Lattice form adaptive infinite impulse response filtering algorithm for active noise control Jing Lu,a) Chunhua Shen, Xiaojun Qiu, and Boling Xu State Key Laboratory of Modern Acoustics and Institute of Acoustics, Nanjing University, Nanjing 210093, China

共Received 24 May 2002; revised 18 October 2002; accepted 23 October 2002兲 In some situations of active noise control, infinite impulse response 共IIR兲 filters are more suitable than finite impulse response 共FIR兲 filters owing to the poles in the transfer function. A number of algorithms have been derived for applying IIR filters in active noise control; however, most of them use the direct form IIR filter structure, which faces the difficulties of checking stability and relatively slow convergence speed for noise composed of narrow-band components with large power disparity. To overcome these difficulties along with using the direct form IIR filters, a new adaptive algorithm is proposed in this paper, which uses and updates the lattice form adaptive IIR filter in an active noise control system. Full mathematical derivations of the proposed algorithm are presented, and the comparison between the proposed algorithm and the commonly used filtered-u LMS and filtered-v LMS algorithms shows the superiority of the proposed algorithm. © 2003 Acoustical Society of America. 关DOI: 10.1121/1.1529665兴 PACS numbers: 43.50.Ki 关KAC兴

I. INTRODUCTION

Most active noise control systems 共ANC兲 use adaptive FIR filters and filtered-x LMS algorithm 共FXLMS兲 due to their simplicity and inherent stability.1 However, there are some situations where adaptive IIR filters may be more suitable.1–3 For example, when there are poles in the primary plant transfer function, or when there is feedback from the control output to the reference sensor, if an FIR filter is used in such a system, very long taps are needed. However, with an IIR filter, much fewer taps can be used, resulting in less computation load. In general, an IIR filter with sufficient order can exactly match poles as well as zeros of the physical system, resulting in a lower residual mean squared error. Although the algorithms using adaptive IIR filters for active noise control have been proposed for many years, they still have not been widely used in the application of the active noise control system due to the following disadvantages.1– 6 First, IIR filters are not unconditionally stable due to the possibility that some poles of the filters might move outside of the unit circle during the weights update. Second, the existing adaptive algorithms have a lower convergence speed and may converge to a local minimum. Therefore, it is recommended that whenever possible, adaptive FIR filters should be used.7 The adaptive IIR filters used in active noise control are usually in the direct form, for example, the filtered-u LMS 共FULMS兲 algorithm,4 filtered-v LMS 共FVLMS兲 algorithm,5 and the ‘‘correct algorithm’’ proposed by Snyder.3 All these adaptive algorithms use the direct form IIR filter, hence having the same problems of possible instability and slow convergence. The lattice structure is an alternative form of a digital filter, which possesses the advantages of inherent stability and greatly reduced sensitivity to the eigenvalue spread a兲

Electronic mail: [email protected]

J. Acoust. Soc. Am. 113 (1), January 2003

of the reference signal.8,9 Many algorithms have been proposed to make the lattice form adaptive IIR filter.9–13 This paper will propose a new adaptive algorithm for using the lattice form adaptive IIR filter in active noise control. Full mathematical derivations of the lattice gradient descent algorithm and a simplified gradient lattice algorithm will be presented, and the performance of the proposed algorithm will be compared with the FUVLMS and FVLMS algorithm. The idea of using lattice filters in active noise control is not new. However, it is usually used as a preprocessor followed by an FIR filter.14 –20 The preprocessor 共lattice filter兲 decorrelates the reference signal to produce uncorrelated backward prediction error signals based on the Gram– Schmidt orthogonalization process.21 Then, the FIR filter operates on these uncorrelated signals; thus, the convergence of the adaptive filter does not suffer from eigenvalue disparity problems. It was shown that this form of the active noise control system converges significantly faster than the traditional transversal filter when the primary noise consists of sinusoidal components with widely differing power. Recent application of lattice filters can be found in Ref. 22, where an active noise control algorithm based on multivariable gradient lattice filters was proposed. However, the authors still treated the lattice structure and FIR filter separately by just using the decorrelation property of lattice filters. The primary difference of our proposed lattice algorithm is that the lattice filter is used as the control filter, not just as a preprocessor; thus, not only the benefits of adaptive IIR filter are held, but also the problem of slow convergence and possibility of instability is avoided. II. THE LATTICE GRADIENT DESCENT ALGORITHM FOR ANC

Figure 1 shows the flowgraph of the tapped state normalized lattice form IIR filter for active noise control for the case in which the filter order M is set to 3. In this figure, the

0001-4966/2003/113(1)/327/9/$19.00

© 2003 Acoustical Society of America

327

FIG. 1. Tapped state normalized lattice filter for active noise control for M⫽3.

primary path transfer function P(z) represents the transfer function from the noise source 共reference signal is assumed to be picked there兲 to the error sensor; the cancellation path transfer function C(z) represents the acoustic path from the secondary source to the error sensor. 兵 ␨ (•) 其 is some kind of noise which is statistically independent of the reference signal 兵 x(•) 其 . Driven by the reference signal 兵 x(•) 其 with x(n) the most recent input sample, the output of the lattice filter y(n) at time n passes through the cancellation path C(z) and produces the control signal s(n) at the location of the error sensor. The error sensor, which is the sum of the primary noise p(n), uncorrelated noise ␨ (n), and control signal s(n), will be picked up by the error sensor and be used by the adaptive algorithm to update the lattice filter parameters. The filter parameters are the rotation angles 兵 ␪ 1 ,..., ␪ M 其 plus the tap parameters 兵 ␯ 0 ,... ␯ M 其 , which are related to the direct form filter parameters in a nonlinear manner, and may be converted to the direct form filter parameters and vice versa.8,21 As shown in Fig. 1, the cascade structure in the lattice filter propagates a forward signal f k (n) and a backward signal b k (n) at time n and section number k. By adapting 兵 ␪ k 其 in such a way that 兩 sin ␪k兩⬍1, the stability of the lattice filter is ensured.9 The output of the lattice filter is M

