Anda di halaman 1dari 4

IEEE SIGNAL PROCESSING LETTERS, VOL. 17, NO.

4, APRIL 2010 327

An Affine Projection Sign Algorithm Robust


Against Impulsive Interferences
Tiange Shao, Yahong Rosa Zheng, Senior Member, IEEE, and Jacob Benesty, Senior Member, IEEE

Abstract—A new affine projection sign algorithm (APSA) is pro- This is why computational efficient methods have also been de-
posed, which is robust against non-Gaussian impulsive interfer- veloped to reduce the computational cost, such as the fast affine
ences and has fast convergence. The conventional affine projection projection (FAP) algorithm [5]. In addition to the drawback of
algorithm (APA) converges fast at a high cost in terms of compu- computational complexity, the APA also suffers performance
tational complexity and it also suffers performance degradation in degradation in non-Gaussian interference due to the nature of
the presence of impulsive interferences. The family of sign algo-
rithms (SAs) stands out due to its low complexity and robustness
the -norm optimization. Interfering signals with heavy-tailed
against impulsive noise. The proposed APSA combines the bene- distributions produce more outliers than Gaussian models and
fits of the APA and SA by updating its weight vector according the -norm minimization criterion is no longer a proper choice.
to the 1 -norm optimization criterion while using multiple pro- Many studies have shown that lower-order norms lead to ro-
jections. The features of the APA and the 1 -norm minimization bustness against impulsive and intensive interference. The least
guarantee the APSA an excellent candidate for combatting impul- mean -norm (LMP) algorithm based on the norm is pro-
sive interference and speeding up the convergence rate for colored posed in [2]. Among all the lower-order algorithms, the family
inputs at a low computational complexity. Simulations in a system of sign algorithms based on the -norm minimization has at-
identification context show that the proposed APSA outperforms tracted more attention due to its considerably low computational
the normalized least-mean-square (NLMS) algorithm, APA, and cost and easy implementation. Only the sign of the error signal
normalized sign algorithm (NSA) in terms of convergence rate and
steady-state error. The robustness of the APSA against impulsive
is involved in the updating process. Many variants of the sign
interference is also demonstrated. algorithm have been developed, including the normalized sign
algorithm (NSA) [6], dual sign algorithm (DSA) [7], and vari-
Index Terms—Adaptive filter, affine projection, sign algorithm. able step-size sign algorithm [8], [9]. The mixed-norm algo-
rithm based on the weighted combination of the and
norms is proposed in [10]. The switched-norm algorithm is pro-
I. INTRODUCTION posed in [11] which switches between and norms. Al-
though the sign algorithms achieve good performance in many

