A PSO-MLPANN Hybrid Approach for Estimation of Human Joint Torques from sEMG Signals

Abstract

Mapping from the electrical activity of muscles to the joint torques is required in many applications like gait analysis, exoskeleton robots, automated rehabilitations and human-machine interactions. This paper investigates the application of a multilayer perceptron artificial neural network (MLPANN) for the estimation of the human knee torques from the surface electromyography (sEMG) signals of the corresponding muscles. Some experiments are performed on a human subject wearing an exoskeleton robot and repeating a special1-DOF motion called vertical sit-to-stand (VSTS) motion. The human knee angle is recorded from the knee encoder of the exoskeleton robot. The sEMG signals of four related lower limb muscles are also recorded at the same time. Then, the inverse dynamic model of the human in VSTS motion is used to compute the corresponding knee torque of the human. The recorded sEMG signals and the calculated torques are then used to form the input-output training set for the MLPANN. To find the best neuron weights for the MLP ANN, particle swarm optimization (PSO) is utilized. Results show that the EMG signals from lower limb muscles contain important information about the knee torques in a VSTS motion. Comparing the performance of the optimized MLP ANN with conventional MLP ANN and radial basis function artificial neural network (RBF ANN) indicates that the proposed method is more efficient in the estimation of joint torques.