偏最小二乘回归模型在EEG特征选择的应用

Application of Partial Least Squares Regression Model in EEG Feature Selection

为了克服主成分分析(PCA)对共空间模式(CSP)提取脑电信号特征进行降维时,仅考虑主成分对输入变量的表征能力,而忽略了对输出变量进行解释的这一个缺点,提出偏最小二乘回归(PLS)进行降维,通过CSP对数据增强后的信号进行特征提取,采用PLS进行降维,将提取的主成分信息包含对因变量解释程度高的特征作为特征向量,使用PSO-SVM进行分类,用2005 BCI竞赛的数据集IIIa进行分类测试,结果得到3位被试的想象运动平均分类正确率91.71%,通过与CSP-LDS、WL-CSP和CSP等算法的比较,3位被试的平均分类正确率最高,验证了该算法的有效性。

When the principal component analysis(PCA)algorithm reduces the dimensionality of the EEG signal features extracted from the common spatial pattern(CSP),it only considers the representation ability of the principal component to the input variables,and ignores the interpretation of the output variables.To overcome this shortcoming,this paper proposes a dimensionality reduction method of partial least squares regression(PLS).Firstly,CSP extracts the feature of signal dealt with data augmentation,and then PLS performs dimensionality reduction on the extracted features.In particular,the paper takes the features that includes a higher level interpretation for the dependent variable in principal component as feature vectors.Finally,PSO-SVM is used to classify.The proposed algorithm is applied on the data set IIIa of the 2005 BCI competition,and the average classification accuracy rate of the three subjects’imagination movement is 91.71%.The classification result has reached the highest average classification accuracy rate in comparison with algorithms such as CSP-LDS,WL-CSP and CSP,which verifies the effectiveness of the algorithm.