Towards a real-time railway monitoring system based on Distributed Acoustic Sensing and a Convolutional Neural Net

Distributed Acoustic Sensing (DAS) enables cost-efficient retrieval of wavefield information alongside large linear infrastructure elements, such as pipelines, cables or railways. The massive datasets require automated approaches to detect and classify anomalies which can trigger further investigation through an operator, if necessary.Within this work we exploit one week of DAS data from a fiber-optic cable co-located with a 50 km long railway line south of Trondheim, in the center of Norway. The data were acquired with a temporal sampling of 2 kHz and channel spacing of 4 m, resulting in 12500 channels. Treating the DAS time-space domain matrices like images we can make use of well-established techniques from the field of computer vision. We compute sliding RMS windows of 60 s and 1.5 km with 50 percent overlap and use them as input images for a Convolutional Neural Network. The network classifies events such as trains, cars, unknown events as well as different noise classes and artefacts. In order to thoroughly train the network, we labeled approximately 1000 RMS images per class and further applied a variety of data augmentation techniques to finally obtain about 5000 labeled images per class. Once trained, we can simply apply a forward pass through the network every 30 s for all the 1.5 km overlapping segments to obtain a live-classification of events along the entire railway line.We present our workflow as well as initial results and discuss the potential of DAS for future railway monitoring and the challenges that we encounter. If successful, these methods can open up an opportunity to exploit a large amount of fibers co-located with railway lines enabling automatization of real-time railway monitoring.Acknowledgements:We acknowledge Bane NOR and Alcatel Submarine Networks for conducting the data acquisition for this project. This work is supported by the SFI Centre for Geophysical Forecasting under grant No. 309960.