自适应变步长的动量项盲源分离方法

Adaptive Variable-step Blind Source Separation with Momentum Factor

针对传统盲源分离算法采用单一步长而无法同时兼顾收敛速度与稳态性以及动量因子选取的问题,介绍了一种盲源分离优化方法。该方法依据自然梯度算法(Natural Gradient Algorithm,NGA)的收敛条件,通过输出信号建立一种新的表示信号分离程度的度量指标,通过此度量指标构造非线性单调函数,使步长与动量因子参数自适应调节,从而可以合理、准确地选择参数。仿真表明了在平稳和非平稳环境下所提分离指标的正确性,且该指标可有效监测信号分离程度;针对步长及动量因子参数选取所设计的优化策略能够有效地缓解固定值对算法性能的约束,在有无噪声的情况下,均获得了优良的分离效果。

For the problem that the traditional blind source separation algorithm can not well balance the convergence rate and the steady-state error by using single step size and the selection of the momentum factor,a blind source separation optimization algorithm is introduced.Based on adaptive natural gradient algorithm(NGA),a new separation indicator is constructed to reveal the separation degree through output signal.Next,a nonlinear monotone function is structured through the separation indicator.Then,step parameters and momentum factors are adaptively updated so that parameters can be chosen reasonably and accurately.The simulation results show that the proposed method is correct in both stationary and non-stationary environments,and it can effectively track the degree of signal separation.The proposed optimization method for choosing step size and the momentum factor parameters can effectively alleviate the restriction of fixed value to performance.In the presence of noise,the modified algorithm can get good separation effect.