Continuous Joint Angle Estimation by Least Support Vector Machine from Time-Delayed sEMG Features

The main works of estimating continuous human kinematic information from surface electromyography (sEMG) are grounded on the extraction of useful information and the construction of proper estimation model. This paper proposes the least square support vector machine ((LSSVM) with time-delayed features (TDF) of sEMG for the continuous wrist palmar flexion-extension angle estimation. The performance of proposed method is verified via an experimental platform for sEMG and angle data records. The feasibility of introducing a time delay value into two machine learning models namely, LSSVM, back propagation neural network (BP), for sEMG-angle estimation is proved. The average correlation coefficients and the average root mean square error (RMSE) of LSSVM with TDF are 9.36±2.5 degree and 0.96±0.02 respectively. This paper obtains three conclusions: 1.The estimation performance of LLSVM and BP neural network have been much improved after a proper single time delay is introduced into the models; 2. An empirical calculation of the optimal time delay value can be obtained in which the optimal time delay is linear with the size of window N; 3.The LLSVM with TDF exhibits the best estimation performance in continuous wrist angle prediction from the sEMG features.