING LETTERS, VOL. 11, NO.

6, JUNE 2004 557

Adaptive Filtering With Averaging-Based Algorithm

for Feedforward Active Noise Control Systems
Muhammad Tahir Akhtar, Student Member, IEEE, Masahide Abe, Member, IEEE, and
Masayuki Kawamata, Senior Member, IEEE

Abstract—This letter proposes an adaptive filtering with

averaging-based algorithm for active noise control (ANC) systems.
This algorithm uses a similar structure as that of the FxLMS-based
ANC system. The proposed algorithm, called Filtered-x Adaptive
Filtering with Averaging (FxAFA) algorithm, uses averages of
both data and correction terms to find the updated values of the
tap weights of the ANC controller. The computer simulations
are conducted for single-channel feedforward ANC systems. It is
shown that the proposed algorithm gives fast convergence as com-
pared with the FxLMS algorithm and achieves better performance
in the presence of the measurement noise. The comparison with
the FxRLS algorithm shows that the proposed FxAFA algorithm
is a better choice for low computational complexity and stable
performance.
Index Terms—Active noise control, adaptive filters, averaging,
FxLMS algorithm.

I. INTRODUCTION

A CTIVE NOISE CONTROL (ANC) [1] is based on the

simple principle of destructive interference of propagating
acoustic waves. The most popular adaptation algorithm used for
Fig. 1. FxLMS algorithm based single-channel feedforward ANC system.
(a) Schematic and (b) block diagram.
ANC applications is the FxLMS algorithm which is a modified
version of the LMS algorithm [2]. The schematic diagram for a
single-channel feedforward ANC system using the FxLMS al- improved convergence properties, have been proposed, viz.,
gorithm is shown in Fig. 1(a). Here, is primary acoustic 1) lattice-ANC systems [3]; 2) IIR-filter-based LMS algorithms
path between the reference noise source and the error micro- called Filtered-u Recursive LMS (FuRLMS) [4], and filtered-v
phone. The reference noise signal is filtered through and algorithms [5]; 3) RLS-based algorithms called Filtered-x RLS
appears as a primary noise signal at the error microphone. The (FxRLS) [1] and Filtered-x Fast-Transversal-Filter (FxFTF)
objective of the adaptive filter is to generate an appro- [6]; and 4) frequency-domain-ANC systems (see [7] and ref-
priate antinoise signal propagated by the secondary loud- erences there in). There are the following problems with these
speaker. This antinoise signal combines with the primary noise approaches. 1) IIR-based structures have inherent stability
signal to create a zone of silence in the vicinity of the error problems; 2) other approaches mentioned above increase the
microphone. The error microphone measures the residual noise computational burden substantially; and 3) RLS-based ANC
, which is used by for its adaptation to minimize the systems have numerical instability problems. These reasons
sound pressure at error microphone. Here accounts for the make FxLMS still a good choice for ANC applications.
model of the secondary path between the output of Another important issue that often arises in the ANC systems
the controller and the output of the error microphone. The is measurement noise in the reference signal (and error signal as
filtering of the reference signal through is demanded well) present due, for example, to airflow over the microphone
by the fact that the output of the adaptive filter is filtered in the duct. It can be shown [1] that due to the measurement
through [2]. noise in the reference signal, the controller does not converge
The FxLMS algorithm is computationally simple, but its to the optimal solution. Further more, due to the measurement
convergence speed is slow. Different ANC algorithms, with noise in the error signal, the overall convergence speed is de-
graded. Therefore, it is necessary to find an efficient method to
improve the performance of the ANC systems in the presence
Manuscript received July 16, 2003; revised October 3, 2003. This work was
supported by the Monbokagakusho Government of Japan. The associate editor of the measurement noise.
coordinating the review of this manuscript and approving it for publication was The main idea in this letter is to accelerate the convergence
Dr. Darren B. Ward. speed of the FxLMS algorithm, and to improve the performance
The authors are with the Graduate School of Engineering, Tohoku University,
Sendai-shi, 980-8579 Japan (e-mail: [email protected]). in the presence of the measurement noise. The method, we use
Digital Object Identifier 10.1109/LSP.2004.827958 for improving the performance, is averaging. The idea of using
1070-9908/04$20.00 © 2004 IEEE
558 IEEE SIGNAL PROCESSING LETTERS, VOL. 11, NO. 6, JUNE 2004

