Keywords: fatigue crack propagation, railroad fatigue, machine learning, predictive analysis.

The railroad tracks are frequented by thousands of rolling stocks every day. Depending on the type of the rolling stocks (fret, passenger, etc.) and the corresponding conveyed weight, the different parts of railroad tracks are under a constant stress. In recent years, thanks to the technological advances, more data are collected using automatic inspections of railroad tracks and infrastructure and have been acquired by the French National Railroad Company (SNCF). In this article, the objective is to analyse the fatigue crack propagation on subsurface of rails (squat defects) with the aim to avoid the potential rail fractures. As a considerable amount of data is provided in this work, we propose the use of data-driven techniques for prediction of the evolution of crack lengths over time. These models have the advantage of considering a number of influent factors in the modelling unlike the mechanical models, e.g., infrastructure and traffic related factors, climatic variables, etc. However, the calibration of the hyperparameters of data-driven models is of utmost importance. We have conducted an analysis to study the effect of hyperparameters on the predictive capacity of models. Finally, a number of state-of-the-art machine learning techniques are evaluated for the prediction of fatigue crack length and their performances are compared. The neural network based models obtain the promising results and could be investigated in more depth in future works. We give also some insights of models which consider temporal dependency between observations.

download the full-text of this paper (PDF, 7 pages, 606 Kb)

go to the previous paper
go to the next paper
return to the table of contents
return to the volume description