Research On Blind Source Separation Of Nonlinear Mixed Signals

Blind source separation is an important part of blind signal procession. At present, most algorithms based on linear blind source separation, but many signals are obtained by nonlinear mixture in real-world, linear algorithms do not work in this condition. Consequently, the research on nonlinear blind source separation not only has great significance for practical applications, but also has great significance for enlarging application fields.This paper explained fundamental theories of existing nonlinear blind source separation algorithms, introduced mathematical models and cost functions which needed by nonlinear blind source separation, indicated advantages and disadvantages of nonlinear blind source separation. To the problems of nonlinear blind source separation, this paper carried on research from two aspects as follo wings:Firstly, nonlinear blind source separation that bases on post non-linear model algorithm needs to compute inverse function. Particle swarm optimization algorithm can be used to compute inverse function, but this algorithm has the problem of premature convergence. To solve this problem, we improve this algorithm and obtain chaos particle swarm optimization by utilizing fitness variance, satisfactory solution and chaos, which can make particles get out of local optimum quickly and solve premature problem. Then we used improved particle swarm optimization algorithm to compute inverse function, this algorithm can improve convergence speed and separating accuracy. At last, we proved the feasibility and effectiveness of algorithms by simulation results, took speech signals as input signals, this simulation results show that the separation affect of chaos particle swarm optimization algorithm is better that of particle swarm optimization algorithm and this algorithm also can separate multiple speech signals.Secondly, algorithms based on post non-linear model didn't work when there are nonlinear cross mixture among channels, we used radial basis function neural network to solve this problem. At beginning, we introduced fundamental theories of radial basis function neural network, which has local approximation property; this property can be used to compute inverse process of nonlinear blind source separation. The key point of radial basis function neural network is to compute its centers; we used to make use of K-means cluster algorithm to obtain these centers. However, this algorithm is sensitive to initial values and often falls into local optimum, these problems make centers and separation signals are inaccurate. To solve these problems, we used particle swarm optimization algorithm to computer centers by clustering and obtain improved radial basis function neural network. At last, we use this neural network to separate signals. We took modulated signal and cosine signal as input signals, this simulation results show that the separation affect of nonlinear blind source separation algorithms based on improved radial basis function neural network is better.