Bridge elements' weight determination and components' condition prediction using machine learning approach
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Managing the deterioration of the bridge components and elements has continued to be one of the major concerns for the Departments of Transportation (DOTs) in the United States due to the huge cost needed for constructing new bridges. To avert disasters that could lead to severe losses, deterioration models have been created to predict the future condition of the deck while attributing the deterioration to different factors elicited by engineering and statistical techniques. Previous deterioration models have been more like linear regression models and did not predict the discretized value of the deck condition rating. A prediction of 6.51 was approximated to be a condition rating of 7, which is inappropriate for a discrete data type. This research project uniquely combines all the bridge features identified in the literature by applying principal component analysis (PCA) to capture the variance in the dataset needed to predict the condition of the bridge components. Feedforward Artificial Neural Network (ANN) models were created using different numbers of principal components and the performances were compared with that of the base model that uses all the features collected from the literature. It was observed that 9, 9, and 10 principal components are needed to create a deterioration prediction model that gives a better prediction accuracy than the base model that uses all the bridge features in the Deck, Superstructure, and Substructure respectively. The deterioration of the bridge elements is also known to influence the condition rating of the bridge components and the overall condition of the bridge. The weight or importance of the bridge elements influences the maintenance, repair, and replacement (MRR) schedule of the Departments of Transportation (DOTs) and the resource allocation to the structures. The DOTs currently use a cost-based approach to assign weight to bridge elements which can be in terms of the loss accrued during downtime or the amount needed for the replacement of the element. However, this approach does not consider the bridge element's structural relevance to the bridge's overall performance. This research also uniquely uses the Random Forest (RF) algorithm, an ensemble of decision trees, to evaluate the importance of different elements to the condition of the bridge components and the overall condition of the bridge. The analysis focused on 15 bridge design types in Delaware, Maryland, Pennsylvania, Virginia, and West Virginia and discovered that the weight of bridge elements is not constant as insinuated by the cost-based approach but varies based on its relevance to the bridge's structural performance. The resultant bridge elements’ weight can be used to construct the Bridge Health Index (BHI) equations for the different bridge types. The novel approach herein provides the DOT personnel with data-driven evidence to determine which set of bridge elements to prioritize in their maintenance actions to improve the components' condition, and the overall condition of their bridge inventory and to ascertain if the elements receiving the highest priority in the MRR schedule and budget allocation are also the same set of elements that bridge inspectors regard as needing attention. Furthermore, the technique presented also serves as an approach for synthesizing the bridge component and element-level data and aids the conversion process between the two important datasets.
Description
Keywords
Bridge components, Bridge deterioration, Bridge elements, Machine learning, Prediction models