非监督特征学习方法在脑电身份识别中的应用

设计并实现了采用非监督特征学习方法对模拟阅读事件相关电位实验中多名受试者脑电信号的特征提取,并对提取的特征向量进行了模式分类.实验中共采集5名受试者脑电信号,每名受试者的特征样本集由其接受模拟阅读靶视觉刺激后100⁴00ms在通道PO3、O1、Oz、O2、PO4、P4、P8、CP6的脑电信号样本组成,各受试者样本集均含400个试次样本.非监督特征学习过程由含6个神经节的BP神经网络完成,后选用支持向量机作为分类器.对比了1试次,2试次、5试次、10试次样本叠加等几种不同情况下采用非监督特征学习方法提取特征的分类正确率.实验结果表明:采用多神经节人工神经网络对5名使用者5试次叠加信号样本提取的特征向量的分类正确率高于90%,显著优于对各单通道时域特征向量的分类正确率,该方法可为以脑电信号为特征的身份识别系统提供一种可行的特征提取方式.

Unsupervised Feature Learning Method with Application to EEG signal based Personal Identification

The multi-ganglion BP neural network based feature learning method, a kind of unsupervised methods, is applied to the feature extraction procedure of Imitating-Reading EEG based personal identification system.Five subjects participated in the Imitating-Reading ERP experiments.The dataset of each subject contains 400 trials of eight channel （ PO3, O1, Oz, O2, PO4, P4, P8, CP6 ） EEG signals ranging from 100ms to 400ms after the subject receiving target stimuli.The multi-ganglion BP neural network, which consists of six relative small-scale auto-encoders, is applied to extract the feature vectors from single-trial EEG signals and two, five, ten-trial averaging EEG signals respectively.The classification procedure is performed by support vector machine and the classification accuracy of the subjects exceeds 90%, when using five-trial averaging samples, considerably higher than using single-channel temporal feature extraction method.This study provides an unsupervised feature learning method for the application of EEG based personal identification system.