Skip to main content
Passa alla visualizzazione normale.

CLARA CELAURO

Prediction of Airport Pavement Moduli by Machine Learning Methodology Using Non-destructive Field Testing Data Augmentation

  • Authors: Baldo N.; Rondinella F.; Celauro C.
  • Publication year: 2023
  • Type: Contributo in atti di convegno pubblicato in volume
  • OA Link: http://hdl.handle.net/10447/594819

Abstract

For the purpose of the Airport Pavement Management System (APMS), in order to optimize the maintenance strategies, it is fundamental monitoring the pavement conditions’ deterioration with time. In this way, the most damaged areas can be detected and intervention can be prioritized. The conventional approach consists in performing non-destructive tests by means of a Heavy Weight Deflectometer (HWD). This equipment allows the measurement of the pavement deflections induced by a defined impact load. This is a quite expensive and time-consuming procedure, therefore, the points to be investigated are usually limited to the center points of a very large mesh grid. Starting from the measured deflections at the impact points, the layers’ stiffness moduli can be backcalculated. This paper outlines a methodology for predicting such stiffness moduli, even at unsampled locations, based on Machine Learning approach, specifically on a feedforward backpropagation Shallow Neural Network (SNN). Such goal is achieved by processing HWD investigation and backcalculation results along with other variables related to the location of the investigation points and the underlying stratigraphy. Bayesian regularization algorithm and k-fold cross-validation procedure were both implemented to train the neural model. To enhance the training, a data analysis technique commonly referred to as data augmentation was used in order to increase the dataset by generating additional data from the existing ones. The results obtained during the model testing phase are characterized by a very satisfactory correlation coefficient, thus suggesting that the proposed Machine Learning approach is highly reliable. Notably, the proposed methodology can be implemented to evaluate the performance of every paved area.