A DAPTIVE filters have been commonly used in various


applications of system identification, such as channel es-
timation, noise cancelation, echo cancelation, image restora-
applications due to their low complexity and robustness against
impulsive noise, they suffer from slow convergence rate, espe-
cially for highly correlated input signals.
tion, and seismic system identification [1]. The most popular We proposes an affine projection sign algorithm (APSA)
adaptive filters are the least-mean-square (LMS) and normal- which updates the weight vector with the -norm optimization
ized LMS (NLMS) algorithms due to their simplicity. How- criterion by using multiple input vectors. The combination of
ever, their major drawbacks are slow convergence and perfor- the benefit of affine projection and -norm minimization
mance degradation with colored input signals or in the presence improves performance on combatting impulsive interference,
of heavy-tailed impulsive interferences [2]. speeding up the convergence rate with colored input signals,
To overcome the deterioration of convergence performance and lowering the computational complexity. The weight adap-
caused by colored input signals, an affine projection algorithm tation of the proposed algorithm does not involve any matrix
(APA), which is based on affine subspace projections, has been inversion but only uses the sign operation of the error vector.
proposed in [3]. Many variants of the APA have been devel- The increase of the computational burden caused by high
oped in recent years [4]. The family of APAs updates the weight projection orders is much lower than the conventional APA.
coefficients by multiple, most recent input vectors instead of a The performance of the proposed APSA is evaluated in the
single, current data vector used in the LMS and NLMS algo- context of system identification and compared with the NSA,
rithms. As the projection order of the APA increases, the conver- APA, and NLMS algorithm. Simulation results with Bernoulli-
gence rate increases and so does the computational complexity. Gaussian (BG) interference and colored input signals demon-
strate the robustness of the APSA against impulsive interfer-
Manuscript received November 01, 2009; revised December 17, 2009. First ence, outperforming the APA and NLMS algorithm. The APSA
published January 12, 2010; current version published February 05, 2010. The also converges much faster and reaches a smaller steady-state
work of T. Shao and Y. R. Zheng was supported by the AFOSR under Grant misalignment than the NSA algorithm.
FA9550-07-1-0336. The associate editor coordinating the review of this manu-
script and approving it for publication was Dr. Ricardo Merched. II. CONVENTIONAL AFFINE PROJECTION ALGORITHM
T. Shao and Y. R. Zheng are with the Department of Electrical and Computer
Engineering, Missouri University of Science and Technology, Rolla, MO 65409 Consider a system identification problem where all signals
USA (e-mail: (ts4g2@mst.edu; zhengyr@mst.edu). are real. The output signal from an unknown system with
J. Benesty is with the INRS-EMT, University of Quebec, Montreal, QC H5A a weight coefficients vector is ,
1K6 Canada (e-mail: benesty@emt.inrs.ca).
Color versions of one or more of the figures in this paper are available online where is the
at http://ieeexplore.ieee.org. input signal vector of length . The variable represents
Digital Object Identifier 10.1109/LSP.2010.2040203 the background noise plus interference signal. The superscript
1070-9908/$26.00 © 2010 IEEE
328 IEEE SIGNAL PROCESSING LETTERS, VOL. 17, NO. 4, APRIL 2010

denotes vector transpose operation. Let be an es- Substituting (11) into the constraint (8), we obtain
timate of at iteration . The a priori error is defined as
, while the a posteriori error is (12)
defined as . Grouping the
recent input vectors together gives the input signal matrix:
. We define the a Substituting (12) into (11), the update equation for the weight
priori and a posteriori error vectors as vector is then:
(1)
(2)
and they can be computed as (13)
Since the a posteriori error vector depends on
(3) which is not accessible before the current update, it is rea-
(4) sonable to approximate it with the a priori error vector
. The minimum disturbance controls the convergence
where is the output vector defined as level of the algorithm and it should be much smaller than
. The classical APA [3] is obtained by one to guarantee convergence. It serves the similar pur-
pose as the step-size parameter in conventional adaptive
algorithms. Following the conventions, we replace by the
step-size parameter . Defining
(5)
with , we
The APA updates the weight coefficients vector as obtain the APSA:

(6) (14)
where the step size and regularization (both are positive
numbers) have been added in (6) for a better control of the al- where represents the regularization parameter which should
gorithm. be a positive number. Since comes from the minimum dis-
turbance constraint , we should choose to ensure
III. AFFINE PROJECTION SIGN ALGORITHM the stability of the algorithm and to achieve a small steady-state
The proposed affine projection sign algorithm is obtained by misalignment.
minimizing the -norm of the a posteriori error vector with a As shown in (14), no matrix inversion is needed for the
constraint on the filter coefficients, proposed APSA and it only requires multiplications at
each iteration for the normalization. In comparison, the com-
(7) putational complexities of the APA and FAP algorithm are
and [5] multiplications respec-
(8) tively, where is the factor associated with the complexity
required in matrix inversion. The proposed APSA is much
where is a parameter ensuring that the updating weight coef- simpler in implementation than the APA and even the FAP.
ficients vector does not change dramatically [11]. We can also Besides, it does not have the numerical problems that the FAP
view (8) as the minimum disturbance constraint. The minimum exhibits. It is also worth mentioning that the APSA with
disturbance controls the convergence level of the algorithm reduces to a new kind of normalized sign algorithm, whose
and it shall be as small as possible. Using the method of La- normalization is based on the Euclidean norm of the input
grange multipliers, the unconstrained cost function can be ob- vector. This is different from the normalized least-mean-abso-
tained by combining (7) and (8), lute deviation (NLMAD) algorithm in [6] which is normalized
by the norm of the input vector.
(9)
where is a Lagrange multiplier. The derivative of the cost ALGORITHM PERFORMANCE
function (9) with respect to the weight vector is The proposed APSA is compared to the NLMS, APA, and
NSA via system identification applications. The adaptive filter
has a length taps. The input signal is chosen to be a
colored Gaussian process. This input is generated by filtering a
white Gaussian noise through a first order system with a pole
at 0.8 or 0.95. An independent white Gaussian noise is added
to the system background with a 30 dB signal-to-noise ratio
(10) (SNR). In addition, a strong interference signal is also added
to the system output with a signal-to-interference ratio
where denotes the sign function and
(SIR) of to 10 dB. The Bernoulli-Gaussian (BG) distribu-
. Setting the derivative tion [11] is used for modeling the interference signal, which is
of equal to zero, we get generated as the product of a Bernoulli process and a Gaussian
process, i.e., , where is a white Gaussian
(11) random sequence with zero mean and variance , and is
SHAO et al.: AFFINE PROJECTION SIGN ALGORITHM 329

