Research On Natural Gradient Algorithm With Optional Weighted Factor

For a long time, the blind source separation is a hot spot in the field of signal processing. The essence of blind source separation is to use the mixed signals received by the receivers to realize the separation of the source signals. Because of many unknown factors in the separation process, the so-called separated signals are actually the estimation of the real signals. Blind source separation technology has been widely used in many fields, such as biomedicine, image processing, radar positioning, communication transmission and seismic exploration.Firstly, this paper introduces the background of blind source separation problem, developing process, as well as the domestic and foreign breakthrough. Then, the basic principles of blind source separation are discussed, including the mathematical model of blind source separation, specific classification and pre-procession. Specially, adaptive algorithms are focused involving the non Gaussian maximization and the minimization of the mutual information theory, also the indexes about the separation performance are given and a theoretical guarantee for the following study is provided.Secondly, the basic principles of the natural gradient algorithm and the momentum term natural gradient algorithm of the blind source separation are studied. The momentum term natural gradient algorithm can effectively improve the convergence speed of the algorithm. Through MATLAB simulation experiments and analysis for the separation performance of the natural gradient algorithm and the momentum based natural gradient algorithm, we can find that the convergence rate of the latter is better than the former's, but the steady-state error of the latter has not diminished and even slightly worsen than the former.Finally, in view of the defects of the steady-state error of momentum natural gradient algorithm, the traditional combined momentum natural gradient algorithm of blind source separation system(including two sub-systems) can resolve the problem, but due to the weighted factor must be limited from 0 to 1, then this algorithm is just to maintain the momentum natural gradient algorithm for steady-state performance. According to that, this paper put forwards a new type of system structure and propose a novel on momentum term based natural gradient algorithm with optional combined weighted factors. MATLAB simulations verified that the new algorithm proposed in this paper is better than the traditional combinated momentum term natural gradient algorithm in both convergence speed and steady-state error performance.