Deep Learning for Automatic Railway Maintenance

We propose a computer vision system which detects defects on rails using images. Banedanmark, the Danish railway manager, conducts regular inspections of their railway infrastructure to assess the safety and efficiency of operations. Indications of defects on rails are detected by a recording car which has a system based on an eddy current system for investigating the top part of the rail and an ultra-sound system for the lower part.

The proposed computer vision system is a complementary and convenient method to identify and monitor defects on rails. The computer vision system is based on the deep learning framework called YOLO v3. The detected defects are represented by bounding boxes. The system can automatically detect surface defects as well as problematic isolation joints from images of the railway track, including complex geometric sections in turnouts.

On an independent test set, it is shown that the computer vision system finds defects, with classification rates of 84% for surface defects, 71% for problematic isolation joints and 74% for non-problematic isolation joints, corresponding to a macro F1-score of 81.5%. This could potentially reduce maintenance cost through automatized fault detection from the developed computer vision system.