Sustainable bridge management using refined feature selection for machine learning-aided bridge condition prediction: Incorporation of Pareto distribution in MRMR method

One of the critical phases of bridge management is condition prediction, which was enhanced after machine learning emerged. Researchers have used feature selection (FS) methods to reduce the predictors and optimise the prediction models. Prior studies have either explored a unified feature set for the overall bridge condition or focused solely on the deck, leading to limited predictions. This study proposes a refined feature selection method highlighting the importance of specific predictors for different bridge elements’ conditions using the California State inspection database from 1983 to 2021. The implemented FS approach consists of a verified Minimum-Redundancy Maximum-Relevance (MRMR) method followed by a Pareto analysis that identifies the most contributory factors in predicting the condition of bridges’ decks, superstructures, and substructures. The applied experiment on the US National Bridge Inventory database reveals that 28–33 predictor variables, out of more than 140 available features, contribute the most to each component's health prediction with a cumulative importance score of over 95 %. Additionally, 22 mutual data items among the selected features are proposed as the minimum required predictors to be gathered by the asset management authorities. This study's achievements help both researchers reduce the running costs of their prediction models and asset managers with data gathering and registration optimisation and, consequently, whole-of-life cycle cost reduction for sustainable asset management.