Xinyue Xu

The timely identification of rail internal defects and the application of corresponding preventive measures would greatly reduce catastrophic failures, such as rail breakage. Ultrasonic rail defect detection is the current mainstream rail defect detection method thanks to its advantages of strong penetration, high accuracy, and ease to deploy. The 2D B-scan image output by ultrasonic detectors contains rich features of defects; however, rail engineers manually identify and localize the defect image, which can be time-consuming, and the image may be subject to missing identification or mistakes. This paper adopted state-of-the-art deep learning algorithms for novel B-scan images for the automatic identification and localization of rail internal tracks. First, through image pre-processing of classification, denoising, and augmentation, four categories of defect image datasets were established, namely crescent-shaped fatigue cracks, fishbolt hole cracks, rail web cracks, and rail base transverse cracks; then, four representatives of deep learning object detection networks, YOLOv8, YOLOv5, DETR, and Faster R-CNN, were trained with the defects dataset and further applied to the testing dataset for defect identification; finally, the performances of the three detection networks were compared and evaluated at the data level, the network structure level, and the interference adaptability level, respectively. The results show that the YOLOv8 network can effectively classify and localize four categories of internal rail defects in B-scan images with a 93.3% mean average precision at three images per second, and the detection time is 58.9%, 376.8%, and 123.2% faster than YOLO v5, DETR, and Faster R-CNN, respectively. The proposed approach could ensure the real-time, accurate, and efficient detection and analysis of internal defects to a rail.

The automated detection of distress such as cracks or potholes is a key basis for assessing the condition of pavements and deciding on their maintenance. A fine-grained pavement distress-detection algorithm based on integrated data using a digital twin is proposed to solve the challenges of the insufficiency of high-quality negative samples in specific scenarios An asphalt pavement background model is created based on UAV-captured images, and a lightweight physical engine is used to randomly render 5 types of distress and 3 specific scenarios to the background model, generating a digital twin model that can provide virtual distress data. The virtual data are combined with real data in different virtual-to-real ratios (0:1 to 5:1) to form an integrated dataset and used to fully train deep object detection networks for fine-grained detection. The results show that the YOLOv5 network with the virtual-to-real ratio of 3:1 achieves the best average precision for 5 types of distress (asphalt pavement MAP: 75.40%), with a 2-fold and 1.5-fold improvement compared to models developed without virtual data and with traditional data augmentation, respectively, and achieves over 40% recall in shadow, occlusion and blur. The proposed approach could provide a more reliable and refined automated method for pavement analysis in complex scenarios.