sEMG-Based Estimation of Human Arm Endpoint Stiffness Using Long Short-Term Memory Neural Networks and Autoencoders

Human upper limb impedance parameters are important in the smooth contact between stroke patients and the upper limb rehabilitation robot. Surface electromyography (sEMG) reflects the activation state of muscle and the movement intention of human body. It can be used to estimate the dynamic parameters of human body. In this study, we propose an estimation model combining long short-term memory (LSTM) neural network and autoencoders to estimate the endpoint stiffness of human arm from sEMG and elbow angle. The sEMG signal is a time varying nonlinear signal. Extracting key features is critical for fitting models. As an unsupervised neural network, autoencoders can select the proper features of sEMG for the estimation. LSTM neural network has good performance in dealing with time series problems. Through a 4-layer LSTM neural network, the mapping relationship between sEMG features and endpoint stiffness is constructed. To prove the superiority of the proposed model, the correlation coefficient between theoretical stiffness calculated by Cartesian impedance model and estimated stiffness and root mean square error (RMSE) is used as the evaluation standard. Compared with two other common models by experiments, the proposed model has better performance on root mean square error and correlation coefficient. The root mean square error and correlation coefficient of proposed model are 0.9621 and 1.732.