一种基于时–空–频联合选择与相关向量机的 运动想象脑电信号分析算法
A motor imagery analysis algorithm based on spatio-temporal-frequency joint selection and relevance vector machine

研究表明: 不同受试者由于个体差异, 会引起在执行相同运动想象任务时, 产生与受试者关联的特定脑电信号 特征, 这是设计脑机接口系统面临的一个实际问题. 为解决这个问题, 本文提出了一种基于时–空–频联合特征的提取方 法. 首先, 对原始118导联的EEG进行空间特征分析, 从中提取出与运动想象相关脑区对应的55导联EEG信号. 进一步, 在训练集上, 通过7–折交叉验证, 训练出与受试者匹配的时间窗和频带. 其次, 利用8个共空域滤波器进行特征提取. 最 后, 将获得基于样本的运动想象特征, 采用相关向量机进行分类. 仿真结果表明: 该算法在第3届脑机接口竞赛数据 集Data IVa分类上获得5位受试者平均分类精度为94.49%, 结果优于当年第1名94.17%. 此外, 与其他3种常用的方法比 较亦具有明显优势. 本文提出的基于样本的时–空–频特征提取方法和相关向量机的结合, 该算法整体性能优越, 为基于 运动想象的脑机接口在线系统设计提供了一种新方法.
Convergent studies have reported inter-subject variability in EEG representation when subjects performed same cognitive tasks, yielding a significant drawback for developing a practical BCI system. In order to address this problem, we have introduced a subject-dependent specio-temporal-frequecy joint feature selection method. Specifically, we first selected 55-channel EEG signals among the original 118-channel recordings according to the close relevance of the signals in motor-related areas. A 7-fold cross validation approach was applied to select the optimal time-window and frequency bands, which match individual subject based upon the training data set. Then motor imagery related features were determined via the common spatial pattern method. The obtained subject-dependent features were feeded to a relevance vector machine for motor imagery classification. The experiment results show that our framework demonstrated superior performance as showing in the higher classification accuracy (94.49% in comparison with the highest classification accuracy 94.17%) in the competition III. Compared with the other three existing methods, our method also has obvious advantages. In summary, we provided feasible framework to account for inter-subject variability, which would be a new method for the designing of the online motor imagery brain computer interface system.