Deep Learning Based Classification of Rail Defects Using On-board Monitoring in the Stockholm Underground

Abstract: The purpose of this work is to find out if an artificial neural network can be useful purpose of this work is to find out if an artificial neural network can be useful in order to detect rail squats with the existing Quiet Track Measurement System (QTMS). Squats are surface-initiated rail defects which arise due to rolling contact fatigue. The monitoring system, installed on seven trains running on the green line in the Stockholm underground, aims to improve the maintenance process. The early detection and surveillance of defects helps to extend the service life of the tracks and reduce operating costs. An artificial neural network is used to analyse the the continuously recorded measurements, which consist of vertical bogie acceleration and surrounding noise, each sampled with a frequency of 22 kHz.In particular, the power spectral density as input for multi-layer Fully-connected Neural Network (FNN) has proven to be promising for accurate squat predictions. The supervised learning was carried out according to the one-vs-all principle, i.e. squats versus all other events. A two-hidden-layer FNN has finally been chosen to complement the QTMS. The usage of the full available frequency range from almost DC up to 11kHz, but minimum 7 kHz, allows good prediction with only low false prediction rates. When concatenating all six measurement channels to a single classifier input, an accuracy of over 96% for the squat class and up to 99.98% can in total be achieved. The chosen network type also showed high stability despite quite strong parameter variations and a massive under-representation of squat observations in the measurement data.However, since limited maintenance information about actual squats is available for labelling and testing, more evaluation is needed. The correct identification of mis-labelled squats indicates the high potentials of artificial neural networks.