Long Short-Term Memory Neural Network Knee- Joint Angle Estimation Using sEMG Signal and Time- Advanced Feature

Abstract

The ability of wearable robots to work based on the wearer’s intension is a major challenge in the control and design of the robots. However, this can be improved by continuously estimating the knee-joint angle using an estimator based on surface electromyography (sEMG) signal. In this research, Long Short-Term Memory (LSTM) Neural Network is utilized to continuously estimate the knee-joint angle from sEMG data so as to determine the amount of compensation needed, with Back Propagation Neural Network (BPNN) as comparison model. The sEMG signals Time Domain (TD) and Time-Advanced Domain (TAD) features were extracted with the later taking into consideration that the production of sEMG signal starts 20-200 ms before the corresponding muscle action. Root-mean-square (RMS), Mean-Absolute-Value (MAV), Variance (VAR) and Standard Deviation (STD) were extracted from each domain and were used separately to train the LSTM model as well as the BPNN model. The performance of each model was evaluated based on Root-Mean-Squared Error (RMSE). The sEMG of the two selected muscles were used as input and the knee-joint angle as output. Overall results show that LSTM using TAD across all features extracted shows improved performance (lower RMSE) by an average of 3.2566⁰ when compared to BPNN model with an average lower RMSE of 6.1945⁰.