Blind Signal Separation Algorithm Based On Least Squares Support Vector Machines

Blind Signal Processing involves a wide range of knowledge, such as artificial neural networks, statistical signal processing and information theory and so on. It has very important theoretical value and became the main development direction of the artificial neural network. Blind source separation is a new technology, which is used to isolate source signals from mixed observed data and has been widely applied in many areas.This thesis studies the state of blind signal separation and the development status of support vector machine, and then discusses blind source separation problems about the signal model and the performance index of the algorithm, and introduces the principle of ICA and related theories knowledge, the basics of support vector machines.FastICA algorithm based on the maximum negative entropy needs to select the appropriate nonlinear function, which it had been selected by human experience in the final signals separation processing. There are some defects on the separation effect and the performance index. In order to solve this problem, the FastICA algorithm based on LS-SVM is proposed to avoid human selection of the nonlinear function. Firstly, it constructs the empirical distribution function of observed signals as its approximation. Secondly, it uses LS-SVM to estimate the density function of observed signal. Finally, it uses knowledge of probability to get the density function of the output vector, then puts it into the iterative formula of the separation matrix. Some simulation about signal separation, speech separation, images separation show that the effects of the improved FastICA algorithm for blind signal separation is better than the former.Finally, since weight iterative formula of extended Infomax algorithm based on information maximization uses the natural gradient method, it has the shortage of slow convergence, large computation, time and storage space. In order to solve this problem, the extended Infomax algorithm based on conjugate gradient is proposed in this thesis. It uses the conjugate gradient direction to replace the negative direction of the natural gradient. Namely, it uses one-dimension minimum point of conjugate gradient to find the optimal value weight iterative formula. Lastly, some simulation results about signal separation, speech separation, images separation show that the effects of the extended Infomax algorithm based on conjugate gradient for blind signal separation is better than the former.