本文基于一个全连接递归网络结构，给出一种新的信息理论的盲源信号分离准则。该准则在保证最大化信息传输的同时最小化输出互信息，从而使网络输出相互独立。由此准则推导的算法对传输中的信息损失引入了Hebb项。该算法既可以分离超高斯源的混迭，又可以分离亚高斯源的混迭信号。计算机仿真结果表明了该算法的良好分离性能。 Based on a fully connected recursive network structure, this paper presents a new information theory blind source separation criterion. This guideline minimizes the output of mutual information while guaranteeing maximum information transmission, thus making the network outputs independent of each other. The algorithm derived from this criterion introduces Hebb terms for information loss in transmission. The algorithm can separate the aliasing of the super-Gaussian source and separate the aliasing signals of the Gaussian source. Computer simulation results show the good performance of the algorithm.