基于非线性动力学和GA-MLPNN的ECoG信号分类

摘要

In order to classify electrocorticogram (ECoG) signals of different mental tasks in a brain-computer interface (BCI) system, a method based on the combination of genetic algorithm (GA) and multilayer perceptron neural network (MLPNN) was presented. The GA approach was used to optimize ECoG channels selection, which minimized the number of channels while maximizing the classification performance. Error back-propagation (EBP) algorithm was used as the learning mechanism of MLPNN. The nonlinear dynamics features (e.g. permutation entropy (PE) and Hurst exponent (HE)) were chosen for the channel selection and classification because the two nonlinear parameters gave high calculation performance and great discriminative ability. The results show that the average classification rate of 87% was obtained using the 15 selected channels as opposed to only 79% by using all 64 channels.%为了对脑-计算机接口(BCI)中不同思维任务下的皮层脑电(ECoG)信号进行分类,提出了基于遗传算法(GA)和多层感知器神经网络(MLPNN)的混合方法.用GA方法优化ECoG通道选择,使得选择通道数最小而分类性能最大.使用误差反馈传播(EBP)算法作为MLPNN的学习机制.实验表明,用排列熵(PE)和Hurst指数刻画ECoG的非线性动力学特征具有较好的计算性能和区分能力,故选择这两个特征量进行通道选择和分类处理.分析结果显示,通过使用选择的15个通道进行分析所得的平均分类率为87%,而使用全部的64个通道的结果仅为79%.