A multi-dimensional attention-based framework for personalized driver fatigue assessment for proactive commercial-vehicle crash prevention.

Abstract

Driver fatigue is a major safety concern in commercial vehicle operations, yet conventional detection approaches often rely on fixed driving-time thresholds that overlook inter-driver variability, cumulative effects across trips, and contextual differences. This study proposes a multi-dimensional, context-aware fatigue assessment framework that combines a driver-specific Bayesian baseline and cross-trip cumulative fatigue debt with six behavioral, spectral, contextual, and event-based indicators through a time-varying attention mechanism. The framework was evaluated using 43.8 million Digital Tachograph (DTG) records from 1,813 commercial drivers in Korea. The estimated fatigue score exhibits a clear diurnal pattern, increases consistently with cumulative fatigue debt, and reveals heterogeneous driver-level profiles. Spatial aggregation identified heterogeneous fatigue-prone road sections, while correspondence with independently identified accident-prone locations was sensitive to the spatial unit of analysis. Because direct physiological measurements and crash-event labels are unavailable, the framework is positioned as an operational fatigue-management tool rather than a clinically validated fatigue detector or crash-prediction model. The proposed approach provides a scalable basis for individualized driver monitoring, fleet scheduling, recovery management, and identification of fatigue-prone road sections, supporting proactive fatigue-risk management in commercial transport operations.