BACKCALCULATION OF FLEXIBLE PAVEMENT MODULI USING ARTIFICIAL NEURAL NETWORKS

Abstract

Artificial neural networks provide a fundamentally new approach to backcalculation of pavement layer moduli from falling-weight deflectometer deflection basins. An artificial neural network is a highly interconnected collection of simple processing elements that can be trained to approximate a complex, nonlinear function through repeated exposure to examples of the function. In the context of backcalculation, a neural network can be trained to approximate the inverse function by repeatedly showing it forward problem solutions. The single most important advantage of using neural networks for backcalculation is speed. Neural networks trained in this study are more than three orders of magnitude faster than conventional gradient search algorithms. Such speed makes real-time backcalculation of moduli possible. Two backpropagation neural networks were trained to backcalculate pavement moduli for three-layer flexible pavement profiles. Synthetic deflection basins with a wide variety of layer moduli and thicknesses were used to train both networks. One network was trained using ideal deflection basins. Subsequent testing showed that the network could backcalculate pavement layer moduli accurately. A second network was trained using basins, with random noise added to simulate measurement errors. When tested using similarly noisy deflection basins, that network did a reasonably good job of predicting moduli, although it exhibited much more scatter in the results. That same network performed very well on experimental data from two pavement test sections of the Strategic Highway Research Program.