This paper describes how one of the more complex aspects of the spectral-analysis-of-surface-waves (SASW) method for determining the stiffness parameters of pavement layers is the inversion procedure. The artificial neural networks (ANN) have been recently advocated for this purpose. In this paper the results from the evaluation of a number of training strategies were used to determine the feasibility of completely substituting the inversion process with the ANN models. The models were evaluated using on a number of well-characterized pavement sections. The results illustrated that most artificial neural networks were capable of generating reasonably robust a priori information for a well-constraint formal inversion. In particular, the upper layers' parameters from most ANN models are so well estimated that they can be removed from the optimization. However, in order to estimate the parameters of the intermediate layers, a formal inversion has to be used.
Abstract