Adaptive noise cancellation using modi ed simulated annealing algorithm

Abstract

Adaptive Noise Cancellation (ANC) entails estimation of signals corrupted by additive noise or other interference. ANC utilizes a \reference" signal correlated in some way with the \primary noise" in the noise cancellation process. In ANC, the reference signal is adaptively ltered and thereafter subtracted from the \pri- mary" input to obtain the desired signal estimate. Adaptive ltering before the subtraction process allows for handling of inputs that are either deterministic or stochastic, stationary or time varying. ANC has been widely applied in the elds of telecommunication, radar and sonar signal processing. The performance and e ciency of ANC schemes is based on how well the ltering algorithm can adapt to the changing signal and noise conditions. It is worthwhile focusing on developing better variants of AI algorithms from the point of view of ANC. This thesis is focused on: development of a modi ed version of the Simulated Annealing (SA) algorithm and its application in ANC. This is alongside an anal- ysis of the e ectiveness of the standard and modi ed SA algorithms in ANC in comparison to standard Least Mean Square (LMS) and Normalized Least Mean Square (NLMS) algorithms. Signals utilized in this study include: sinusoidal signals, fetal electrocardiogram signals and randomly generated signals. The modi ed SA algorithm has been developed on the basis of making modi ca- tions to the control parameters of the standard SA on the basis of the acceptance probability and the cooling schedule. A low complexity acceptance probability scheme has been proposed. The proposed cooling schedule is iteration-adaptive to improve on algorithm convergence. The ANC problem is formulated as a min- imization problem entailing the minimization of the di erence between a noise contaminated signal and a weighted estimate of the noise content. This is achieved through optimal ANC tap-weight adjustment. The algorithms under study are applied in the weight generation process with the expected outcome as ideally a noise free signal. In this evaluation, performance measures analyzed in the study are mis-adjustment and convergence rate. To evaluate these, Euclidean distances and the correlation factors between the desired signal and the ltered signal are applied. In the said analysis the improved SA is found to generate the minimal error and fast execution speed in ANC compared to standard SA, LMS and Normalized LMS. The main contribution done in this study is the validation of the application of modi ed SA algorithm in adaptive lters. This has been done through a series of simulations involving the SA algorithm in a MATLAB environment. In addition, through improvements made on the standard SA algorithm, the convergence rate of SA has been increased alongside the overall solution accuracy.