A Study On Blind Source Separation In Non-underdetermined System

Signals received by communications equipment always are the mixture of signals in time domain or frequency domain, and signals are affected by multipath channel during transmission. In some cases, the prior knowledge of the source signal and the transmission system is not known. The traditional signal processing method shows insufficient gradually for non-stationary time-varying channel. Therefore, the development of new type of blind source separation technology is becoming a challenging topic in communication field, and its flexibility, stability and efficiency will become a brand-new concept of communication technologyBased on the introduce of the basic theory of Blind Source Separation(BSS), this paper focuses on the application of the algorithm on separating hybrid fixes signals and Time-delay convolution speech signals. The main works of this paper can be summarized as follows.In chapter1, we introduce the historical background, classification, research status and development trend of the BSS. And establish BSS basic model of the instantaneous mixed and convolution mixed signals.In chapter2, the basic information theory involved in the BSS is summarized, the structure of the control algorithm and the performance index of corresponding algorithm functions, namely string sound interference index and coherent coefficient index, are introduced to present the qualitative and quantitative comparison of the constructing algorithm.In Chapter3, a new blind source separation algorithm is proposed which is based on the kernel function. Reference to the signal’s kurtosis, the signals can be divided into sub-gauss or sup-gauss mixture signals.Firstly, the proposed algorithm will bring in a nonlinear kernel function and mapping the separated signals to a nonlinear kernel space at the same time, what’s more, by optimizing the smoothing parameter and updating the mixed separation matrix step by step to separate the signals effectively. According to the simulation results, the new algorithm shows better significantly performances both in convergence rate and steady state aspects comparing with the EASI, Whitening, Natural Gradient method, meanwhile, it can be adapt in non-stationary environment.In Chapter4, to separate the time-delayed convolution speech signals, a frequency BSS algorithm which is based on the multi-channel noise offset is studied, the proposed method combines the BSS technology and the array processing technology, evenmore, the airspace information in frequency domain adequately, first, by using the spectrum abruption technology, the mixtured signals is divided into several independent signals in frequency domain, there is one principal component signal, and the others which are leak into it is considered as disturb signals. Then, the noise offset technology is used to separate the mixed signals independently, as a result, the frequency domain recovery signals can be gained. To show the good solution, the computer simulation result is given, a higher correlation coefficient can be gained by this algorithm, meanwhile, watching from the speech spectrogram, the spectrogram of the recovery speech signals can be observed clearly which indicated that the new algorithm is a good solution to such mixed signals.The last chapter is a summary of the dissertation, mainly depicts the disadvantage and some study directions in BSS fields which make an important role for the later research.