一种基于自然梯度的两步盲源分离算法  

Blind Source Separation Based on Two-Segment Natural Gradient Algorithm

作　　者：李思怡[1] 王永威[1] 黄琰[1] 陈惠娟[1] 

机构地区：[1]西北工业大学,陕西西安710129

出　　处：《微电子学与计算机》2013年第6期169-172,共4页Microelectronics & Computer

摘　　要：基于自然梯度的盲源分离算法通常有固定步长和自适应变步长两种算法.固定步长算法在求解初阶段具有较快的收敛性,但是随迭代的进行,稳定性较差;自适应步长算法步长可调有较好的全局收敛性,但降低算法的收敛速度.针对算法收敛速度和算法稳定性之间的矛盾,提出了基于自然梯度的两步盲分离算法.首先,在自适应步长公式中引入高阶相关系数;其次,综合固定步长以及改进的自适应步长算法,提出两步盲源分离算法.仿真实验证明,两步盲分离算法可以有效提高盲分离算法的稳定性和收敛速度.There are two common algorithms based on natural gradient method, which are fix-step-size algorithm and adaptive step-size algorithm, respectively. Fix-step-size algorithm has a very fast convergence rate in the initial stage of Blind Source Separation. As iteration goes on, however, the step size gradually leads to poor robust performance. On the other hand, for adaptive method, the relatively small initial step size inevitably leads to more times of iteration. To further reduce the contradiction between convergence speed and convergence robust, as well as redundancy computation, two-segment method is therefore proposed in this paper. In addition, the adaptive step formula is modified by the related coefficient with high order. Computer simulation result confirms the effectiveness in improving convergence rate and robustness, as well as reduction of computation in BSS problem.

关 键 词：盲信号分离 自然梯度算法 分阶段盲分离算法 

分 类 号：TP315[自动化与计算机技术—计算机软件与理论]