Machine Learning Models for Predicting Permanent Deformation in Railway Tracks

To enhance track geometry maintenance planning and reduce infrastructure costs, accurate predictions of accumulated permanent track deformation (settlement) caused by cyclic loading of ballast and subgrade is crucial for railway infrastructure managers. This paper proposes a novel approach to predict long-term settlement with reduced computational cost, based on an extensive parameter study using a hybrid methodology to evaluate both short- and long-term track performance. Various machine learning techniques are compared and employed to develop predictive models, which are validated using measured results from a filed demonstrator of ballasted track. The performance and accuracy of each model are assessed using multiple metrics, and a sensitivity analysis is conducted to identify influential explanatory variables. Notably, the developed random forest model demonstrates good agreement with field measured settlement data. This approach bridges the gap between numerical simulation and empirical data, offering an improved holistic understanding of permanent track deformation. The methodology holds potential for implementation in a computational decision support system for railway track maintenance and renewal management.