Adaptive Bridge Condition Forecasting through a Cluster-Based Exploration

Abstract

Transportation networks in the United States contain more than 615,000 bridges. Federal regulations require routine inspections of these infrastructure systems at least every 24 months. However, recent studies have suggested that this approach is costly, inefficient, and, in some cases, risky. An essential element in developing customizable alternatives for condition assessment plans is the design of precise and robust forecasting models for bridge deterioration, capable of sustaining acceptable predictive accuracy even when operating with small subsets of data. The overarching objective of this research is to introduce a novel forecasting method for bridge deterioration conditions, augmented with an exploration mechanism to identify a subset of bridges crucial for maintaining the predictive power of the model in subsequent years. The proposed method comprises two main components. Firstly, a forecasting model for bridge deterioration conditions was designed, formulated, and developed based on a novel framework for ordinal extreme gradient boosting. Secondly, an exploration mechanism is employed using a new multidimensional clustering method that integrates K-means clustering and dynamic time wrapping. The proposed method is applied using historical condition assessment data from bridges in the three neighboring states of New York, New Jersey, and Connecticut. The outcomes validate the performance of the proposed method, demonstrating that the forecasting model accurately predicts future bridge conditions. Additionally, the clustering component successfully identifies a small subset of bridges essential for maintaining the predictive power of the forecasting model. The findings of this study will assist transportation agencies in utilizing their bridge inspection resources more efficiently and in customizing their condition assessment operations based on bridge characteristics and expected deterioration levels.