Uncertainty-aware detection of early damage in a full-scale steel railway bridge

Reliable detection of early structural damage is essential for safe and efficient management of existing bridges, yet validation of unsupervised damage detection methods is often limited by the scarcity of validation data, particularly for early damage. Recent developments in uncertainty-aware vibration-based damage detection show that autoregressive coefficients and their estimator covariances can enable a reliable computation of Mahalanobis squared distances, even in high-dimensional settings. This paper, which builds on these developments, evaluates the capability of detecting damage states representing early damage. To this end, a new experimental measurement campaign has been conducted in which 22 gradual damage states, representing different progressions of four realistic damage types, were imposed on a full-scale steel truss bridge under controlled excitation and moderate environmental variability. The results show clear separation between undamaged and damaged conditions across all damage states, including the earliest damage progressions. Near-perfect classification is achieved using the uncertainty-aware Mahalanobis distance, with balanced accuracies above 99%. Moreover, the damage-index distributions exhibit low dispersion and consistent behavior across temperature regimes. These findings demonstrate that subtle damage-related changes in the structural response of steel bridges can be detected supporting potential integration of such work into real-world SHM systems after further validation under operational conditions.