A data-driven maintenance policy for railway wheelset based on survival analysis and Markov decision process

Wheelsets absorb a significant part of the maintenance budget of any train operating company. Although wheel wear has been an extensively discussed topic in the literature, wear rates are very rarely characterized by using degradation data in a real-world case study aimed at identifying optimal maintenance policies including both degradation and recovery modeling. Furthermore, wheel defects, which impose an additional challenge to the modeling of the lifecycle of the wheels, are usually considered separately in the literature. In this study, conducted at a Portuguese train operating company, 17 years of inspection data are used to estimate wheel wear rates and survival curves, which are further incorporated into a Markov decision process (MDP) model. A bidimensional framework considering discrete intervals of wheel diameter along with a quantitative variable (kilometers since last turning/renewal) is used to represent the possible wheel states, while the probability of a defect interfering with the wheel maintenance schedule is modeled by contemplating survival curves derived from a Cox proportional-hazards model. Optimal results in terms of minimal cost policy are discussed in the context of the MDP, but a more realistic and easy-to-implement policy fixing one of the parameters is compared with the optimal policy. Results showed that in practice train operating companies might benefit from using the easy-to-implement policy, which has an associated long-run average cost only about 1% higher than the one suggested by the optimal decision map.