Transformer-based network with spatial correlation change and multi-segment attention for sequential EMG recognition

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Surface electromyography (sEMG) is crucial for motion recognition but faces challenges in complex sequential movements.
• Existing methods struggle with sequential feature extraction and generalization across applications.

Purpose of the Study:

• To introduce a novel Transformer-based architecture, the Multi-Interval Driven Transformer (MIDT), for improved sEMG motion recognition.
• To enhance sequential feature modeling and address limitations in current sEMG analysis.

Main Methods:

• Developed MIDT with a muscle correlation-guided adaptive segmentation module.
• Implemented hierarchical self-attention mechanisms for capturing sub-movement features and long-range dependencies.
• Validated MIDT on a new upper-limb sequential movement dataset (ULSE) and a public dataset.

Main Results:

• MIDT achieved 92.56% classification accuracy on the ULSE dataset, outperforming state-of-the-art by 5.07%.
• Demonstrated superior robustness with 46.14% lower cross-subject variance.
• Attained 80.93% top accuracy on a public dataset, surpassing mainstream methods.

Conclusions:

• MIDT effectively decodes sub-movement execution and motion state transitions.
• Provides quantitative support for personalized motion control and rehabilitation assessment.
• Highlights broad application potential in wearable human-machine interfaces and neurorehabilitation engineering.