Simultaneous and Continuous Motion Estimation of Upper Limb Based on SEMG and LSTM

Abstract

Continuous joint motion estimation based on surface electromyography (SEMG) signal plays an important role in human-machine interaction (HMI). Due to the requirements of several possible applications in which simultaneous control of multiple joints is needed, such as intelligent limbs and exoskeletons, simultaneous and continuous joint angle estimation of multiple joints is of great significance. In this paper, long short-term memory network (LSTM) was used to simultaneously estimate continuous elbow and wrist joint angles using time-domain features extracted from the SEMG. Nine healthy subjects participated in the experiment and their six muscle (i.e., biceps brachii (BB), triceps brachii (TB), flexor carpi radialis (FCR), extensor carpi radialis (ECR), flexor carpi ulnaris (FCU), and extensor carpi ulnaris (ECU)) SEMG signals were taken as algorithm inputs. SEMG features of each channel including mean absolute value (MAV) and root mean square (RMS) are extracted. The experimental results demonstrate that LSTM with 12 dimensional features as input presents the best estimation performance. Compared with estimation of single joint angle using genetic algorithm (GA) optimized back propagation neural network (BPNN), the average root mean square error (RMSE) for elbow and wrist were respectively reduced by 23.26% and 5.68% while the average coefficient of determination (${R}^{2}$) for elbow and wrist were respectively increased by 4.90% and 2.25%.When compared with LSTM with 8-dimensional features as input, the average RMSE for elbow and wrist were respectively reduced by 29.38% and 1.88% while the average ${R}^{2}$ for elbow and wrist were respectively increased by 15.62% and 17.73%.