y 共 n 兲⫽



k⫽0

b k共 n 兲 ␷ k ,

共1兲

兵 b k (n) 其 for k⫽M ,M ⫺1,...,1 are obtained by the Schur recursion9



e 共 n 兲 ⫽ p 共 n 兲 ⫹s 共 n 兲 ⫹ ␨ 共 n 兲 ⫽ 关 P 共 z 兲 ⫹W 共 z 兲 C 共 z 兲兴 x 共 n 兲 ⫹ ␨ 共 n 兲 , 共3兲 where W(z) is the transfer function of the lattice filter. The parametric derivatives of this error signal are given by

⳵e共 n 兲 ⳵s共 n 兲 ⳵W共 z 兲 ⫽ ⫽ C共 z 兲x共 n 兲 ⳵␷k ⳵␷k ⳵␷k ⳵e共 n 兲 ⳵s共 n 兲 ⳵W共 z 兲 ⫽ ⫽ C共 z 兲x共 n 兲. ⳵␪k ⳵␪k ⳵␪k

共4兲

The derivative with respect to the tap parameters 兵 ␯ k 其 is straightforward. In the lattice form, there is M

册冋

cos ␪ k f k⫺1 共 n 兲 ⫽ b k共 n 兲 sin ␪ k

⫺sin ␪ k cos ␪ k

册冋



f k共 n 兲 , b k⫺1 共 n⫺1 兲

W共 z 兲⫽ 共2兲

where f M (n)⫽x(n) and b 0 (n)⫽ f 0 (n). Development of the lattice version of the gradient descent algorithm for active noise control follows the same methodology as for the direct form: the output error is differentiated with respect to the filter parameters to obtain negative gradient signals.9 The convergence properties of the direct form and lattice algorithms are theoretically equiva328

lent: both algorithms seek the minimum points of the cost function E 关 e 2 (n) 兴 , but in different parameter spaces. The key advantage of the lattice over the direct form concerns filter stability: the lattice filter is inherently stable in timevarying environments while the direct form is not.8,9 The following mathematical derivations are similar to those in Ref. 9, except where in Ref. 9, the algorithm is derived for normal adaptive filtering without taking into the account of the cancellation path. Set z as the unit delay operator, that means for any input sequence 兵 u(n) 其 , zu(n)⫽u(n⫺1); therefore, the transfer function that will be used in the following derivations can be regarded as the rational model of the unit delay operator. As with the direct form algorithm, the output error signal is

J. Acoust. Soc. Am., Vol. 113, No. 1, January 2003



k⫽0

␷ kB k共 z 兲 ,

共5兲

so that

⳵e共 n 兲 ⫽B k 共 z 兲 C 共 z 兲 x 共 n 兲 , ⳵␷k

共6兲

where B k (z) is the transfer function of the lattice filter corresponding to the kth backward signal. The signals obtained from Eq. 共6兲 are called filtered regressor signals, as they are formed by filtering the input with the cancellation path transLu et al.: Lattice filter for active noise control

FIG. 2. Filtered regressor signals for the tap parameters.

fer function and the lattice filter. The signals can be obtained with an auxiliary lattice filter, as shown in Fig. 2 for the case M ⫽3, where the filtered regressor signal for the tap parameters 兵 ␯ k 其 is 兵 b ck (n) 其 . The instantaneous estimate of the gradient signal of the cost function E 关 e 2 (n) 兴 corresponding to the tap parameter ␯ k is

⳵e共 n 兲 ⫽2e 共 n 兲 b ck 共 n 兲 . ⵜ␷ k 共 n 兲 ⫽2e 共 n 兲 ⳵␷k

共7兲

With the negative gradient direction ⫺ⵜ ␯ k (n), the corresponding gradient descent form algorithm can be easily constructed as shown in Appendix A. For the case of C(z) ⫽1 where the cancellation path is ideal, the algorithm is simplified to the case of normal lattice form adaptive filtering; the filtered regressor signals for the tap parameters become 兵 b k (n) 其 in Fig. 1. In this case, it is not necessary to use a separate auxiliary lattice filter to obtain the filtered regressor signals. Obtaining derivative signals with respect to the rotation angles 兵 ␪ k 其 is more complicated. Via Eqs. 共3兲 and 共5兲, there is



and



册冋

⫺sin ␪ l

册冋

⫺sin ␪ l

cos ␪ l F ␪ k,l⫺1 共 z 兲 ⫽ B ␪ k,l 共 z 兲 sin ␪ l

cos ␪ l F ␪ k,l⫺1 共 z 兲 ⫽ B ␪ k,l 共 z 兲 sin ␪ l ⫹



which requires ⳵ B l (z)/ ⳵ ␪ k . Set now

obtaining

⳵ B l共 z 兲 ⌬ = B ␪ k,l 共 z 兲 , ⳵␪k

the

sensitivity

cos ␪ l

⫺sin ␪ l

By applying differential operator ⳵ / ⳵ ␪ k to the z transform of Eq. 共2兲, there are

