A 2014 Bayesian semi-parametric study of Swedish locomotive wheel degradation (piecewise-constant hazard with shared gamma frailty) closed with three unresolved threads: confirming whether wheel/bogie position reliably affects lifetime, explaining the between-locomotive heterogeneity captured by the gamma frailty through measured operating conditions, and optimising maintenance strategies from the derived reliability functions. Later literature has substantially settled the position-effect thread and the maintenance-optimisation thread. The surviving open core is the middle thread: identify and quantify which measured operational covariates (for example climate, track geometry, suspension configuration, driver behaviour and maintenance policy, as the available fleet records permit) explain the between-locomotive frailty in wheel degradation, and propagate the explained heterogeneity together with parameter uncertainty into refined reliability predictions and maintenance decisions.

Background

Railway locomotive wheels degrade through wear and rolling contact fatigue (surface damage driven by the wheel–rail contact), typically tracked by flange-thickness and wheel-diameter loss (the flange being the inward-projecting rim on the running tread that keeps the wheel guided on the rail), and are periodically restored by re-profiling (turning the wheel on a lathe to restore the profile, which consumes wheel diameter). Because re-profiling and replacement dominate wheel life-cycle cost, fleet operators want statistical models that convert degradation measurements into reliability estimates and maintenance intervals. In survival analysis the hazard rate is the instantaneous failure rate at a given mileage or age; a piecewise-constant hazard model lets this rate change stepwise across mileage bands, giving a semi-parametric, flexible lifetime model. A frailty is a random multiplicative factor attached to a group of related observations; a shared gamma frailty (gamma-distributed with mean 1 and variance measuring heterogeneity) captures the tendency of wheels on the same locomotive to degrade alike because of unmeasured shared conditions. In the Bayesian framework the model parameters receive posterior distributions estimated by Markov chain Monte Carlo (MCMC) sampling, and effects are judged by highest-posterior-density (HPD) intervals, the Bayesian analogue of confidence intervals. Lin, Pulido and Asplund (RAMS 2014; journal versions in Reliability Engineering & System Safety 2015 and Proc. IMechE Part F 2015) applied a piecewise-constant hazard regression with shared gamma frailty, estimated by Bayesian and classical methods, to flange-degradation-derived lifetimes of wheels on Swedish locomotives (a bogie is the rigid frame carrying a group of wheelsets — wheel-and-axle assemblies — beneath the vehicle). They found an exponential degradation path adequate, reliability dropping sharply after roughly 180,000 km, a positive but statistically insignificant bogie-position coefficient (β1>0\beta_1>0 at the 95% HPD level), and significant between-locomotive heterogeneity in the frailty, which they attributed to unmeasured operating conditions such as climate, track geometry, suspension configuration, driver behaviour and maintenance policy. Subsequent work resolved much of this agenda: axle-position-dependent wear rates are now demonstrated on large maintenance datasets (e.g., guide wheels — those on the outer, curve-steering axles of a three-axle bogie — of three-axle bogie locomotives wearing about 22% faster than middle wheels), and maintenance optimisation from degradation/reliability models is a developed subfield — dynamic inspection/replacement policies on the same Swedish locomotive wheel-set data (Liu, Lin, Zhang and Xie 2018), semi-Markov (state-transition models in which the time between transitions is random) and Markov-decision-process (sequential decision models optimising a cost criterion) re-profiling policies for high-speed train wheels (E, Li, Jiang and Li 2018; Zeng et al. 2022), re-profiling optimisation for three-axle bogie locomotives validated against measured data (Zeng, Chen, Chi and Zhang 2025), and cost-optimal group maintenance (jointly scheduling maintenance for a set of components) for metro wheels (Shu, Luo and Dai 2026). Frailty methodology itself advanced (an inverse Gaussian process — a degradation-path model whose random increments follow an inverse Gaussian distribution — extended with a frailty term; Morita et al. 2021), and mixed-effect models (regressions combining population-level fixed effects with unit-specific random effects) with Bayesian updating were applied to high-speed train wheels (Li et al. 2019). No published study, however, moves from unexplained frailty to explained heterogeneity: incorporating measured operational covariates into the locomotive-wheel frailty model, quantifying how much heterogeneity they remove, and carrying that — with parameter uncertainty, which follow-up maintenance models still treat as fixed point estimates — into the maintenance decision.

Problem Statement

For a locomotive (or comparable powered rolling-stock) fleet with per-wheel degradation histories and per-locomotive records of operating conditions, determine whether and how measured operational covariates — such as climate/season, track geometry characteristics, suspension configuration, driver behaviour and maintenance policy — explain the between-locomotive frailty, using an appropriate hazard/frailty extension of the Bayesian piecewise-constant hazard model with shared gamma frailty of Lin, Pulido & Asplund. The study is to source or construct the fleet it analyses (the source study's locomotive setting or a comparable powered rolling-stock fleet), documenting its construction and its identification limits. A model-generated fleet — one whose degradation histories and operating-condition records are produced by the study rather than measured on real wheels, whether described as simulated, synthetic or otherwise — may be used only to validate the method, and its covariate results are a method-validation sub-result that does not by itself establish that measured operational covariates explain the real between-locomotive heterogeneity; that explanatory claim requires a fleet with measured records. Determine which of these covariates are statistically supported, and quantify how much of the between-locomotive heterogeneity (frailty variance) they explain, defined as the fractional reduction of the estimated frailty variance relative to the unexplained-frailty baseline, reporting both the support determination and this quantity under one clearly declared model specification. Then propagate the resulting refined reliability model, including parameter uncertainty (posterior distributions rather than point estimates), into maintenance decisions derived from it — for example re-profiling and inspection intervals — and evaluate how these decisions and reliability predictions change compared with the unexplained-frailty model. The problem covers both the identification and quantification of the supported covariates together with their decision impact — the propagation of parameter uncertainty into the maintenance decisions and the comparison against the unexplained-frailty model — and the case in which the available measured covariates do not explain the observed frailty, a finding bounded by the identification limits of the fleet data that leaves the maintenance-decision step with no covariate-explained heterogeneity to propagate.

