Improving the Robustness and Adaptability of sEMG-Based Pattern Recognition Using Deep Domain Adaptation.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Rehabilitation Technology

Background:

• Surface electromyography (sEMG) based pattern recognition (PR) offers potential for improving amputees' quality of life.
• Current sEMG PR systems lack the robustness and adaptability required for widespread clinical application.

Purpose of the Study:

• To develop a novel deep learning network, the multi-task dual-stream supervised domain adaptation (MDSDA), for robust and adaptable sEMG-based PR.
• To simultaneously achieve long-term reliability and user adaptability in sEMG pattern recognition systems.

Main Methods:

• A convolutional neural network (CNN) forms the basis of the proposed MDSDA network.
• Long-term, multi-subject sEMG signal acquisition was performed with 12 able-bodied subjects performing 30 distinct gestures (static and dynamic).
• Four train-test estimations were employed to rigorously evaluate MDSDA's robustness and adaptability against conventional CNN and fine-tuning methods.

Main Results:

• The MDSDA network demonstrated superior performance compared to standard CNN and fine-tuning approaches.
• Analysis revealed high separability between static and dynamic gestures performing similar actions, suggesting a reduced signal collection burden.
• The proposed MDSDA system achieved robust and generalized pattern recognition.

Conclusions:

• The MDSDA network shows significant potential for enhancing the reliability and adaptability of sEMG-based PR systems.
• This advancement could lead to more effective and user-friendly prosthetic devices for amputees.
• The findings suggest MDSDA is a promising solution for clinical applications requiring robust and generalized PR.