A weighted multi-output neural network model for the prediction of rigid pavement deterioration

Abstract

ABSTRACT A novel weighted multi-output neural network (NN) model is proposed for predicting the deterioration of rigid pavements based on Iowa pavement management system data. This first-of-a-kind model simultaneously predicts four pavement condition metrics concerning rigid pavements, including IRI, faulting, longitudinal crack and transverse crack. It provides an opportunity to efficiently evaluate pavement conditions and to make treatment decisions based on multi-condition metrics, such as the pavement condition index (PCI) for budget allocation models. Compared to traditional single-output NN models, this multi-output model is capable of incorporating correlations among different condition metrics. During model training, each condition metric is assigned a weight to reflect its relative importance. When the weights equal to those in the formula for the multi-condition metric, the prediction performance for PCI is optimal (13% lower MSE than optimal, single-output models). The multi-output model improves on the prediction performance for three of the four individual condition metrics compared to optimal single-output models. Results show that the consideration of correlations could improve the prediction performance for single and multi-condition metrics. Finally, variable weighting is critical for achieving the optimal balance of prediction performance among the various metrics as dictated by the needs of the decisionmaker.