Design and Development of a Wayside AI-Assisted Vision System for Online Train Wheel Inspection

Abstract

Wayside inspection of rolling stock has been around for some time and wheel impact load and fiber-grating sensors are actively explored for getting high-fidelity data. Visual inspection from wayside provides the opportunity to gain high-resolution data, which can help in the early diagnosis of potential faults. It is rarely explored due to complexities associated with calibration, moving and rotating targets, and difficulties associated with data acquisition. This paper explores and presents an in-depth design and development strategy for such systems. It presents the development steps, implementation, and results of a vision inspection system for regular and automated inspection of train wheels. First, various configurations for positioning of the cameras in a three-dimensional setting are considered and discussed, followed by online data acquisition for establishing a data set. Later, a comprehensive comparative analysis was conducted on several object detection algorithms for wheel segmentation task. Different algorithms are evaluated using COCO evaluation metrics, and the best-performing model, YOLOv9, achieves a mAP50 of 0.94 and a recall of 0.91. The developed system has produced satisfactory results in acquiring proper wheel tread images and segmenting the wheel. Further avenues for countering lighting issues and defect detection are provided.