Study On The Algorithm Of PSO Instantaneous And Frequency Domain Convolution Blind Source Separation

Blind Source Separation(Blind Source Separation,BSS)is a signal processing method that uses a corresponding algorithm to reconstruct or extract the original signal from the received multi-path mixed signal.Among them,the instantaneous blind source separation has the problems of convergence speed and stability,especially the blind source separation based on the improved intelligent algorithm.Therefore,how to solve this problem becomes the current research hotspot.In practical applications,most of the received mixed signals are more in line with the convolutional hybrid model.The complexity of convolution blind source separation,the unsatisfactory separation effect and the practical operation makes the research still a difficult point at home and abroad.This thesis will study the particle swarm optimization(Particle Swarm Optimization,PSO)algorithm of the instantaneous model and the convolution model frequency domain algorithm,including the following aspects:(1)Aiming at the problem that the intelligent algorithm has slow convergence speed and large steady-state error in the instantaneous blind source separation,a blind source separation algorithm based on inertia weight and learning factor is developed.The algorithm is based on the iteration of inertia weight,and the learning factor makes corresponding linear or nonlinear changes,which enhances the interaction of two parameters,which not only accelerates the convergence speed,but also strengthens the steady-state error,and achieves blind source separation of multiple BOC mixed signals is.(2)Aiming at the shortcomings of intelligent algorithms which are easy to fall into local optimum when implementing instantaneous blind source separation,a blind source separation algorithm based on Givens transform and introducing second-order oscillations is proposed.The algorithm introduces the second-order oscillating link based on combination of inertia weight and learning factor,so that the diversity of the particle is not reduced by iteration,effectively avoiding the algorithm falling into local extreme.At the same time,the Givens transform is used to change the separation matrix rotation parameter into the representation of the rotation angle.Reduce the complexity of the algorithm.The stability of the algorithm is verified by the blind source separation of speech signals.(3)Aiming at the problems of sorting ambiguity and low separation precision of convolution blind source separation frequency domain algorithm in practical application,this thesis studies a frequency domain blind source separation sorting algorithm based on region growth and power ratio correlation.The algorithm first converts the convolution mixed signal into the frequency domain by short-time Fourier transform,and establishes an instantaneous model at each frequency point for separation.Secondly,the separated signal power is compared with the correlation by point-by-point sorting,and then the result is divided according to the threshold.It is a small number of small areas,and is replaced and sorted by regional growth;finally,the correct separated signal obtained in the frequency domain is converted to the time domain.The algorithm minimizes the phenomenon of frequency point sorting error diffusion and realizes the convolution blind source separation of speech signals.In this thesis,the particle swarm optimization algorithm is researched and improved and applied in the instantaneous blind source separation,which improves the separation performance to some extent,and provides more research ideas for the application of intelligent algorithm in blind source separation.On the other hand,this thesis analyzes and experimentally simulates the improved algorithm of convolution blind source separation,and compares and analyzes the different convolution blind source separation algorithm by using different performance indicators.Compared with other algorithm,the improved algorithm has more good performance.