Machine learning-based future performance prediction model for bridge inspection and performance data in South Korea

Abstract

Bridges are vital components of modern transportation infrastructure. The collapse or deterioration of bridges due to aging can disrupt local transportation networks, underscoring the importance of proper maintenance. To address this issue, numerous studies in various countries have focused on predicting the future performance of bridges using inspection data and applying probabilistic, statistical, machine learning, and deep learning techniques. Given that each country has distinct bridge inspection and performance evaluation standards and dataset formats, prediction techniques must be tailored to the specific needs of each country. Following the collapse of the Seongsu Bridge in 1994, South Korea has evaluated bridge performance based on a grading system ranging from A to E, with datasets managed through a Bridge Management System (BMS). In this study, we trained several models, including linear regression, random forest, LightGBM, and deep neural networks, on South Korea’s BMS dataset to develop a component-level grading prediction model. After evaluating their performance, LightGBM, an ensemble model, was selected as the optimal model. This model, tailored to the structure of South Korea’s BMS dataset, demonstrated high performance, with an average accuracy of 80–96% for each component. Using this model, we predicted the future performance of bridges over the next 3 years and found that the number of bridges requiring maintenance, graded as C, tended to increase by 20%. These results provide an intuitive understanding of the changes in bridge performance and grades throughout the bridge lifecycle, contributing to more efficient budget allocation for bridge maintenance in South Korea.