References

  1. Jing Lin, Julio Pulido, Matthias Asplund (2014), 'Analysis for locomotive wheels' degradation', 2014 Annual Reliability and Maintainability Symposium (RAMS), pp. 1–7, DOI 10.1109/RAMS.2014.6798521, https://doi.org/10.1109/rams.2014.6798521
  2. Jing Lin, Julio Pulido, Matthias Asplund (2015), 'Reliability analysis for preventive maintenance based on classical and Bayesian semi-parametric degradation approaches using locomotive wheel-sets as a case study', Reliability Engineering & System Safety 134:143–156, DOI 10.1016/j.ress.2014.10.011, https://doi.org/10.1016/j.ress.2014.10.011
  3. Jing Lin, Matthias Asplund (2015), 'Bayesian semi-parametric analysis for locomotive wheel degradation using gamma frailties', Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit 229(3):237–247, DOI 10.1177/0954409713508759, https://doi.org/10.1177/0954409713508759
  4. Bin Liu, Jing Lin, Liangwei Zhang, Min Xie (2018), 'A Dynamic Maintenance Strategy for Prognostics and Health Management of Degrading Systems: Application in Locomotive Wheel-sets', 2018 IEEE International Conference on Prognostics and Health Management (ICPHM), pp. 1–5, DOI 10.1109/ICPHM.2018.8448740, https://doi.org/10.1109/icphm.2018.8448740
  5. E. Mingcheng, Bing Li, Zengqiang Jiang, Qi Li (2018), 'An Optimal Reprofiling Policy for High-Speed Train Wheels Subject to Wear and External Shocks Using a Semi-Markov Decision Process', IEEE Transactions on Reliability 67(4):1468–1481, DOI 10.1109/TR.2018.2870276, https://doi.org/10.1109/tr.2018.2870276
  6. Yuanchen Zeng, Dongli Song, Weihua Zhang, Bin Zhou, Mingyuan Xie, Xu Tang (2022), 'An Optimal Life Cycle Reprofiling Strategy of Train Wheels Based on Markov Decision Process of Wheel Degradation', IEEE Transactions on Intelligent Transportation Systems 23(8):10354–10364, DOI 10.1109/TITS.2021.3093019, https://doi.org/10.1109/tits.2021.3093019
  7. Yuanchen Zeng, Dongli Song, Weihua Zhang, Bin Zhou, Mingyuan Xie, Xu Tang (2020), 'A new physics-based data-driven guideline for wear modelling and prediction of train wheels', Wear 456–457:203355, DOI 10.1016/j.wear.2020.203355, https://doi.org/10.1016/j.wear.2020.203355
  8. Yi Zeng, Bingzhi Chen, Qingguang Chi, Xu Zhang (2025; first published online 24 July 2024), 'Wheel Wear Modeling and Re-Profiling Strategy Optimization Based on the Maintenance Data of a Three-Axle Bogie Locomotive', Transportation Research Record: Journal of the Transportation Research Board 2679(2):602–618, DOI 10.1177/03611981241263351, https://doi.org/10.1177/03611981241263351
  9. Ke Shu, Min Luo, Jinzhen Dai (2026), 'Cost-optimal group maintenance strategy for metro wheels using degradation data and particle swarm optimization', Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit 240(7):809–822, DOI 10.1177/09544097261426681, https://doi.org/10.1177/09544097261426681
  10. Qi Li, Jiayu Zhou, Zengqiang Jiang, Mingcheng E, Jing Ma (2019), 'Train Wheel Degradation Modeling and Remaining Useful Life Prediction Based on Mixed Effect Model Considering Dependent Measurement Errors', IEEE Access 7:159058–159068, DOI 10.1109/ACCESS.2019.2950696, https://doi.org/10.1109/access.2019.2950696
  11. Lia H. M. Morita, Vera L. Tomazella, Narayanaswamy Balakrishnan, Pedro L. Ramos, Paulo H. Ferreira, Francisco Louzada (2021; first published online 18 September 2020), 'Inverse Gaussian process model with frailty term in reliability analysis', Quality and Reliability Engineering International 37(2):763–784, DOI 10.1002/qre.2762, https://doi.org/10.1002/qre.2762
  12. Matthias Asplund, Stephen M. Famurewa, Matti Rantatalo (2016; first published online 15 December 2015), 'Data Quality Assessment of Automatic Wheel Profile Measurement Systems', Lecture Notes in Mechanical Engineering (Springer), pp. 717–738, DOI 10.1007/978-3-319-23597-4_53, https://doi.org/10.1007/978-3-319-23597-4_53