Advances in predictive maintenance planning of roads by empirical models

Roads constitute one of the most valuable asset group of a country. Road asset value is better preserved by preventive instead of corrective maintenance. Traditionally, road maintenance has been planned using mechanistic or hybrid models for estimation, prediction and optimisation. This thesis proposes a framework for road maintenance planning, where Road User Costs and Agency Costs are estimated and forecasted, road condition forecasted and future maintenance works optimised with empirical models. The concept is called predictive maintenance planning. Various empirical methods were combined and applied in sub-tasks of road maintenance planning. The methods included Sequential Input Selection Algorithm for variable selection, k-means++ for clustering, Principal Component Analysis for dimension reduction, Markov Chains, Ordinary Least Squares regression, Radial Basis Functions and Least Squares Support Vector Regression for forecasting and Genetic Algorithms and Variable Neighbourhood Search for optimisation. The research showed that accuracy of road condition forecasting can be increased with non-linear empirical models using collected data from the roads. The best method was Least Squares Support Vector Regression for multi-step ahead forecasting. The best applied optimisation method combined Parallel Genetic Algorithms with Variable Neighbourhood Search. The Thesis shows that accuracy can be increased and cost saved in road maintenance by a paradigm shift from mechanistic to empirical models in road maintenance planning.