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）， a lightweight MI-EEG decoding neural network with multi-domain feature fusion was proposed. In the proposed network， lightweight modules were introduced to extract multi-domain features， including the use of SincNet for frequency-domain feature extraction and the use of Temporal Convolutional Network （TCN） for time-domain feature extraction. Additionally， after extracting time-frequency domain features， the Squeeze-and-Excitation （SE） attention was incorporated to calibrate feature maps adaptively， thereby emphasizing important features and suppressing redundant information. Finally， separable convolution was employed to fuse time-frequency features effectively， thereby addressing the limitation of single-domain feature information. Furthermore， a joint loss function combining cross-entropy and center loss was adopted to constrain network training， thereby optimizing both intra-class and inter-class classification performance. Experimental results showed that on the Motor Imagery （MI） public datasets BCI 2a， SMR-BCI， and OpenBMI， the proposed network had the parameter counts of 6 870， 5 690， and 6 870， respectively， the average accuracies of 74.78%， 71.93%， and 65.40%， respectively， and the average Kappa values of 0.70， 0.66， and 0.59， respectively； Compared to Deep Convolutional Network （DeepConvNet）， lightweight EEG convolutional neural Network （EEGNet）， and Temporal Convolutional Network-based EEG Recognition （EEG-TCNet）， the proposed network achieved average accuracy improvements of 11.06， 8.85， and 6.36 percentage points on the BCI 2a dataset； 10.53， 4.17， and 3.57 percentage points on the SMR-BCI dataset； and 5.09， 4.99， and 2.33 percentage points on the OpenBMI dataset， respectively. The results demonstrate that the proposed network ensures lightweight design while maintaining robust decoding performance.

Key words: Motor Imagery (MI), ElectroEncephaloGram (EEG), lightweighting, multi-domain feature fusion, Squeeze-and-Excitation (SE) attention

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，提出一个多域特征融合的轻量化MI-EEG解码神经网络。该网络通过轻量化模块提取多域特征，包括利用SincNet提取频域特征以及利用时间卷积网络（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，以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，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的训练过程，以优化类内和类间的分类性能。实验结果表明，该网络在BCI 2a、SMR-BCI和OpenBMI这3个运动想象（MI）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。与深层卷积网络（DeepConvNet）、轻量级脑电信号卷积神经网络（EEGNet）和基于时间卷积的脑电信号识别网络（EEG-TCNet）相比：在BCI 2a数据集上，所提网络的平均准确率分别提升了11.06、8.85和6.36个百分点；在SMR-BCI数据集上，所提网络的平均准确率分别提升了10.53、4.17和3.57个百分点；在OpenBMI数据集上，所提网络的平均准确率分别提升了5.09、4.99和2.33个百分点。可见，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