SEMG-based consecutive estimation of human lower limb movement by using multi-branch neural network

Abstract

Surface electromyography (sEMG) has the potential for human lower limb movement analysis, including gait phases recognition and joint angle estimation, which can provide a great level of human interaction with the exoskeleton orthotic devices. In this paper, a method based on deep learning is proposed, which maps the multichannel sEMG signals to human lower limb movement, including 4 different gait phases and 3 flexion/extension joint angles. First, five time-domain features and spectrogram data as frequency domain features are extracted from the sEMG data from 8 muscles of right legs. Then, a multi-branch neural network (MBNN) with convolutional neural layers and recurrent neural layers is constructed, which uses both the extracted features and raw data as input to analyze human movement. Experimental results show that the mean accuracy of classification of our proposed methods can reach high level (90.92 ± 3.58% for speed dependent and 85.04 ± 5.14% for speed independent). Meanwhile, average of the root mean square error between estimated and real joint angles is (3.75 ± 1.52 degree for speed dependent and 6.12 ± 2.54 degree for speed independent). These results indicate that the proposed method can be used to facilitate adoption of exoskeleton orthotic device in real-life applications, with gait phases determining impedance characteristic of devices and angles estimating joint movement.