A YOLOv3-based Deep Learning Application Research for Condition Monitoring of Rail Thermite Welded Joints

Abstract

Thermite welded joints are commonly applied in running rails and any failures related to it will result in delay of time schedule or safety issues. The detection and monitoring of these joints are essential for pre-emptive maintenance and therefore to improve the structural integrity of the railway system. This study aims to produce an application prototype to detect and monitor the rail thermite welded joints based on YOLOv3 deep learning algorithm. By using YOLOv3 (You Only Look Once) as the detector, the classification and location boundary box can be determined for each image of thermite welded joint. Experiment on the training and validation of this algorithm has achieved a good result and this application prototype, to be integrated with the proposed camera monitoring system, is capable of detecting and monitoring crucial components of rail track system.