Electromyography spectrogram-driven Vision Transformers for automated muscle fatigue assessment in construction operations

Construction workers face significant physical strain and musculoskeletal risks. While electromyography (EMG) has emerged as a robust assessment tool, current frameworks struggle with generalizability across different muscle groups. To solve this, we propose a Vision Transformer framework that applies self-attention to EMG spectrograms, modeling spatial-temporal dependencies to capture localized fatigue patterns. Experimental evaluations conducted on material lifting and hammering tasks targeting the biceps brachii, lumbar-erector spinae, and rectus femoris muscle groups demonstrate that the developed framework achieves an average classification accuracy of 82.7% (biceps brachii: 83.7%, lumbar-erector spinae: 84.4%, rectus femoris: 79.9%). These findings support the feasibility of a unified transformer-based spectrogram framework for localized fatigue monitoring across muscle groups with different biomechanical roles. The study contributes a task-grounded EMG fatigue assessment pipeline and provides evidence for its applicability in construction-relevant conditions, supporting the development of scalable fatigue-monitoring systems for physically demanding occupational environments.