Using Artificial Neural Networks as a Forward Approach to Backcalculation

Abstract

In recent years, artificial neural networks have successfully been trained to backcalculate pavement layer moduli from the results of falling weight deflectometer (FWD) tests. These neural networks provide the same solutions as existing programs, only thousands of times faster. Unfortunately, their use is constrained to the test conditions assumed during network training. These limitations arise from practical aspects of neural network training and cannot be circumvented easily. The goal of this research was to develop a backcalculation program combining the speed of neural networks and the flexibility of conventional programs to produce the same solutions as existing programs. This was accomplished by forgoing neural network backcalculation in favor of neural network forward-calculation, that is, using neural networks in place of complex numerical models for computing the forward-problem solutions used by the conventional backcalculation programs. A suite of neural networks, covering a range of flexible pavement structures, was trained using data generated by WESLEA, the forward-problem solver used in the WESDEF backcalculation program. When tested on 110 experimental FWD results, a version of WESDEF augmented by the neural networks provided statistically identical answers 42 times faster, on average, than the original. Provisions have been made for periodic upgrades as additional networks are trained for other pavement types and test conditions. Meanwhile, the original WESLEA can still be used when an appropriate network is unavailable. This preserves the flexibility of the original program while taking maximum advantage of the speed gains afforded by the neural networks.