Developing Artificial Neural Network Models to Automate Spectral Analysis of Surface Wave Method in Pavements

Abstract

The spectral analysis of surface waves (SASW) method is a nondestructive testing method of pavements based on the dispersive characteristic of seismic surface waves. The method can provide the thickness and stiffness of pavement layers. One of the more complex aspects of the SASW method is an iterative process to estimate the pavement parameters, called the inversion procedure. In this paper, the feasibility of completely automating the inversion process and substituting it with artificial neural network (ANN) models was explored. A number of different ANN models were developed using various ANN training strategies. To improve the performance of some ANN models, a sequential modeling technique was implemented. In the sequential modeling, some pavement parameters are estimated first from an initial set of ANN models, which is then the input to subsequent ANN models to estimate other parameters of interest. Furthermore, the performance of the ANN models was evaluated using a number of well-characterized pavement sections. The results illustrated that ANN models could estimate the upper layers’ parameters so well that they could replace the inversion process. The ANN models for other layers were capable of generating robust initial estimates for a well-constrained formal inversion that can be readily automated.