Steel bridge girders are often subjected to strikes by over-height vehicles that exceed the allowable vertical clearance underneath a bridge. Such strikes may affect the serviceability of a bridge structure and can significantly reduce the load-carrying capacity of damaged girders. Although finite element analyses are commonly used to estimate the residual capacity of damaged girders, developing and analyzing such models for every strike incident is computationally intensive and time-consuming. To address these limitations, this paper presents a new explainable machine learning methodology that can accurately and efficiently predict the residual load-carrying capacity of damaged steel I-girders. The methodology is conducted in five main stages: (1) data collection, (2) data analysis and preprocessing, (3) model training, (4) model validation, and (5) model interpretation. Five machine learning models were developed and trained to predict the residual capacity of damaged girders. Results show that the eXtreme Gradient Boosting (XGBoost) algorithm outperformed all other models, achieving R 2 and MAPE scores of 95.4% and 3.21%, respectively, on the unseen test set. Moreover, the SHapley Additive exPlanations framework (SHAP) was used to explain the global performance of the XGBoost model and interpret its individual predictions. SHAP values revealed that predicted residual capacity is reduced by increases in girder span length and observed horizontal and vertical damage deflections. The proposed machine learning models are expected to provide bridge engineering researchers and practitioners with an efficient and explainable method for assessing the residual capacity of damaged steel I-girders without relying on time-consuming and computationally intensive simulations. • A dataset of simulated damaged steel I-girders was created using FE analysis. • Relationships between girder attributes and load-carrying capacity were investigated. • Several ML models were trained to predict residual capacity of damaged steel girders. • K-fold cross-validation was used to confirm the generalizability of ML models. • SHAP values were employed to interpret XGBoost model outputs of predicted capacities.
Ibrahim et al. (Thu,) studied this question.