Fig. 3. Misalignment comparison of the APSA, NLMS, APA, and NSA. Other
M = 1; 2; 5; 10; 20 parameters are the same as those in Figs. 1 and 2.
 = 0:01
Fig. 1. Misalignment of the APSA with varying values of
and same step size of . The input is an AR(1) with a pole at 0.8. The
SNR = 30 dB
030 dB Pr = 0:001
background noise is Gaussian with . The interference is a BG
with SIR= and .

Fig. 4. Misalignment comparison of the APSA, NLMS, APA, and NSA.


SIR = 010 dB . Other parameters are the same as those in Fig. 3.

 =
0:1; 0:01; 0:0025; 0:001 M = 2
Fig. 2. Misalignment of the APSA with varying step sizes of
and same projection order of . Other
parameters are the same as those in Fig. 1.

a Bernoulli process with the probability mass function given as


for , and for . The
average power of a BG process is . Keeping the average
power constant, a BG process is spikier when is smaller. It
reduces to a Gaussian process when .
The convergence is evaluated by the normalized misalign-
ment [1] defined as .
The ensemble average of 20 trails is used for . The regu-
larization parameter is set to 0.0001 for the APA and 0 for the
APSA.
This work first examines the performance of the APSA with
different projection orders , as shown in Fig. 1, where
, and the interference is a BG with SIR
and is used. The APSA with higher projection order
achieves both faster convergence and lower misalignment for Fig. 5. Misalignment comparison of the APSA, NLMS, APA, and NSA.
to 10. When is larger than a certain value (in this SIR = 10 dB . Other parameters are the same as those in Fig. 3.
case is 10), the convergence is faster with a larger but the
330 IEEE SIGNAL PROCESSING LETTERS, VOL. 17, NO. 4, APRIL 2010

The case of no impulsive interference is shown in Fig. 6,


where only background noise with an SNR of 30 dB is present.
The input is a higher correlated AR(1) process with a pole at
0.95. The APA performs best and the NSA performs worst. The
APSA performs in the middle, but no worse than the conven-
tional NLMS.
Finally, this work studies the effect of the impulsiveness of
the BG interference on the performance of the APSA and NSA,
as shown in Fig. 7. The SIR is fixed at and the values
of are selected to be 0.001, 0.01, and 0.1. The more spikier
the interference, the better the APSA and NSA perform. In all
the interference scenarios, the APSA outperforms the NSA and
the performance gain of the APSA are greater than that of the
NSA.

