Explainable Machine Learning‐Assisted Progressive Collapse Resistance Prediction of Reinforced Concrete Beam‐Column Substructures

ABSTRACT

Reinforced concrete (RC) beam‐column substructures are the key supports of frame structures, which experiences compressive arch action (CAA) and tensile catenary action (TCA) to resist the progressive collapse. Existing expression models involve complicated inference processes, which struggle with generalized application due to specific simplified assumptions. This paper proposes a new progressive collapse resistance prediction model of RC substructures driven by interpretable hybrid machine learning methods. A high‐quality database integrated with 10‐dimensional features is established, including CAA and TCA stages during progressive collapse. Three different machine learning models are proposed to conduct a comparison including random forest (RF), least square boosting (LSBoost), and generalized additive models (GAM). Particle swarm optimization (PSO) algorithm is adopted herein to obtain the optimal machine learning model parameters. The Shapley (SHAP) analysis is performed to reveal the prediction process of machine learners involving global interpretable, local interpretable and feature dependence. The results show that PSO‐RF model has a better peak prediction accuracy than PSO‐LSBoost and PSO‐GAM models during CAA and TCA stages. Based on SHAP learning, it is known that the span‐to‐height ratio and cross‐sectional area have significant impacts on the model's prediction results. The proposed PSO‐RF model has superior computational accuracy than existing physical models, which provides a key support for the codes revision of RC structures to resist progressive collapse.