Evaluating Deterioration of Tunnels Using Computational Machine Learning Algorithms

Abstract

Tunnels are an integrated part of the transportation infrastructure. Structural evaluation and inspection of tunnels are vital tasks to assess the deterioration of tunnels and maintain their level of service. Researchers have developed many predictive models that describe the deterioration of various infrastructure systems using data from formal inspections. However, there is a lack of research that developed predictive models of deterioration of tunnels in the US. Therefore, this paper investigated the feasibility of using various machine learning techniques to develop a computational data-driven decision support tool that predicts the deterioration of tunnels in the US. An ex ante framework for predicting the deterioration of tunnels in the US was developed. The research methodology comprised (1) collecting, cleaning, and standardizing data for tunnels in the US from the Federal Highway Administration (FHWA); (2) identifying the best subset of variables that allow predicting the deterioration of tunnels; (3) utilizing existing machine learning algorithms, namely k-nearest neighbors (KNN), random forest (RF), artificial neural networks (ANNs), and support vector machine (SVM), to develop classification models that predict the deterioration of tunnels; (4) optimizing the accuracy of the developed models by determining the best set of hyperparameters that result in the most accurate performance; (5) comparing the performance of the developed models and selecting the best performing model to be used as a decision support tool; and (6) evaluating and validating the performance of the selected model. The results identified 18 variables that greatly affect the deterioration of tunnels, with the tunnel width having the greatest impact on the prediction of deterioration of tunnels. Results indicated that the RF algorithm reached an accuracy of 85.38%, which was the highest accuracy, compared with KNN, ANN, and SVM, which reached an accuracy of 80.12%, 56.14%, and 56.73%, respectively. In addition, the entropy criterion function with a maximum of five features and 500 estimators successfully constructed the best hyperparameters for the selected RF model. Therefore, the developed decision support tool can be used by transportation entities to estimate the overall condition of tunnels based on specific tunnel parameters with reasonable prediction accuracy. It also can aid decision makers in developing, optimizing, and prioritizing maintenance plans and allocation of funding.