Classification of Rail Irregularities from Axle Box Accelerations Using Random Forests and Convolutional Neural Networks

Abstract

AbstractThe continuously increasing demand for mobility results in increased loading of the Swiss railway network, which is further associated with higher wear and deterioration of the rail infrastructure. Safety relevant surface defects on railway tracks, such as squats, have acted as an important driver of rail replacements in Europe. The early detection of such defects can support the planning of appropriate maintenance measures, such as grinding, which prolong the remaining life of the rails. On-board monitoring has redefined the paradigm of railway infrastructure monitoring, via use of in-service vehicles as mobile sensing systems. Such vehicles are equipped with sensors, e.g. axle box accelerometers in order to continuously collect information on the track and vehicle condition, and support the monitoring of railway assets and infrastructure. Acceleration-based monitoring has been shown to bear tremendous potential for offering temporally and spatially dense diagnostics of railway infrastructure. While the potential of such a monitoring scheme has been proven, the generalization has been limited due to the small sample sizes in existing studies.We propose a methodology to recognize and classify between the most common rail irregularities, namely surface defects, insulated joints and welds, by exclusively relying on the availability of on-board acceleration measurements. We combine labeled information, stemming from rail-head image-based detection, with acceleration measurements. Two classification approaches are compared in this work. The first methodology exploits Convolutional Neural Networks (CNNs) that are applied to the Fourier coefficients, which are computed from acceleration time-series data. The second methodology relies on a more classical machine learning approach, applied on features that are extracted from the acceleration time series, which are then classified using Random Forests. Finally, the uncertainty of the acceleration metrics and of the ground-truth labels is analyzed, motivating the application of acceleration-based detection for improvement of rail condition monitoring. The resulting classifiers can be deployed on regular passenger trains for enabling the continuous and automated monitoring of the rail condition.KeywordsRail and track dynamicsMachine learningBig dataData-driven diagnosticsCondition monitoringDamage assessment