Movement recognition via channel-activation-wise sEMG attention

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Surface electromyography (sEMG) signals offer rich data for understanding muscle activity and user intent.
• Current sEMG feature extraction methods face challenges due to signal stochasticity, transiency, and non-stationarity.
• Effective feature extraction is crucial for applications in rehabilitation, clinical diagnosis, and human-computer interaction.

Purpose of the Study:

• To overcome limitations in current sEMG signal processing.
• To develop a robust method for extracting representative features for movement recognition.
• To enhance the accuracy and generalizability of sEMG-based gesture classification.

Main Methods:

• A novel 3-axis view of sEMG features integrating temporal, spatial, and channel-wise summaries.
• Utilizing a state-of-the-art Transformer architecture with an attention-based module.
• Enabling efficient parallel search and extraction of global contextual relevance among channels.

Main Results:

• The proposed method achieved state-of-the-art (SOTA) accuracy on two Ninapro datasets.
• Demonstrated superior performance compared to existing methods for sEMG gesture classification.
• Showcased strong generalization ability through pretraining and fine-tuning on different datasets.

Conclusions:

• The novel 3-axis sEMG feature extraction method significantly improves movement recognition accuracy.
• The Transformer-based approach effectively captures inter-channel relationships for enhanced performance.
• The method exhibits excellent generalization, paving the way for broader clinical and engineering applications.