文献类型：期刊文献

英文题名：Online Incremental Learning Based on Mechanomyography for Upper-Limb Action Pattern Recognition in Long-Term Rehabilitation Processes

作者：Cao, Gangsheng[1]; Xia, Chunming[1]; Chang, Yukun[1]

机构：[1] East China University of Science and Technology, School of Mechanical and Power Engineering, Shanghai, China

年份：2025

起止页码：208

外文期刊名：2025 IEEE 3rd International Conference on Sensors, Electronics and Computer Engineering, ICSECE 2025

收录：EI(收录号：20260820131260)

语种：英文

外文关键词：Convolutional neural networks - E-learning - Exoskeleton (Robotics) - Learning algorithms - Learning systems - Multilayer neural networks - Online systems - Pattern recognition systems

摘要：This paper proposes an online incremental learning (OIL) pattern recognition method based on mechanomyography (MMG), aiming to address the issue of reduced recognition accuracy in traditional MMG-based action pattern recognition methods during long-term rehabilitation training processes. The experiment simulated rehabilitation training frequency by collecting MMG signals generated during four types of hand movements (wrist flexion, extension, ulnar flexion, and radial flexion) from subjects over 12 consecutive days, resulting in 12 groups of MMG datasets. Three neural network algorithms, namely perceptron algorithm (PA), multilayer perceptron (MLP), and convolutional neural network (CNN), were employed for action pattern recognition research. The first six groups of MMG were used for basic model training, while the last six groups were utilized for online incremental learning and recognition performance testing of the models. Experimental results demonstrate that the online incremental learning method significantly enhances the models' recognition performance on new data. When the proportion of the incremental training set is 20%, the recognition rates of PA-OIL, MLP-OIL, and CNN-OIL reach 94.02%, 94.04%, and 96.02%, respectively, representing a comprehensive improvement of over 10% compared to models without online incremental learning. This online incremental learning approach enables pattern recognition of non-identically distributed MMG data while enhancing model recognition accuracy and generalization capabilities. It is expected to promote the application of MMG-based upper-limb rehabilitation exoskeletons and advance the development of upper-limb rehabilitation systems. ? 2025 IEEE.