A Physics-Informed Low-Shot Adversarial Learning for sEMG-Based Estimation of Muscle Force and Joint Kinematics

Abstract

Muscle force and joint kinematics estimation1 from surface electromyography (sEMG) are essential for 2 real-time biomechanical analysis of the dynamic interplay 3 among neural muscle stimulation, muscle dynamics, and 4 kinetics. Recent advances in deep neural networks (DNNs) 5 have shown the potential to improve biomechanical anal-6 ysis in a fully automated and reproducible manner. How-7 ever, the small sample nature and physical interpretability 8 of biomechanical analysis limit the applications of DNNs. 9 This paper presents a novel physics-informed low-shot 10 adversarial learning method for sEMG-based estimation of 11 muscle force and joint kinematics. This method seamlessly 12 integrates Lagrange's equation of motion and inverse dy-13 namic muscle model into the generative adversarial net-14 work (GAN) framework for structured feature decoding and 15 extrapolated estimation from the small sample data. Specif-16 ically, Lagrange's equation of motion is introduced into 17 the generative model to restrain the structured decoding 18 of the high-level features following the laws of physics. A 19 physics-informed policy gradient is designed to improve 20 the adversarial learning efficiency by rewarding the consis-21 tent physical representation of the extrapolated estimations 22 and the physical references. Experimental validations are 23 conducted on two scenarios (i.e. the walking trials and 24 wrist motion trials). Results indicate that the estimations 25 of the muscle forces and joint kinematics are unbiased 26 compared to the physics-based inverse dynamics, which 27 outperforms the selected benchmark methods, including 28 physics-informed convolution neural network (PI-CNN), val-29 lina generative adversarial network (GAN), and multi-layer 30 extreme learning machine (ML-ELM). 31 Index Terms-muscle force and joint kinematics, surface 32 Electromyographic, low-shot learning, generative adversar-33 ial network, physics-informed optimization, mode collapse 34 35 I. INTRODUCTION 36 H UMAN movements involve complex interactions within 37 the neuromuscular system. The estimation of muscle 38 force and joint kinematics dynamics provides detailed biome-39 chanical analysis to understand the human neuromuscular 40 system [1], [2], which benefits high-level exoskeleton control