Optimizing maintenance decisions in railway wheelsets: A Markov decision process approach

Abstract

This article models the decision problem of maintaining railway wheelsets as a Markov decision process, with the aim to provide a way to support condition-based maintenance for railway wheelsets. A discussion on the role of the railway wheelsets is provided, as well as some background on the technical standards that guide maintenance decisions. A practical example is explored with the estimation of Markov transition matrices for different condition states that depend on the wheelset diameter, its mileage since last turning action (or renewal) and the damage occurrence. Bearing in mind all the possible maintenance actions, an optimal strategy is achieved, providing a map of best actions depending on the current state of the wheelset.

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

“…In an SMDP framework, the same bidimensional model considering the wheel diameter and flange thickness is proposed, although it accounts for shock‐based failures in the wheel re‐profiling policy. Braga and Andrade 27 explore the wheel re‐profiling problem through an MDP approach by considering a model that uses the wheel diameter along with a quantitative variable (mileage since last turning) and also incorporates the transition to states with damage. As a result, maintenance guidelines based on mileage since turning are compared across discretized values of wheel diameter.…”

“…The parametric approach results in exponential failure times. In the study by Braga and Andrade, 27 the transition probabilities to damaged states are obtained through an empirical logit model. Hence, the first work relies on a parametric approach with resulting exponential failure times, whereas the second one uses an empirical approach that completely relies on the data acquired, which could potentially lead to underestimated probabilities due to disparities found in the ratio of damaged and non‐damaged wheels.…”

“…The Portuguese train operating company database does not distinguish between turning situations with damage from without damage. Due to this database limitation, the diameter loss due to turning (Δ D T ) probability distributions for damaged and undamaged wheels was considered the same as in Braga and Andrade 27 …”

“…The probabilities represented in Figure 10 withdrawn from Braga and Andrade 27 were calculated using the relative frequency from past samples as an approximation of the transition probabilities, that is, in which n j is the number of wheelsets that transit to a class j of diameter loss and N is the total number of wheelsets.…”

“…For deriving the damage probabilities, a Cox proportional-hazards model (CPHM) is implemented. 30,31 Due to the database limitation of providing reliable information regarding the damage occurrence, another database from previous references, 27,[32][33][34] and from a different train operating company is used for deriving the CPHM.…”