Data-Driven Solutions for Backcalculating Elastic Moduli of Flexible Pavements from FWD Test

Abstract

Traditional methods for calculating pavement layers elastic moduli from falling weight deflectometer (FWD) tests often rely on computationally intensive iterative processes and lack struggle to capture complex variable relationships. This article highlights the utilization of machine learning (ML) algorithms, which include artificial neural networks (ANN), long-short-term memory (LSTM), and random forests (RF), to predict the elastic moduli of multi-layered flexible pavement based on FWD test. All ML algorithms were developed using synthetic databases derived from the exact stiffness matrix scheme, which was employed for the analysis of multi-layered pavements under axisymmetric surface loading. The development of ML models involves preprocessing of data, hyperparameter optimization, and performance evaluation.  The input variables consist of the FWD surface deflections, the magnitude of applied loading, and the layer thicknesses, while the output variables represent the predicted layered elastic moduli of the pavement structure. The ANN and LSTM models capture complicated relations more effectively than the RF model in the backcalculation of the layered elastic modulus based on the FWD test. Among the two, LSTM achieves higher accuracy, with the average values across all layer moduli of R² and MAPE being 99.04% and 2.41%, respectively, in the test set. The applicability of LSTM model is further demonstrated by comparing with the backcalculated elastic modulus based on the FWD field experiments performed on the infrastructure of roads in Thailand. Furthermore, a sensitivity analysis reveals that deflections near the center of loading predominantly impact the predictions of upper layer moduli, while the moduli of lower layers are influenced by deflections across all geophones.