Elbow Joint Angle Estimation in Wavelet Neural Network Using ReliefF Selected Features of sEMG and Post Filter

In the control of electromyography-driven exoskeleton, the algorithms must estimate the human motion with fast speed and high accuracy. In this paper, we propose a wavelet neural network (WNN) to estimate the continuous wrist joint angle using the reliefF selected features of surface electromyography (sEMG) and post filter (WNN using RSF&PF). The working process of WNN using RSF&PT can be described as three steps. The candidate sEMG features extracted in time domain are selected according to reliefF theory, and then the selected features are relayed to WNN for angle estimation. In the end, the moving average filter, a post filter, is used to smooth the output curve of the WNN. The estimation performance of WNN using RSF&PF is evaluated via an experimental platform and compared with that of support vector machine as well as that of radial basis function neural network. The average execution time of WNN using RSF&PT in processing 20-s of the sEMG data is 1.104s. The root mean square error and correlation coefficient are 8.73 degree, 0.978 respectively. The comparison results suggest that WNN using RSF&PF achieves the best estimation in terms of accuracy and speed in this work.