Pavement performance prediction by artificial neural networks

There is a growing need for accurate prediction of pavement condition for use in many modules of a Pavement Management System (PMS). Traditionally regression models have been employed for this purpose, and development of such models has not been an easy task, one reason being that the causal factors of pavement deterioration are not well understood. With the advent of PMS, large databases of pavement performance have evolved. What is desired, therefore, is a technique adaptable to solve functional relationships that are often nonlinear for which many of the regression models are not very well suited. As an alternative, artificial neural networks, which offer a flexible form of mapping between inputs and outputs consistent with any underlying relationship, are utilized for prediction modeling. Artificial neural networks are trained to predict Pavement Condition Rating, a composite index indicating the condition of a pavement. Flexible (original and overlaid), composite, jointed concrete, and continuously reinforced concrete pavements are the four families of pavements considered. A brief description of artificial neural networks is included. In order to examine the predictive capability of neural networks, the results obtained are compared with those of regression models developed with the same data The paper concludes that artificial neural network is a viable alternative for prediction modeling.