averaging to accelerate the convergence of the stochastic gra- through the modeling filter , and is
dient algorithms was originally proposed in the stochastic and the reference signal picked by reference microphone. Equation
optimization literature [8], [9]. Later, this idea was extended to (3) shows that the adaptation process is perturbed by an un-
adaptive filters [10], [11]. Recently, averaging-based adaptive desired term . Assuming that is represented
algorithms have been proposed for blind multiuser detections by a FIR filter of tap-weight length , the filtered reference
in Direct Sequence/Code division multiple access (DS/CDMA) signal is obtained as
systems [12], [13].
(4)
In this letter, we explore the realization of an ANC algorithm
based on adaptive filtering with averaging (AFA). The proposed where is the impulse response of the
algorithm is called filtered- AFA, FxAFA, algorithm. The modeling filter and
FxAFA algorithm has slightly increased computational com- is an —sample reference signal vector. Taking the
plexity as compared with the FxLMS algorithm. The computer -transform of (2), we get
simulations are conducted for single-channel feedforward
(5)
ANC systems. The simulation results show that the proposed
FxAFA algorithm provides better performance than the FxLMS When the adaptive filter converges, the residual error is (ide-
algorithm. ally) zero, i.e., , which requires to realize the
The organization of this letter is as follows. In Section II, the optimal transfer function
proposed FxAFA algorithm is explained in connection with the
FxLMS algorithm and some properties of the proposed algo- (6)
rithm are discussed. In Sections III details of the computer ex-
periments are given, and in Section IV concluding remarks are This equation shows that the optimal solution is in-
presented. dependent of the measurement noise associated with the
error microphone. However the optimal solution for ideal case
II. AVERAGING BASED FILTERED- ALGORITHM ( for ) is distorted by the refer-
ence input measurement noise, .
A. Algorithm Development Since we have assumed that the both and are
Fig. 1(b) shows the block diagram of the feedforward ANC white Gaussian noise signals, we can use averaging to remove
system of Fig. 1(a). Here, and are measurement their effects. In [11] two averaging-based adaptive filtering
noise signals associated with the reference and error micro- algorithms are proposed. The first algorithm uses averaging
phones, respectively. We make the following assumptions for in iterates only and in the second algorithm averaging is
and . A.1) They are uncorrelated with each other incorporated with both iterates and observations. Motivated by
and with the reference and error signals as well. A.2) They are the second approach, we incorporate averaging with both the
zero mean white Gaussian noise signals with variances and iterates, , and the correction term (the observation vector),
. , of the FxLMS algorithm and propose
Here, A.1) is evident from the nature of the system [Fig. 1(a)] the following algorithm:
and A.2) comes from the fact that and are produced
(7)
by the turbulent air flow (random in nature) over the micro-
phones [14]. where
In Fig. 1, is obtained offline and kept fixed during the
online operation of ANC. Assuming that is a FIR filter of (8)
tap-weight length , the secondary signal is expressed as
(1)
where is the tap-weight
vector and is an (9)
sample reference signal vector. The residual error signal Here, computing the running average of the data does not put
is given as so much computational burden since averages can be calculated
(2) recursively. For example, (8) can be recursively computed as
where is the primary disturbance signal, (10)
is the reference noise signal, is the
secondary canceling signal , denotes linear convolution, Similarly, the averaged gradient vector (9) can be com-
and and are impulse responses of the primary path puted as
and secondary path , respectively. The FxLMS update
equation for the coefficients of is given as
(3) (11)
where is the step size parameter Equations (1), (4), (7), (10), and (11) are combined to give
, is the reference signal filtered the proposed FxAFA algorithm. We see that the introduction of
AKHTAR et al.: ADAPTIVE FILTERING WITH AVERAGING-BASED ALGORITHM 559

averaging in the FxLMS update equation results in a multistep TABLE I

algorithm (proposed FxAFA algorithm). Hence, an increased COMPUTATIONAL COMPLEXITY COMPARISON BETWEEN FxLMS
AND FxAFA ALGORITHMS
computational burden is expected as discussed later in this
section.

B. Choice of the Parameter

Here, we present some comments on the choice of the param-
eter in the proposed algorithm. For convenience, we rewrite
the proposed algorithm in a compact form

(12)

