Investigating improvements to neural network based EMG to joint torque estimation

Abstract

AbstractAlthough surface electromyography (sEMG) has a high correlation to muscle force, an accurate model that can estimate joint torque from sEMG is still elusive. Artificial neural networks (NN), renowned as universal approximators, have been employed to capture this complex nonlinear relation. This work focuses on investigating possible improvements to the NN methodology and algorithm that would consistently produce reliable sEMG-to-knee-joint torque mapping for any individual. This includes improvements in number of inputs, data normalization techniques, NN architecture and training algorithms. Data (sEMG) from five knee extensor and flexor muscle from one subject were recorded on 10 random days over a period of 3 weeks whilst subject performed both isometric and isokinetic movements. The results indicate that incorporating more muscles into the NN and normalizing the data at each isometric angle prior to NN training improves torque estimation. The mean lowest estimation error achieved for isometric motion was 10.461% (1.792), whereas the lowest estimation errors for isokinetic motion were larger than 20%.