册冋



if k⫽l 共9兲



F l共 z 兲 , zB l⫺1 共 z 兲

if k⫽l

by using F l⫺1 共 z 兲 ⫽cos ␪ l •F l 共 z 兲 ⫺sin ␪ •zB l⫺1 共 z 兲

共11兲

B l 共 z 兲 ⫽sin ␪ l •F l 共 z 兲 ⫹cos ␪ •zB l⫺1 共 z 兲 . For k⫽l, there is



册冋

cos ␪ l F ␪ k,l⫺1 共 z 兲 ⫽ B ␪ k,l 共 z 兲 sin ␪ l



⫺sin ␪ l cos ␪ l



册冋

F ␪ k,l 共 z 兲 zB ␪ k,l⫺1 共 z 兲



⫺B l 共 z 兲 . F l⫺1 共 z 兲

共12兲

Because F M (z)⫽1 and B 0 (z)⫽F 0 (z) for all 兵 ␪ k 其 , the boundary conditions for completing the recursion are F ␪ k,M 共 z 兲 ⫽0,

⳵ F l共 z 兲 ⌬ = F ␪ k,l 共 z 兲 . ⳵␪k

F ␪ k,l 共 z 兲 zB ␪ k,l⫺1 共 z 兲

⫺cos ␪ l



function

册冋



共10兲



共8兲

cos ␪ l

F ␪ k,l 共 z 兲 , zB ␪ k,l⫺1 共 z 兲

⫺sin ␪ l

M

⳵e共 n 兲 ⳵W共 z 兲 ⳵ B l共 z 兲 ⫽ C共 z 兲x共 n 兲⫽ ␷l C共 z 兲x共 n 兲, ⳵␪k ⳵␪k ⳵␪k l⫽0

cos ␪ l

册冋

B ␪ k,0共 z 兲 ⫽F ␪ k,0共 z 兲 .

共13兲

For illustration purposes, Fig. 3 shows the filtered regressor signal corresponding to the rotation parameter ␪ 2 for the filter order of M ⫽3. The instantaneous estimate of the gradient signal of the cost function E 关 e 2 (n) 兴 corresponding to the rotation parameter ␪ k is

FIG. 3. Filtered regressor signal for the rotation parameter.

J. Acoust. Soc. Am., Vol. 113, No. 1, January 2003

Lu et al.: Lattice filter for active noise control

329

ⵜ␪ k 共 n 兲 ⫽2e 共 n 兲

⳵e共 n 兲 ⫽2e 共 n 兲 b ␪ k 共 n 兲 . ⳵␪k

共14兲

An overall algorithm list is given in Appendix A. Note that the ‘‘Test’’ step in the algorithm not only guarantees the stability of the adaptive process but also ensures the uniqueness of the mapping from the transfer function space to the parameter space.9 It should be noted that M additional lattice filters are required to obtain the filtered regressor signals 兵 ⫺ⵜ ␪ k (n) 其 corresponding to the rotation parameters. Thus, the complexity is of the order M 2 , both for computation and storage. Considering that the normally used direct form IIR filters such as filtered-u4 algorithm and simplified filtered-v algorithm5 required only order M computation and storage, the increased complexity of the lattice form is an obvious disadvantage. A simplified gradient lattice algorithm with a computation complexity order of M is described below.

III. THE SIMPLIFIED GRADIENT LATTICE ALGORITHM FOR ANC

After examining the behavior of the filtered regressor signals 兵 ⵜ ␪ k 其 corresponding to the rotation parameters along the reduced error surface with the tap parameters 兵 ␯ k 其 being optimized, a partial gradient algorithm of order M complexity can be derived. The derivation is quite complex, and the details are omitted for brevity. However, a similar deviation of adaptive lattice algorithm that used in system identification can be found in Ref. 9, which does not take into account the cancellation path as in an ANC system. Consider the ideal update formula

␷ k 共 n⫹1 兲 ⫽ ␷ k 共 n 兲 ⫺

␮ ⳵ E 关 e 2 共 n 兲兴 , 2 ⳵␷k

␪ k 共 n⫹1 兲 ⫽ ␪ k 共 n 兲 ⫺

␮ ⳵ E 关 e 共 n 兲兴 , 2 ⳵␪k

共15兲

where the final correction terms may be written as the inner product

␮ ⳵ E 关 e 2 共 n 兲兴 ⳵W共 z 兲 ⫽⫺ ␮ C 共 z 兲 ,S x 共 z 兲 ⫺ 2 ⳵␷k ⳵␷k



⫻ 关 P 共 z 兲 ⫹W 共 z 兲 C 共 z 兲兴 ,



␮ ⳵ E 关 e 2 共 n 兲兴 ⳵W共 z 兲 ⫽⫺ ␮ C 共 z 兲 ,S x 共 z 兲 ⫺ 2 ⳵␪k ⳵␪k

共16兲



⫻ 关 P 共 z 兲 ⫹W 共 z 兲 C 共 z 兲兴 , where S x (z) is the spectral density function associated with the reference signal 兵 x(•) 其 and the inner product is defined as 具 F(z),G(z) 典 ⫽ (1/2␲ j)养 兩 z 兩 ⫽1 F(z)G(z ⫺1 )(dz/z) . With tap parameters 兵 ␯ k 其 being optimized, there is



330

⫺␮



⳵W共 z 兲 C 共 z 兲 ,S x 共 z 兲关 P 共 z 兲 ⫹W 共 z 兲 C 共 z 兲兴 ⳵␪k

⫽␮





⳵D M共 z 兲 1 W 共 z 兲 C 共 z 兲 ,S x 共 z 兲 ⳵␪k D M共 z 兲



⫻ 关 P 共 z 兲 ⫹W 共 z 兲 C 共 z 兲兴 . Thus

␪ k 共 n⫹1 兲 ⫽ ␪ k 共 n 兲 ⫹ ␮

⳵D M共 z 兲 1 W共 z 兲C共 z 兲x共 n 兲, ⳵␪k D M共 z 兲 共18兲

where D M (z) is equal to H(z) of the corresponding equivalent direct form IIR filter with a transfer function of G(z)/H(z). After further deviation and approximation, it can be shown

⳵D M共 z 兲 1 ⬇ ␥ k zB k⫺1 共 z 兲 , ⳵␪k D M共 z 兲

共19兲

where M

␥ k⫽



l⫽k⫹1

cos ␪ l , ␥ M ⫽1.

