一种改进的自适应混合神经网络盲分离算法

传统的前馈神经网络盲源分离算法由于步长固定存在许多缺点,而基于Sigmoid函数的自适应步长算法虽然能够克服固定步长算法的缺陷,但其稳态性能较差。针对这个问题,提出一种改进的自适应步长算法,该算法可灵活地控制步长因子函数的形状,在近零点处变化较Sigmoid函数更加缓慢,性能更加优越;同时针对前馈神经网络的不足,在前馈神经网络结构中引入递归结构,利用改进的自适应步长算法控制学习速率。仿真分析表明该算法具有更快的分离速度和更加优越的分离效果。

Improved self-adaptive mixing neural network algorithm forblind source separation

The traditional feedforward neural network blind source separation algorithm is imperfect because of its fixed learning step. Although the self-adaptive step size algorithm based on Sigmoid-function can overcome the shortcomings of fixed step, its steady-state performance is poor. According to this problem, this paper proposed an improved self-adaptive step algorithm, which could flexibly control the shape of the step curve and the shape changed more slowly near the zeros than Sigmoid-function, the performance was more excellent. Secondly, considering the insufficient of feedforward neural network structure, this paper added a recursive structure into the whole model, adjusting learning step size with the improved self-adaptive step algorithm control algorithm. The simulation analysis shows that the algorithm has a faster separation speed and a better performance in convergence.