Continuous Prediction of Lower-Limb Joint Torque Based on IPSO-LSTM

Abstract

Predicting joint torque is increasingly important for wearable devices, especially exoskeleton robots. Continuous joint torque prediction based on surface electromyography (sEMG) signals and joint angles can be used for human-machine cooperative control of exoskeletons. Improved particle swarm optimization (IPSO) algorithm was proposed to optimize long short-term memory (LSTM) neural network, which was trained with lower-limb joint angles and sEMG signals of ten muscles to predict hip flexion/extension, knee flexion/extension and ankle dorsiflexion/plantarflexion torques. We used root mean square error (RMSE) and coefficient of determination between predicted and measured joint torques to evaluate the prediction performance. According to the results, compared with LSTM and PSO-LSTM (Particle Swarm Optimization-based LSTM) model, the mean RMSE of IPSO-LSTM (Improved Particle Swarm Optimization-based LSTM) decreases by 21.5% and 12.7%, respectively, and the mean coefficient of determination increases 0.013 and 0.0057, respectively. Therefore, IPSO-LSTM has higher accuracy in continuous prediction of joint torque of lower limbs.