Online detection of character and 3D surface defect in steel rail production

Abstract

The surface quality of steel rail is connected to the safety and service life of high-speed rail transportation, and its rail waist character is essential for logistical monitoring and quality traceability. At the moment, it is difficult to use the same set of equipment to recognize features and defects in three dimensions on the complicated surface of the rail. The ring stroboscopic illumination system was devised in this study based on the features of the complicated surface of the rail, and the whole surface image of the rail was gathered by seven linear scan cameras. Create a point cloud model of the rail surface, then re-calibrate the light source's direction based on the rail's fundamental geometry. The normal vector of the rail surface is then calculated to appropriately recreate the 3D surface of the rail. This research provides a method for eliminating gradient error in the direction of motion by using point cloud registration to increase the accuracy of 3D rail surface reconstruction. The breadth and depth of surface defects were assessed using the rail surface's rebuilt 3D model, and the average relative inaccuracy was 7.23%. The Yolo deep learning algorithm is utilized for character identification, and recognition accuracy may reach more than 99%. Experiments suggest that the approach may be used to detect rail surface defects in three dimensions on time. The method not only could be beneficial to the monitoring and optimization of the quality in the rail manufacturing process, but also establish a solid foundation for increasing the safety of high-speed trains.