加入动量项的改进盲分离算法

Improved Blind Separation Algorithm by Adding Momentum Term

在盲信号分离算法的推导过程中,常采用最速下降法、自然梯度和牛顿法等对代价函数进行最小化,推导过程复杂.文中仿照在BP神经网络算法中加入动量项使算法得到改进这一方法,提出在互累积量迫零算法的推导中加入动量项.加入动量项的改进算法尽可能地保持了输出分量之间的独立,并在保持和原算法一样简单迭代的前提下,提高了收敛速度,且使训练避免陷入局部极小.仿真结果表明该算法的分离误差减小,能有效分离源信号.

During the deduction of blind signal separation algorithms, the commonly adopted steepest descent method, natural gradient method and Newton＇ s method to minimize the contrast function may result in the complexity of the deduction. So, the momentum term is added in the derivation of the Cross-Cumulant-Zero-Forcing algorithm by imitating the addition of the momentum term in BP neural network to improve the algorithm, which keeps the independence of the outputs to the limit, and is of a higher convergence speed than the original algorithm with the same briefness. Moreover, the local minimization in training can be avoided. Simulated results indicate that the proposed algorithm is of little separation error and is helpful to the effective separation of source signal.