目的 针对目标用户的脑电识别问题，提出了基于共同空间模式（CSP）与迁移学习的脑电识别方法。 方法 首先对脑电数据进行预处理，选择事件相关去同步（ERD）现象存在明显差异的时间窗0.5～2.5 s及包含α和β波的宽频带8～30 Hz脑电，然后对多用户利用CSP分别提取特征得到特征向量，最后利用迁移学习进行目标用户的脑电识别。 结果 用户aa在Cz通道处想象右手的ERD高于想象双脚的ERD。用户aa、al、av、aw、ay的分类正确率分别为93.8%、100.0%、84.2%、94.6%、94.4%，平均分类正确率为92.4%，优于常用分类器SVM和期望最大化EM，仅比竞赛第一名清华大学所采用的方法低1.8%。 结论 基于CSP与迁移学习的脑电识别方法能够利用非目标用户提高目标用户的脑电识别性能，对运动想象脑-机接口系统的研究具有重要意义。 Objective Aiming at the problem of target user electroencephalogram (EEG) recognition, an EEG recognition method was presented based on common spatial patterns (CSP) and transfer learning. Methods Firstly, preprocess was adopted on the original EEG data, and time windows 0.5~2.5 s and broad frequency band 8~30 Hz EEG signals, which contained α and β wave, were selected. Here event-related desynchronization (ERD) phenomenon existed significant differences. Afterwards, by CSP preprocessed EEG signals of multi-user were conducted to extract feature and feature vectors were obtained, respectively. Finally, by transfer learning target user EEG recognition was completed. Results In channel Cz, ERD of right hand motor imagery was higher than ERD of foot motor imagery. The classification accuracy of users aa, al, av, aw, and ay were 93.8%, 100.0%, 84.2%, 94.6%, and 94.4%, respectively. The average classification accuracy was 92.4%, which was better than the commonly used classifiers SVM and EM. The method was only lower than the method of the first winner in the competition adopted by Tsinghua University 1.8%. Conclusions EEG recognition method based on CSP and transfer learning increased target user EEG recognition performance by using non-target users and had important implications for the study of motor imagery brain-computer interface.
Research on EEG recognition method based on common spatial patterns and transfer learning
Objective Aiming at the problem of target user electroencephalogram (EEG) recognition, an EEG recognition method was presented based on common spatial patterns (CSP) and transfer learning. Methods Firstly, preprocess was adopted on the original EEG data, and time windows 0.5~2.5 s and broad frequency band 8~30 Hz EEG signals, which contained α and β wave, were selected. Here event-related desynchronization (ERD) phenomenon existed significant differences. Afterwards, by CSP preprocessed EEG signals of multi-user were conducted to extract feature and feature vectors were obtained, respectively. Finally, by transfer learning target user EEG recognition was completed. Results In channel Cz, ERD of right hand motor imagery was higher than ERD of foot motor imagery. The classification accuracy of users aa, al, av, aw, and ay were 93.8%, 100.0%, 84.2%, 94.6%, and 94.4%, respectively. The average classification accuracy was 92.4%, which was better than the commonly used classifiers SVM and EM. The method was only lower than the method of the first winner in the competition adopted by Tsinghua University 1.8%. Conclusions EEG recognition method based on CSP and transfer learning increased target user EEG recognition performance by using non-target users and had important implications for the study of motor imagery brain-computer interface.
Transfer learningBrain-computer interfaceMotor imageryTarget userCommon spatial patterns
万方数据
