ABSTRACT
A heuristic approach based on visual inspections cannot provide an adequate and unbiased evaluation of bridges’ structural state. Fatigue assessment methods, based on the structural dynamic response, turn out to be crucial to consider the current structural conditions. Conventionally, strain data under traffic loading are collected for fatigue assessment. Although direct measurements of strain signals allow effective structural health monitoring strategies, they would induce a considerable rise in costs and errors, specifically where an extensive sensor network is needed. To pursue higher reliability of the monitoring system, a machine learning approach to convert acceleration to strain signals is proposed from a life cycle assessment perspective. Graph neural networks (GNNs), a class of machine learning models recently emerging in the civil engineering field, are employed to achieve this task. The GNN enables the encoding of structural information to describe the relationship between its nodes, representing the structural joints, and creating a geometry-based architecture in this application. The approach has been validated on a numerical simulation and promising results have been obtained.