共20兲

The resulting algorithm would appear as v k 共 n⫹1 兲 ⫽ v k 共 n 兲 ⫺ ␮ e 共 n 兲 •B k 共 z 兲 C 共 z 兲 x 共 n 兲 ,

2



Using the above condition, it can be shown that if the parameters are held stationary, the following equation can be obtained:



⳵W共 z 兲 C 共 z 兲 ,S x 共 z 兲关 P 共 z 兲 ⫹W 共 z 兲 C 共 z 兲兴 ⫽0. ⳵␷k J. Acoust. Soc. Am., Vol. 113, No. 1, January 2003

共17兲

k⫽0,1,...,M , 共21兲 ␪ k 共 n⫹1 兲 ⫽ ␪ k 共 n 兲 ⫹ ␮ e 共 n 兲 • ␥ k zB k⫺1 共 z 兲 W 共 z 兲 C 共 z 兲 x 共 n 兲 , k⫽1,2,...,M . Figure 4 shows the flowgraph for generating the necessary filtered regressor signals. The algorithm listing appears in Appendix B. Note that the technique used to determine the values of 兵 sin ␪k(n⫹1)其 and 兵 cos ␪k(n⫹1)其 is called ‘‘annihilation operations,’’ 9 where the need for computing trigonometric functions in every step is avoided and thus the efficiency for the algorithm’s implementation is improved. Accordingly the ‘‘Test’’ step in the algorithm has been modified and appears different from that in Appendix A. It can be found that the complexity of the algorithm is reduced to the order of M . Note, although it can be shown that the stationary points of the above simplified algorithm are indeed the stationary points of the cost function E 关 e 2 (n) 兴 , 9 the possibility that the convergence points are the saddle points cannot be excluded because the expected value of the update term concerning the rotation angles 兵 ␪ k 其 is not indeed a negative gradient vector of the cost function. However, the following simulations suggest that this algorithm tends towards a local minimum. Lu et al.: Lattice filter for active noise control

FIG. 4. Generation of filtered regressor signals in simplified gradient algorithm.

IV. SIMULATION RESULTS A. Description of the simulations

In this section, several illustrative results are presented on comparisons between the proposed algorithm and commonly used FULMS algorithm4 and FVLMS algorithm.5 Only the simplified gradient lattice algorithm described in Sec. III will be used for simulations, which will be called LFRLMS 共lattice filtered reference LMS兲 algorithm in the following context. All the simulations were conducted using the acoustic transfer functions of a single input and output active noise control system measured in the anechoic room with a sampling rate of 8000 Hz. Figure 5 shows a schematic diagram of the system, where two identical loudspeakers were placed 20 cm from each other, one acted as the noise source and the other acted as the control source. An error microphone was placed 1.3 m away from the center of two loudspeakers. The impulse responses corresponding to the primary path and the cancellation path are shown in Fig. 6. Note, the optimum filter form for ANC is somewhat like ⫺ P(z)/C(z) and the filter order of the adaptive controller should be equal to or higher than that of ⫺ P(z)/C(z) 共which is 128 in our simulation兲 to yield optimum control as far as ANC for wideband noise is concerned. However, in most realistic ANC systems, it is not possible to make the adaptive filter order satisfy the above condition, so the order of the IIR filter was set to 64 for all the algorithms to make the simulations more realistic while still holding quite good performance. The step size parameters of the adaptive algo-

FIG. 5. Schematic diagram of the simulation system. J. Acoust. Soc. Am., Vol. 113, No. 1, January 2003

rithms were adjusted to the extent that any increase to the parameter would cause the control process to be unstable.

B. Convergence speed

The white Gaussian noise with 4000-Hz bandwidth generated by the computer was used as the noise source first, and the convergence speeds of the three algorithms are shown in Fig. 7. It can be seen that there is no dramatic difference among these three algorithms as far as for attenuating white noise. FVLMS algorithm and FULMS algorithm perform almost the same and the proposed LFRLMS algorithm gives slightly better performance than the other two algorithms. Theoretically the convergence performance of the lattice form and direct form IIR filters should be similar for the noise signal with quite flat power spectrum. However, because lattice form adaptive IIR filters are more stable, the convergence coefficient may be set a little larger, resulting in faster convergence speed.

FIG. 6. Impulse responses used for primary path and secondary path with 共a兲 primary path and 共b兲 secondary path. Lu et al.: Lattice filter for active noise control

331

FIG. 7. Convergence comparisons between different algorithms with 共a兲 LFRLMS algorithm; 共b兲 FULMS algorithm; and 共c兲 FVLMS algorithm.

FIG. 9. Learning curves for the FVLMS algorithm with 共a兲 perfect model of the secondary path; 共b兲 secondary path model with a SNR of 5 dB; and 共c兲 secondary path model with a SNR of 0 dB.

