The Research Of Blind Source Separation Algorithm Based On Adaptive Variable Step Size

The signal processing can obtain the desired signals quickly and efficiently,as a powerful method of obtaining signals.But, inevitably, scholars usually gather the ingredients who do not want to see. If they know the source of information and the transmission channel signal, the interference signal can be filtered to obtain a useful signal. But more often, scholars get mixed signals observed without any prior information in real life, therefore, the conventional signal processing technology cannot recover the source signals. Blind source separation appeared and got attention, it can well isolate each source signal from mixed sources so as to solve the problem of signal processing field.The dissertation describes the basic theory of blind source separation, and studied the natural gradient adaptive algorithm in detail. Traditional natural gradient algorithm, there is an inherent contradiction between the convergence rate and steady-state error of these two indicators. To overcome this drawback, it is proposed variable step size natural gradient algorithm to improve the convergence speed while greatly reducing the steady-state error. Based on the traditional variable step size natural gradient algorithm,this paper combine it with simulated annealing, and propose an improved variable step size algorithm.The separation process is divided into two phases, firstly, this paper use a fixed step size convergence speed to reduce the amount of computation, in the second phase, establish a function of between step and characterization parameters stateto reduce the steady-state error.The process of blind source separation includes pretreatment and optimization objective function. Pretreatment commonly uses pre-whitening,mixed signals become mutually orthogonal after whitening treatment, and the separating matrix has become orthogonal matrix, too. Thus simplifying the algorithm, but it limits the application of the algorithm so that the algorithm does not have equivariant property. And in the face of ill mixing matrix or the source signal is weak, the method separation performance will be poor.Therefore, this paper study the natural constraints by using orthogonal gradient algorithm, it can not only reach the effect of pre-whitening and may also have equivariant property, so as to make separation more stable.