Research On Nonlinear Active Noise Control System Based On Function Link Artificial Neural Networks

The rapid development of industrialization and urbanization has brought along with the serious noise pollution,which opens up an interesting research area of noise control.Technically,the field of noise control can be broadly classified into two domains: passive and active.Passive noise control(PNC)technique using special materials to absorb or/and isolate unwanted acoustic noise is usually bulky,difficult to install,expensive,and inefficient at low-frequency noise.In contrast,active noise control(ANC)technique can overcome the disadvantages of PNC method and meet the requirements of effectively eliminating or reducing the low-frequency noise.With the development of electronic technology and adaptive processing theory,ANC technique is getting more and more attention since it permits improvements in noise control with potential benefits in weight,size and cost.This dissertation mainly focuses on the new design of nonlinear adaptive controller using FLANN for ANC system to improve noise cancellation performance and reduce computational complexity,in the presence of strong nonlinearity.It involves the following aspects: Firstly,we investigate the performance of FLANN-based ANC system,including analyzing the nonlinear characteristics in the ANC system,analyzing the nonlinear modeling capability of FLANN for the ANC system;Secondly,based on these analyses,various controllers structures based on FLANN with cross-terms for the ANC system are proposed.In addition,novel algorithms based on filtered-error technique and data-dependent partial update strategy are designed to further reduce the computational burden.The main contributions of this thesis are given as follows:(1)Based on the analysis of the nonlinear characteristics in the components of the ANCsystem,various types of nonlinear distortions commonly encountered in actual ANCsystems are identified.At the same time,analysis of the nonlinear model capability ofFLANN pointed out that the trigonometric nonlinear functional expansion is a set ofmemory-less functional links(i.e.,each functional link depends only on the currentinput sample),and lacks the basic functions to represent the cross-terms.Thereby,atheoretical basis is offered to understand better the behavior of the FLANN-based ANCsystem for nonlinear distortion types caused by components in the ANC system.Thecauses for such behavior have also been analyzed.Simulation experiments are alsoperformed to confirm the validity of this analysis.(2)By introducing suitable cross-terms into the trigonometric nonlinear functionalexpansion and using a simplified channel-diagonal method,a novel simplifiedgeneralized FLANN(SG-FLANN)filter is proposed for the nonlinear ANC system toimprove the noise cancellation performance.Furthermore,to reduce the computationalcomplexity,an M-max simplified generalized filtered-error least mean square(MmSGFE-LMS)algorithm is designed for SG-FLANN controller.The algorithm usesthe filtered-error technique to avoid the computational cost of filtering the signalsthrough the secondary path,and employs the data-dependent partial update strategy toreduce the computational cost of updating filtering weights.Simulation results showthat the proposed system achieves good noise cancellation performance in the presenceof strong nonlinearity.(3)Inspired by the bilinear filtering approach,a novel bilinear FLANN(BFLANN)filterfor the NANC system is proposed.The BFLANN exploits cross-terms based on bothfeedback and feedforward polynomials,thus it can accurately model nonlinear systemswith shorter memory length than that of the SG-FLANN.To avoid the instabilityproblem as in the case of bilinear filters,a sufficient condition BIBO(bounded inputbounded output)is offered to BFLANN.Furthermore,in virtue of the leaky techniqueand data-dependence partial update strategy,the M-max leaky bilinear filter-error LMS(MmLBFE-LMS)algorithm is designed for BFLANN,which can reduce thecomputational burden and increase the algorithm stability.Simulation results andcomputational analysis show that the ANC system using BFLANN outperforms theANC systems using SG-FLANN.(4)By incorporating the concept of exponential sinusoidal models and cross-terms into thetrigonometric nonlinear functional expansion,and using channel-reduced diagonal(CRD)structure,a novel generalized exponential FLANN with CRD structure(GE-FLANN-CRD)filter for the nonlinear ANC system is proposed.Based on theexponential nonlinearity,sinusoidal nonlinearity,power nonlinearity and memorynonlinearty,the nonlinear processing capability of the GE-FLANN-CRD filter isenhanced in nonlinear ANC.Furthermore,to reduce the computational cost of updatingthe exponential factor,the M-max generalized exponential filter-error LMS(MmGEFE-LMS)algorithm is derived for GE-FLANN-CRD controller.To guaranteeconvergence for the algorithm,the stability conditions are also discussed.Even thoughthe GE-FLANN-CRD controller requires a higher computational complexity than theSG-FLANN and BFLANN controllers,it exhibits a better noise-canceling performance.(5)Based on the effectiveness of the pipelined architecture,a novel pipelined generalizedFLANN(P-GFLANN)filter is proposed to reduce the computational complexity for thegeneralized FLANN(GFLANN)in ANC applications.Thanks to the engineeringprinciple of divide-and-conquer and biological principle of modules,the proposedP-GFLANN can significatly reduce the computational burden and can be furtherimprove the performance.However,due to the coupling mechanism between the linearand nonlinear parts of pipelined architecture,the computational complexity of theadaptive algorithm for P-GFLANN filter is quite large.To avoid this disadvantage,ahierarchical update P-GFLANN(HUP-GFLANN)filter for ANC is presented.Data-dependent hierarchical M-Max filtered-error LMS(HMmFE-LMS)algorithm isderived to selectively update coefficients of the HPU-GFLANN filter,which canfurther reduce the computational complexity.Theoretical analysis and simulationresults both demonstrate that the HUP-GFLANN controller outperforms theP-GFLANN and GFLANN controllers.