Cross-User Electromyography Pattern Recognition Based on a Novel Spatial-Temporal Graph Convolutional Network.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Myoelectric control technology development relies on accurate electromyography (EMG) pattern recognition.
• High-density surface EMG (HD-sEMG) signals offer rich spatiotemporal information crucial for advanced pattern recognition.
• Existing methods often struggle with individual differences and user training burdens.

Purpose of the Study:

• To develop a robust electromyography (EMG) pattern recognition solution using graph neural networks (GNNs).
• To integrate a convolutional neural network (CNN) feature extraction module with a multi-view spatial-temporal graph convolutional network (MSTGCN).
• To evaluate the proposed CNN-MSTGCN model's performance in user-independent and transfer learning scenarios.

Main Methods:

• A novel CNN-MSTGCN model was designed by integrating a CNN feature extraction module into an MSTGCN classifier.
• Experiments were conducted using HD-sEMG data from 17 gestures across 11 subjects.
• Ablation studies, user-independent recognition, and transfer learning-based cross-user recognition experiments were performed.

Main Results:

• The CNN-MSTGCN model demonstrated significant improvements in EMG pattern recognition compared to ResNet50 and LSTM.
• In user-independent tests, CNN-MSTGCN achieved a 68% recognition rate (vs. ResNet50: 47.5%, LSTM: 57.1%).
• Transfer learning experiments (TL-CMSTGCN) yielded a 92.3% recognition rate, outperforming TL-ResNet50 (84.6%) and TL-LSTM (85.3%).

Conclusions:

• The proposed CNN-MSTGCN model effectively enhances EMG pattern recognition accuracy.
• GNNs show promise in mitigating the impact of individual differences in EMG signals.
• The developed model offers a viable solution for robust myoelectric control and reduced user training.
• The research highlights GNNs' potential for advancing myoelectric control technology.