Machine Learning Techniques for Pattern Recognition in Railway Switches: A Real Case Study

Abstract

Railway switches are formed by different electro-mechanical components that operate railway directions. These systems play a key role on railway networks and any malfunctioning or failure may cause train delays, reducing the operational safety and increasing the maintenance costs. Railway points require condition monitoring systems and new machine learning techniques to analyze the real state of the switches, reducing maintenance activities and ensuring proper behavior of the system. The novelty proposed in this paper is the application of machine learning techniques for pattern recognition. The duration of the opening and closing movements of railway switches has been defined as main variable for this work, since elevated periods may indicate the presence of failures or issues. The case study proposed in this paper presents real data of several switches analyzed with different machine learning techniques: Shapelets and k-Nearest Neighbors. These algorithms are used to detect movements of longer duration and the results provide accuracies above 90%.