Attention-based multi-views and multi-scale temporal information fusion convolutional neural network for motor imagery EEG decoding

In this paper, we propose an innovative deep learning architecture called Attention-based Multi-view and Multi-scale Temporal Information Fusion Convolutional Neural Network (AMMFCNN) for decoding motor imagery EEG signals. AMMFCNN effectively captures global dependencies and multichannel, multi-temporal information. The model applies spectral filtering to the EEG data and uses channel attention mechanisms to redistribute feature weights across frequency bands. Multi-scale temporal and spatial convolutions are then applied to extract both temporal and spatial features, which are further processed using multi-head self-attention modules. These features are fused using element-wise addition before being passed to a fully connected layer for classification. Our method was validated on the BCI Competition IV 2a dataset and achieved superior performance compared to state-of-the-art methods, demonstrating its effectiveness for motor imagery EEG decoding. The results show that AMMFCNN outperforms previous models, such as EEGNet, FBCNet, IFNet, and Conformer by $\mathbf{9. 0 2 \%}$, $\mathbf{6. 3 2 \%}$, $\mathbf{2. 6 6 \%}$, and 3.51% respectively. The source code can be found at the following link: https://github.com/Qiantoulv/ADMMFCNN.