RNN-based pavement moduli prediction for flexible pavement design enhancement

• Published in: Case Studies in Construction Materials Vol. 20; p. e02811
• Main Authors: Al-Qaili, Abdulraaof H., Al-Mansour, Abdullah I., Al-Solieman, Hamad, AlSharabi, Khalil
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.07.2024; Elsevier
• Subjects: Artificial neural network; Falling weight deflectometer; Layer moduli backcalculation; Pavement design; Recurrent neural network; Pavement design; Falling weight deflectometer; Layer moduli backcalculation; Artificial neural network; Recurrent neural network
• ISSN: 2214-5095; 2214-5095
• DOI: 10.1016/j.cscm.2023.e02811

In order to facilitate the effective implementation of the MEPDG, researchers concentrate on quantifying local material properties, with a particular emphasis on pavement layer moduli. The layer modulus is a critical parameter necessary for calculating pavement responses (stress, strain, and deflections) resulting from traffic loading. Accurately determining the layer modulus is crucial for enhancing pavement design as it directly impacts the required pavement layer thicknesses and associated costs. Backcalculation is a commonly used method for analyzing Falling Weight Deflectometer (FWD) data to determine pavement layer moduli, with Artificial Neural Networks (ANNs) being the traditional choice. However, ANNs have limitations in terms of convergence accuracy and generalization capability. The aim of this study is to improve the backcalculation of layer moduli to enhance pavement design. By utilizing FWD data, Recurrent Neural Network (RNN) was employed to address the limitations of conventional ANN. Both ANN and RNN networks were developed and trained using identical properties. The findings demonstrate that RNN achieved faster convergence and higher convergence accuracy compared to ANN. The RNN network generated reasonable and precise layer moduli values, exhibiting a determination coefficient (R) of 0.95 in comparison to the measured values, while the ANN network had an R-value of 0.79. The results indicate that the RNN network can learn the continuity pattern between deflection basin points, thereby enhancing the accuracy of FWD backcalculation.