Identification of bridge curvature profiles from dynamic responses induced by moving vehicles using autoregressive models

This paper introduces a novel methodology for structural damage localization in beam-like single- and multi-span bridge structures based on curvature profiles extracted from acceleration measurements. The approach builds on the rationale that autoregressive models, when trained on ambient vibration data, fail to reconstruct the quasi-static response induced by moving loads. This limitation produces a reconstruction residual that, under suitable conditions, corresponds to a shifted and scaled version of the curvature profile of the structure generated by a point load applied at the sensor location. The proposed method enables the calculation of this residual with extremely sparse sensor networks that do not require synchronization. A damage index is then defined from variations in the estimated curvature profile, enabling localization of stiffness reductions. To eliminate the need for manual parameter tuning, a model order selection criterion is proposed, which makes the method fully automated, unlike existing approaches that rely on prior knowledge of the monitored structure. The methodology is validated through numerical simulations that incorporate vehicle-bridge interaction phenomena and road roughness, as well as experimental data from a full-scale truss bridge. The results demonstrate that the proposed method achieves damage localization performance comparable to established filter-based techniques, while offering improved spatial resolution and requiring no tuning parameters.