IV. CONCLUSIONS
Fig. 6. Misalignment comparison of APSA, NLMS, APA, and NSA. No im- This paper has proposed an affine projection sign algorithm
pulsive noise. The input is an AR(1) with a pole at 0.95.
(APSA) that updates its weight vector according to the sign of
the a priori error vector based on the -norm optimization. A
constraint is applied on variations of the weight vector, leading
to normalization based on the correlation matrix of the input
signal. The proposed APSA combines the benefits of the APA
and sign algorithm. The affine projection makes the APSA con-
verge fast with colored input signals while the optimiza-
tion guarantees its robustness against impulsive interference. In
addition, the APSA has much lower computational complexity
than the conventional APA because its adaptation only involves
the sign operation. As a result, a large projection order can be
selected to achieve faster convergence rate with affordable com-
putational cost. Simulations have also confirmed that the pro-
posed APSA algorithm improves the capability of combatting
impulsive interferences and accelerating the convergence rate
with colored input signals. Its performance in Gaussian noise is
also better than that of the conventional sign algorithm.

REFERENCES
[1] S. Haykin, Adaptive Filter Theory, 4th ed. Upper Saddle River, NJ:
Fig. 7. Misalignment comparison of APSA and NSA under a BG interference
0
Prentice-Hall, 2002.
with various levels of impulsiveness. SIR = 10 dB, the values of P r are [2] M. Shao and C. L. Nikias, “Signal processing with fractional lower
chosen as 0.001, 0.01, and 0.1. The input is an AR(1) with a pole at 0.8. order moments: Stable processes and their applications,” Proc. IEEE,
vol. 81, pp. 986–1010, Jul. 1993.
[3] K. Ozeki and T. Umeda, “An adaptive filtering algorithm using an or-
thogonal projection to an affine subspace and its properties,” Electron.
steady-state misalignment level is higher. Increasing the pro- Comm. Jpn., vol. 67-A, no. 5, pp. 19–27, May 1984.
jection order also means increased computational complexity. [4] H. C. Shin, A. H. Sayed, and W. J. Song, “Variable step-size NLMS
Therefore a proper selection of provides good tradeoff be- and affine projection algorithms,” IEEE Signal Process. Lett., vol. 11,
no. 2, pp. 132–135, Feb. 2004.
tween convergence rate and computational complexity. [5] S. L. Gay and S. Tavathia, “The fast affine projection algorithm,” in
This work also examines the effect of the step size on the Proc. IEEE ICASSP, 1995, vol. 5, pp. 3023–3026.
misalignment of the APSA, as shown in Fig. 2, where four [6] O. Arikan, A. E. Cetin, and E. Erzin, “Adaptive filtering for
step sizes, , are used in the sim- non-Gaussian stable processes,” IEEE Signal Process. Lett., vol.
1, no. 11, pp. 163–165, Nov. 1994.
ulations. A small step size slows down the convergence rate [7] C. Kwong, “Dual sign algorithm for adaptive filtering,” IEEE Trans.
but also lowers the steady-state misalignment. In contrast, a Commun., vol. 34, no. 12, pp. 1272–1275, Dec. 1986.
large step size speeds up the convergence rate but gives a higher [8] T. Shao and Y. R. Zheng, “A new variable step-size fractional lower-
steady-state misalignment. order moment algorithm for non-Gaussian interference environments,”
The comparisons of the proposed APSA with the conven- in Proc. IEEE ISCAS, May 2009, pp. 2065–2068.
[9] Y. R. Zheng and T. Shao, “A variable step-size LMP algorithm for
tional NLMS, APA, and NSA are shown in Figs. 3–5 under a BG heavy-tailed interference suppression in phased array radar,” in Proc.
interference with a fixed value of but different SIRs IEEE Aerospace Conf., Mar. 2009, pp. 1–6.
( , , and 10 dB). The APA and NLMS are more [10] J. Chambers and A. Avlonitis, “A robust mixed-norm adaptive filter
likely to diverge in more intensive interference while the two algorithm,” IEEE Signal Process. Lett., vol. 4, no. 2, pp. 46–48, Feb.
1997.
sign algorithms are robust against the impulsive interference. [11] L. R. Vega, H. Rey, J. Benesty, and S. Tressens, “A new robust variable
The APSA converges faster and achieves smaller steady-state step-size NLMS algorithm,” IEEE Trans. Signal Process., vol. 56, no.
misalignment than the NSA. 5, pp. 1878–1893, May 2008.