Abstract: Abstract: Aiming at the multi-classification problem for motor imagery electroencephalography (MI-EEG), This paper improved on the basis of existing research, and constructed a lightweight convolutional neural network (L-Net) and a lightweight hybrid network (LH-Net) based on deep separable convolution. Experiments and analyses are carried out on the BCI competition IV-2a data set. The results show that L-Net can fit the data faster than LH-Net, and the training time is shorter. However, the stability of LH-Net is better than that of L-Net, and the classification performance on the test set has better robustness. Its average accuracy and kappa are about 5% higher than that of L-Net. In order to further improve the classification performance of the model, a new method of adding Gaussian noise based on the time-frequency domain is adopted, and data augmentation (DA) is performed on the training samples. The simulation verification of the noise intensity is carried out, and the optimal noise intensity range of the two models is inferred. The simulation results show that after using the data augmentation method in this paper, the four classification accuracy of the two models can reach more than 93%, and the classification effect has been significantly improved.

Key words: Keywords: electroencephalography, motor imagery, deep learning, data augmentation

摘要： 摘 要: 针对运动想象脑电（Motor imagery electroencephalography , MI-EEG）多分类问题，本文在已有研究的基础上进行改进，构建了基于深度可分离卷积的轻量级卷积神经网络(Lightweight convolutional neural network, L-Net）和轻量级混合网络（Lightweight hybrid network, LH-Net）。并在BCI竞赛Ⅳ-2a四分类数据集上进行了实验和分析，结果表明L-Net比LH-Net可以更快地拟合数据，训练时间更短，但是LH-Net的稳定性比L-Net更好，在测试集上的分类性能具有更好的稳健性，平均准确率和平均kappa比L-Net高出4%左右。为了进一步提升模型分类性能，采用了基于时频域的高斯噪声添加新方法，对训练样本进行了数据增强(Data augmentation, DA)。并针对噪声的强度进行了仿真验证，推测出两种模型的最优噪声强度的取值范围。仿真结果表明使用了本文的数据增强方法后，两种模型的四分类准确率最高可达93%以上，分类效果得到了明显提高。

关键词: 关键词: 脑电信号, 运动想象, 深度学习, 数据增强