Академический Документы
Профессиональный Документы
Культура Документы
Signal Processing
journal homepage: www.elsevier.com/locate/sigpro
Short communication
art ic l e i nf o a b s t r a c t
Article history: The main drawback of the sign subband adaptive filter algorithm (SSAF) is its large steady-state error. To
Received 4 December 2015 improve the performance, this paper proposes a novel normalized logarithmic subband adaptive filter
Received in revised form algorithm (NLSAF), which is derived from a new normalized logarithmic cost function. Due to possess the
4 March 2016
advantages of the normalized subband adaptive filter (NSAF) and SSAF algorithms in the proposed NLSAF
Accepted 6 March 2016
Available online 26 March 2016
algorithm, it achieves a low steady-state error and the robustness performance against impulsive noise.
Then, by using the energy conservation method, the mean-square convergence performance of the
Keywords: proposed NLSAF algorithm is presented. Simulations on system identifications demonstrate that the
logarithmic cost function proposed subband adaptive filter performs better than the conventional SSAF algorithm in an impulsive-
Subband adaptive filtering
noise scenario.
Impulsive noise
& 2016 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.sigpro.2016.03.006
0165-1684/& 2016 Elsevier B.V. All rights reserved.
P. Wen et al. / Signal Processing 127 (2016) 282–287 283
where ui(k)¼[ui(kN),ui(kN 1),…,ui(kN M þ1)]T and ηi(k ) is a ith ei, D(k ) ⎛ ei, D(k ) ⎞
1 ⎜ ⎟
subband noise. The ith subband output error is defined as (
J ei, D(k ) =) A(k )
− ln⎜ 1 + α
α ⎝ A(k )
⎟
⎠ (9)
ei, D(k ) = di, D(k ) − yi, D (k ). (4)
and α > 0 is the design parameter. When ei, D(k ) / A(k ) < 1, we
The original SSAF algorithm can be derived by minimizing the expand (9) with a Taylor series for small perturbations of the error
following cost function:
⎛ ⎛ ⎞2 ⎞
ei, D(k ) 1 ⎜ ei, D(k ) α 2 ⎜ ei, D(k )
eD(k ) (
J ei, D(k ) =) − α − ⎟ + …⎟
1 α⎜⎜ 2 ⎜⎝ A(k ) ⎟ ⎟⎟
J (k ) = A(k ) A(k ) ⎠
N−1 ⎝ ⎠
∑i = 0 uiT (k )ui (k ) (5)
⎛ ⎞2 3 ⎛ e (k ) ⎞3
α ⎜ ei, D(k ) ⎟ − α ⎜ i, D ⎟ + ….
where eD(k)¼ [e0,D(k),e1,D(k),…eN 1(k)]T. =
2 ⎜⎝ A(k ) ⎟
⎠ 2 ⎜⎝ A(k ) ⎟
⎠ (10)
Applying the gradient descent method, the conventional SSAF
algorithm for updating the tap-weight vector is expressed as
We note that the expand cost function includes the higher-
^ ( k + 1) = w
^ (k ) + μ U (k )sgn⎡⎣ eD(k )⎤⎦ order measures of the error. Thus, the proposed algorithm can
w
N−1 achieve smaller steady-state mean square errors through the use
∑i = 0 uiT (k )ui (k ) (6) of the higher-order statistics for small perturbations of the error.
where μ is the step-size, U (k ) = [u0(k ) , u1(k ) , ... , uN − 1(k )]. In the Taking the derivative of (9) with respect to the tap-weight vector
following, we define ^ (k )
w
Table 1 0.3
Summary of the computational complexity.
0.2
Algorithm Multiplication Addition
0.1
NLMS 3Mþ 1 3M
SSAF Mþ M/N þ 3NLþ 2N 2MNþ 1 0
VRP-SSAF Mþ M/N þ 3NLþ (2þ 2N)N 3MN N M þ 2
NLSAF 2Mþ M/N þ 3NLþ 2N þ 2 2MNþ 2 -0.1
-0.2
-0.3
(
αf ei, D(k ) )
( ) (
∇w J ei, D(k ) = ∇w f ei, D(k ) ) 1 + αf -0.4
( ei,D(k)) (11)
-0.5
where
-0.6
ei, D(k ) 0 200 400 600 800 1000
( )
f ei, D(k ) =
A(k )
. Samples
(12)
Fig. 2. Measured impulse response of the acoustic echo path.
