Comparative Study of Data Mining Models for Prediction of Bridge Future Conditions

Abstract

Highway and bridge agencies use several systematic inspection approaches to ensure an acceptable standard for their assets in terms of safety, convenience, and economic value. The Bridge Condition Index (BCI), used by the Ontario Ministry of Transportation, is defined as the weighted condition of all bridge elements to determine the rehabilitation priority for the bridge. Therefore, accurate forecasting of BCI is essential for bridge rehabilitation budgeting and planning. The large amount of data available about bridge conditions for several years enables the use of different mathematical models to predict future BCI. This research focuses on investigating different classification models developed to predict the BCI in the province of Ontario, Canada, based on the publicly available historical data for 2,802 bridges over a period of more than 10 years. Predictive models used in this study include k-nearest neighbors (k-NN), decision trees (DTs), linear regression (LR), artificial neural networks (ANN), and deep learning neural networks (DLN). These models are compared and statistically validated via cross validation and paired t-test. The decision tree model showed acceptable predictive results (within 0.25% mean relative error) when predicting the future BCI and is the recommended option based on its performance and certainty in posterior maintenance decision making for the selected case study.