Research On Blind Source Separation Based On Artificial Immune Algorithm And Neural Network

Blind source separation (BSS) is one of the hotspot researching in the field of signal processing and neural network. At present, it has been successfully used in speech recognize, digital communication, biomedical signal processing, and so on. Because it is very difficult in the theory, the study of it by now is far from mature and however there are many problems to not resolve.The thesis is mainly concerned with two typical problems of BSS: linear and instantaneous mixed blind separation and nonlinear mixed blind separation. To the question of linear and instantaneous mixed BSS, there are three aspects of the thesis, as follows:1. The deficiencies of existing methods based on neural network are analyzed. And a neural network based on immune genetic optimization for BSS is proposed on the basis of those deficiencies. The network has two layers. The first layer performs pretreatment of data. The second layer performs separation of sources, where the weights of separate network are updated by immune genetic algorithm (IGA) .The effectiveness and advantage of the algorithm is testified by simulation.2. The deficiencies of BBS algorithm based on natural gradient are analyzed. And the algorithm is improved by using conjugate gradient. The simulations show that the improved algorithm is better than the original algorithm at the capability and separate effect. But it is still affected by activation function.3. For overcoming the effect of activation function, a new BSS algorithm isproposed. It is a combination of immune algorithm (IA) and high order cumulates and it can effectively accomplish a separation of sources. The algorithm separates two steps. The first step estimates the dimension of sources and abstracts principal component information by eigenvalue decomposition. The second step performs separation of sources by using IA, where contrast function is based on the transformation of four-order mutual cumulates. The simulations show that the capability and separate effect of the algorithm is improved obviously.Moreover, some researches of the nonlinear mixed are done. In the thesis, Radial Basis Function (RBF) neural network based on IM and the theory of Maximum Entropy (ME) are combined to resolve one kind of nonlinear mixed problems. The unsupervised learning process for optimization of parametric RBF is separated two steps. The first step mainly estimates the center vector of basis function by using an algorithm of immune clustering. The second step performs separation of sources, where the weights of network are updated by ME. The effectiveness of proposed method is demonstrated by simulations.