VI International Telecommunications Symposium (ITS2006), September 3-6, 2006, Fortaleza-CE, Brazil 1 FxLMS Algorithm with Variable Step Size and Variable Leakage Factor for Active Vibration Control Walter A. Gontijo, Orlando J. Tobias, Rui Seara, and Eduardo M. O. Lopes Abstract—This paper discusses a novel adaptive algorithm termed variable step size and variable leakage factor filtered-x least-mean-square (VSSVLFxLMS) algorithm for applications in active vibration control (AVC). For such applications, the leaky filtered-x LMS (LFxLMS) algorithm has been considered for stability purposes. However, when the standard LFxLMS algorithm (with fixed leakage) is used, the control performance is strongly dependent on the selected leakage value. A similar result is noticed concerning the chosen step-size value. Thus, by using the proposed VSSVLFxLMS algorithm better step-size and leakage values are obtained, improving considerably the control performance. The conjectures of this work are confirmed by implementing the referred algorithm in an experimental setup of an AVC system. Controller Error sensor Primary source Secondary source Reference sensor Aluminium beam Support Support 50 mm 50 mm 100 mm 900 mm Index Terms—FXLMS algorithm, leaky FXLMS algorithm, active vibration control. I. INTRODUCTION In recent years, the area of active vibration control (AVC) has experienced a considerable development coming from two important facts. The increasing interest in using that technique to develop smart structures and, secondly, the great advance and wide diffusion of the digital signal processor (DSP) boards. The latter has allowed implementing more complex and better control algorithms. In AVC applications (see Fig. 1), the filtered-x least-mean-square (FxLMS) algorithm is extensively used [1]-[3]. Such popularity is mainly due to its robustness and low complexity [1], [2]. However, the direct implementation of FxLMS algorithm is not always possible. For instance, the finite-precision arithmetic of the DSPs or an insufficient spectral content of the input signal can lead the adaptive algorithm to instability [4], [5]. A widely used solution to overcome such a problem is the use of the leaky FxLMS algorithm. It results from introducing a penalizing term into the standard cost function, giving rise to the leaky FxLMS (LFxLMS) algorithm [4]-[6]. Fig. 1. Experimental setup of the AVC system. When the LFxLMS algorithm is applied to an AVC system, it is noticed that the canceling performance achieved is strongly dependent on the leakage value used. In [4] and [7], it is demonstrated that in the presence of nonlinearities in the adaptation path, there is an optimum value for the leakage factor to attain a maximum cancellation. Such an optimum can be theoretically determined if the system parameters (plant, nonlinearity model, input signal power, among others) are known. In practice, these parameters are obviously unknown; thus, the best leakage factor value must be determined by a trial and error procedure. In addition, since the weight update equation is dependent on the product of the step-size parameter and leakage factor, for a given leakage factor value there is also a particular step-size value that leads to the condition of maximum cancellation. Thus, to achieve an optimal tradeoff between step size and leakage factor for algorithm performance, the novel variable step size and variable leakage factor FxLMS (VSSVLFxLMS) algorithm is proposed in this paper. The advantages of this improved version of the FxLMS algorithm are the following: i) There is no need to know the system parameters. Manuscript received March 31, 2006. This work was supported in part by the Brazilian National Council for Scientific and Technological Development (CNPq). Walter A. Gontijo, Orlando J. Tobias, and Rui Seara are with the LINSE - Circuits and Signal Processing Laboratory of the Department of Electrical Engineering at the Federal University of Santa Catarina, Florianópolis, Santa Catarina, Brazil (e-mails: [email protected], [email protected]; [email protected]). Eduardo M. O. Lopes is with PISA – Group for Integrated Research on Vibrating and Acoustic Systems of the Department of Mechanical Engineering at the Federal University of Santa Catarina, Florianópolis, Santa Catarina, Brazil (e-mail: [email protected]). SBrT © ii) Both step size and leakage factor are self-adjustable, aiming to achieve the maximum vibration cancellation. iii) Ability to accommodate time-varying conditions. The proposed algorithm has been implemented in the EZ-Kit LiteTM 21161, Analog Devices’ DSP platform, with 32 bits floating-point arithmetic. Section III shows some numerical results of the proposed algorithm for performance assessment. 