Because the new cost function (9) is a convex function, it has no
local minima. Then using the gradient descent method, the NLSAF
algorithm updates the weight vector as
N−1
^ ( k + 1) = w
w ^ (k ) + μ ∑ ui(k)
(
∂f ei, D(k ) ) ( )
αf ei, D(k )
i=0 (
∂ei, D(k ) 1 + αf ei, D(k ) )
αui(k)ei, D(k)
N−1
^ (k ) + μ A(k)
=w ∑ ei, D(k)
i=0 1+α
A(k)
N−1
^ (k ) + μ αui (k )ei, D(k )
=w ∑
i=0 A(k ) + α A(k ) ei, D(k ) (13)
lowing assumptions.
Fig. 5. MSDs comparison of the proposed algorithm for different α with impulsive
measurement noise, SNR ¼30 dB ( p = 0.001).
N−1
vT (k )ui (k )ei, D(k ) 4.2. Steady-state analysis
‖v( k + 1)‖2 = ‖v(k )‖2 − 2μα ∑
i=0 A(k ) + α A(k ) ei, D(k )
In this section, the mean square derivation of the proposed
N−1
2 2 ui (k )T ui (k )ei2, D(k ) algorithm at steady state will be performed. When an adaptive
+μ α ∑ 2 filter operates in steady state, the MSD satisfies c (k + 1) = c (k ), as
i=0 ( A(k) + α A(k ) ei, D(k ) ) (16) k → ∞, we will get an equation on the steady-state MSD from (22)
i.e., c (∞), as shown in (24)
Taking the mathematical expectation of (16), we get
N −1 σ u2 (∞) ⎛ ei2, D (∞) ⎞
N−1 ⎛ ⎞ 2μα ∑ i
E⎜ ⎟c (∞)
ea, i (k )ei, D(k ) 2 2 ⎜ ⎟
E⎜⎜ ⎟ ( ) ( ) ( )
pκ + 1 σ υ /N + σ u (∞)c (∞) ⎝ E A(k ) + α E A(k ) ei, D (∞) ⎠
c ( k + 1) = c (k ) − 2μα ∑ ⎟
i =0 i
i=0 ⎝ A(k ) + α A(k ) ei, D(k ) ⎠ ⎛ ⎞
N −1 ⎜ ei2, D (∞) ⎟
N−1
⎛ ⎞ =μ 2α 2M ∑ σ u2 (∞)E⎜ ⎟
⎜ ui (k )T ui (k )ei2, D(k ) ⎟ i =0
i ⎜⎜
(
2⎟
E ( A(k )) + α E ( A(k )) ei, D (∞) ⎟⎠ )
+ μ2 α 2 ∑ E⎜ 2⎟
⎝ (24)
i = 0 ⎜ A(k ) + α A(k ) e (k )
( )⎟
⎝ i, D ⎠ (17) As can be seen, it is difficult to find a closed-form solution of
2 c (∞) from (24), because it is a transcendental equation. However,
where c (k ) = E[‖v(k )‖ ] represents the MSD at kth iteration. De-
we can use some numerical computation methods to obtain its
noting the a priori subband error ea, i = vT (k )ui(k ).
numerical solution, such as bisection method [18]. Besides, when
Assuming that ea, i(k ) is a zero-mean Gaussian process for suf- step-size μ is very small, we can use an assumption to solve (24),
ficiently long filters. To simplify the evaluation of the above ex- i.e., c (∞)σu2i(∞) < < συ2/N [12]. In this case, the steady-state MSD of
pectation, we use the Price's theorem and Assumption 1 in (17) the proposed algorithm is approximated as
and have the following approximation.
