基于卷积神经网络和支持向量机的事件相关电位识别方法

An Event-Related Potential Recognition Method Based on Convolutional Neural Network and Support Vector Machine

针对事件相关电位(ERP)脑电信号存在个体差异性强、信噪比低等特点而导致其识别困难以及传统的卷积神经网络(CNN)对小样本分类易产生过拟合等问题,在CNN和支持向量机(SVM)融合模型的基础上,提出一种用于ERP信号分类识别的CNN-SVM组合分类器方法。该方法以滤波后的原始多通道ERP信号作为输入,首先使用一维时间卷积核对信号的时域进行卷积操作;然后使用一维空间卷积核进行空堿卷积,从信号的时间信息和空间信息中学习特征;最后通过降采样、全连接等操作完成CNN模型的训练;接着再次将信号导入训练好的CNN模型来提取信号的降采样层特征,最后利用SVM对信号降采样层特征进行分类识别。实验结果表明:利用该组合分类器方法可以在少量重复的视觉刺激下实现对受试者P300信号的有效识别;在重复4次以上刺激后,该组合分类器的平均识别准确率为93.49%,相比于传统的CNN方法,其平均识别准确率提升了4.36%;与用于ERP信号识别的经典算法——逐步线性判别分析和贝叶斯线性判别分析相比,该组合分类器的平均识别正确率分别提高了6.38%和4.16%。该组合分类器只需要少量的重复实验便可获得较高的目标识别准确率,有效提高了ERP信号的识别效果。

In view of the difficulties in identifying the event-related potential(ERP)signals due to their characteristics of strong individual differences,and the overfitting problem of traditional convolutional neural network(CNN)for small samples,based on a fusion model of CNN and support vector machine(SVM),a CNN and SVM combined classifier for ERP signal classification and recognition is proposed.This method takes the filtered original multi-channel ERP signals as input.Firstly,a one-dimensional time convolution kernel is used to convolve the time domain of the signals,then the spatial convolution is performed by using a one-dimensional spatial convolution kernel to learn features from temporal and spatial information of signals.Finally,the training of the CNN model is accomplished by such operations as down-sampling and full connection.After that,the signals are imported into the trained model again to extract the down-sampling layer features of the signals,and SVM is finally used to classify and identify the features.Classification results show that the proposed combined classifier can effectively recognize the P300 signal component under a small number of repeated visual stimuli.The average recognition accuracy of the proposed method is 94.08%after the repetition of more than four times of visual stimuli,and it has an average accuracy improvement of 4.36%compared with the traditional CNN method.Compared with the classical algorithms of stepwise linear discriminant analysis(SWLDA)and Bayesian linear discriminant analysis(BLDA),the average recognition accuracy of the combined classifier is improved by 6.83%and 4.16%,respectively.The combined classifier only needs a small number of repeated experiments to obtain higher target recognition accuracy,and effectively improves the recognition effect of ERP signals.