基于深度学习框架的多类运动想象脑电分类研究

Research of multi-class motor imagery EEG classification based on deep learning framework

模式识别是脑机接口(brain computer interface,BCI)系统的核心部分,其中特征提取和分类方法对最终分类结果有着决定性作用.针对多类运动想象脑电识别过程中特征提取困难,识别准确率低的问题,文中提出了一种新的基于深度学习框架的多类别运动想象脑电分类方法.首先,为了满足深度学习方法的大批量样本的需要,使用自编码器(auto-encoder,AE)对训练样本进行扩充;其次,针对脑电信号的特点,设计4个巴特沃斯带通滤波器提取脑电的θ、α、β和γ波段,并对每一波段的信号进行傅里叶变换,然后计算幅值的均值和方差;最后,通过深度信念网络(deep belief network,DBN)对脑电信号进行分类识别.文中使用BCIⅢ的竞赛数据集对所提出的方法进行验证,实验结果表明,文中方法能够有效地提高多类运动想象脑电的分类准确率,分类结果的平均kappa系数达到了0.802 4.

Pattern recognition is the key module in brain computer interface(BCI)system,which contains feature extraction and classifier design.Aiming at the difficulty of feature extraction and low accuracy during the process of multi-class motor imagery(MI)electroencephalography(EEG)recognition,this paper presents a new approach for multi-class EEG classification based on deep learning framework.Firstly,auto-encoder(AE)is employed to expand training samples,in order to satisfy the needs of large numbers of samples of deep learning method.Secondly,according to the characteristics of EEG,four Butterworth band pass filters are designed to extractθ,α,βandγbands from original EEG signal,then FFT is used to convert these four bands into frequency domain,at the same time,mean value and variance are calculated.Finally,deep belief network(DBN)is utilized to classify MI EEG,which has a strong ability to learn feature from inputs.In this paper,the BCIⅢcompetition dataset is used to verify the proposed method,and experimental results show that this method can effectively improve the recognition accuracy of multi-class EEG,and the mean kappa coefficient of classification results reaches 0.802 4.