Research And Realization On Genetic Algorithms Based Blind Signal Separation Methods

As a class of signal processing technique ,blind signal separation (BSS) is used to recover the original signals from the observed ones with little prior-information about sources and mixture system. So its research becomes important both theoretically and practically in the fields of biomedical, speech recognition, communication, etc. By now BSS has been a topical subject in modern signal processing fields.According to the different mixture models of sources, linear BSS can be divided into two classes: instantaneous mixture BSS and convolutive mixture BSS. In this dissertation, some classical algorithms of recent BSS have been studied. The main contributions are as follows:Reviewing the research situation of BSS over recent decades systematically and choosing some popular BSS algorithms to study and analyze by simulated experiments. Two common defects are founded in BSS algorithms based on independent component analysis (ICA):①The selection of none-linear function (NLF)in ICA depends on the kurtosis of original signals, which degrades the performance of separation seriously when the observed signals are the mixture of Super-Gaussian and Sub-Gaussian signals.②As the optimization approach in ICA for searching separation matrix(or separation matrixes), Gradient has shortcomings that configuration of initial value and the length of pace will make the separation algorithms strap into local optimum values .In this thesis, two new BSS algorithms are proposed respectively based on two different mixture models. One is the BSS based on kernel density estimation (KDE) and genetic algorithm (GA), the other is the blind deconvolution based on high order cross cumulants and GA. Without NLF, the performance of separation in both algorithms is independent with the kurtosis of the sources. On the other hand, global-searching GA avoids the defect of gradient so that the real optimum values can be found in the end. The correctness and validation of the two algorithms are testified by theory analysis and simulated experiments. The simulated results also show that the proposed methods are superior to the traditional ICA.