Improving cross-session motor imagery decoding performance with data augmentation and domain adaptation

Abstract

Recent research has increasingly utilized deep learning (DL) to decode electroencephalogram (EEG) signals, enhancing the accuracy of motor imagery (MI) classification. While DL has improved MI decoding performance, challenges persist as distribution variances in MI-EEG data across different sessions. Additionally, the collection of EEG signals for MI tasks presents significant challenges, particularly in terms of time and economic costs. Data collection not only requires professional equipment and controlled environments but also demands the cooperation of a large number of participants to obtain sufficient sample size and diversity. To address these issues, this study proposes two methods to improve the decoding performance of MI-EEG signals based on an improved lightweight network. Firstly, a recombination-based data augmentation method leveraging channel knowledge is proposed to expand the training dataset and enhance model classification generalization, without the need for additional experiments to collect new data. Secondly, an improved domain adaptation network is introduced to align feature distributions between different domains, minimizing domain gaps. The proposed domain adaptation method aligns the target EEG domain with the corresponding class centers using pseudo-labeling. Extensive experiments are conducted using a cross-session training strategy on the BCIC IV 2a and BCIC IV 2b datasets. The results demonstrate that the proposed data augmentation method and improved domain adaptation method effectively enhance classification accuracy, providing a novel perspective for the practical application of MI-EEG.