Physics-informed neural networks for early-warning fatigue damage detection and localization in a steel railway bridge using sparse ambient monitoring data

Abstract

Aging steel railway bridges accumulate fatigue damage at locations that periodic inspection rarely reaches and that conventional vibration monitoring detects only after severe stiffness loss. Existing physics-informed learning methods for structural health monitoring (SHM) are mostly validated on synthetic data and single snapshots, leaving their performance on continuous field data with verified damage untested. This study develops a strain-based monitoring approach, supported by a physics-informed neural network (PINN), for early detection of fatigue damage in a steel railway bridge. The measured strain is compensated for temperature using a model calibrated on healthy-state data, the resulting anomaly is tracked through a statistical control chart, and a PINN embedding the Euler–Bernoulli equation and the strain–stiffness relationship converts the sparse measurements into a continuous, physically valid flexural-stiffness field, identified through two-stage training. The framework was evaluated on 56 ambient recordings spanning 19 days across an inspection-confirmed fatigue crack in the Vänersborg railway bridge, Sweden, using only three strain gauges (SG9, SG10, SG11) and one accelerometer (A5). The main findings are: the framework issued a warning 15 days before the in-service monitoring system; the identified stiffness reduction concentrated in the region of the crack-adjacent gauge, consistent with the inspection finding; the identified stiffness field tracked the progression of the damage as a normalized reduction relative to the healthy baseline; and perfect detection was achieved under a persistence rule with zero false alarms. A co-located accelerometer recorded a 94-fold spike at crack propagation but gave no prior warning, confirming that the early detection derives from the temperature-corrected strain identification. The framework provides a reproducible, low-cost, and interpretable basis for proactive bridge maintenance.