578 VI International Telecommunications Symposium (ITS2006), September 3-6, 2006, Fortaleza-CE, Brazil II. VSSVLFXLMS ALGORITHM In this section, the weight updating equation for the proposed algorithm is derived. To this end, the block diagram of the AVC experimental setup (Fig. 1), shown in Fig. 2, is considered. In Fig. 2, variables d (n) and e(n) denote the primary and error signals, respectively. Vector x(n) represents the reference signal and yl (n) is the output signal of the secondary source. The filter impulse response cascaded with the adaptive filter, termed secondary path, is given by s [ s0 s1 " sM 1 ]T . The estimate of the secondary proposing an algorithm in which both the step-size and leakage parameters are variable. In this way, (3) and (4) must be modified as follows: w (n 1) x(n) d(n) wo + w(n) y(n) 6 Q ( n ) w ( n ) P ( n )e( n ) x f ( n ) , Q ( n ) 1 P ( n) J ( n ) . (7) In next the section, the proposed procedure to adjust the leakage factor will be derived. B. Variable Leakage Factor To obtain a variable leakage factor we use the stochastic gradient rule. In this way, the leakage factor is adjusted in accordance with the negative gradient of the quadratic error. Thus, J ( n) yl (n) s (6) and path s, denoted by sˆ [ sˆ0 sˆ1 " sˆM 1 ]T , is obtained by using a system identification procedure as, for instance, the one suggested in [1]. 2 J (n 1) U w e 2 ( n) , 2 w J (n 1) (8) where U is a positive constant with 0 U 1 . The second term on the r.h.s of (8), by considering (6), can be written as s^ (n) f LMS we2 (n) wJ (n 1) e(n) we2 (n) w w ( n) Let us consider the stochastic gradient rule ww (n) wJ (n 1) (1) where J (n) denotes the cost function, given by [8] J ( n) e2 (n) JwT (n)w (n) , (2) and J t 0 is the leakage factor. Qw (n) Pe(n)xf (n) , (4) w{[1 P(n) J (n 1)]w (n 1) P(n)e(n 1)xf (n 1)} wJ (n 1) P(n)w (n 1). (11) J ( n) (3) where Q 1 PJ , (10) Finally, from (9), and substituting (10) and (11) into (8), the expression for adjusting the leakage factor is given by Now, determining the gradient of (2) and substituting it into (1), one obtains w (n 1) 2e(n)xf (n) . Now, the second term on the r.h.s. of (9) is obtained from (6), as follows: A. Weight Update and Leakage Adjust Equations P w ( n) J ( n ) , 2 (9) where Fig. 2. Block diagram of the AVC system using the FxLMS algorithm. w (n 1) T ª we 2 (n) º w w (n) , « » ¬« w w (n) ¼» wJ (n 1) J (n 1) UP(n)e(n)w T (n 1)xf (n) . (12) C. Variable Step Size The step-size value is adjusted by using a similar approach to (8), resulting in the following expression adapted from [9]: and P( n ) P(n 1) Ue(n)xTf (n)e(n 1)xf (n 1) , (13) M 1 x f ( n) ¦ sˆi x(n i) . (5) i 0 Vector xf (n) represents the reference signal filtered by the secondary path estimate. In the different approaches of the LFxLMS algorithm found in the open literature, the parameter J is considered constant [2], [8]. On the other hand, in our case, we are SBrT © where xf (n) is obtained from (5). Note that by monitoring the error signal, when the adaptive filter weights have converged (error signal approaches to zero) the adjustment process, defined by (12) and (13), is stopped. In practice an upper and lower limit, for (12) and (13), should be specified. Such a procedure is to avoid that both parameters do not become too small or too large, making the adaptive algorithm very slow or even unstable, respectively. 579 VI International Telecommunications Symposium (ITS2006), September 3-6, 2006, Fortaleza-CE, Brazil 1.0 Firstly, experimental results illustrating the behavior of the leaky FxLMS algorithm (with fixed leakage factor) are presented, considering several leakage values. The aim of this experiment is to support the development of the variable-leakage factor algorithm. The experimental results are obtained by measuring the steady-state error value of the LFxLMS algorithm as a function of the leakage factor value (see Fig. 3). Note from this figure that the behavior of the steady-state error is a convex function, evidencing the existence of an optimum leakage value. The process to obtain such an optimum is manual, being cumbersome and time-consuming. Thus, the possibility of implementing an automatic procedure is of great interest from a practical point of view. 0.8 Signal from the error sensor III. EXPERIMENTAL RESULTS 3 