基于神经网络的含噪动态源分离算法  

The separation algorithm for dynamic sources in the noisy case based on neural network

作　　者：苏巧[1] 魏以民[1] 沈越泓[1] 

机构地区：[1]解放军理工大学通信工程学院,江苏南京210007

出　　处：《电子技术应用》2018年第2期88-91,96,共5页Application of Electronic Technique

基　　金：国家自然科学基金(61172061)

摘　　要：针对含噪声且源信号数目动态变化条件下的混合信号分离问题进行了研究,提出了一种新型在线盲源分离算法,该算法包括两部分:一是基于最小描述长度(MDL)的动态源数目估计算法,该算法能实时精确地估计出瞬时源数目;另一个是基于偏差去除的变步长神经网络算法,该算法采用前馈神经网络结构,在学习准则中加入了相应于噪声的偏差去除项,并在此基础上给出了变步长策略。仿真实验表明,新型算法在含噪静态源和动态源中都具有优异的分离性能,并且优于现存的针对动态源的盲源分离算法。In view of the dynamic sources in the noisy case, this paper proposes a novel online blind source separation（BSS） algo- rithm which contains two parts. One is a dynamic source number estimation algorithm based on Rissanen＇s minimum description length（MDL）. It can estimate the instantaneous source number accurately. The other is a variable step-size neural network algorithm based on bias removal. It employs the structure of feed-forward neural network, introduces the bias removal term corresponding to the noise into the learning criterion and provides the variable step-size scheme for the new learning criterion. Simulation results show that the proposed algorithm has the superior separation performance for both static sources and dynamic sources in the noisy case in comparison with the existing BSS algorithm for the dynamic sources.

关 键 词：盲源分离 含噪动态源 自适应神经网络 偏差去除 

分 类 号：TN911[电子电信—通信与信息系统]