The Application And Research On Blind Source Separation Technology Based On Neural Network

Blind Source Separation (BSS) based on neural network is the combination of artificialneural network, information theory and statistical signal processing, since1990s, it has beenrapidly developed and made considerable progress. Under the consideration of instantaneouslinear combination model, this paper makes an in-depth study of Independent ComponentAnalysis (ICA) based on neural network, and optimizes the neural network algorithm in theaspects of learning step size and network structure. The major work is summerized asfollowed:1. On the basis of natural gradient learning algorithm of ICA neural network, thisdissertation presents the general steps of neural network BSS algorithm, introduce afeedforward network model including two parts: whiting network and separation network;,and combining the engineering application confirms the rule of realtime adaptive learning tothe two structs. As the critical issue to solve BSS problem, several commonly used Nonlinearactivation functions and adjustment activation function for separationg complex mixed signalare discussed.the simulation analysis shows that the algorithm of ICA neural network hascertain limitation.2. In the research of the steady-state solution to the ICA algorithm, comparing withadvantages and disadvantages of self-adaptive step size algorithm, in order to overcomedefect of LMS algorithm, an improved self-adaptive step algorithm is proposed, which canflexibly control the shape of the step curve and changes slowly near the Zeros.the improvedalgorithm could accelerate convergence and prevent the steady-state disorder in the same time,while, because of lack of prior information, a replace function and its parameters are proposed.This improved algorithm adjusts learning step size in accordance with real-time status ofSeparation network, and could separate mixed signalsin high speed even when constant stepsize algorithm is unperfect, improve convergence speed with the steady-state stability;meanwhile, under the condition of mutation transmission channel the performance of theimproved algorithm is excellent.3. Considering the insufficient of feed-forward neural network that be easily trapped inthe local minimum or convergence platforms, a recursive structure is added into the whole BPmodel, adjusting learning step size with the improved self-adaptive step algorithm. Under thecondition of real the ship radiation noise mixed model, whether through constant channel ormutate channel this improved self-adaptive mixing neural network algorithm for blind source separation has quick response, good trackability and perfecg separation performance, lay agood foundation for target recognition, orientation and fixed position, and the simulationresult proves the feasibility of this algorithm to process of mixed underwater acoustic signalin real-time linear model.