A Deep Neural Network Approach to Predict Overlay Thickness of Asphalt Pavements Using Deflection Parameters and Estimated Traffic

The objective of this study was to develop a deep neural network (DNN)-based approach to predict the overlay thickness of asphalt pavements using deflection bowl parameters measured with a falling weight deflectometer and the estimated traffic. The scope of the effort was two-fold: (1) Develop a DNN to determine the overlay thickness using deflection and traffic parameters; and (2) train and test the model’s performance. Over 1,300 datapoints from datasets collected from different geographical locations, such as the USA, Canada, and India, were used to train, validate, and test the performance of the model, so that the insufficiency of the historical data could be overcome. The developed network architecture was efficient in predicting the overlay thickness with a reasonably high coefficient of determination (${\mathrm{R}}^2>80\%$). The Morris method of sensitivity analysis was performed to understand the importance of each input parameter in predicting the asphalt overlay thickness. The absolute mean and standard deviation of elementary effects of individual parameters were in close approximation, indicating that each input variable contributed to the overlay thickness prediction. It is noteworthy that the developed model eliminates resource intensive methods of quantifying the pavement thickness, such as cutting and coring of the pavement and rigorous back-calculation processes, thus helping in the prediction of overlay thickness at the project level. Overall, the developed DNN model can help roadway agencies in making rapid and appropriate decisions pertinent to pavement maintenance, rendering it as one of the quality control toolkits easily adoptable during pavement design and operation phases.