Implementation of Deep Neural Networks for Pavement Condition Index Prediction

Abstract

Pavement condition index (PCI) is commonly used in pavement management systems (PMS) for indicating the extent of the distresses on the pavement surface. PCI values are a function of distress type, severity, and density. Artificial neural network (ANN) techniques have successfully modeled the performance of in-service pavements, due to their efficiency in predicting and solving nonlinear relationships and dealing with uncertain large amounts of data. Aiming to investigate and examine the efficiency and reliability of ANN models, this paper develops and trains a deep ANN (DNN) model to predict the PCI values and compares the DNN model performance against conventional prediction methods, such as linear and nonlinear regression. Several models with different hyperparameters and architecture were developed and trained using 536,848 samples and tested on 134,212 samples. The root mean square error (RSME) of the tested DNN model is significantly superior to the best fitted linear and nonlinear regression models. In line with the literature, the most influencing variables for PCI prediction are distresses related to alligator cracking, swelling, rutting, and potholes. These findings suggest that DNN models could be incorporated into the PMS for PCI determination.