Study On Blind Source Separation Methods Based On Post-nonlinear Hybrid Model

In recent years,Blind Source Separation(BSS),as a new method in the field of intelligent signal processing,has been favored by many scholars because it can restore the source signals only according to the characteristics of the observed signals.BSS has outstanding application prospects in biometrics,data mining,mechanical and electrical fault diagnosis and communication.At present,most of the research on BSS problems focuses on the linear mixing case,while the researches on the nonlinear mixing model are relatively few.But due to the linear mixing situation exists only in ideal model,in a real environment signals are usually mixed with nonlinearity.In the nonlinear mixed models,the post-nonlinear mixed model(PNL),retaining the signal characteristics of both linear and nonlinear transmission processes,is more close to the actual environment than the other nonlinear models,and has been proved to be relatively simple and separable.In this paper,on the basis of BSS theory and related algorithms,the BSS problem based on the post-nonlinear mixing model is researched,and two optimized BSS algorithms are proposed.The main work is as follows:(1)Aiming at the gradient disappear problem of the natural gradient algorithm,a multilayer neural network is constructed with the signal KL divergence as the separate evaluation function and the PReLU function as the activation function,to estimate the probability density function of the source signal and finally achieve the separation of signals.The simulation is designed to separate the mixtures with Gauss signal and artificial Sub-Gaussian signals.The signal waveforms which are separated by the natural gradient algorithm with either Sigmoid function or PReLU function as activation function are compared.Finally,by comparing the PI performance of the two algorithms and the evaluation criteria of the iteration number,it is proved that the proposed algorithm has certain improvement in the aspect of the gradient vanishing,and has remarkable improvement in the convergence speed and the separation effect.(2)Aiming at the initial value sensitive problem of FastICA algorithm,a new algorithm which combines the NMF algorithm with the FastICA algorithm is proposed.This algorithm takes the negative entropy as the objective function and initializes the NMF matrix with orthogonal constraints to obtain the initial basis matrix and coefficient matrix.The optimal basis matrix and coefficient matrix are obtained by updating the NMF algorithm iteratively,and the optimal basis matrix is used as the input matrix of the observation signal in the FastICA algorithm.The simulation is designed to separate image signals,and the performance between the proposed algorithm and FastICA in the aspect of the convergence speed and SNR is compared.The results show that the separation images of the proposed algorithm looks closer to the source images,and the sensitivity problem of FastICA to the initial value(input matrix)is solved to some extent.In SNR respect,the performance of the proposed algorithm is also improved obviously.