Network-Level Bridge Condition Degradation Prediction and Interpretation Analysis Based on a Hybrid Ensemble Model

Abstract

Accurate prediction of network-level bridge conditions is crucial for informed maintenance decision-making. Traditional single algorithms struggle to extract bridge degradation features, leading to limited prediction accuracy. This study proposes a hybrid ensemble model approach that combines eXtreme Gradient Boosting (XGBoost), support vector regression (SVR), and artificial neural network (ANN) algorithms, significantly improving prediction performance. In addition, the SHapley Additive exPlanations (SHAP) method is utilized to analyze key influencing factors on bridge degradation, providing interpretability to the model’s predictions and offering valuable references for maintenance strategies. First, this study integrates inspection reports, design drawings, maintenance records, and technical condition ratings of 800 bridges to construct a comprehensive database. In total, 12 features, including bridge type, age, maximum span, maintenance frequency, and traffic volume, are extracted as inputs, with condition ratings as the output. Second, XGBoost, SVR, and ANN models are employed, and two hybrid ensemble models: (1) Inverse-Variance Weighted Hybrid Ensemble Prediction (HEP-IV), using inverse-variance weighting; and (2) artificial neural network-based Hybrid Ensemble Prediction (HEP-ANN), a meta-learner, are developed and tested for prediction accuracy. These models are tested for prediction accuracy. Finally, the SHAP method is applied to identify important factors, such as bridge age, maintenance frequency, and deck width, as well as their interactions. Experimental results indicate that the hybrid ensemble model (HEP-IV) outperforms other models, showing superior prediction accuracy and better generalization ability, with HEP-IV achieving the best performance across all evaluation metrics. The coefficient of determination for the test set is 0.982, the root mean square error is 0.066, and the mean absolute error is 0.04. The model’s interpretability quantifies the effect of key factors, enabling precise prioritization of maintenance interventions, which supports optimized budget allocation and policy-making for bridge network management.