基于复合型神经网络的非线性ICA及其在SCP少次提取中应用研究

该文提出一种基于MISEP和NLFA方法的复合无监督多层感知神经网络模型解决非线性独立分量分析(ICA)的解混问题,并对MISEP神经网络中用到的两种Sigmoid函数及新引入的径向基函数(RBF)作了信号分离性能的对比分析。实验结果表明,本文方法可以更好地从非线性混合信号中复现源信号,稳定性高,同时应用于慢皮层电位(SCP)的少次提取,经与相干平均法比较,波形的整体提取效果明显。

Nonlinear ICA Based on a Combined Neural Network and Its Application to Single-Trial Extraction of SCP

A combined, unsupervised, multilayer perceptron neural network model based on MISEP and NLFA is presented to resolve the separation problem of nonlinear ICA, the separation performances of signals are compared between two sigmoid functions used in the latent layers of MISEP and introduced RBF. Experimental results show this algorithm can recover sources from nonlinear mixtures better and has good stabilization, it is also applied to single-trial extraction of SCP, the whole effect is evident compared with the averaged method.