Research On Blind Sources Separation Algorithm Based On Nonholonomic Natural Gradient

Blind Source Separation (BSS) is intended to separate or recover a set of unobservable source from another set of mixtures, without having access to the mixing matrix coefficients. BBS has become the basic and important issue in signal processing and communication areas and takes a key role in speech recognition, medical signal processing, remote sensing, radar and communication systems and seismic exploration.Firstly, this thesis introduces the researching background and domestic and international development of blind source separation, and summaries the basic theory of blind source separation, including probability theory, matrix analysis, information theory and the mathematical framework model of BSS. In addition, this theisi introduces the natural gradient algorithm and the non-holonomic natural gradient algorithm of blind source separation algorithms simply, and analyses the deficiencies of the algorithms.Then, in view of the deficiencies of the natural gradient algorithm and the non-holonomic natural gradient algorithm, the variable step-size non-holonomic natural gradient algorithm based on sign operator is proposed in this theisi. The new algorithm combines the non-holonomic natural gradient with the sign operation which is derived from a generalized dynamic separation model, and the variable step-size which is based on the gradient of cost function is also applied to the proposed algorithm. The sign operator improves the performance of the algorithm and lead to much faster convergence speed, and the variable step-size balances the contradiction between convergence speed and the error in steady state. In addition, this theisi proposes a new constrained algorithm based on non-holonomic natural gradient to process instantaneous blind source separation problem in the non-stationary environments. At the same time, an adaptive means is adopted to the constrain factor. It could do the adjustment adaptively according to the separation situation. Computer simulation results show that the approved algorithms performance is superior to the usual algorithm.Finally, summarizes the work of this theisi, and points out that there are still some problems to be further research in blind source separation, including noise problem, underdetermined problems and so on, and the theory and application still has great space for development.