C. Robustness to cancellation path model errors

As the most often used FXLMS algorithm in ANC, all three of these algorithms use the model of the cancellation path transfer function. In convergence comparison, the simulation results were calculated with models that were exactly the same as the plant, which may not be the case in practical situations. In order to evaluate the effect of the model errors on the performance of the algorithms, simulations of the same ANC system with noisy cancellation path models were performed. Note that if the errors in the cancellation path model are too large, none of the algorithms will converge. The performance of all these algorithms using exact and noisy plant models are shown in Figs. 8 –10, respectively. It can be seen that even when the cancellation model error becomes quite large 共with SNR 0 dB兲, the proposed LFRLMS algorithm still gives quite good performance while the FVLMS algorithm and the FULMS both deteriorate greatly. This can partly be explained by the following reasons: the lattice structure bears the ability of orthogonalizing the input

signal8,21 and this ability compensates the influence of the cancellation path model error partly and makes the lattice form adaptive IIR filtering algorithm more robust. It also can be seen that FVLMS algorithm deteriorates most when cancellation path model error was added. This is probably caused by the fact that the complex calculation of gradient vector corresponding to the parameters of direct form IIR filter used in FVLMS algorithm makes it more sensitive to the model error. Although simulations here provide some indications for the different robustness of the three algorithms, further theoretical work is ongoing to fully characterize the behavior of all the adaptive IIR filters used in ANC. Another important property that should be noted is that with the addition of cancellation path model error, there is almost no necessity to modify the step size of the proposed LFRLMS algorithm, which makes this algorithm more attractive in practice. This is also owed to the superiority of stability of the proposed algorithm.

FIG. 8. Learning curves for the FULMS algorithm with 共a兲 perfect model of the secondary path; 共b兲 secondary path model with a SNR of 5 dB; and 共c兲 secondary path model with a SNR of 0 dB.

FIG. 10. Learning curves for the LFRLMS algorithm with 共a兲 perfect model of the secondary path; 共b兲 secondary path model with a SNR of 5 dB; and 共c兲 secondary path model with a SNR of 0 dB.

332

J. Acoust. Soc. Am., Vol. 113, No. 1, January 2003

Lu et al.: Lattice filter for active noise control

white-noise signal; and from the learning curve of 50 000 iterations in Fig. 12, it can be seen that LFRLMS algorithm converges on a level approximately 2 dB below the other two algorithms. It should also be noted that the FVLMS algorithm suffers most from the change of noise source probably also due to the complex calculation of gradient vector corresponding to the parameters of direct form IIR filter. Although all the above simulations are based on the impulse responses measured in the anechoic room, similar results can be obtained by using the impulses responses measured in a normal room with room dynamics. The main differences are the longer filter length and heavier computation burden. V. CONCLUSIONS FIG. 11. Learning curves for different algorithms of 5000 iterations with noise source of large power disparity. 共a兲 LFRLMS algorithm; 共b兲 FULMS algorithm; and 共c兲 FVLMS algorithm.

D. Test for noise with large power disparity

In an actual ANC system, the noise signal to be controlled sometimes contains narrow-band components with large power disparity such as fan noise and babble noise. This results in a large eigenvalue spread of the input autocorrelation matrix and will cause the convergence rate of normal LMS algorithm decrease significantly.1,8 To test the efficiency of different algorithms for the controlling of more ‘‘real’’ noise, the summation of 100 sinusoid signals with random frequency between 0 and 3000 Hz were used as the noise for the simulations and all the sinusoid components have random amplitudes between 0 and 0.5 and random initialization phases between 0 and 360 deg. The learning curves shown in Figs. 11 and 12 were obtained from an ensemble of 20 trials. Comparing Fig. 11 with Fig. 7, it can be seen that the convergence rate of LFRLMS algorithms only decreases slightly while both FVLMS and FULMS algorithms converge much slower than that for controlling

In this paper, the full and simplified gradient IIR lattice algorithms for ANC were mathematically derived. Then, the simplified gradient IIR lattice algorithm was tested by using the measured transfer functions from an active noise control system. The simulation results demonstrated that the proposed lattice form adaptive IIR filtering algorithm not only converges faster than the commonly used FULMS and FVLMS algorithms when the noise source consists of sinusoid components with wide power disparity, but also converges to a smaller mean squared error. It also showed that the proposed algorithm is far less sensitive to the cancellation path modeling error, which possibly results in a more robust system in practice. Theoretical analysis of the stability of the proposed algorithm and the implementation of the algorithm in a real-time DSP ANC system are ongoing. ACKNOWLEDGMENTS

The authors are sincerely grateful to Professor P.A. Regalia 共Institut National des Te´ le´ communications, France兲 and Professor Martin Bouchard 共School of Information Technology and Engineering, University of Ottawa兲 for their helpful advice on using the adaptive IIR filters and some calculations concerning simulations. This work was supported by National Natural Science Foundation No. 60272037. APPENDIX A: THE GRADIENT LATTICE ALGORITHM

Initialization: Set the order of the lattice filter M and stepsize ␮ All the filter coefficients and states are set to 0. Lattice filter computation: • Let f M (n)⫽x(n). • for k⫽M , M ⫺1,...,1 do f k⫺1共n兲 cos ␪k共n兲 ⫺sin ␪k共n兲 f k共n兲 ⫽ bk共n兲 sin ␪k共n兲 cos ␪k共n兲 bk⫺1共n⫺1兲



册冋

册冋



end for • b 0 (n)⫽ f 0 (n). • Lattice filter output: FIG. 12. Learning curves for different algorithms of 50 000 iterations with noise source of large power disparity. 共a兲 LFRLMS algorithm; 共b兲 FULMS algorithm; and 共c兲 FVLMS algorithm. J. Acoust. Soc. Am., Vol. 113, No. 1, January 2003

M

y 共 n 兲⫽



k⫽0

b k共 n 兲 ␷ k共 n 兲 .

Lu et al.: Lattice filter for active noise control

333

APPENDIX B: THE SIMPLIFIED GRADIENT LATTICE ALGORITHM

Post filter computation: • Let f cM (n)⫽c(n), where

Initialization: Set the order of the lattice filter M and stepsize ␮ All the filter coefficients and states are set to 0.

N

c共 n 兲⫽

兺 c wi共 i 兲 x 共 n⫺i 兲 .

i⫽0

(c wi (n) (i⫽0,...,n) are the estimated cancellation path impulse response with order n⫹1.)



