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

Abstract

Electromyography (EMG) based Human-Computer Interaction (HCI) through wearable devices frequently encounter variability in body postures, which can modify the amplitude and frequency features of surface EMG (sEMG) signals. This variability often results in reduced gesture recognition accuracy. To enhance the robustness of sEMG-based gesture interfaces, mitigating the effects of body position variability is essential. In this paper, we proposed a Dynamic Balanced Single-Source Domain Generalization (DBSS-DG) transfer learning framework, which only used sEMG signal data from one posture as source domain for model training but can also generate good performance under different body postures as target domain. Validation was performed on the sEMG dataset from 16 subjects across four postures: standing, sitting, walking, and lying. With standing as the source domain, the model achieved gesture recognition accuracies of 90.79 ± 0.09%, 88.78 ± 0.06%, and 90.87 ± 0.1% for sitting, walking, and lying as the target domains, respectively, producing an average improvement of 4.71% over non-transfer learning approaches. Furthermore, the performance of our model exceeded that of many well-known single-source domain generalization methods, establishing its effectiveness in practical applications.