Abstract: To address the issues of slow decoding speed in large-scale networks and insufficient utilization of feature information in decoding motor imagery electroencephalogram (MI-EEG) signals, a lightweight MI-EEG decoding neural network based on multi-domain feature fusion was proposed. The proposed network introduced lightweight modules to extract multi-domain features, including the use of SincNet for frequency-domain feature extraction and temporal convolutional networks (TCN) for time-domain feature extraction. Additionally, after extracting time-frequency domain features, the squeeze-and-excitation (SE) attention was incorporated to adaptively recalibrate feature maps, emphasizing important features and suppressing redundant information. Finally, separable convolution was employed to effectively fuse time-frequency features, addressing the limitation of single-domain feature information. Furthermore, a joint loss function combining cross-entropy and center loss was adopted to optimize both intra-class and inter-class classification performance during network training. The proposed network has parameter counts of 6,870, 5,690, and 6,870 on the BCI 2a, SMR-BCI, and OpenBMI motor imagery public datasets, respectively, achieving average accuracies of 74.78%, 71.93%, and 65.40%, with average Kappa values of 0.70, 0.66, and 0.59. The experimental results show that on the BCI 2a dataset, compared with networks such as DeepConvNet, EEGNet and EEGTCNet, the improvement is 11.06, 8.85 and 6.36 percentage points, respectively; on the SMR-BCI dataset, compared with networks such as DeepConvNet, EEGNet and EEGTCNet, the improvement is 10.53, 4.17, and 3.57 percentage points; and on the OpenBMI dataset, 5.09, 4.99, and 2.33 percentage points compared to networks such as DeepConvNet, EEGNet, and EEGTCNet, respectively. The proposed network balances decoding performance while ensuring lightweight.

Key words: Motor Imagery &#40, MI&#41, Electroencephalogram &#40, EEG&#41

摘要： 针对解码运动想象脑电信号(MI-EEG)时大规模网络解码速度慢，以及特征信息未能充分利用等问题，提出了一个多域特征融合的轻量化MI-EEG解码神经网络。所提网络通过轻量化模块提取多域特征，包括利用SincNet提取频域特征、时序卷积网络(TCN)提取时域特征；在时频域特征提取之后引入挤压激励(SE)注意力对特征图进行自适应校准，突出关注重要特征并抑制冗余信息。最后，利用可分离卷积实现时频特征的有效融合，以解决单域特征信息不足的问题。此外，采用交叉熵和中心损失函数联合约束网络训练过程，以优化类内和类间的分类性能。该网络在BCI 2a、SMR-BCI和OpenBMI三个运动想象公开数据集上的参数量分别为6,870、5,690和6,870，平均准确率分别达到74.78%、71.93%和65.40%，平均Kappa值分别达到0.70、0.66和0.59。实验结果表明，在BCI 2a数据集上，与DeepConvNet、EEGNet和EEGTCNet等网络相比分别提升了11.06、8.85和6.36个百分点；在SMR-BCI数据集上，与DeepConvNet、EEGNet和EEGTCNet等网络相比分别提升了10.53、4.17和3.57个百分点；在OpenBMI数据集上，与DeepConvNet、EEGNet和EEGTCNet等网络相比分别提升了5.09、4.99和2.33个百分点。所提网络在保证轻量化的同时兼顾了解码性能。

关键词: 运动想象, 脑电信号, 轻量化, 多域特征融合, 挤压激励注意力