TDLNet: Transfer Data Learning Network for Cross-Subject Classification Based on Multiclass Upper Limb Motor Imagery...

Area of Science:

• Neuroscience
• Biomedical Engineering
• Machine Learning

Background:

• Current motor imagery brain-computer interfaces (MI-BCI) lack diverse instruction sets for single-limb movements, hindering practical applications.
• Classifying brain activity across different individuals remains a significant challenge in MI-BCI development.

Purpose of the Study:

• To propose a novel transfer data learning network (TDLNet) for cross-subject intention recognition in multi-class upper limb motor imagery.
• To effectively decode multi-category motor imagery for single-limb movements to advance MI-BCI technology.

Main Methods:

• Developed TDLNet, incorporating a Transfer Data Module (TDM) for processing cross-subject electroencephalogram (EEG) signals and fusing channel features.
• Integrated a Residual Attention Mechanism Module (RAMM) to dynamically focus on relevant EEG signal channels.
• Utilized a feature visualization algorithm based on occlusion signal frequency for qualitative analysis.

Main Results:

• TDLNet achieved superior classification results compared to CNN-based and transfer learning methods on two datasets.
• In a 6-class scenario, TDLNet reached 65%±0.05 accuracy on the UML6 dataset and 63%±0.06 on the GRAZ dataset.
• Visualization confirmed TDLNet's ability to generate distinct classifier patterns for various upper limb motor imagery tasks using different signal frequencies.

Conclusions:

• TDLNet effectively addresses the challenge of cross-subject classification for multi-class upper limb motor imagery.
• The proposed framework demonstrates significant potential for advancing the practical applications of MI-BCI.