A Data-Driven Fault Diagnosis Method for Railway Turnouts

Abstract

Turnout systems on railways are crucial for safety protection and improvements in efficiency. The statistics show that the most common faults in railway system are turnout system faults. Therefore, many railway systems have adopted the microcomputer monitoring system (MMS) to monitor their health and performance in real time. However, in practice, existing turnout fault diagnosis methods depend largely on human experience. In this paper, we propose a data-driven fault diagnosis method that monitors data from point machines collected using MMS. First, based on a derivative method, data features are extracted by segmenting the original sample. Then, we apply two methods for feature reduction: principal component analysis (PCA) and linear discriminant analysis (LDA). The results show that LDA gave a better performance in the cases studied. A problem that cannot be overlooked is that the imbalanced quantity of rare fault samples and abundant normal samples will reduce the accuracy of classic fault diagnosis models. To deal with this problem of imbalanced data, we propose a modified support vector machine (SVM) method. Finally, an experiment using real data collected from the Guangzhou Railway Line is presented, which demonstrates that our method is reliable and feasible in fault diagnosis. It can further assist engineers to perform timely repairs and maintenance work in the future.