Research On Blind Source Separation Based On Improved FastICA

With the development of blind source separation(BSS)technology,its applications will be extensive increasingly.At the same time,some problems will be discovered when BSS technology was used.The first one is how do you estimate the number of source signals in the BSS.In other words,if the relationship between the number of sensors and the number of source signals cannot be judged,the type of BSS cannot be determined.Therefore,it is necessary to estimate the number of source signals.The second one is separation performance of BSS algorithm in noisy environment.Now BSS with noise we concerned is that noise is to be solved treated as a kind of independent signal.This technique has been very mature and has been successfully applied in many fields.However,in practical engineering application,noise is not mixed into the source signal as an independent signal,so the blind source separation of mixed signal with noise is worth studying.Finally,the stability problem of BSS algorithm including the disorder of the separation signal,the uncertain amplitude of the separation signal and selecting the initial value of the random iterative matrix when we optimize the target function,these problems may lead to the instability of BSS algorithm itself.The thesis mainly focuses on the problems mentioned above.Firstly,based on empirical mode decomposition and singular value decomposition(EMD-SVD)approach to estimate the number of source signals is studied,the method will be out of work when the experiment is repeated many times.The reason for the above problems is that the decomposition scale of the empirical mode decomposition(EMD)is inconsistent and the modal mixing when the signal is decomposed.As a result,the algorithm cannot accurately estimate the number of source signals.Therefore,the thesis improves the original algorithm and proposes ensemble empirical mode decomposition and singular value decomposition(EEMD-SVD)to solve the shortcomings of EMD-SVD.Secondly,the thesis studies the performance of the BSS algorithm with noisy audio signals and proposes a scheme of double noise reduction that it can improve the separation effect of the BSS algorithm in the noise environment,the reliability of the scheme is proved by simulation experiment.Finally,the thesis found that the performance of the algorithm is greatly affected by the initial value of the random iteration matrix and step size by the study of the gradient algorithm,the ICA algorithm,equivariant adaptive separation via independence(EASI)algorithm,kurtosis algorithm and fast independent component analysis(FastICA)algorithm.Consequently,the steepest descent method is proposed to improve the sensitivity of the random iteration matrix to the initial value of FastICA algorithm by the thesis,simulation results show that the improved algorithm is correct.