The Research Of Blind Separation For Temporally Correlated Sources

Blind Source Separation(BSS) is a technology that can recover the sources signal only according to several sensor signals from a unkown mixing system. In recent years, BSS is becoming an attractive problem of signal processing area, and it has been used in many domains such as Biomedical Engineering, wirless communication and data mining.BSS includes two methods that respectively based on high-order statistics(HOS) and second-order statistics(SOS). Most of BSS algorithms suppose the independence of the source signals, such as ICA by maximization of nongaussianity and by maximum likelihood estimation. These algorithms always use the high-order statistics. However, if the source signals have temporal structures, only the second-order statistics can work well.In this thesis, the principles and algorithms of BSS are discussed. Firstly, several Independent Component Analysis algorithms are introduced. Secondly, we discuss the algorithms based on SOS and some preprocessing methods. Finally, we research the Blind Source Extraction(BSE) of temporally correlated sources. BSE is an effective method of BSS which can extracte the interesting signals.We present a blind extration algorithm of nonstationary source based on the SOS. The proposed algorithm has a low computed complication and is robust in noisy environment. It can be used to the separation of nonstationary signals such as speech and music. Computer simulations show the better performance of the porposed method. In this thesis, we try to use the auditory model in BSS. We use a group Gammatone filers which can simulate the cochlear to replace the linear predictor. This application can improve the performance in noisy environment.