基于CSP的多类运动想象脑电特征自动选择算法

Automatic selection algorithm for multi-class motor imagery of EEG eigenvalues based on CSP

目的在基于协方差矩阵近似联合对角化(joint approximation diagonalization,JAD)的多类共空间模式(common spatial pattern,CSP)运动想象检测滤波器的设计过程中,需要对关键特征向量进行选择。较常用的基于"最高得分特征值准则"的特征向量选择方法会出现不同类数据的最高得分特征值对应同一个特征向量,因此导致无效CSP滤波器的出现,进而影响系统识别率。本文在传统JAD方法上提出一种特征值自动选择方法以解决特征值选择无效问题。方法基于BCI Competition 2005data IIIa(BCI2005)和实验室自主采集三类运动想象脑电(EEG)数据集,对不同想象类别数据对应同一个特征向量的异常现象进行实验分析。结果在两个数据集自测试下,本方法的三类运动想象平均识别率分别达到82.78%和85.92%,比传统JAD提高3.44%和3.25%。结论基于CSP的多类运动想象脑电特征自动选择算法能够有效解决特征值选择无效问题,进而提升运动想象BCI系统的分类识别率。

Objective The joint approximation diagonalization （JAD） of the eovariance matrix extends the common spatial pattern （CSP） algorithm to the muhi-class motor imagery, in which the key feature vectors should be chosen appropriately. The most common method is to select the eigenvectors corresponding to the highest score eigenvalues. However, according to these choice criteria, the same eigenvectors are often just selected for the datasets of different classes, which may cause the failure of CSP spatial fihering and the decline of the classification accuracy. A method with the new choice criterion is proposed in this paper, which can automatically select the effective eigenvectors based on the traditional JAD algorithm. Methods The three-class motor imagery signals of two datasets （ BCI Competition 2005 dataset IIIa and our own recorded experiment dataset） were used to testify the validity of the algorithm. Results The mean classification accuracies of the three-class motor imagery were calculated with the self-testing of the two datasets. The accuracies calculated by our proposed algorithm achieved 82.78% and 85.92%, which were improved by 3.44% and 3.25% respectively, compared to the traditional JAD algorithm. Conclusions This algorithm can automatically select the effective features based on CSP, and avoid selecting the useless features for classification, which can greatly improve the classification accuracies of motor imagery BCI system.