Instant Motion Recognition of Lower Limb Using Surface Electromyography Signal Based on Convolutional Neural Network

The field of rehabilitation robotics has emerged as a prominent area of research in the medical community in recent years, offering innovative and promising solutions for patients with physical disabilities. The existing lower limb rehabilitation robots have some shortcomings in human-machine compatibility and motion synergy due to the limitation of the real-time human motion intention recognition. In this paper, an instant motion recognition of the lower limb was proposed using surface electromyography (sEMG) for significantly shortening the time delay between the robot motion and the actual human motion. The first 200ms data of the complete sEMG signal for a single motion cycle was employed in the recognition algorithm. The proposed recognition algorithm demonstrated superior accuracy compared to traditional machine learning methods. The accuracy of 9-class recognition was 94.32%, while 12-class recognition achieved that of 89.32%. A data augmentation method based on magnitude warping followed by shifting was employed to enhance recognition accuracy by 11.37% in 9-class recognition and 19.23% in 12-class recognition when comparing with no augmentation. The optimal results for motion recognition were observed when the first 200ms data of the complete sEMG signal for a single motion cycle with a duration of 1.0s were utilized, as evidenced by comparative results. The proposed method offers a novel approach to enhancing human-machine compatibility and comfort in the application of lower limb rehabilitation robots in the future.