Optimizing Machine Learning Algorithms for Improving Prediction of Bridge Deck Deterioration: A Case Study of Ohio Bridges

Abstract

The deterioration of a bridge’s deck endangers its safety and serviceability. Ohio has approximately 45,000 bridges that need to be monitored to ensure their structural integrity. Adequate prediction of the deterioration of bridges at an early stage is critical to preventing failures. The objective of this research was to develop an accurate model for predicting bridge deck conditions in Ohio. A comprehensive literature review has revealed that past researchers have utilized different algorithms and features when developing models for predicting bridge deck deterioration. Since, there is no guarantee that the use of features and algorithms utilized by past researchers would lead to accurate results for Ohio’s bridges, this research proposes a framework for optimizing the use of machine learning (ML) algorithms to more accurately predict bridge deck deterioration. The framework aims to first determine “optimal” features that can be related to deck deterioration conditions, specifically in the case of Ohio’s bridges by using various feature-selection methods. Two feature-selection models used were XGboost and random forest, which have been confirmed by the Boruta algorithm, in order to determine the features most relevant to deck conditions. Different ML algorithms were then used, based on the “optimal” features, to select the most accurate algorithm. Seven machine learning algorithms, including single models such as decision tree (DT), artificial neural networks (ANNs), k-nearest neighbors (k-NNs), logistic regression (LR), and support vector machines (SVRs), as well as ensemble models such as Random Forest (RF) and eXtreme gradient boosting (XGboost), have been implemented to classify deck conditions. To validate the framework, results from the ML algorithms that used the “optimal” features as input were compared to results from the same ML algorithms that used the “most common” features that have been used in previous studies. On a dataset obtained from the Ohio Department of Transportation (ODOT), the results indicated that the ensemble ML algorithms were able to predict deck conditions significantly more accurately than single models when the “optimal” features were utilized. Although the framework was implemented using data obtained from ODOT, it can be successfully utilized by other transportation agencies to more accurately predict the deterioration of bridge components.