Continuous Angle Prediction of Lower Limb Knee Joint Based on sEMG

Abstract

With the development of rehabilitation robots in the field of medical rehabilitation, the human action classification based on surface electromyography (sEMG) has been widely used in robot human-computer interaction. Since the estimation of continuous joint angle can be employed to improve the performance of human-machine coordination, the accurate extraction of continuous joint angle from sEMG is one of the research difficulties in the field. This paper presents an Attention-LSTM model based on the electromyographic fusion features. The sEMG of the lower limb in four channels and the corresponding knee angles of the lower limb during normal walking are collected by the sEMG acquisition system and the motion capture system. The Butterworth band-pass filter is used to reduce the out-of-band high-frequency and low-frequency noise of the data, and then the data is normalized. The sliding window is utilized to extract the time advance features and time delay features of sEMG. After the smoothing, the data is input into the Attention-LSTM model. The mean square error of prediction accuracy is used as the evaluation index. Finally, the data are analyzed by one-way analysis of variance (ANOVA) for significant differences $(\mathrm{p}\lt0.05)$. The experimental results show that the RMSE difference of the proposed fusion feature for predicting the knee angle is significantly smaller than that of the control group, and can accurately predict the knee angle of the lower extremity in the future from 75ms to 200ms (RMSE=$1.7444 \pm 0.2031)$, thus achieving the better human-computer interaction.