Prediction of the crack condition of highway pavements using machine learning models

Abstract

Departments of Transportation regularly evaluate the condition of pavements through visual inspections, nondestructive evaluations, image recognition models and learning algorithms. The above methodologies, though efficient, have drawn attention due to their subjective errors, uncertainties, noise effects and overfitting. To improve on the outcomes of the shallow learning models already used in pavement crack prediction, this paper reports on an investigation of the use of recursive partitioning and artificial neural networks (ANN; deep learning frameworks) in predicting the crack rating of pavements. Explanatory variables such as the average daily traffic and truck factor, roadway functional class, asphalt thickness, and pavement condition time series data are employed in the model formulation. Overall, it is observed that the recursive partitioning (regression tree – R² > 0.8 and classification tree – R² > 0.6) and ANN (continuous response – R² > 0.8 and categorical response – R² > 0.6) are compelling machine learning models for the prediction of the crack ratings based on their goodness-of-fit statistics, mean absolute deviation (MAD < 0.4) and the root mean square errors (RMSE between 0.30 and 0.65).

Keywords:

• Pavement crack prediction
• recursive partitioning
• artificial neural network
• classification and regression tree
• machine learning
• infrastructure performance modelling

Acknowledgments

The authors would like to appreciate the support of the Florida Department of Transportation for providing the authors with the highway pavement survey data and hurricane damage records on civil infrastructure in Florida.

Disclosure statement

The authors would like to state that there is no potential conflict of interest.