Abstract: In order to improve classification accuracy, this paper describes a novel method based on Wavelet Packet Decomposition（WPD） and voting combination of multiple classifiers to classify the Motor Imagery（MI） Electroencephalogram（EEG） signals. First, the pre-processed MI data is decomposed into wavelet coefficients using WPD, from which features based on Relative Wavelet Packet Energy（RWPE） in all sub-bands are extracted; then, the L-2 norm of wavelet packet coefficients in special sub-bands at channels C3 and C4 is obtained according to the diversity of the hemispheric brainwave in different mental tasks; finally, features are combined to feed into a multiple classifier combination based on majority voting strategy for classification and correct rate of 92.85% is achieved. The experimental results indicate that the RWPE and L-2 norm effectively reflect the event-related desynchronization and synchronization（ERD and ERS） characteristics of left and right MI；the combined classifier improves the classification performance, which is superior to the single classifier.

Key words: Brain-Computer Interface（BCI）, feature extraction, Wavelet Packet Decomposition（WPD）, multiple classifiers, voting combination

摘要： 为了提高脑思维任务分类精度，提出了一种基于小波包分解和多分类器投票组合的运动想象任务分类方法。该方法利用小波包分解对经过预处理的脑电信号进行分解，提取所有频带上的相对小波包能量特征；根据不同脑思维任务下左右半脑各通道间的差异性对C3、C4两通道求取特定频带上的小波包系数的L-2范数作为特征；采用基于投票策略的组合分类器对两种联合特征进行分类，得到了92.85%的识别精度。实验结果表明，联合特征向量较好地反映了左右手运动想象脑电信号的事件相关去同步（ERD）和事件相关同步（ERS）的本质特性；组合分类器识别效果优于单一分类器。

关键词: 脑-机接口, 特征提取, 小波包分解, 组合分类器, 投票组合