Abstract:

Aiming at the problem of low recognition accuracy of multi-class motion imaging EEG signals, this paper proposes a convolutional neural network model that incorporates attention modules. The model utilizes the attention module to fully mine the channels and spatial features of the EEG signals, and establishes the relationship between the importance and the recognition tasks, thereby improving the recognition accuracy of the motor imaging EEG signals. The signal undergoes a common space mode to improve the signal-to-noise ratio. The wavelet transform is used to convert the signal into a two-dimensional time-frequency map. The feature adjustment is carried out through the two dimensions of the channel and space in the attention module to strengthen the weakening of useful features. The useless features enable the convolutional network to fully extract higher-level abstract features and finally perform the recognition of the motion imaging task. The method in this paper has been evaluated on the BCI Competition IV Datasets 2a and BCI Competition III-IIIa datasets respectively, and compared with convolutional neural networks and other algorithms. The experimental results show that the method proposed in this paper can achieve good accuracy and can effectively improve the recognition accuracy of EEG signal motor imaging tasks.

Key words: motor imagination, common spatial pattern, wavelet transform, convolutional neural network, convolutional attention module

摘要：

针对多类别运动想象脑电信号识别精度不高的问题，提出了一种融合注意力模块的卷积神经网络模型。该模型利用注意力模块充分挖掘脑电信号的通道和空间特征，建立其与识别任务之间的重要程度关系，从而提高运动想象脑电信号的识别准确率。信号经过共空间模式提高信噪比，利用小波变换将信号转换成二维时频图，通过注意力模块中通道和空间两个维度进行特征的调整，以强化有用特征弱化无用特征，使卷积网络充分提取更高层次的抽象特征，并最终执行运动想象任务的识别。分别在BCI竞赛IV Datasets 2a和BCI竞赛III-IIIa数据集上进行了有效性评价，并与卷积神经网络以及其他算法进行了比较。实验结果表明，提出的方法可达到良好的准确率，能够有效提高脑电信号运动想象任务的识别准确率。

关键词: 运动想象, 共空间模式, 小波变换, 卷积神经网络, 卷积注意力模块