Estimation of continuous and constraint-free 3 DoF wrist movements from surface electromyogram signal using kernel recursive least square tracker. (September 2018)

• Record Type: Journal Article
• Authors: Bakshi, Koushik; Manjunatha, M.; Kumar, C.S.
• Is Part Of: Biomedical signal processing and control. Volume 46(2018)
• Page Start: 104
• Page End: 115
• Publication Date: 2018-09
• Subjects: Surface electromyogram (sEMG) -- Upper limb prosthesis -- Kernel ridge regression (KRR) -- Kernel recursive least square tracker (KRLS-T) -- Simultaneous and proportional control (SPC); Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2018.06.012
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Forearm Electromyogram signal based wrist motion tracking. The dynamic constraint-free 3 dimensional wrist motion profiles considered, are very much similar to various daily life activities or vocational activities. Use of nonlinear regression algorithm KRLS-T and performance comparison with artificial neural network, kernel ridge regression. Pseudo-online simulation of KRLS-T based wrist motion estimator showing average accuracy of 90% or more. It can be realized in real time prosthesis controller. Abstract: We have employed Kernel Least Square Tracker (KRLS-T), a nonlinear kernel based recursive algorithm, to estimate 3 dimensional wrist kinematics from sEMG signals of forearm muscle groups. KRLS-T combines the advantage of kernel techniques and adaptive estimation and hence has been considered for predicting 3 dimensional wrist angles from nonlinear and non-stationary sEMG. We have been able to successfully predict 6 basic and 2 dynamic, continuous and constraint-free wrist motions for 10 normal subjects in an offline mode with more than 90% accuracy. The continuous wrist motion profiles, considered here, resemble the complex and dexterous wrist motions involved in various activities of daily life. Statistical significance analysis shows that KRLS-T performs better than Kernel Ridge Regression (KRR) and a feed-forward back propagation neural network during a 10-fold cross validation stage. Subsequently, a real-life scenario has been emulated for the KRLS-T based motion predictor where 2 different trials' data are combined and given sequentially as input to the estimator. Its fast adaptation capability to the nonstationary sEMG-wrist angle relationship, as reported here, makes it a promising option for implementing intuitive prosthesis control.