Broken rails remain one of the main causes of railway accidents despite improved rail quality and inspections. Today, signalling system with track circuits are often used to detect broken rails, however new signalling systems such as the European Train Control System (ETCS) with axle counters are replacing track circuits and therefore new methods are needed to detect broken rails. One promising technology is distributed acoustic sensing (DAS). The best DAS analytical techniques to show the position of the rail break are gradient analysis, convolution using the train consist and 2 D sinusoidal convolution. This paper describes field tests carried out on DAS to evaluate its ability to detect broken rails and the moment when rails break. The testing showed that DAS has good potential for detecting broken rails in both scenarios. Especially helpful is the ability of DAS to identify the moment of rail breakage and its potential for determining the break’s location on open lines. Recommendations for further research include testing these qualities in more detail and considering how DAS could be efficiently combined with new signalling systems.

This paper presents a comprehensive investigation of machine learning-based intrusion detection methods to reveal cyber attacks in railway axle counting networks. In contrast to the state-of-the-art works, our experimental results are validated with testbed-based real-world axle counting components. Furthermore, we aimed to detect targeted attacks on axle counting systems, which have higher impacts than conventional network attacks. We present a comprehensive investigation of machine learning-based intrusion detection methods to reveal cyber attacks in railway axle counting networks. According to our findings, the proposed machine learning-based models were able to categorize six different network states (normal and under attack). The overall accuracy of the initial models was ca. 70–100% for the test data set in laboratory conditions. In operational conditions, the accuracy decreased to under 50%. To increase the accuracy, we introduce a novel input data-preprocessing method with the denoted gamma parameter. This increased the accuracy of the deep neural network model to 69.52% for six labels, 85.11% for five labels, and 92.02% for two labels. The gamma parameter also removed the dependence on the time series, enabled relevant classification of data in the real network, and increased the accuracy of the model in real operations. This parameter is influenced by simulated attacks and, thus, allows the classification of traffic into specified classes.