A Dynamic Balanced Single-Source Domain Generalization Model for Cross-Posture Myoelectric Control.

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Machine Learning

Background:

• Wearable electromyography (EMG) based human-computer interaction (HCI) faces challenges due to body posture variations affecting surface EMG (sEMG) signal features.
• This variability significantly reduces the accuracy of gesture recognition in sEMG interfaces.
• Robustness in sEMG-based gesture recognition requires effective mitigation of body position variability.

Purpose of the Study:

• To develop a transfer learning framework for robust sEMG-based gesture recognition.
• To address the challenge of reduced gesture recognition accuracy caused by body posture variability.
• To enable models trained on one posture to generalize well to others.

Main Methods:

• Proposed a Dynamic Balanced Single-Source Domain Generalization (DBSS-DG) transfer learning framework.
• Utilized sEMG signal data from a single posture (standing) as the source domain for model training.
• Validated the framework on a dataset comprising 16 subjects across four postures: standing, sitting, walking, and lying.

Main Results:

• The DBSS-DG model achieved high gesture recognition accuracies: 90.79% (sitting), 88.78% (walking), and 90.87% (lying) when trained on the standing posture.
• Demonstrated an average accuracy improvement of 4.71% compared to non-transfer learning approaches.
• Outperformed several established single-source domain generalization methods.

Conclusions:

• The proposed DBSS-DG framework effectively enhances the robustness of sEMG-based gesture recognition across different body postures.
• This approach offers a significant improvement over traditional methods, showing promise for practical HCI applications.
• The study highlights the potential of domain generalization techniques in overcoming signal variability in wearable sensor data.