基于注意力时间卷积的运动想象脑电分类方法

Method of MI⁃EEG classification based on attentional temporal convolution

以运动想象为基础的脑机接口技术有助于运动障碍患者的康复,因而被广泛应用于康复医疗领域。针对目前脑电信号的信噪比导致深度学习方法在运动想象数据集上解码精度不高的问题,提出一种基于注意力时间卷积的运动想象脑电分类方法。首先利用深度卷积模块初步提取脑电信号中的时间与空间信息,采用多尺度卷积模块中三个不同大小的卷积块进一步提取MI-EEG(运动想象脑电)数据中整体和细节特征;再经过多头注意力模块突出数据中最有价值的特征,利用时间卷积网络提取高级时间特征;最后,经过全连接网络和softmax层输出分类结果。实验结果表明,在BCI竞赛IV-2b数据集上,所提模型对运动想象二分类任务的平均分类准确率达到了84.26%,与已有的基准模型相比,该方法的准确率有显著提高。

Brain-computer interface(BCI)technology based on motor imagery is helpful for the rehabilitation of patients with movement disorders,and thus it is widely used in the field of rehabilitation medicine.In allusion to the problem that the current signal-to-noise ratio of EEG signals leads to poor decoding accuracy of deep learning methods on motor imagery datasets,a method of MI-EEG(motor imagery electroencephalogram)classification based on attentional temporal convolution is proposed.The deep convolution module is used to initially extract the temporal and spatial information in the EEG signals,and three convolutional blocks with different size in the multi-scale convolution module are used to further extract the global and detailed features from the MI-EEG data.The most valuable features in the data are highlighted by means of multi-head attention module,and the advanced temporal features are extracted by means of temporal convolution network,and then the classification results are output by connected network and softmax layer.The experimental results show that,on the BCI competition IV-2b dataset,the proposed model can realize an average classification accuracy of 84.26%for the motor imagery binary classification task,and this method has significantly improved accuracy compared with existing benchmark models.