Data-driven reliability design of transverse impacted concrete-filled steel tube columns

Abstract

There is an elevated likelihood that concrete-filled steel tube columns will be struck by transverse vehicle collisions together with the global acceleration of traffic and urbanization. This work intends to implement data-driven techniques to determine column deformation and quantify column damage upon lateral impact. Six novel machine learning models (gaussian process regression, support vector regression, least-squares boosting regression, multilayer perceptron, decision trees and random forests) were incorporated alongside Bayes optimization based on 351 samples. With a predicted deflection correlation of 0.93, least-squares boosting regression delivered the strongest training effect comprehensively. The SHapley Additive exPlanations approach was adopted to further foster user trust and comprehension. The first seven features were examined via parameter investigation according to feature importance ranking, results suggested that the best strategy for reducing lateral deformation is boosting steel tube thickness and outer diameter along the gradient downwards of deformable contour map. A reliability-based design procedure that serves as an appropriate guideline for impact protection was laid out by Monte Carlo sampling to determine failure probability of three damage levels in concrete-filled steel tube columns.