• for k⫽M ,M ⫺1,...,1 do



册冋

cos ␪ k 共 n 兲 f ck⫺1 共 n 兲 ⫽ b ck 共 n 兲 sin ␪ k 共 n 兲

⫺sin ␪ k 共 n 兲 cos ␪ k 共 n 兲

册冋

f ck 共 n 兲 b ck⫺1 共 n⫺1 兲



⫺sin ␪ k 共 n 兲 cos ␪ k 共 n 兲

册冋

f k共 n 兲 b k⫺1 共 n⫺1 兲



• b 0 (n)⫽ f 0 (n). • Lattice filter output:

• b c0 (n)⫽ f c0 (n). 共Filtered regressor signal corresponding to ␯ k is b ck (n).) • for k⫽1,...,M do Let f ␪ k,M ⫽0. for l⫽M ,M ⫺1,...,1 do if l⫽k

册冋

cos ␪ l 共 n 兲 f ␪ k,l⫺1 共 n 兲 ⫽ b ␪ k,l 共 n 兲 sin ␪ l 共 n 兲 ⫹



⫺b cl 共 n 兲 f cl⫺1 共 n 兲

⫺sin ␪ l 共 n 兲



cos ␪ l 共 n 兲

册冋

f ␪ k,l 共 n 兲 b ␪ k,l⫺1 共 n⫺1 兲

册冋

cos ␪ l 共 n 兲 f ␪ k,l⫺1 共 n 兲 ⫽ b ␪ k,l 共 n 兲 sin ␪ l 共 n 兲



M

y 共 n 兲⫽



k⫽0

b k共 n 兲 ␷ k共 n 兲

Post filter computation: • Let f cM (n)⫽c(n), where N

c共 n 兲⫽

兺 c wi共 i 兲 x 共 n⫺i 兲 .

i⫽0

(c wi (n) (i⫽0,...,n) are the estimated cancellation path impulse response with order n⫹1.) • for k⫽M ,M ⫺1,...,1 do

else



册冋

cos ␪ k 共 n 兲 f k⫺1 共 n 兲 ⫽ b k共 n 兲 sin ␪ k 共 n 兲

end for

end for



Lattice filter computation: • Let f M (n)⫽x(n). • for k⫽M ,M ⫺1,...,1 do

⫺sin ␪ l 共 n 兲 cos ␪ l 共 n 兲

册冋

f ␪ k,l 共 n 兲 b ␪ k,l⫺1 共 n⫺1 兲





册冋

cos ␪ k 共 n 兲 f ck⫺1 共 n 兲 ⫽ b ck 共 n 兲 sin ␪ k 共 n 兲

cos ␪ k 共 n 兲

册冋

f ck 共 n 兲 b ck⫺1 共 n⫺1 兲



end for • b c0 (n)⫽ f c0 (n). 共Filtered regressor signal corresponding to ␯ k is b ck (n).)

end if b ␪ k,0共 n 兲 ⫽ f ␪ k,0共 n 兲 .

Filter regressor:

end l loop.

• Let ␥ M ⫽1.

Filtered regressor signal corresponding to ␪ k : M

b␪k共n兲⫽

⫺sin ␪ k 共 n 兲

兺 ␷lb␪k,l共n兲. l⫽0

• for k⫽M ,M ⫺1,...,1 do Filtered regressor signal corresponding to ␪ k : b ␪ k 共 n 兲 ⫽⫺ ␥ k b yk⫺1 共 n 兲

␥ k⫺1 ⫽ ␥ k cos ␪ k 共 n 兲

end k loop.

end for

Filter coefficient updates:

␷ k(n⫹1)⫽ ␷ k(n)⫹ ␮ e(n)bck(n) ␪ k 共 n⫹1 兲 ⫽ ␪ k 共 n 兲 ⫹ ␮ e 共 n 兲 b ␪ k 共 n 兲

Filter coefficient updates:

␷ k 共 n⫹1 兲 ⫽ ␷ k 共 n 兲 ⫺ ␮ e 共 n 兲 b ck 共 n 兲 • Let g M ⫽1

Test:

• For k⫽M ,M ⫺1,...,1 do

for k⫽1,...,M do if 兩 ␪ k 共 n⫹1 兲 兩 ⬎ ␲ /2

set ␪ k 共 n⫹1 兲 ⫽ ␪ k 共 n 兲 .

冋 册冋

cos ␪ k 共 n 兲 g k⫺1 ⫽ qk sin ␪ k 共 n 兲

⫺sin ␪ k 共 n 兲 cos ␪ k 共 n 兲

册冋

gk ␮ e 共 n 兲 b ␪k共 n 兲



Test: end for 334

J. Acoust. Soc. Am., Vol. 113, No. 1, January 2003

if g k⫺1 ⬍0, set g k⫺1 ⫽g k cos ␪k(n) Lu et al.: Lattice filter for active noise control

