On the use of machine learning for fragility analysis of multi-span girder bridges

Abstract

Companies and enterprises managing the transportation networks are in charge of performing seismic risk assessments for a sensible number of bridges designed, over the past decades, without fulfilling requirements of anti-seismic codes. Recent research studies have focused on leveraging machine learning strategies to enhance the efficiency of probabilistic seismic assessments. This study tackles the integration of machine learning algorithms in the case of bridge-specific applications to predict probabilistic seismic demand models for substructure components such as piers, accounting for knowledge-based uncertainties.The machine-learning-based methodology employed builds on the results of nonlinear time history analyses and considers features related to seismic excitation and structural parameters. Random Forest algorithm is employed to investigate the methodology for a multi-span simply supported girder bridge, which is a representative example of the most prevalent bridge class in Europe. Aiming to reduce the need for extensive nonlinear time history analyses in the risk assessment of bridges belonging to this specific class, the methodology proposed shows promising potential highlighted in the results.