⎛ ⎞
ei2, D(∞)
⎛ ea, i (k )ei, D (k )
⎞
(
E ea, i1(k )e i1, D (k ) ⎜
⎛
) e i2, D (k ) ⎞
σu2i(∞)E⎜ ⎟
E⎜⎜ ⎟≈p E⎜ 1 ⎟ ⎜ 2⎟
⎟ ⎟ μαM
N−1
⎝( E ( A( k ) ) + α E ( A ( k )) e (∞) )⎠
⎝ A(k ) + α A(k ) ei, D (k ) ⎠ (
E e i2, D (k )
1 ) ⎝ A(k ) + α A(k ) e i1, D (k ) ⎠ c (∞) = ∑ ⎛
i, D
⎞
2 σu2 (∞) ei2, D(∞)
E ea, i2(k )e i2, D (k ) ⎛⎜
( ) e i2 , D (k ) ⎞
⎟
i=0 i
E⎜ ⎟
2 2
+ ( 1 − p) E⎜ ⎟ ( pκ + 1)συ / N ⎝ E( A(k)) + α E( A(k)) ei, D(∞) ⎠ (25)
(
E e i2 , D (k )
2 )
⎝ A(k ) + α A(k ) e i2, D (k ) ⎠ (18)
The relation (25) reveals that the steady-state MSD of the
Likewise, assuming ‖ui(k )‖2 is uncorrelated with f (ei, D(k )) for NLSAF algorithm will increase as the step-size increases.
sufficiently long filters [17], we have
⎛ ⎞ ⎛ ⎞
⎜ u i(k )T u i(k )ei2, D(k ) ⎟ ⎡ 2⎤
⎜ ei2, D(k ) ⎟ 5. Simulation results
E⎜ 2
⎟ ≈ E⎣ ‖u i(k )‖ ⎦E⎜ 2
⎟
⎜ A(k ) + α A(k ) e (k ) ⎟ ⎜ A(k ) + α A(k ) e (k ) ⎟
⎝ ( i, D ) ⎠ ⎝ ( i, D ) ⎠ (19) In this section, the performance of the proposed algorithm is
evaluated through Monte Carlo (MC) simulations in the system
where
identification. The cosine modulated filter bank is used in all the
ei2, D(k ) subband filter algorithms. We use the measured impulse response
(
f ei, D(k ) =) 2 of the acoustic echo path as the unknown system, which is de-
( A(k) + α A(k ) ei, D(k ) ) (20) picted in Fig. 2. In the exact-modeling scenario, the echo path is
2
truncated to the first 32 or 512 tap weights. The number of sub-
Assuming that the fluctuation of ‖ui(k )‖ from one iteration to bands N is 4 for prototype filter length L¼32. The colored input
the next is small enough [4] for sufficiently long filters, we can use signal is obtained by filtering a zero-mean white Gaussian random
the approximation sequence through a first-order system ϕ(z ) = 1/(1 − 0.9z −1). All the
simulation results are obtained by ensemble averaging over 50
A(k ) ≈ E( A(k )) (21)
independent trials. In all simulations, the measurement noise υ(n)
Substituting (18)–(21) into (17) yields is added to y(n) = woT (n)u(n) with 20 dB or 30 dB signal-to-noise
ratios (SNR).
N −1 σ u2 c (k ) ⎛ ei2, D (k ) ⎞
In Fig. 3, we selected different values for the parameter
c ( k + 1) = c (k ) − 2μα ∑ i
E⎜ ⎟
2 2 ⎜ ⎟ κ (κ = 1, 10, 100, 500 and 1000). The adaptive filter tap-length is
i =0 ( p κ + 1 )υ
σ / N + σ uic ( k ) ⎝ (
E A ( k ) ) + α E ( A(k )) ei, D (k ) ⎠
⎛ ⎞ 512. As can be seen, the proposed algorithm for different κ has the
N −1 ⎜ ei2, D (k ) ⎟
+ μ 2α 2M ∑ σ u2 E⎜ ⎟ same initial convergence speed. By decreasing this parameter the
i ⎜
⎜ E ( A(k )) + α E ( A(k )) e (k ) ⎟⎟
2
i =0
⎝ ( i, D ⎠) (22) steady-state MSD reduces. However, the steady-state MSDs for
different κ only exist a small deviation. The impact of the para-
For the algorithm to be stable, the mean-square deviation must meter can be ignored. In the following simulations, the value of κ
decrease iteratively, implying that c (k + 1) − c (k ) < 0. Thus the is chosen as 1000.
