Momentum Algorithm Based On Double Adaptive Blind Source Separation Coupling

Blind Source Separation is the process to restore the observed signals received bytransducer after transiting unknown hybrid system, based solely on the condition of the statistical independence of input source signals. It contains lots of subject knowledgesuch as signal processing, artificial neural network and information theory, etc. Moreover, it has become significance research subject in many fields, for instance, biomedicine,medical picture, radar and communication system, voice separation, data mining, physical geography, earthquake detection, weather forecast and climatic prediction, etc.In this thesis, the author reviews the theory of blind source separation and focuseson the research of problems on the adaptive natural gradient algorithm of instantaneous composite signal, especially the MATLAB simulation and analysis on separation performance of natural gradient algorithm and natural gradient algorithm with momentum term. With regard to the defeats of these algorithms, the author makes improvement and optimization. At last, an effective algorithm is put forward which can ameliorate thewhole performance of the algorithm. The details are as follows:First, the author reviews the research background, significance, research history, current situation as well as the application of the blind source separation; expounds the basic theory of blind source separation including its mathematical model, separability anduncertainty; makes an brief introduction of mathematics and information theory used during resolving the problems of blind source separation; at last proposes construction method and optimization criterion of cost function as well as algorithmic separation performance metrics during blind source separation.Secondly, the present thesis introduces batch natural gradient algorithm and the adaptivenatural gradient algorithm, mainly introducing Natural Gradient Algorithm in the adaptivenatural gradient algorithm. Then the author analyzes the advantages and disadvantages of thealgorithm by the means of computer simulation. In addition, Natural Gradient Algorithm with Momentum Term is introduced in detail considering that the Algorithm can’t involve bothconvergence speed and steady state error. This can make separation matrix reach its optimalvalue by less iterations. The simulation experiment proves that Algorithm can achieve the mixedsignal’s separation whether the environment is steady or not.Thirdly, this paper inspires the momentum derived from the neural network, andintroduces a smoothing factor. A fire-new framework was propased by adaptively combiningtwo separation systemes based on Natural Gradient Algorithm with Momentum Term. Asmoothing factor was used to adjust the proportion of of the two components. It can speed upconvergence speed and reduce the misadjustment error in the steady state simultaneously. Andeffectively improve the overall performance of the blind source separation algorithms. The newalgorithm proposed in this paper has superior convergence speed,stability. However, thisalgorithm has the disadvantage of two pieces of convergence. Therefore, Natural GradientAlgorithm with Momentum Term base on two separation system coupling is optimized andfurther improves the separating property of the algorithm.Finally, the author summarizes the main contents of this thesis and points out thebottle-neck and the future prospect of the development of blind source separation technology.