Research On Frequency Domain Algorithm For Convolutive Blind Source Separation

The blind source separation is a new signal processing technology that estimates the source signal and the mixed channel when only the mixed signal is known.Among them,the blind source separation of convolutive mixtures considering reflection,refraction and noise interference has become a research focus in recent years because the model is closer to the real signal transmission environment.The process of frequency-domain blind source separation of convolutive mixtures is to transform the convolutional mixed signal in the time domain into instantaneous mixed signal at each frequency bin in the frequency domain by Short Time Fourier Transform,and perform instantaneous mixed source separation at each frequency bin.But before restored to the time-domain signal by inverse Short Time Fourier Transform,the scaling and permutation ambiguity must be solved.Due to its low computational complexity and wide applicable environment,the frequency domain algorithm has a wider range of applications.The focus of this thesis is on the research of frequency-domain algorithms for convolutive blind source separation in speech signals environment and radar signals environment.Firstly,two types of ambiguity in the frequency-domain convolutive blind source separation is studied.In this thesis,the minimum distortion method is used to solve the scaling ambiguity.Aiming at the permutation ambiguity,two traditional sorting algorithms are first introduced,namely,the amplitude correlation algorithm(Murata algorithm)and the direction of arrival estimation algorithm(DOA algorithm).After introducing the principles,advantages and disadvantages of the two permutation algorithms,we propose an improved algorithm based on Murata algorithm,which is named the Murata algorithm based on influence weights(SP-Murata algorithm).The algorithm is aimed at the problem of poor robustness of the Murata algorithm.By introducing spacing influence weight and performance influence weight,the effect of frequency space and separation performance at each frequency bin are used as reference to sort.The improved algorithm controls the influence of the frequency bins sorted in the neighborhood on the frequency bins unsorted.Simulation experiments show that under the convolutive model of speech signals and radar signals,the improved algorithm has significantly improved separation performance over the traditional amplitude correlation algorithm.Secondly,in order to solve permutation ambiguity in the frequency-domain convolutive blind source separation,another method is described in this thesis named independent vector analysis algorithm(IVA algorithm).This algorithm can complete the separation process while avoiding permutation ambiguity in frequency domain.For the independent vector analysis algorithm,we first introduce its algorithm principle and mathematical model.For computational complexity and complex iterative optimization of the IVA algorithm,the variable step size is introduced to improve the time efficiency of the algorithm.Then,the influence of the step size change rate in variable step sizes on the separation performance and time complexity of the algorithm is researched.Based on this,an independent vector analysis algorithm based on iterative step parameter optimization(Fast-IVA algorithm)is proposed.Simulation experiments show that the speech and radar signals in the convolution environment correspond to their respective optimal step change rates.At the same time,the Fast-IVA algorithm under the effect of the respective optimal step size change rate can take into account the algorithm performance and time efficiency.Finally,through simulation experiments,two frequency-domain algorithms of convolutive blind source separation proposed in this thesis are compared,and the applicable conditions of the two algorithms are given in the speech and radar convolution environments,respectively.The influence of the order of the impulse response function on the two algorithms is discussed in the radar environment,which provides new ideas for future research.