Research On Overdetermined Blind Source Separation Technology And Application

As a powerful branch of signal processing,blind source separation has been successfully applied in the fields of speech signal processing,communication signal processing,biomedical and image signal processing.Overdetermined blind source separation refers to the process of estimating each source signal based on only the observed mixed signals without knowing any prior information of the source signals and the mixing matrix of channel when the number of received signals is greater than the number of source signals.For different mixing models in blind source separation,this thesis mainly investigates on two blind separation algorithms under instantaneous mixing and convolutive mixing.The specific investigations and achievements are listed as follows:(1)For the convolutive mixed source signals,based on the existing non-orthogonal joint block diagonalization algorithm,a convolutive blind separation algorithm based on filter order estimation is proposed and applied into separating convolutive mixed speech signals in the near-field environment.By using the eigenvalue decomposition of the mixed signal autocorrelation matrix and setting the threshold value of the adjacent eigenvalues to obtain the number of source signals and the filter order,this algorithm effectively solves the initial value selection problem in the non-orthogonal joint block diagonalization algorithm.By introducing pre-whitening process for the dimension reduction of the non-square equivalent mixing matrix,the limitation that the hybrid matrix must be a square matrix in the fast nonorthogonal joint block diagonalization algorithm is effectively eliminated.Furthermore,the experimental results show that,the proposed method improves performance on separating mixed signals over the existing methods.(2)For the instantaneous mixed source signals,according to the vandermonde characteristic of mixed frequency-hopping signal matrix,a single source point(SSP)detection algorithm based on two time-frequency ratios is proposed for the network station sorting application of frequency-hopping mixed signals in the far-field environment.The proposed detection method effectively avoids the large complexity of the existing SSP detection algorithms,whose simulation time is long when the number of simulated points is large.The simulation results show that,the estimation error of the hybrid matrix is never decreased but with half simulating time by using the proposed algorithm.For the frequency-hopping signal with noises and interferences,since the non-single source point is easily misjudged as a SSP,the estimation error of the direction of arrival(DOA)then increases.A SSP correction algorithm based on the start and stop position detection of SSP is proposed,which effectively removes the influence of the pseudo SSP on the estimation of DOA,by utilizing the detected start and stop position detection of true SSP for DOA estimation of true single source points.(3)For the poor separation of mixed frequency hopping signals in complex frequencyhopping networks,a blind separation algorithm for synchronous orthogonal frequencyhopping signals under multiple frequency-hopping periods,and a blind separation algorithm for asynchronous non-orthogonal frequency-hopping signals based on time-domain segmentation are proposed.Specifically,the former mainly utilizes the time-frequency information extraction algorithm of frequency-hopping signals to segment the mixed frequency-hopping signals according to the hopping period,and separately processes the each segmented signals for mixing matrix estimation and signal separation,finally concatenates them according to the clustering centers of DOAs.The simulation results show that,although in the undesired case that signal-to-noise ratio(SNR)is low and the number of source signals increases,the proposed method estimate the DOA and the mixing matrix effectively,and ensure better separation.Besides,the latter segments the time-domain mixed frequency-hopping signals by using the estimation of hopping times,uses different DOA estimation strategies and blind separation algorithms according to different numbers of each part of source signals.The simulation results show that the proposed methods can effectively estimate the DOA of each source signal and ensure better separation performance under low SNR.