Research On Blind Source Separation Algorithms And Its Application

Blind source separation (BSS) is a newly presented and lively area of signal processing domain in recent years, which estimates sources given mixed signals without prior knowledge such as sources number, location and mixing process. BSS is a powerful tool of data analysis and signal processing, and it has been applied in many areas such as biological and medicine signal processing, data mining, speech signal processing, pattern recognition, wireless communication and so on.Considering its wide application and ability of data processing, Researchers have paid a great deal of attention to BSS and made a great progress in the past twenty years. This thesis researches the separation algorithms of speech signals. Firstly, a brief introduction of the development history and current research status are summarized, and several applications of the BSS are given. Then, this thesis introduces several kinds of important algorithms of BSS.This thesis discusses the theories and algorithms of BSS, and especially researches the underdetermined blind source separation problem. The researches concentrate on the following topics:The blind separation problem of convolutive mixtures and delayed mixtures is discussed in this thesis. The separation when there are more sources than mixtures is researched emphatically. The main works of this thesis are as follows:The permutation indetermination problem in the frequency domain blind source separation is discussed. By utilizing the characteristic that amplitude correlation between neighbor bins of the same signal is better than that of different signals, an improved method based on the amplitude correlation between neighbor bins to eliminate the permutation indetermination is proposed. It is feasible to implement ICA algorithm and permutation method simultaneously.The underdetermined blind source separation algorithms in time-frequency domain are researched. A novel blind source separation algorithm based on computational auditory scene analysis (CASA) is proposed, which can separate several sources from two sensor signals by clustering to interaural time differences (ITD) and interaural intensity differences (IID). A speech blind separation algorithm for convolute mixture is proposed, which can separate convolved mixtures in underdetermined case by clustering in time-frequency domain of mixtures.The blind separation problem for sources that are sparse insufficiently is researched. A novel two-step underdetermined blind source separation algorithm is presented, which estimates mixing coefficient more efficiently using clustering algorithm based on grid and density, and it estimates the mixing matrix better. When recovering source signals, a simpler method is used to get l1, norm minimization solution.An algorithm of sparse sources recovery based on matching pursuit (MP) is proposed. Considering the application of MP in sparse sources recovery in blind source separation, this paper improves classical MP algorithm and has obtained a better performance. Proposed method works well even the mixing matrix is ill-conditioned by reduce the error when match failed.