摘要： 分析了信源数未知或信源数目动态变化时，超定盲信号分离(源信号个数n<观测信号个数m)问题中自然梯度算法发散的原因。在此基础上，构造了一种新的自适应盲分离算法。该算法不仅克服了已有算法不能稳定收敛的缺点，而且在信源数动态变化的情况下，无须根据输出分量间的关系去除冗余分量，大大简化了算法的计算量与复杂度。仿真结果验证了该算法的收敛稳定性与分离的有效性。

关键词: 盲信号分离, 自然梯度算法, 神经网络

Abstract: The analysis of the cause which cased the natural gradient algorithm divergence in over-determined BSS problem with an unknown or dynamical changing source number is made. A new adaptive blind separation algorithm is constructed. The new algorithm overcomes the drawback of being unable to convergence stably, and can delete the redundant components without basing on correction among output components, which simplifies the computation and complexity. The simulation results illustrate the convergent stability and the separating efficiency.

Key words: Blind source separation, Natural gradient algorithm, Neural network