Reliable Source-Free Domain Adaptation for Cross-User Myoelectric Pattern Recognition

• Published in: IEEE sensors journal Vol. 24; no. 23; pp. 39363 - 39372
• Main Authors: Zhang, Xuan, Wu, Le, Zhang, Xu, Chen, Xiang, Li, Chang, Chen, Xun
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.12.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation; Adaptation models; Algorithms; Cross-subject; Data models; Electrodes; Electromyography; EMG control; Fingers; Human motion; Model updating; Muscles; Myoelectricity; Parameters; Pattern recognition; Reliability; Sensors; Solid modeling; source-free domain adaptation (SFDA); Teachers; transfer learning
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2024.3475818

Surface electromyographic (sEMG) signals are widely used for human-machine interaction (HMI) control, providing information about user movement intent. However, interindividual differences in muscle anatomy pose a challenge for cross-user myoelectric pattern recognition (MPR) algorithms. Existing cross-user MPR algorithms rely on domain adaptation (DA) using data from source and target users for model updating. However, using historical user data in commercial HMI devices risks disclosing user health information and biometric privacy. Therefore, enabling MPR algorithms to update models quickly and solely based on target user data in a source-free manner is crucial. With this aim, this article proposes a reliable source-free DA (RSFDA) framework that enables rapid cross-user application of myoelectric algorithms. Specifically, the proposed FSFDA framework employs a teacher-student framework. Both the teacher and student models are initialized with the source model. During the update of model parameters, the teacher framework utilizes historical network parameters to prevent knowledge forgetting, while the student model continuously updates parameters while ensuring consistency with the teacher model output. As a result, the final student model demonstrates increased stability and reliability in classifying gestures from new users. The experimental results demonstrate that the proposed RSFDA approach achieves a recognition accuracy of <inline-formula> <tex-math notation="LaTeX">94.44\%~\pm ~5.68\% </tex-math></inline-formula>, which outperforms the state-of-the-art methods on a high-density sEMG dataset using only five samples per gesture. Furthermore, this framework is effective even when only one sample is provided or when gesture categories are missing. This study provides a faster and safer strategy for cross-user MPR, enabling multiuser control.