Post-earthquake seismic capacity estimation of reinforced concrete bridge piers using Machine learning techniques

Abstract

This study analytically explored a parameterized set of 20 reinforced concrete bridge piers which share several geometrical and material properties commonly found in Canadian bridges. A nonlinear fiber-based modelling approach with a proposed material strength degradation scheme is developed using the OpenSEES platform. A multiple conditional mean spectra approach is used to select a suite of 50 mainshock-aftershock (MS-AS) ground motion records for the selected site in Vancouver, British Columbia. Nonlinear time history analysis is performed for mainshock-only and mainshock-aftershock excitations, and static pushover analysis is also performed in lateral and axial directions for the intact columns, as well as in their respective post-MS and post-AS damaged states. Using the resulting data, a framework for post-earthquake seismic capacity estimation of the bridge piers is developed using machine learning regression methods, where several candidate models are tuned using an exhaustive grid search algorithm approach and k-fold crossvalidation. The tuned models are fitted and evaluated against a test set of data to determine a single best performing model using a multiple scorer performance index as the metric. The resulting performance index suggests that the decision tree model is the most suitable regressor for capacity estimation due to this model exhibiting the highest accuracy as well as lowest residual error.