Using support vector machines to predict the probability of pavement failure

This paper presents a method to predict the probability of structural failure of road pavements using information contained in road data sets. Expert knowledge was used to develop failure charts to identify the potential factors that may contribute towards pavement failure. A computational technique (a support vector machine) was built to use this information to determine, from the data sets, the probability of failure of individual road sections. With this prediction comes an indication of the predominant failure types, the causes of structural failure and the risk profile of a road network. The usefulness of the approach was demonstrated on a data set taken from the New Zealand long-term pavement performance study of state highways. Analysis of the data set showed that the network was in good condition, but a small number of pavement sections with a high likelihood of failure were identified. Furthermore, the application of the failure paths examined the three predominant failure types occurring on the network and identified their possible causes. Rutting appears to be significantly influenced by the road pavement strength, fatigue cracking seems to be affected notably by the environment (i.e. water ingress) and shear failure is caused primarily by the combination of traffic, pavement composition and strength. In addition, it was confirmed that measured functional pavement condition alone is not a good identifier of failure and that the inclusion of a parameter related to strength, such as pavement deflection, is essential.