Preliminary fast assessment of bridge risk by neural network

The conservation, maintenance, and monitoring of bridges have received growing attention worldwide over the past decade. Recent approaches to this problem highlight the lack of comprehensive multi-risk assessment methodologies that incorporate all risk components, i.e., hazard, vulnerability, and exposure as well as consider multiple risk sources. Moreover, existing risk assessment approaches are often time and resource consuming. Therefore, fast procedures that allow for a preliminary identification of bridges at high-risk situations are fundamental tools for managing large bridges stocks to prioritize the visual inspections and subsequent insights. This paper proposes a framework for developing machine learning models to predict the risk levels of bridges using a reduced set of selected parameters that can be easily obtained from available census data or archive documents. The study exploits the potentialities of Artificial Neural Networks and provides insights into effective choices for input data selection, methods for handling inhomogeneous data, and optimization of model parameters. Finally, the developed framework has been applied to the Italian highway network, assessing the risk levels of a large number of bridges, which have been classified according to the multi-level and multi-risk procedure recently enacted by the Italian national government.