Improving the Adaptability of sEMG-Based Gesture Recognition Using Variational Mode Decomposition and Transfer Learning
Hao Wu, Feng Wang, Juan Zhao, Seiichi Kawata, Jinhua She · 2024
Surface electromyography (sEMG) based gesture recognition has received broad attention and application in rehabilitation areas. sEMG signals exhibit strong user dependence properties among users with different physiology, causing the inapplicability of the recognition model on new users. Transfer learning (TL) is a representative method to reducing user gaps by utilizing features already learned by pre-trained models. However, TL uses a large amount of training data due to the discrepancy of sEMG among different users, which increases the training burden. In this paper, a multi-user adaptive network (MUAN) is devised to decompose the insensitive features among different users to improve gesture recognition accuracy for new users, which is based on variational modal decomposition (VMD), convolutional neural network (CNN), and TL. Ninapro dataset is used to evaluate the adaptability of MUAN and the training burden on new users. Experimental results show that MUAN outperforms CNN and TL, and reduces the training burden for new users. MUAN has the potential to provide a robust and generalized HMI system for clinical applications.