Variable Step Size Blind Source Separation Based On Natural Gradient Algorithm

Blind Source Separation is a hot research field of signal processing, attracting the interest of academic research of many scholars of the neural network academia and signal processing in recent years. In short, the blind source separation is to recover an independent source only by the mixed system output data with unknown transmission channel characteristics, input information, or little priori knowledge. At present, the blind source separation has been widely applied to communication systems, speech recognition, signaldenoising, the sonar problem, biomedical signal processing and other fields.Traditional blind source separation assume that signals are stationary, but in many cases, the statistical characteristics of signals can change over time, showing the cyclostationary features. Taking advantage of such characteristics, on the one hand, can more accurately reflect the nature of the signal relative to the traditional method; on the other hand, compared with the general non-stationary signal, cyclostationary signal, with a spectrum redundancy feature, has a potential anti-jamming capability. Therefore, the analysis of the cyclostationary feature of signals is of great significance to the signal processing under the noisy background. This paper illustrates the current situation of the research at home and abroad and application areas of the blind source separation technique. As well as that, on the basis of the blind source separation and the knowledge of cyclostationary theory, this paper proposes a new blind source separation algorithm, which fulfills the mutual information minimization through the natural gradient algorithm thereby achieving the best result of the blind source separation. Same as above uses a fixed step size, so that the convergence speed and steady-state error cannot be balanced. In order to overcome this weakness, the paper improves the algorithm, and proposes a variable step size blind source separation algorithm based on the natural gradient, which can not only acquire a faster convergence rate, but also reduce the steady state error and avoid the influence of the step selection to separation effects.