step-size has to fulfill the condition Fig. 4 shows the performance of the proposed algorithm using
different number of subbands. The adaptive filter tap-length is 512.
⎧ ⎫
⎪ N−1 Similar to the SSAF algorithm, with the number of subbands increas-
c (k)σu2 ⎛ ei2, D(k) ⎞⎪
⎪ ∑i = 0 i
E⎜ ⎟⎪ ing, the convergence speed of the proposed algorithm becomes slow
⎪ ( pκ + 1)συ2 / N + σu2ic (k) ⎝ E( A(k)) + α E( A(k)) ei, D(k) ⎠ ⎪ and the steady-state MSD reduces. Simultaneously, with the number
0 < μ < 2⎨ ⎬
⎪ ⎛ ⎞ ⎪ of subbands increasing, the computational complexity also increases.
N−1 2 ⎜ ei2, D(k)
⎪ αM ∑i = 0 σuiE ⎟ ⎪
⎜ 2⎟ Thus, a suitable number of subbands are determined as a trade-off
⎪ ⎝ ( E( A(k)) + α E( A(k)) ei, D(k) ) ⎠ ⎪
⎩ ⎭ (23) between convergence speed and steady-state MSD.
P. Wen et al. / Signal Processing 127 (2016) 282–287 287
In Fig. 5, we selected different values for the parameter E⎡⎣ ei21, D(k )⎤⎦ = E⎡⎣ ea2, i (k )⎤⎦ + ( κ + 1)συ2/N (A3)
α (α = 100, 500, 1000, 1500 and 2000). The length of the adaptive
filter is 512. As can be seen, with a small parameter α , the pro-
posed algorithm yields a reduced steady-state MSD at the expense E⎡⎣ ei22, D(k )⎤⎦ = E⎡⎣ ea2, i (k )⎤⎦ + συ2/N (A4)
of a poor convergence speed. Therefore, the parameter α is de-
termined as a trade-off between convergence speed and steady-
E⎡⎣ ei2, D(k )⎤⎦ = pE⎡⎣ ei21, D(k )⎤⎦ + ( 1 − p)E⎡⎣ ei22, D(k )⎤⎦
state misalignment. To obtain a small misalignment and fast
convergence rate, the value of α is chosen as 1500 in the following
simulations. =E⎡⎣ ea2, i (k )⎤⎦ + ( pκ + 1)συ2/N (A5)
Fig. 6 compares the MSDs of the NLMS, conventional SSAF, VRP-
According to the definitions of the previous MSD(k ) and ea, i(k ),
SSAF and the proposed algorithm with the fixed step-size and
we have the following relation for each subband
different SNRs. The length of the adaptive filter is 512. As can be
seen, the proposed algorithm has the robust performance against E⎡⎣ ea2, i (k )⎤⎦ = σu2i(k )c (k ) (A6)
impulsive noise. And it achieves a faster convergence rate and a
smaller steady-state error, compared to the conventional SSAF and Hence, combining (A3)–(A6), (17) can be derived.
VRP-SSAF.
Fig. 7 shows that the steady-state MSD curve of theoretical and
simulation results for different M, where the dotted lines represent References
the theoretical steady-state MSDs. As can be seen, the proposed
theoretical results calculated by (25) are in good agreement with [1] B. Widrow, S.D. Sterns, Adaptive Signal Processing, Prentice-Hall, Englewood
simulation results. Therefore, it is concluded that (25) is an ap- Cliffs, NJ, 1985.
proximate solution of the transcendental Eq. (24). [2] S. Zhang, J. Zhang, Modified variable step-size affine projection sign algorithm,
Electron. Lett. 49 (20) (2013) 1264–1265.
