| Blind Source Separation (BSS) problem is to separate or extract individual source signal from a set of mixed signals, assuming the source signals are independent and no a-prior information is known about the mixed signals.This thesis is mainly concerned with two problems of BSS: linear instantaneous mixtures and nonlinear instantaneous mixtures, and the objective criterions are mainly based on two theories : information theory and stochastic theory.1) Researches on Independent Component Analysis, Information Maximization, Minimization of Mutual Information and Maximization Likelihood have been done. Combining the FastICA algorithm with the Maximization Likelihood, the paper improves robustness and convergence of the original algorithm for the linear separation. Neural network is an effective approach to BSS. According to maximum output entropy, a new separation algorithm using Radial basis function (RBF) neural network for nonlinear mixtures is proposed. Since the instinctunsupervised learning of the RBF network and blind signal processing are in essence unsupervised learning procedures, therefore the algorithm based on RBF seems rational.2) According to signals high order cumulant. BSS problems are transformed into diagnolization of special matrix. This method simplified the complexity of linear separating algorithms. This thesis construct an characteristic-like function and find a special matrix, transforming BSS into eigenvalue decomposing problems to solve a class problem of linear mixture.Moreover, some elementary research on BSE has been done, the extractioncondition was discussed and the results show that BSE can solve singular linearmixture problems to some extent. Finally an algorithm for multi-sources blindextraction based on information theory was studied.The experiments with speech signals and carrier wave signals have beendone to validate various BSS methods proposed in the thesis and simulationresults show that the algorithms are effective.
|
PDF Full Text Request