Metodología no supervisada para identificar cambios estructurales en puentes ferroviarios bajo cargas dinámicas

Abstract

CONDITION FOR PUBLICATION OF PROJECT. This study centers in the application of unsupervised data-driven structural health monitoring (SHM) of a single-span railway bridge. A numerical model with finite elements (FE) with train loads as moving concentrated loads was developed incorporating OpenSees to measure structural response, taking advantage of the large-magnitude loads for enhancing sensivity to structural changes. The proposed methodology identify damage by introducing train speed, axle loadings, and stiffness reduction by 5% to 20% in localized critical sections of the bridge. To ensure a continuous assessment, a hybrid combination of multivariate autoregressive (MAR) model, principal components analysis, mahalanobis distance and k-means clustering algorithm was applied to the monitored data, as a unsupervised machine learning algorithm. The strategy proved to be robust when detecting undamaged and damaged scenarios across varying train speed and train loads. Its sensivity to smaller damaged levels that consists of small stiffness reductions, and independence from labelled damaged training data, make this a reliable solution for long-term continuous monitoring at instrumented railway bridges. Keywords: Supervised learning, SHM, Damage detection, MAR model