An investigation into data driven modelling of rail degradation due to rolling contact fatigue

• Main Author: Riley, Christina Marie
• Format: Dissertation
• Language: English
• Published: University of Southampton 2023

One of the major problems affecting the UK rail network is a family of defects known as Rolling Contact Fatigue (RCF). RCF is a phenomena which arises from repeated contact stresses at the wheel-rail interface resulting in cracks forming at the rail surface, which if left unmanaged can lead to rail fracture. Management of RCF is largely performed using re-profiling methods such as rail grinding and milling. The objectives of such techniques are to restore rail profiles, remove minor cracks, and stall cracks in their early stages of growth, and therefore these activities have typically been performed cyclically at time (or traffic) based intervals. In recent years, the advances in monitoring technologies has dramatically increased the data available to the network operator, in particular Eddy Current technology, which is capable of identifying the depths of RCF cracks in their early stages. This data set is previously unexplored, and presents the opportunity for investigating modern data mining methods to discover insights that may better inform RCF maintenance strategies. Real, operational data however are often noisy, and if the noise is not accounted for can have significant implications on the accuracy of subsequent analysis and modelling. This thesis thus investigates the use of numerous data pre-processing techniques which enable Eddy Current data to be reliably used for information extraction and data-driven modelling. In particular, we address the difficulties in spatially aligning low frequency, sparse data by incorporating data partitioning, cross correlation and optimisation methods. Additionally, the successful preparation of the data enables two main approaches to be explored. Firstly, simple analytical techniques are applied to derive degradation patterns which can augment the current preventive and corrective maintenance decision making processes. Secondly, we demonstrate a methodology for developing a RCF prediction model using several machine learning algorithms for regression analysis. Whilst the resulting models show excellent function fitting capabilities, particularly in the case of ensemble, tree-based methods, we also highlight the potential problems that may arise when using these methods. Despite this, future developments of these models could present excellent opportunities for modelling these complex relationships. At the same time, the data processing and analytical techniques could be presently incorporated into existing RCF management strategies.