Investigation And Application Of Blind Sources Separation Based On Neural Networks

Recently, Blind Sources Separation (BSS) have made rapid progress in signal processing fields, and it has received great application in many fields. However, many problems on its theories and applications need to be investigated and improved further.In this thesis, we first address an introduction to the research status of BSS. Then, preliminary knowledge of BSS algorithms is given. At last, based on the InforMax algorithm, we introduce the BSS algorithms based on the neural networks and higher-order statistics, and investigation the BSS application to gene microarray data and voltage frequency flicker. The main works in this thesis can be introduced as follows:1. According the estimation of probability density function based on neural networks, we obtain the score estimation of function. Then, the BSS algorithms based on the neural networks are presented. Finally, we simulate the signal separations of sub-Gauss and sub-Gauss, sup-Gauss and sup-Gauss, sub-Gauss and sup-Gauss, and image mix, and investigate computation complexity and convergent characteristic. Simulations indicate its better performances.2. Using the estimation theory of probability density function based on high-order neural networks, we obtain the polynomial expansion of score function. Then, we substitute the polynomial to the InforMax BSS algorithm and present the BSS algorithm based on higher-order statistic. Moreover, we separate the mixtures of sub-Gauss and sub-Gauss and sub-Gauss and sup-Gauss. Simulation results show its better performances.3. We do the heterogeneity Correction using the BSS technology, instead of the traditional biomedicine methods. On the other hand, the detection of the voltage frequency flicker is a key technology in voltage fluctuation and frequency flicker measurement; it can affect the precision of voltage fluctuation. We extract the signal frequency by an appropriate embedding matrix is constructed as the input signals using the BSS algorithm. The simulation results show its effectiveness.4. The number of unknown source signals is unknown in the present world. In this situation, to estimate the number of sources, we first estimate the noise covariance matrix by the factor analysis approach, and then estimate the number of nonzero eigen-value and therefore the number of sources. The simulation results show the approach's efficiency.