0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8 -1.0 0 1000 2000 3000 4000 5000 6000 Iterations (a) 0.26 Leaky FXLMS -20 Error signal spectrum 0.24 Steady-state MSE Control off Control on 0.22 0.20 Optimum leakage value 0.18 0.16 0.05 -40 -60 -80 -100 0.10 0.15 0.20 0.25 0.30 0.35 -120 Leakage factor 0 50 100 200 250 300 350 400 450 Frequency [Hz] Fig. 3. Variation of the steady-state error as a function of the leakage factor value. (b) -5 Zoom around 300 Hz The VSSVLFxLMS algorithm has been implemented in the EZ-Kit LiteTM 21161 platform from Analog Devices. In Figs. 4 and 5, the obtained experimental results are illustrated using the active vibration control setup given in Fig. 1. These figures show the evolution of the error signal as well as the power spectrum of the steady-state error signal with and without control, for two frequencies of the perturbation signal (80 Hz and 300 Hz). In Fig. 4, the system is disturbed by a sinusoidal signal of 300 Hz. Fig. 4(a) shows the evolution of the error signal coming from the error sensor (see Fig. 1), illustrating the algorithm convergence. In Fig. 4(b), the spectrum of the error signal after algorithm convergence is plotted. A detail of Fig. 4(b) around the frequency of 300 Hz is shown in Fig. 4(c). From these figures, we can observe the controlling effect of the proposed algorithm. The attenuation obtained for the 300 Hz case is 11 dB. In Fig. 5, the obtained results for a disturbance signal of 80 Hz are shown following the same presentation pattern as in Fig. 4. Here, the same observations drawn from Fig. 4 are also verified for Fig. 5, except that now the obtained attenuation is 7 dB. We attribute such a difference to the mechanical characteristics of the system under control. Further studies are being carried out aiming to improve the attenuation characteristics of the proposed algorithm. SBrT © 150 Control off Control on -10 -15 -20 -25 270 280 290 300 310 320 330 340 Frequency [Hz] (c) Fig. 4. Experimental results for 300 Hz. (a) Error signal. (b) Error signal spectrum with and without control. (c) Detailed plot around 300 Hz. IV. CONCLUDING REMARKS In this work, the variable step size and variable leakage factor FxLMS algorithm has been derived and tested in a real AVC application. The proposed algorithm exhibits a good performance at the expense of a small computational 580 VI International Telecommunications Symposium (ITS2006), September 3-6, 2006, Fortaleza-CE, Brazil complexity increase. Since different procedures for step size and leakage factor adjustment can be performed, further studies for improving the behavior of the VSSVLFxLMS algorithm are currently underway. REFERENCES [1] [2] [3] 0.8 [4] Signal from the error sensor 0.6 0.4 [5] 0.2 [6] 0 -0.2 [7] -0.4 -0.6 -0.8 [8] 0 1000 2000 3000 4000 5000 6000 [9] Iterations (a) Control off Control on Error signal spectrum -20 -40 -60 -80 -100 -120 0 50 100 150 200 Frequency [Hz] (b) Control off Control on Zoom around 80 Hz -8 -10 -12 -14 -16 -18 50 60 70 80 90 100 Frequency [Hz] (c) Fig. 5. Experimental results for 80 Hz. (a) Error signal. (b) Error signal spectrum with and without control. (c) Detailed plot around 80 Hz. SBrT © 4 581 S. Kuo and D. R. Morgan, Active Noise Control Systems. John Wiley & Sons, 1996. S. J. Elliott, Signal Processing for Active Control, Academic Press, 2001. C. Hansen, “Active noise control – from laboratory to industrial implementation,” in Proc. of Noise-Con’97, Pennsylvania, USA, June 1997, pp. 3-38. O. J. Tobias and R. Seara, “On the LMS algorithm with constant and variable leakage factor in a nonlinear environment,” (to be published) IEEE Trans. Signal Processing, vol. 54, pp. 1-10, 2006. J. M. Cioffi, “Limited-precision effects in adaptive filtering,” IEEE Trans. Circuits and Syst., vol. 34, no. 7, pp. 821-833, July 1987. O. J. Tobias and R. Seara, “Performance comparison of the FXLMS, nonlinear FXLMS and leaky FXLMS algorithms in nonlinear active control applications,” in Proc. European Signal Processing Conf. (EUSIPCO), Toulouse, France, Sep. 2002, pp. 1-4. F. Heinle, R. Rabenstein, and A. Stenger, “A measurement method for the linear and nonlinear properties of electroacoustic transmission systems,” Elsevier Signal Processing, vol. 64, no. 1, pp. 49-60, Jan. 1998. P. Darlington, “Performance surfaces of minimum effort estimators and controllers,” IEEE Trans. Signal Processing, vol. 43, no. 2, pp. 536-539, Feb. 1995. B. Farhang-Boroujeny, Adaptive Filters: Theory and Application, John Wiley & Sons, 1998.
© Copyright 2026 Paperzz