Prediction of Exoskeleton-Induced Upper-Limb Movement Intention from Electromyography Signals Using Convolutional Neural Networks

Abstract

Surface electromyography (EMG) provides a non-invasive interface for decoding human motor intent and has been widely studied for rehabilitation and assistive robotics; however, reliable upper-limb motion intent classification remains challenging in rehabilitation scenarios, particularly for exoskeleton-guided movements where EMG activity is weaker and less discriminative than during voluntary motion. This paper proposes a lightweight convolutional neural network (CNN) framework for classifying upper-limb motions directly from raw EMG signals acquired during passive movement, targeting three representative rehabilitation tasks: elbow flexion–extension, shoulder flexion–extension, and shoulder abduction–adduction. Six-channel surface EMG signals were acquired using the CLEVERarm upper-limb exoskeleton developed by our research team, segmented into overlapping windows, and evaluated using trial-wise cross-validation. Class imbalance due to differences in movement duration was addressed using a weighted loss function. Experimental results demonstrate that the proposed framework achieved a mean classification accuracy of 92.01% (±5.8%) and an F1-score of 92.01% across trials, with strong class discrimination measured by the area under the receiver operating characteristic curve (AUC: 0.97–0.98), while maintaining low computational complexity. The findings demonstrate that subject specific CNN-based decoding of raw EMG enables motion classification during exoskeleton-driven passive upper-limb movements, which is relevant for rehabilitation applications.