Improved artificial neural networks for predicting the response of unbonded concrete overlays in a faulting prediction model

Abstract

ABSTRACT Transverse joint faulting is a common distress that develops in unbonded concrete overlays (UBOLs). In the previous predictive faulting model for UBOLs, artificial neural networks (ANNs) were trained to predict the critical response (deflections) resulting from traffic and environmental loads. These ANNs were trained using an extensive factorial of finite element runs of a field-validated structural model. While the ANNs were developed to predict the critical response of UBOLs with asphalt and fabric interlayers, the predictive models were not able to differentiate the critical response as a function of interlayer type. In this study, the ANNs were improved to address this prior limitation. A separate set of ANNs was produced for UBOLs with asphalt interlayers and UBOLs with fabric interlayers. Several enhancements were also performed, including modification of the ANN architecture and the inclusion of an extensive sensitivity analysis in the validation process. The newly developed ANNs are incorporated into the Pitt UBOL-ME faulting prediction model and design guide.