PREDICTION OF TRACK STRUCTURE DEFECTS BY MACHINE LEARNING METHODS

Abstract

This paper proposes methods of defect prediction for railroad track superstructure objects using machine learning methods. In the railroad sector, most maintenance work is performed on a schedule, which can lead to system failure between two adjacent checks. In recent years, much attention has been paid to new technologies and “smart” approaches based on machine learning techniques, to build a predictive maintenance system. The problem of defect detection from a machine learning perspective is a classification problem with two classes. The initial observation data for the state of the superstructure of the railway track of the problem are unbalanced. This is due to the fact that one of the classes, on the objects of which a track structure defect has been registered, occurs much less frequently. Therefore, when solving the problem, an important parameter is the binarization threshold for machine learning algorithm responses. Modern methods for solving classification problems for tabular data were used to solve the problem. In addition to classical machine learning methods, such as gradient boosting, recurrent neural networks of different architectures were used. The results suggest that a practical threshold has been reached for the accuracy of model predictions, taking into account the noisiness of the input data. The practical significance of this work is that the proposed set of methods can be considered as part of a track maintenance decision-making system. It can be easily adapted for online operation and integrated with an automated measuring system based on a track geometry “recording” car.