ADFCNN: Attention-Based Dual-Scale Fusion Convolutional Neural Network for Motor Imagery Brain–Computer Interface

Convolutional neural networks (CNNs) have been successfully applied to motor imagery (MI)-based brain–computer interface (BCI). Nevertheless, single-scale CNN fail to extract abundant information over a wide spectrum from EEG signals, while typical multi-scale CNNs cannot effectively fuse information from different scales with concatenation-based methods. To overcome these challenges, we propose a new scheme equipped with attention-based dual-scale fusion convolutional neural network (ADFCNN), which jointly extracts and fuses EEG spectral and spatial information at different scales. This scheme also provides novel insight through self-attention for effective information fusion from different scales. Specifically, temporal convolutions with two different kernel sizes identify EEG $\boldsymbol{\mu }$ and $\boldsymbol{\beta }$ rhythms, while spatial convolutions at two different scales generate global and detailed spatial information, respectively, and the self-attention mechanism performs feature fusion based on the internal similarity of the concatenated features extracted by the dual-scale CNN. The proposed scheme achieves the superior performance compared with state-of-the-art methods in subject-specific motor imagery recognition on BCI Competition IV dataset 2a, 2b and OpenBMI dataset, with the cross-session average classification accuracies of 79.39% and significant improvements of 9.14% on BCI-IV2a, 87.81% and 7.66% on BCI-IV2b, 65.26% and 7.2% on OpenBMI dataset, and the within-session average classification accuracies of 86.87% and significant improvements of 10.89% on BCI-IV2a, 87.26% and 8.07% on BCI-IV2b, 84.29% and 5.17% on OpenBMI dataset, respectively. What is more, ablation experiments are conducted to investigate the mechanism and demonstrate the effectiveness of the dual-scale joint temporal-spatial CNN and self-attention modules. Visualization is also used to reveal the learning process and feature distribution of the model.