Comparison of two strategies for handgrip force prediction based on sEMG

Abstract

The control system of myoelectric prostheses requires high precision and rapid response. Many algorithms have been applied in prosthesis control. In this paper, BackPropagation Neural Network (BPNN) and Multiple Nonlinear Regression (MNLR) are applied to predict handgrip force through surface electromyography (sEMG) signals of forearm muscles. In the following experiments, the root mean square (RMS) data extracted from sEMG signals are randomly separated into training dataset (75%) and testing dataset (25%). When the dataset is trained, the Root Mean Square Error can reach about 1.145kfg (BPNN) and 3.452kfg (MNLR), respectively. BPNN consumes about 21.435s and MNLR spends about 0.013s. During testing the dataset, BPNN and MNLR obtain the Root Mean Square Error about 1.207kfg and 3.620kfg, respectively. BPNN consumes nearly the same time with MNLR. Based on the comparison of BPNN and MNLR, BPNN outperforms MNLR at accuracy, and MNLR is better than BPNN at response time. This study results will provide an important basis for the reasonable selection of prosthetic wrist system. Keywords-handgrip force; sEMG; RMS; BPNN; MNLR