physical model-based predictive maintenance for rail infrastructure
Annemieke Meghoe is a PhD student in the department of Mechanics of Solids, Surfaces and Systems. Her supervisor is prof.dr.ir. T. Tinga from the Faculty of Engineering Technology.
The very first rails in the Netherlands were introduced in around 1839 and the maintenance activities over the past decades have been based mainly on experience from the past and from periodic inspections. With increasing traffic load, introduction of new vehicles and changing environmental conditions the prediction of rail damage becomes inaccurate and railway maintenance planning needs to be optimized. This thesis describes the utilization of physics-based models for rail damage prediction and railway maintenance planning optimization. A physical model is able to calculate the yet unknown future degradation if the current state of the system is known. To achieve this, numerical or mathematical descriptions of dominant failure mechanisms are usually used. However, physical failure models directly coupled with monitoring techniques for varying operating conditions in the rail infrastructure field are not available. Most of the research in this field is conducted at material level and the application in maintenance modelling is limited. Hence, the main objective of the research is the further development and application of the physics-based models within the rail-infrastructure for varying conditions and to use the outcome (e.g. critical parameters and Remaining Useful Life) for monitoring and maintenance purposes.
From the rail failure and maintenance cost database of Strukton Rail and also verified by literature study, it became apparent that wear and Rolling Contact Fatigue (RCF) are the most dominant degradation mechanisms for rail damage. Wear and RCF are caused by high stresses within the wheel-rail contact resulting from heavy loads and vehicle dynamics. Therefore, multi-body dynamics simulations are used to predict the dynamic behaviour and contact forces. However, the prediction for varying operation conditions means performing multiple simulations and can be time-consuming. Therefore, this thesis proposes the use of meta-models for both wear and RCF in order to increase the computational efficiency.
First, a sensitivity analysis is performed to investigate the dominant influence parameters related to operational conditions. The results from the sensitivity analysis show that rail profile, vehicle speed, rail cant, axle load, curve radius, material hardness, friction coefficient and the primary longitudinal and lateral stiffness of the train bogie are relevant. After gaining insight into the most dominant parameters the meta-models are developed that establish the relation between rail wear and the usage profile of railway tracks. The geometry of these tracks also serves as input parameters of the meta-models. The fit of these models is based on a large data set generated from various scenarios. The selection of these scenarios is carried out by means of a Design of Experiments (DOE) method, known as the Latin Hypercube Sampling (LHS). This method randomly generates scenarios with different parameter settings. The best fit of the meta-model is then found by using the Response Surface Methodology.
The output of the meta-models for both wear and RCF are then validated with field measurements. Eddy current measurements were used to validate the performance of the RCF meta-models. The validation results were positive, but the model's only shortcoming was that it was unable to predict the size of a crack. Therefore, in order to fill this gap, data-driven methods were introduced and a hybrid method was developed.
The validation of the wear meta-models was performed with measured rail profiles using portable measuring devices like Railmonitor and MiniProf. Furthermore, due to the large uncertainty in the input parameters a stochastic approach was adopted in order to provide infrastructure managers the predicted rail wear with certain confidence bounds. The predicted mean and variance of the meta-model response, in this case the rail wear area, corresponded with the results of the field measurements.