Simultaneous estimation of multi-finger forces by surface electromyography and accelerometry signals

• Published in: Biomedical signal processing and control Vol. 70; p. 103005
• Main Authors: Mao, He, Fang, Peng, Li, Guanglin
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.09.2021
• Subjects: Electromyography; Finger force estimation; GRNN; Myoelectric control; Myoelectric control; Electromyography; GRNN; Finger force estimation
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2021.103005

•A method of combining sEMG and accelerometry signals to decode finger forces is proposed.•Features are extracted from both signals and are combined into multi-modal feature sets.•GRNN is used to map the multi-modal feature set to 6-DOF finger forces.•Higher estimation accuracy (R2 = 93.33 ± 3.45%, NRMSE = 2.85 ± 0.58%, and MAE = 0.34 ± 0.15) is achieved. Myoelectric prostheses generally use pattern recognition strategies to decode users’ gesture intention; however, they lack intuitive force control. Regression strategy can extract force information of multiple degrees of freedom (DOFs) and has become a research focus in myoelectric control. In this paper, to realize a simultaneous estimation of continuous forces exerted at multi-fingertips, we used a generalized regression neural network (GRNN) to associate both surface electromyography (sEMG) and accelerometry (ACC) signals to finger forces of 6 DOFs (flexion of each finger plus thumb abduction). Totally nine force patterns were tested, including six single-DOF patterns (activation of each DOF) and three multi-DOF patterns of simultaneous activation of two DOFs. We extracted four popular sEMG feature sets from myoelectric signals and combined each of them with the mean feature of ACC signals to construct a multi-modal feature set. The estimation performance was evaluated by the coefficient of determination (R2), the normalized root mean squared error (NRMSE), and the mean absolute error (MAE). By combining the ACC modality, the estimation accuracy of all four sEMG feature sets was significantly improved. For intact subjects, the R2, NRMSE, and MAE values using the optimal feature set were 93.33 ± 3.45%, 2.85 ± 0.58%, and 0.34 ± 0.15, respectively. For amputees, the R2, NRMSE, and MAE values were 73.16 ± 5.79%, 5.43 ± 1.19%, and 0.73 ± 0.39, respectively. Besides, when the number of training samples decreased, the multi-modal feature set showed higher robustness. The results demonstrated the potential application of the proposed method for enabling natural and intuitive force control of dexterous prosthetic hands.