where is a slowly varying gain parameter. In

adaptive algorithms, it is desirable to have a large step gain at from the gradient vector. The second round of the aver-
the startup for fast convergence. As the time increases, the gain aging attempts to remove the effect of from . From
is desirable to slowly decrease so that misadjustment is small. (6), it is clear that the is affecting the optimal solution in a
The time varying gain indeed exhibits these properties and nonlinear fashion, and averaging cannot eliminate it. Neverthe-
. It is seen that will rapidly decrease less we can expect better results as compared with the standard
the gain parameter, and hence adaptation process may be very FxLMS algorithm (see the simulation results presented in Case
slow. Therefore, one may wish to choose [11]. On con- 2).
trary if close to zero is selected then is very slowly de-
creasing. This is also not desirable for large mismatch. Hence III. COMPUTER SIMULATIONS
is the recommended range for the values of .
In this section, the performance of the proposed FxAFA
C. Comparison of Computational Complexity algorithm is demonstrated using computer simulation. Two
In Table I, the computational complexity analysis, on the parameters are used for performance evaluations. They are
basis of computations required for completion of operations 1) the noise reduction, , achieved at the error microphone,
per iteration, is presented. It is seen that the computational which is defined as ; and
burden of the proposed FxAFA algorithm ( 2) the estimation error of , which is defined as
multiplications/iteration) is greater than that of the FxLMS ,
algorithm ( multiplications/iteration), but it is far where is the optimal value of the tap weight vector ob-
less than that of the FxRLS algorithm ( tained under ideal conditions when there is no measurement
multiplications/iteration [6]). noise, and is the order of the control filter .
The data for acoustical paths is adopted from [1] where both
D. Comments on Convergence Behavior the primary acoustical path and the secondary path
are modeled by IIR filters of order 25 (the data is provided on
1) Ideal Condition : We know that the
a disk included with [1]). Since industrial noise often has sig-
method of steepest descent computes a tap-weight vector that
nificant power in the frequency range between 50–250 Hz [16],
moves down the ensemble-average error-performance surface
simulations are carried with signals having frequency falling in
along a deterministic trajectory that terminates on the Wiener
this range. The sampling frequency of 2 kHz is used. The sec-
solution (it takes infinite number of iterations, , to do so). The
ondary-path model is an FIR filter of order 128, and
LMS algorithm, on the other hand behaves differently because
is identified offline. The ANC controller is also an FIR
of the presence of the gradient noise: Rather than terminating
filter of length 128.
on the Wiener solution, the tap-weight vector computed by the
LMS algorithm executes a random motion around the minimum
point of the error performance surface [15, p. 234]. Furthermore, A. Case 1
by assigning a small value to the step size parameter, the adap- It has been shown in [6] and [17] that FxRLS algorithm
tation is made to progress slowly, and the effects of the gradient becomes numerically unstable toward the higher number of
noise on the tap weights are largely filtered out [15, p. 235]. iterations; hence, it is not included in the simulations presented
In the proposed algorithm, the aim is to have the iterations here. The reference noise is a broadband signal and is a
move to the Wiener Solution reasonably fast. With the averaging sinusoid containing five harmonics (of equal power) with
approach of (9), with the estimates from (12) are allowed the fundamental frequency of 50Hz. First we consider the
to approach the vicinity of the true value faster. At the same ideal case when there is no measurement noise present in the
time, averaging removes the random fluctuations in the gradient system, i.e., . The parameters for the FxLMS
vector and ensures that the iterations move toward the optimal algorithm and the FxAFA algorithm are adjusted for fast and
(Wiener) solution (refer to simulation case study Case 1). stable convergence and (by trial-and-error) are found to be
2) Measurement Noise Condition : In and . The
the presence of the measurement noise signals, and , noise reduction curves for the two algorithms are shown in
the first round of averaging is expected to remove the effect of Fig. 2(a). We see that the proposed FxAFA algorithm achieves
560 IEEE SIGNAL PROCESSING LETTERS, VOL. 11, NO. 6, JUNE 2004

hand, a large step size can be selected and hence fast conver-
gence can be realized.

IV. CONCLUDING REMARKS

This letter proposes a new ANC algorithm based on adaptive
filtering with averaging. The main limitation of the proposed
algorithm is its poor tracking properties, which is due to
the running-length averaging process. This problem can be
Fig. 2. Performance comparison between FxAFA and FxLMS algorithms. overcome, for example, by using weighted averaging with
(a) Noise reduction < versus iteration time n in Case 1. (b) Noise reduction < exponential forgetting factor [12], or by re-initializing the
versus iteration time n in Case 2.
averaging process at regular intervals. It would be interesting
to do theoretical analysis of the proposed algorithm on the
similar lines as done for the FxLMS algorithm in [18]. The
development of an ANC system with online secondary-path
modeling, incorporating adaptive filtering with averaging is a
task for future work.

