Abstract
This study aims to classify five typical motion states of the human upper limb based on surface electromyography signals, thereby supporting the real-time control system of an assistive upper limb exoskeleton. We propose a deep learning model combining convolutional neural networks, bidirectional long short-term memory networks, and attention mechanism to enhance the accuracy of motion state recognition in complex scenarios. Surface electromyography data were collected from ten participants for the biceps, triceps, and deltoid muscles, covering five representative states: resting, mild activity, rapid movement, dynamic load-bearing, and static load-bearing. Following the systematic fusion of multi-domain features spanning time, morphological, frequency, and cepstral characteristics, temporal features were structured through sliding window segmentation to serve as inputs for the proposed model. The proposed model achieved a classification accuracy of 97.29% on the test set, with an average accuracy of 88.17 ± 5.39% under leave-one-subject-out cross-validation, outperforming baseline algorithms. These findings highlight the model's potential in motion state classification, facilitating advanced, intelligent control capabilities of human-exoskeleton systems.