Masked Autoencoder with Swin Transformer Network for Mitigating Electrode Shift in HD-EMG-based Gesture Recognition.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• High-density surface electromyography (HD-sEMG) is vital for myoelectric control gesture recognition.
• HD-sEMG models suffer performance degradation due to recording condition variability, such as electrode shift.
• Existing models lack robustness against real-world recording condition changes.

Purpose of the Study:

• To propose a novel framework, Masked Autoencoder with Swin Transformer (MAST), for robust HD-sEMG-based gesture recognition.
• To enhance the generalization capabilities of pattern recognition models for myoelectric control.
• To address the performance degradation caused by electrode shifts in HD-sEMG recordings.

Main Methods:

• Developed the MAST framework utilizing a masked subset of HD-sEMG channels for training.
• Employed four masking strategies: random block, temporal, sensor-wise random, and multi-scale masking.
• Utilized a three-path Swin-Unet encoder-decoder architecture to capture time-domain, frequency-domain, and magnitude-based features.
• Implemented self-supervised pre-training on augmented inputs to improve model generalization.

Main Results:

• The MAST framework demonstrated superior performance compared to existing methods.
• The proposed masking strategies and multi-path architecture effectively learned robust latent representations.
• Self-supervised pre-training significantly improved model generalization across different recording conditions.

Conclusions:

• The MAST framework offers a robust solution for HD-sEMG-based gesture recognition, mitigating performance loss from electrode shifts.
• The combination of masking strategies and Swin Transformer architecture enhances model adaptability and generalization.
• This approach holds significant potential for improving the reliability of myoelectric control systems.