[3] T. Shao, Y.R. Zheng, J. Benesty, An affine projection sign algorithm robust
against impulsive interference, IEEE Signal Process. Lett. 17 (4) (2010)
6. Conclusion 327–330.
[4] K.A. Lee, W.S. Gan, S.M. Kuo, Subband adaptive Filtering: Theory and Im-
plementation, Wiley, Hoboken, NJ, 2009.
In this paper, we propose a novel normalized logarithmic cost [5] K.A. Lee, W.S. Gan, Improving convergence of the NLMS algorithm using
function. Using the general steepest descent method, the NLSAF constrained subband updates, IEEE Signal Process. Lett. 11 (9) (2004) 736–739.
algorithm is derived from the new cost function. With the benefit [6] K.A. Lee, W.S. Gan, Inherent decorrelating and least perturbation properties of
the normalized subband adaptive filter, IEEE Trans. Signal Process. 54 (11)
of the robust logarithmic cost function, the proposed NLSAF al- (2006) 4475–4480.
gorithm is able to suppress the effect of impulsive noise on the [7] J.H. Kim, J.H. Chang, S.W. Nam, Sign subband adaptive filter with l1-norm
filter weight update. Simultaneously, the NLSAF can obtain a low minimisation-based variable step-size, Electron. Lett. 49 (21) (2013)
1325–1326.
steady-state error. The complexity, convergence and steady state
[8] J. Ni, F. Li, Variable regularisation parameter sign subband adaptive filter,
analysis of the NLSAF algorithm was carried out based on the Electron. Lett. 46 (24) (2010) 1605–1607.
energy conservation method. Compared to the SSAF and VRP-SSAF [9] A.H. Sayed, Fundamentals of Adaptive Filtering, Wiley–Interscience, New York,
algorithm, simulations on system identification show that the 2003.
[10] S. Zhang, J. Zhang, Transient analysis of zero attracting NLMS algorithms
proposed algorithm obtains 3 dB attenuation in term of the stea- without Gaussian input signal, Signal Process. 97 (2013) 100–109.
dy-state MSD than the conventional SSAF at the same convergence [11] M.O. Sayin, N.D. Vanli, S.S. Kozat, A novel family of adaptive filtering algo-
rate. With different tap lengths and step sizes, the theoretical re- rithms based on the logarithmic cost, IEEE Trans. Signal Process. 62 (17) (2014)
4411–4424.
sults of NLSAF have supported our theoretical analysis on the [12] Y.P. Li, T.S. Lee, B.F. Wu, A variable step-size sign algorithm for channel esti-
steady-state MSD. mation, Signal Process. 102 (2014) 304–312.
[13] S. Zhang, J. Zhang, Enhancing the tracking capability of recursive least p-norm
algorithm via adaptive gain factor, Digit. Signal Process. 30 (2014) 67–73.
[14] J. Shin, J. Yoo, P. Park, Variable step-size sign subband adaptive filter, Signal
Appendix A Process. Lett. 20 (2) (2013) 173–176.
[15] J.W. Yoo, J.W. Shin, P.G. Park, A band-dependent variable step-size sign sub-
band adaptive filter, Signal Process. 104 (2014) 407–411.
In this appendix, we give a detailed derivation about (17).
[16] J. Ni, X. Chen, J. Yang, Two variants of the sign subband adaptive filter with
ei1, D(k ) = ea, i (k ) + ςi(k ) improved convergence rate, Signal Process. 96 (2014) 325–331.
(A1) [17] T.Y. Al-Naffouri, A.H. Sayed, Transient analysis of adaptive filters with error
nonlinearities, IEEE Trans. Signal Process 51 (2003) 653–663.
ei2, D(k ) = ea, i (k ) + υi (k ) [18] J.F. Traub, Iterative Methods for the Solution of Equations, Am. Math. Soc.
(A2) (1982).