INVERSE ANALYSIS OF PAVEMENT LAYER PROPERTIES FROM FALLING WEIGHT DEFLECTOMETER DATA USING MACHINE LEARNING MODELS

Abstract

Finding the layer thicknesses and pavement moduli are essential to evaluate construction quality and pavement life. Destructive test like core sampling provides limited information about pavement characteristics. Nowadays, Falling Weight Deflectometer (FWD), a nondestructive test, is famous for determining pavement health and predicting layer moduli using back-calculated software. However, the accuracy of that backcalculated software is not always acceptable. The accuracy depends on the exactness of layer moduli's seed values and the layer thicknesses values. This research aims to determine the relationship between FWD deflection basin parameters (DBPs) and asphalt pavement layer properties from a database and to predict the layer properties from that relationship. Machine learning models are developed to find the relationship. Due to the scarcity of field FWD data of Roads and Highways Department (RHD) this study uses layered elastic system-based software, General Analysis of Multilayered Elastic System (GAMES), to simulate FWD test on flexible pavement and generate synthetic data base. Validation of the GAMES software’s capability to simulate FWD test is conducted beforehand. For developing models, a total of two thousand FWD tests were conducted using GAMES software to generate deflection data and DBPs with the randomly selected layer properties of the national highways of Bangladesh. Random selection process is applied within the data range to nullify the issue of biasness. Prediction models for pavement layer property are developed embracing Machine Learning (ML) technique like Support Vector Regression (SVR), Random Forest (RF) and multilinear regression using the FWD data base. Thicknesses and moduli were predicted with reasonable accuracy by all three methods. RF, SVR and multilinear regression models establish good relationship between FWD deflection data, DBPs and layer properties. The accuracy co-efficient Mean Absolute Error (MAE), Root Mean Square (RMSE), and Goodness of fit (R2) show that the proposed formulation can predict the thicknesses and moduli with reasonable accuracy. The thickness of the HMA layer is predicted well with R2 values of 0.98, 0.65, and 0.91 with SVR, RF, and multilinear regression models, respectively. Additionally, the modulus of the HMA layer is predicted with R2 values of 0.95, 0.87, and 0.87 with RF, SVR, and multilinear regression models, respectively. SVR method predicts layer iii parameter with lowest MAE and RMSE values. The results indicate that SVR classification produce more accurate results than Random Forest and linear regression. Developed prediction models utilizing FWD deflection data may be applied to predict layer thicknesses and moduli of pavement layers, minimizing the need for destructive tests in a busy roadway and overcoming the dependencies of the back-calculated software.