A machine learning approach for load rating of bridges

The United States has more than 617,000 highway bridges, of which over 40,000 (7.5%) are classified in poor condition, underscoring the need for efficient and reliable bridge management strategies. Load rating is a fundamental component of the Bridge Management System (BMS), yet current analytical procedures are often time-consuming, costly, or require detailed structural modeling. This study investigates the use of machine learning (ML) techniques as a predictive alternative for estimating bridge rating factors (RFs) and for classifying the level of modeling required for load rating. A database of 268 bridges in Michigan, referred to as Bridge Net (BrNet), was developed and numerically modeled using AASHTOWare Bridge Rating (BrR) software, incorporating 22 geometric, structural, and condition-based attributes. Four regression models; decision trees, boosted decision trees (BDT), support vector regression (SVR), and Gaussian process regression; were trained to predict inventory and operating RFs. The BDT model achieved the best predictive accuracy, with normalized root mean square error (NRMSE) values of 0.17 and mean absolute errors as low as 0.12. In parallel, an SVR classifier correctly identified the required modeling priority for 82% of bridges in the dataset. The results demonstrate that ML-based prediction can substantially reduce rating effort, supporting faster decision-making in bridge maintenance and rehabilitation programs.