Predicting Concrete Bridge Deck Deterioration: A Hyperparameter Optimization Approach
Nour H. Almarahlleh, Hexu Liu, Osama Abudayyeh, Rabia Emhamed Al Mamlook
Abstract
Concrete bridge decks are critical transportation infrastructure components where deterioration can compromise structural integrity and public safety. This study develops machine learning (ML) models using the National Bridge Inventory (NBI) to classify deck conditions and predict deterioration trajectories. Models were tested and trained on inspection records from over 28,786 bridges in Michigan over 23 years, from 1992 to 2015. Eleven approaches were evaluated after hyperparameter optimization, based on 10-fold cross-validation, including logistic regression, gradient boosting, AdaBoost, random forest, extra trees, K-nearest neighbors, naive Bayes, decision tree, LightGBM, CatBoost, and bagging. Model effectiveness was assessed using accuracy, recall, F1-score, and area under the curve. Results indicate the optimized CatBoost classifier achieved 96.66% testing accuracy in rating deck conditions. The incorporation of hyperparameter optimization has significantly enhanced the overall predictive performance of the models, ensuring robust and reliable deterioration forecasting. The research sheds light on crucial factors such as deck age, area, and average daily traffic, contributing to a more comprehensive understanding of the factors influencing bridge deck condition ratings. These insights inform preventative maintenance planning to extend service life. This work pioneers a data-driven framework to forecast concrete deterioration, empowering officials with precise predictions to optimize infrastructure management under budget constraints. The approach provides a promising decision-support tool for sustainable infrastructure.