Onset Forecasting of Hand-Arm Fatigue via Physics-Informed Machine Learning and Wearable IMU Sensors

Abstract

Hand-arm muscle fatigue is a multifaceted phenomenon, encompassing biomechanical, physiological, and perceptual components related to effort, discomfort, and individual tolerance. This study introduces a wearable sensor-based machine learning framework for the early detection of upper-limb fatigue during repetitive tasks. Signals from Inertial Measurement Units (IMUs) are segmented into overlapping time windows and analysed to extract kinematic indicators associated with fatigue-related movement changes, such as increased movement irregularity, tremor-like fluctuations, reduced fluidity, micro-pauses, jerk variations, and alterations in gesture amplitude. These features are normalised relative to an initial baseline, helping to minimise inter-trial variability and enabling a more personalised assessment of fatigue progression. Supervised classification models detect the onset of fatigue using binary labels, with each data point labelled as either fatigued or not fatigued. A physics-informed fatigue index based on root-mean-square (RMS) acceleration, which is a statistical measure of the magnitude of acceleration signals, provides an interpretable measure of the total effort accumulated over time. This index is mapped to a perceived-exertion scale, a numerical system that represents how hard a task feels, to help identify fatigue phases from baseline (initial state) through transitional stages to critical onset (the point at which fatigue becomes critical). The physics-informed part does not replace subjective fatigue assessments, but instead guides the model toward plausible effort and recovery dynamics, meaning how effort builds up and decreases over time. This framework represents an initial step toward real-time, anticipatory monitoring of perceived fatigue. Future validation will require integrating subjective self-assessments, such as Borg or RPE scales, with physiological measurements and multi-subject experimental protocols. Ultimately, the goal is to develop a personalised system capable of estimating fatigue levels in real time and predicting the approach to critical thresholds in advance, with potential applications spanning occupational ergonomics, rehabilitation, sports, and human-machine interaction.