基于频带能量归一化和SVM-RFE的ECoG分类

ECoG classification based on band power normalization and SVM-RFE

针对基于运动想象(左手小手指和舌头)的皮层脑电(electrocorticographic,ECoG)信号的分类问题,对BCI2005竞赛数据集I中的ECoG信号使用频带能量(band power,BP)归一化算法提取运动相关电位(movement related potential,MRP)、μ节律和β节律的频带能量作为特征。针对特征提取后维数较高的问题,使用基于支持向量机的回归特征消去(support vector machine recursive feature elimination,SVM-RFE)算法进行特征选择,通过对训练数据集使用10段交叉验证(cross validation,CV)的方法寻找最佳特征组合,确定特征在维数为6时具有最低平均识别错误率,对测试数据集采用同样的方法和同样的组合进行特征提取,并使用线性支持向量机进行分类,分类正确率可以达到93%。

Aiming at the ECoG classification of different imagined movements of left little finger and tongue,BP（band power） normalization algorithm was used to extract MRP（movement related potential）,μ rhythm and β rhythm from the ECoG in BCI2005 competitive dataset I.In order to cut down the high dimensions of the extracted features,SVM-RFE was used to select the features that are more suitable for the classification.And a 10-fold cross validation of the training dataset was carried out to search the optimum feature combination,the best feature dimensions of 6 was determined according to the mean classification accuracy of each dimension.Then,all the selected features were fed into a linear SVM to train a model,which could be used to predict the labels of the features also selected by the same way from the testing dataset.And the final classification accuracy is 93%.