Predicting Worker Fatigue Using Wearable Sensor Data: Comparative Analysis of Classical and Deep Learning Models

Abstract

Ergonomic risk assessment is critical for preventing musculoskeletal disorders (MSDs) in occupational settings, yet high-dimensional data from wearable sensors poses challenges in capturing complex, non-linear posture dynamics. This study proposes a novel hybrid framework integrating Graph Autoencoders (GAE) with Principal Component Analysis (PCA) for dimensionality reduction and feature selection in ergonomic risk classification. By modeling human postures as skeletal graphs—joints as nodes and biomechanical connections as edges—the approach preserves structural dependencies, addressing limitations of linear methods like PCA. Applied to a synthetic dataset of 500 samples (including posture angles, task errors, workload, and machine parameters), the framework employs Chi-Square Tests and Correlation Analysis to identify key risk factors, with VC Dimension guiding model complexity to balance bias-variance tradeoffs. Results demonstrate that the hybrid GAE-PCA retains 99.7% variance in three components and achieves 94% classification accuracy, outperforming PCA alone by 14%. This graph-based innovation, novel as of October 2025, enhances interpretability and supports real-time ergonomic interventions, offering a scalable solution for Industry 5.0 workplace safety.