7

end for • Let ␣ 0 ⫽g 0 • For k⫽1,2,...,M do

冋 册冋

cos ␪ k 共 n⫹1 兲 ␣k ⫽ 0 ⫺sin ␪ k 共 n⫹1 兲

册冋 册

sin ␪ k 共 n⫹1 兲 ␣ k⫺1 cos ␪ k 共 n⫹1 兲 q k

end for Post filter computation: M b ck (n) ␷ k (n) • y c (n)⫽ 兺 k⫽0

• Let f y M (n)⫽y c (n) • for k⫽M ,M ⫺1,...,1 do



册冋

cos ␪ k 共 n⫹1 兲 f yk⫺1 共 n⫹1 兲 ⫽ b yk 共 n⫹1 兲 sin ␪ k 共 n⫹1 兲 ⫻



f yk 共 n 兲 b yk⫺1 共 n 兲



⫺sin ␪ k 共 n⫹1 兲 cos ␪ k 共 n⫹1 兲



end for • b y0 (n)⫽ f y0 (n). 1

S. M. Kuo and D. R. Morgan, Active Noise Control Systems—Algorithms and DSP Implementations 共Wiley, New York, 1996兲. 2 B. L. Olsen, R. W. Jones, B. R. Mace and C. R. Halkyard, ‘‘Increasing the Convergence Rate of Adaptive Feedforward ANC,’’ in Proceedings of International Symposium on Active Control of Sound and Vibration, Fort Lauderdale, FL, December 1999. 3 C. H. Hansen and S. D. Snyder, Active Control of Noise and Vibration 共E&FN SPON, 1997兲. 4 L. J. Eriksson, ‘‘Development of the Filtered-U Algorithm for Active Noise Control,’’ J. Acoust. Soc. Am. 89, 257–265 共1991兲. 5 D. H. Crawford and R. W. Stewart, ‘‘Adaptive IIR Filtered-v Algorithms for Active Noise Control,’’ J. Acoust. Soc. Am. 101, 2097–2103 共1997兲. 6 L. J. Eriksson, T. A. Laak, and M. C. Allie, ‘‘On-line Secondary Path Modeling for FIR and IIR Adaptive Control in the Presence of Acoustic Feedback,’’ in Proceedings of International Symposium on Active Control of Sound and Vibration, Fort Lauderdale, FL, December 1999.

J. Acoust. Soc. Am., Vol. 113, No. 1, January 2003

A. P. Liavas and P. A. Regalia, ‘‘Acoustic Echo Cancellation: Do IIR Models Offer Better Modeling Capabilities Than Their FIR Counterparts?’’ IEEE Trans. Signal Process. 46共9兲, 2499–2504 共1998兲. 8 S. Haykin, Adaptive Filter Theory 共Prentice-Hall, Englewood Cliffs, NJ, 1991兲. 9 P. A. Regalia, Adaptive IIR Filtering in Signal Processing and Control 共Dekker, New York, 1995兲. 10 S. Horvath, Jr., ‘‘Lattice Form Adaptive Recursive Digital Filters: Algorithms and Applications,’’ in Proceedings of IEEE Int. Symp. Circuits Syst., pp. 128 –33 共1980兲. 11 P. A. Regalia, ‘‘Stable and Efficient Lattice Algorithms for Adaptive IIR Filtering,’’ IEEE Trans. Signal Process. 40共2兲, 375–388 共1992兲. 12 K. X. Miao, H. Fan, and M. Doroslovaeki, ‘‘Cascade Lattice IIR Adaptive Filters,’’ IEEE Trans. Signal Process. 42共4兲, 721–741 共1994兲. 13 R. Lopez-Valcarce and F. Perez-Gonzalez, ‘‘Adaptive Lattice IIR Filtering Revisited: Convergence Issue and New Algorithms with Improved Stability Properties,’’ IEEE Trans. Signal Process. 49共4兲, 811– 821 共2001兲. 14 D. C. Swanson, ‘‘Lattice Filter Embedding Techniques for Active Noise Control,’’ in Proceedings of International Congress and Exposition on Noise Control Engineering, pp. 165–168 共1991兲. 15 N. C. Mackenzie and C. H. Hansen, ‘‘The Use of an Alternative Adaptive Algorithm with a Lattice Structured Filter for a Multi-channel Active Noise or Vibration Control System,’’ in Proceedings of International Congress and Exposition on Noise Control Engineering, pp. 177–180 共1991兲. 16 K. Char and S. M. Kuo, ‘‘Performance Evaluation of Various Active Noise Control Algorithm,’’ in Proceedings of International Congress and Exposition on Noise Control Engineering, pp. 331–336 共1994兲. 17 S. M. Kuo and J. Luan, ‘‘Cross-coupled Filtered-X LMS Algorithm and Lattice Structure for Active Noise Control Systems,’’ in Proceedings of IEEE Int. Symp. Circuits Syst., pp. 459– 462 共1993兲. 18 H. J. Lee, Y.-C. Park, C. Lee, and D. H. Youn, ‘‘Fast Active Noise Control Algorithm for Car Exhaust Noise Control,’’ IEE Electron. Lett. 36共14兲, 1250–1251 共2000兲. 19 Y. C. Park and S. D. Sommerfeldt, ‘‘A Fast Adaptive Noise Control Algorithm Based on the Lattice Structure,’’ Appl. Acoust. 47共1兲, 1–25 共1996兲. 20 Y. Tu and C. R. Fuller, ‘‘Multiple Reference Feedforward Active Noise Control. II. Reference Preprocessing and Experimental Results,’’ J. Sound Vib. 233共5兲, 761–774 共2000兲. 21 C. F. N. Cowan and P. M. Grant, Adaptive Filters 共Prentice-Hall, Englewood Cliffs, NJ, 1985兲. 22 S. J. Chen and J. S. Gibson, ‘‘Feedforward Adaptive Noise Control with Multivariable Gradient Lattice Filters,’’ IEEE Trans. Signal Process. 49共3兲, 511–520 共2001兲.

Lu et al.: Lattice filter for active noise control

335

Lattice form adaptive infinite impulse response ... - Semantic Scholar

State Key Laboratory of Modern Acoustics and Institute of Acoustics, Nanjing University, Nanjing 210093,. China .... cos l•Fl z sin •zBl 1 z. 11. Bl z sin l•Fl z cos •zBl 1 z . For k l, there is. F k,l 1 z. B k,l z .... erated by the computer was used as the noise source first, and .... nology and Engineering, University of Ottawa for their.

