An Automated Machine Learning approach for the rapid estimation of seismic risk in existing bridges

The assessment of seismic risk for existing bridges is a key element in ensuring the efficiency and safety of infrastructure networks. The Italian Guidelines for existing bridges (GL) propose a multi-level and multi-risk methodology to estimate overall risk through a synthetic indicator known as the Class of Attention (CoA). The parameters required for seismic risk assessment can be obtained from census data, preliminary evaluations, and direct inspections. However, direct inspections are costly and often affected by subjectivity, particularly in evaluating the Level of Defectiveness (LoD). This study introduces an automatic prediction approach for the seismic CoA using Automated Machine Learning (AutoML) techniques, trained on synthetic data generated from the logical structure and decision tables defined in the GL. Each bridge is represented as a numerical configuration, and the corresponding CoA is derived through an algorithm replicating the GL assessment process. The models were trained using AutoGluon, excluding the LoD parameter and focusing on subsets of parameters obtainable without in-depth investigations. Experimental validation was performed using a real dataset provided by the Italian National Agency for Safety of Railways and Roads (ANSFISA), including information from nationwide monitoring activities. The results show that the models trained on synthetic data can predict seismic CoA in real cases with good accuracy, offering a valuable tool to support inspection prioritization and optimize intervention planning under a seismic risk management perspective.