Interpretable Capacity Prediction of RC Bridges Based on Genetic Algorithm-enhanced Support Vector Machine Learning

Abstract

Existing reinforced concrete (RC) bridges are subjected to environmental erosion and vehicle loads. It is becoming an urgent problem to evaluate the safety condition of bridge structures combining inspection data with artificial intelligence methods. This paper proposes a data-driven capacity assessment framework for existing RC bridges. The load capacity limit state (LCLS) and serviceability limit state (SLS) prediction model are established based on the proposed information fusion machine learning model. The genetic algorithm (GA) optimized support vector machine (SVM) learner is established to capture the relationship between the feature variables and the LSLS or SLS. Forty-five samples are obtained by static and dynamic simulation of the ANSYS models. Five-dimensional parameters are adopted as the key input parameters of the model, including the maximum dynamic deflection, crack opening ratio, and crack normal damage ratio at midspan, 1/4 span, and 3/4 span. The Shapley additive explanations method is proposed to conduct parameters sensitivity analysis. The results show that the GA-SVM regression algorithm in LCLS and SLS reduction factor prediction is better than the artificial neural network (ANN) model. The crack opening ratio is the most critical parameter that can considerably affect the outcomes of the LCLS and SLS prediction.