139KB Sizes 0 Downloads 273 Views

Recommend Documents

SVM Optimization for Lattice Kernels - Semantic Scholar
[email protected]. ABSTRACT. This paper presents general techniques for speeding up large- scale SVM training when using sequence kernels. Our tech-.

SVM Optimization for Lattice Kernels - Semantic Scholar
gorithms such as support vector machines (SVMs) [3, 8, 25] or other .... labels of a weighted transducer U results in a weighted au- tomaton A which is said to be ...

RESPONSE CHARACTERISTICS OF RADIATION ... - Semantic Scholar
Acknowledgement: This research was supported by Southern California Edison under contract No. 8T073901 while at San Diego State University. This work became possible with diligent support from. David Deane and Kathryn McCarty while both were at San D

QRD-RLS Adaptive Filtering - Semantic Scholar
compendium, where all concepts were carefully matured and are presented in ... All algorithms are derived using Givens rotations, ..... e-mail: [email protected].

ADAPTIVE KERNEL SELF-ORGANIZING MAPS ... - Semantic Scholar
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT ..... The central idea of the information theoretic learning proposed by Principe et al. ...... Technology degree from Department of Electronics and Communication Engineering.

decentralized set-membership adaptive estimation ... - Semantic Scholar
Jan 21, 2009 - new parameter estimate. Taking advantage of the sparse updates of ..... cursive least-squares using wireless ad hoc sensor networks,”. Proc.

Fractional Order Adaptive Compensation for ... - Semantic Scholar
ing the FO-AC is much smaller than that using the IO-AC. Furthermore, although the ... IEEE Trans. on Ind. Electron., 51:526 – 536, 2004. D. Y. Xue, C. N. Zhao, ...

Adaptive Algorithms Versus Higher Order ... - Semantic Scholar
sponse of these channels blindly except that the input exci- tation is non-Gaussian, with the low calculation cost, com- pared with the adaptive algorithms exploiting the informa- tion of input and output for the impulse response channel estimation.

QRD-RLS Adaptive Filtering - Semantic Scholar
although one chapter deals with implementations using Householder reflections. ...... For comparison purposes, an IQRD-RLS algorithm was also implemented. ..... plications such as broadband beamforming [16], Volterra system identification ...

Fractional Order Adaptive Compensation for ... - Semantic Scholar
1. J. µ + B1)Vd(s). −. µs1−ν. J vd(s)+(. µ. Js. + 1)vd(0). (36). Denote that ν = p q. , sν = s p q , ..... minimization. IEEE Trans. on Ind. Electron., 51:526 – 536, 2004.

ADAPTIVE KERNEL SELF-ORGANIZING MAPS ... - Semantic Scholar
4-5 Negative log likelihood of the input versus the kernel bandwidth. . . . . . . . . 34. 5-1 Kernel ...... He received his Bachelor of. Technology degree from ...

QRD-RLS Adaptive Filtering - Semantic Scholar
useful signal should be carried out according to (compare with (11.27)) ..... plications such as broadband beamforming [16], Volterra system identification [9],.

Hybridization and adaptive radiation - Semantic Scholar
and computer simulations, have demonstrated that homoploid hybrid speciation can be .... predicts a phylogenetic signature that is recoverable with the use of ...

QRD-RLS Adaptive Filtering - Semantic Scholar
Cisco Systems. 170 West Tasman Drive, ... e-mail: [email protected]. Jun Ma ..... where P = PMPM−1 ···P1 is a product of M permutation matrices that moves the.

lattice gas cellular automaton modeling of surface ... - Semantic Scholar
now equal to Lpl, and we define a uniform random vari- able that determines where the adatom, if created, will land. Thus, we require only two random number gen- erations. Once the adatom lands on a particular lattice site r, we increase the height a

Grain Yield Response Of Rice Cultivars Under ... - Semantic Scholar
when evaluated under controlled irrigation condition. ... irrigation condition and b) upland condition with .... cultivar performance for cultivar x location data.

Soil Carbon and Nitrogen Storage in Response to ... - Semantic Scholar
Jul 6, 2006 - prevents invasion of woody species (Axelrod, 1985;. Wright, 1980; Archer et al., 1988, 2001). Therefore, prescribed fire is often employed as a land management tool to suppress the encroachment of woody plants into grass-dominated ecosy

Response to the discussion of “Gaussian Process ... - Semantic Scholar
of a Gaussian process regression model for the calibration of multiple response .... like to acknowledge the financial support from the EPSRC KNOW-HOW ...

WSTF regulates the H2A.X DNA damage response ... - Semantic Scholar
Dec 17, 2008 - response and expands our knowledge of domains that contain intrinsic ..... Full Methods and any associated references are available in the ...

Distributed forensics and incident response in the ... - Semantic Scholar
the investigator complete access to the underlying filesystem and memory. .... connecting to the Internet over a Network Address Trans- lation device, or behind a .... tions create a dedicated console process that waits for the client to complete ...

On the Dynamic Nature of Response Criterion in ... - Semantic Scholar
pants study items from different taxonomic categories, with cate- gories studied ... recognition test containing targets and distractors from each class of stimuli.

Ferritin and the response to oxidative stress - Semantic Scholar
*Department of Biochemistry, Wake Forest University School of Medicine, Winston-Salem, NC 27157, U.S.A., †The Comprehensive Cancer Center,. Wake Forest ...

An Adaptive Weighting Approach for Image Color ... - Semantic Scholar
video mobile phones are popular and prevalent. Mobility and ... interpolation (here we call such coalition demosizing). By this .... Left: by AW; center: by. DBW ...

An Adaptive Weighting Approach for Image Color ... - Semantic Scholar
Embedded imaging devices, such as digital cameras and .... enhancement”, Signals, Systems and Computers, 2000, Vol. 2, pp. 1731- ... House (Proposed).