Wearable-IMU Machine Learning for Exertion and Fatigue Assessment with an Occupational Exoskeleton in Manufacturing

Abstract

Exoskeletons are increasingly used to reduce physical strain in manual material handling, yet their true influence on worker exertion and fatigue remains challenging to quantify due to the lack of scalable and objective assessment methods. This study introduces a wearable-sensor framework that links subjective workload perception to quantifiable biomechanical responses during realistic industrial tasks, performed with (Exo) and without (NoExo) passive shoulder exoskeleton assistance. Twenty participants completed three representative tasks including loading, overhead screwing, and palletizing, while wearing ten IMUs positioned on the upper trapezius, deltoids, biceps, lumbar erector spinae, and hamstrings. Accelerometer signals (at 148 Hz) were converted to magnitude, segmented into 0.5 s windows with 50% overlap, and reduced to six biomechanical features per sensor, yielding 60 features per window. The dataset comprised 36,901 NoExo and 38,023 Exo windows, each labeled through task-specific Borg-scale mappings. Machine-learning models were evaluated using Leave-One-Participant-Out cross-validation. Random Forest and XGBoost achieved the highest performance (median accuracy = 0.83, F1-weighted = 0.82), outperforming Logistic Regression and Linear Discriminant Analysis. Exoskeleton use decreased inter-participant variability and reduced misclassifications, with top models reaching near-perfect separability for several individuals ($\mathrm{AUC}>0.95$). Feature-importance analyses consistently identified the lower back and hamstrings as the most informative sensor locations, enabling up to a 70% reduction in sensor count with minimal performance degradation. Critically, the proportion of hard-exertion windows decreased by nearly 75% with exoskeleton assistance, demonstrating measurable ergonomic benefit. These results establish a deployment-ready approach for real-time, objective assessment of physical workload in manufacturing environments. The proposed framework advances scalable evaluation of assistive technologies, supports data-driven ergonomic design and training, and offers the first evidence that wearable IMU-based machine-learning models can reliably quantify exoskeleton effectiveness in complex, real-world occupational settings.