REFERENCES
[1] S. M. Kuo and D. R. Morgan, Active Noise Control Systems-Algorithms
and DSP Implementation. New York: Wiley, 1996.
[2] B. Widrow and S. D. Stearns, Adaptive Signal Processing. Englewood
Fig. 3. Performance comparison between FxAFA and FxLMS algorithms.
1 ()
Cliffs, NJ: Prentice-Hall, 1985.
(a) Estimation error w n versus iteration time n in Case 2. (b) Residual error
()
[3] Y. C. Park and S. D. Sommerfeldt, “A fast adaptive noise control algo-
signal e n in Case 2. rithm based on lattice structure,” Appl. Acoust., vol. 47, no. 1, pp. 1–25,
1996.
[4] L. J. Eriksson, M. C. Allie, and R. A. Greiner, “The selection and appli-
faster convergence and higher noise reduction than the FxLMS cation of an IIR adaptive filter for use in active sound attenuation,” IEEE
algorithm. This performance difference is in agrement with Trans. Acoust., Speech, Signal Processing, vol. ASSP-35, pp. 433–437,
the comments presented earlier. 1987.
[5] D. H. Crawford and R. W. Stewart, “Adaptive IIR filtered-v algorithms
for active noise control,” J. Acoust. Soc. Amer., vol. 101, no. 4, pp.
B. Case 2 2097–2103, 1997.
[6] M. Bouchard and S. Quednau, “Multichannel recursive-least-squares
In the next experiment, we consider the situation when both algorithms and fast-transversal-filter algorithms for active noise con-
the reference microphone and the error microphone introduce trol and sound reproduction systems,” IEEE Trans. Speech Audio Pro-
measurement noise signals and , respectively, with cessing, vol. 8, pp. 606–618, Sept. 2000.
[7] S. M. Kuo and M. Tahernezhadi, “Frequency-domain periodic active
the variances . In order to have a fair com- noise control and equalization,” IEEE Trans. Speech Audio Processing,
parison, the parameters for the two algorithms are kept same as vol. 5, pp. 348–358, July 1997.
before. The noise reduction, , curves are shown in Fig. 2(b). [8] B. T. Polyak and A. B. Juditsky, “Accelaration of stochastic approxima-
tion by averaging,” SIAM J. Contr. Optim., vol. 30, no. 4, pp. 838–355,
We see that the performance of the two algorithms is degraded as July 1992.
compared with that shown in Fig. 2(a). The curves for , [9] H. J. Kushner and J. Yang, “Stochastic approximation with averaging
the estimation error of , are shown in Fig. 3(a). We see of iterates: Optimal symptotic rate of convergence for general process,”
that for the same step size SIAM J. Contr. Optim., vol. 31, no. 4, pp. 1045–1062, 1993.
[10] G. G. Yin and Y. Zhu, “Averaging procedures in adaptive filtering: An
the for the FxLMS algorithm decreases ini- efficient approach,” IEEE Trans. Automat. Contr., vol. 37, pp. 466–475,
tially very fast, but later it starts increasing and then settles at a Apr. 1992.
value higher than that of the FxAFA algorithm. [11] G. Yin, “Adaptive filtering with averaging,” in Adaptive Control, Fil-
tering, and Signal Processing, K. J. Astrom, G. C. Goodwin, and P. R.
As discussed earlier in Section II, due to measurement-noise Kumar, Eds. New York: Springer-Verlog, 1995.
signals, the performance of the two algorithms is degraded. Nev- [12] V. Krishnamurthy, “Averaged stochastic gradient algorithms for
ertheless the FxAFA algorithm, due to averaging process, gives adaptive blind multiuser detection in DS/CDMA systems,” IEEE Trans.
Commun., vol. 48, pp. 125–134, Jan. 2000.
better performance than the FxLMS algorithm. Since, in both [13] G. G. Yin, V. Krishnamurthy, and C. Ion, “Iterate-averaging sign algo-
algorithms, the estimation error, , is still negative, hence rithms for adaptive filtering with applications to blind multiuser detec-
the overall performance of the ANC system is stable, even at tion,” IEEE Trans. Inform. Theory, vol. 49, pp. 657–671, Mar. 2003.
[14] A. Roure, “Self-adaptive broadband active sound control system,” J.
. Sound Vibr., vol. 101, no. 3, pp. 429–441, 1985.
One way to solve this divergence problem is to use small step [15] S. Haykin, Adative Filter Theory, 4th ed. Englewood Cliffs, NJ: Pren-
size so that adaptation takes place slowly. The step size of the tice-Hall, 2002.
two algorithms is reduced on trial-and-error basis. The resulting [16] K. K. Shyu and C. Y. Chang, “Modified FIR filter with phase compen-
sation technique to feedforward active noise controller design,” IEEE
curves for estimation error are shown in Fig. 3(a). The corre- Trans. Ind. Electron., vol. 47, pp. 444–453, Apr. 2000.
sponding curves for residual error signal, , are shown in [17] M. T. Akhtar, M. Abe, and M. Kawamata, “On use of averaging in
Fig. 3(b). We see that to make FxLMS algorithm converge all FxLMS algorithm for single-channel feedforward ANC systems,” in
Proc. IEEE ISCAS 2003, vol. IV, 2003, pp. 389–392.
the time, a very small step size is needed and hence convergence [18] E. Bjarnason, “Analysis of the filtered-x LMS algorithm,” IEEE Trans.
speed is very slow. In case of FxAFA algorithm, on the other Speech Audio Processing, vol. 3, pp. 504–514, Nov. 1995.