一种信号源盲分离的神经网络算法

A Neural Network for Blind Source Separation

本文提出了一种新的盲信号分离的神经网络算法。神经网络的第一层使用奇异值分解(SVD)方法对观测信号进行预白化处理。在传感器的数目不少于源信号的情况下,预白化处理能够估计出源信号的数目,同时压缩掉冗余信息。神经网络的第二层是分离层。分离层的权值矩阵应该是正交矩阵。本文应用一个正交严格受限(SOC)算法调整分离网络的权值。其中,用恢复信号的四阶互累积量的平方构造代价函数。仿真实验验证了算法的有效性。

We present a new neural network algorithm for blind source separation. The first layer performs pre-whitening. An algorithm based on singular value decomposition(SVD) is proposed for pre-whitening. If there are more mixtures than sources, it is possible to use the pre-whitening approach for estimating the number of the sources and compressing information. The second layer performs separation of the sources, where the column vectors of the weight matrix are orthogonal. The associated learning rule is termed Orthogonal Strongly-Constrained(SOC) algorithm, where cost function J is based on fourth order cumulant. The efficiency of the algorithm is verified by simulations.