Remaining Useful Life Prediction for Railway Switch Engines Using Classification Techniques

Abstract

A highly available infrastructure is a premise for capable railway operation of high quality. Therefore, maintenance is necessary to keep railway infrastructure elements available. Railway switches, especially, are critical because they connect different tracks and allow a train to change its moving direction without stopping. Their inspection, maintenance and repair have long been identified as a cost driver. Switch failures, particularly, are responsible for a comparable high number of failures and delay minutes. The reduction of failures would not only save maintenance costs, but also let more trains arrive on time and hence increase the attractiveness of the railway transport. Therefore, upcoming failures need to be revealed early enough to allow an effective planning and execution of failure preventing maintenance activities. Research is exploring ways to predict the remaining useful life of switches.This paper presents an approach to predict the remaining useful life (RUL) of railway switch engine failures. The development is based on measurement data of the electrical power consumption of switch engines. The two year time series of 29 switches of Deutsche Bahn was recorded by a commercial switch diagnostic system leading to roughly 250 000measurement tuples. Since earlier researched showed that the electrical data alone is not sufficient enough, additional data is integrated. It takes into account the dependency of the switch condition data from climatic conditions and certain properties of the switch construction type.Predicting a RUL is quite challenging in many PHM applications. To avoid common problems with uncertainty in measurement data, a long prediction horizon (month) of small time units (hours) and to stabilise end user acceptance the approach transforms the RUL prediction problem into a classification problem of multiple classes. It, then, uses two different supervised classification techniques, Artificial Neural Networks (aNN) and Support Vector Machines (SVM), to predict the RUL in the form of classes. However, as known from the no free lunch-theorem of classification, there is no ultimately best performing technique. The success depends on the problem and data structure as well as on the parametrisation of the technique or the selected algorithm respectively. Especially aNN and SVM have a high number of possible parametrisations. They can fail the task or result in a very good performance under the heavy influence of their parametrisation. Hence, it is an important aspect of this paper to share how the different parameters effect the RUL prediction and which parameters result in maximum performance. In order to compare the performance, two metrics are chosen, the Matthews Correlation Coefficient (MCC) as single value metric and a visualisation of the confusion matrix as more comprehensible metric. Finally, deriving those parameters maximising the RUL prediction results enables one of the two classification techniques to reveal upcoming failures of the switch engine early enough to prevent them.