Data-Driven Approaches for Predicting Bridge Defectiveness: A Machine Learning-Based Assessment of Structural and Contextual Factors

The conservation, maintenance, and monitoring of bridges have gained increasing importance worldwide, particularly in Italy, where a vast number of aged structures exist, and several bridge collapses have recently occurred. To address this issue, the Italian Ministry of Transport introduced national Guidelines in 2020. Their application to a large sample of bridges revealed critical challenges in asset management, highlighting the need for predictive models to enhance inspection efficiency. Machine learning (ML) techniques, including decision trees, random forests, and neural networks, offer a promising solution. Traditional approaches in bridge conservation typically rely on risk classification models, using pre-processed parameters —such as the defectiveness level— to achieve highly accurate predictions. However, using the defectiveness level as an input feature poses a conceptual issue: rather than serving as an independent explanatory variable, it acts as a dependent variable. While this approach enhances prediction accuracy, it ultimately compromises the interpretability of the model’s features. Based on these observations, this study adopts a data-driven approach, testing a decision tree classifier, random forest, and artificial neural network to predict the defectiveness level. The proposed approach is tested and validated on a database of inspected bridges located in Italy.