Optimized machine learning for predicting bridge condition through hybrid Bayesian-grid search

Abstract

Purpose This study aims to develop machine learning (ML) models to predict the sufficiency rating (SR) of bridges, thereby supporting decision-making in infrastructure management, including repair, maintenance and reconstruction. Design/methodology/approach The National Bridge Inventory data was cleaned, labeled and normalized, and then the Boruta algorithm was used to select 22 significant features out of 40 original variables. A hybrid Bayesian-grid (HBG) technique was developed to optimize the hyperparameters of three predictive models: artificial neural networks (ANNs), support vector regression and k-nearest neighbors. Additionally, a multiple linear regression model was developed as a baseline for comparison. Model performance was evaluated using R², adjusted R², mean absolute error and root mean square error. Findings The ANN model outperformed the others. For classification tasks, the ANN model achieved 88.11% accuracy in categorizing bridges into three condition levels: good (SR = 80), need maintenance (80 > SR = 50) and need reconstruction (SR < 50). The HBG approach significantly improved predictive accuracy while reducing computational cost and processing time. The integration of the Boruta feature selection method enhanced both model accuracy and interpretability. Originality/value The HBG technique proposed in this study represents a significant advancement in the field by offering a robust and flexible framework for hyperparameter tuning that can be applied across a wide range of ML applications. These findings offer practical tools and insights for bridge management agencies to improve evaluation and maintenance decision-making, ultimately enhancing infrastructure sustainability and resource efficiency.