A CNN-Based Approach Combining Temporal Shift Module and Attention for Motor Imagery Classification

This paper addresses the challenge of extracting spatial features from EEG signals and integrating temporal features when reorganizing electrodes into 3D data. The proposed method utilizes Convolutional Neural Networks (CNN) with Temporal Shift Module (TSM) and attention mechanism to capture spatio-temporal features of motor imagery EEG signals. Traditional EEG classification methods struggle with temporal dynamics and noise interference, affecting accuracy. To overcome this, an end-to-end classification network is designed, incorporating TSM for interacting with EEG information at different time points and attention mechanism for emphasizing important information. The model is validated using the BCI Competition IV 2a dataset and compared against methods like FBCSP and EEGNet. Experimental results demonstrate the model's superior performance, achieving approximately 10% improvement over FBCSP and an average accuracy increase of 4% compared to EEGNet. Ablation experiments verify the effectiveness of the TSM module. This method provides a stable and effective solution for classifying motor imagery EEG signals, benefiting the development and optimization of brain-computer interface systems.