Extracting Control Information from Multi-channel Surface EMG Signals to Drive a Musculoskeletal Model

Abstract

Electromyography (EMG) signals have been used for multiple degrees-of-freedom (DOFs) continuous control of myoelectric prostheses. Recently, musculoskeletal models (MMs) have been studied to decode hand and wrist movements from EMG signals. In this study, we proposed an MM combined with a non-negative matrix factorization algorithm (NMF-MM) to predict wrist Pro/Sup movements. 8-channel EMG signals were acquired from one able-bodied subject. We extracted a pair of virtual antagonistic muscles corresponding to wrist Pro/Sup movements from the 8-channel EMG signals using the NMF. It addressed the difficulties of acquiring the EMG signals of supinator from the skin surface. The model included two Hill-type muscle actuators, each with a contractile unit and a parallel elastic unit. Five parameters were optimized for each of the two muscles. The NMF-MM was compared with artificial neural network (ANN) and linear regression (LR). The Pearson’s correlation coefficient (r) and normalized root mean square error (NRMSE) were calculated between measured and predicted joint angles. The prediction performance of NMF-MM (r=0.88, NRMSE=0.17) was slightly better than that of the ANN (r=0.86, NRMSE=0.19) and the LR (r=0.83, NRMSE=0.17). The results demonstrated the feasibility of extracting control information from the multi-channel surface EMG signals by NMF to drive the MM. The outcomes provided a promising way for prosthetic hand control and slave-hand control of minimally invasive surgical robots.