A Novel Dual-Model Adaptive Continuous Learning Strategy for Wrist-sEMG Real-Time Gesture Recognition

Abstract

Surface electromyography (sEMG) is a promising technology for hand gesture recognition, yet faces challenges in user mobility and individual calibration. This paper introduces a novel dual-model adaptive continuous learning (DM-ACL) strategy for wrist-based sEMG real-time gesture recognition. The core of the DM-ACL strategy is a semi-supervised online learning algorithm that uses the kNN model to provide auxiliary labels for real-time sEMG signals, enhancing the robustness and adaptability of the deep learning model. Experimental results show that the DM-ACL strategy outperforms conventional transfer learning (TL) methods. Using the CNN-LSTM model as the baseline, the DM-ACL method achieved a recognition accuracy of 95.33% with an average of 33.6 s of sEMG data per gesture, while the conventional TL method attained an accuracy of 82.82%. With the CNN model as the baseline, the DM-ACL method achieved a recognition accuracy of 92.37% with an average of 48 s of sEMG data per gesture, while the conventional TL method attained an accuracy of 84.59%. The DM-ACL strategy efficiently improves performance for new users and maintains high accuracy across sessions, even in the presence of inter-session domain shifts. This enhances the practical usability of sEMG-based gesture recognition systems, particularly in real-time applications.