Wheel and axle defect detection based on deep learning

• Published in: Research and Review Journal of Nondestructive Testing Vol. 1; no. 1
• Main Authors: Peng, Jian ping, Zhang, Qian, Zhao, Bo
• Format: Journal Article
• Language: English; German
• Published: NDT.net 01.08.2023
• ISSN: 2941-4989; 2941-4989
• DOI: 10.58286/28166

With technological innovations in the world of high-speed railways, railways have become an indispensable and important part of life. As a key part of the train, the safety of the wheels and axles cannot be ignored. Industry often uses non-destructive testing (NDT) methods, and because of the special structure of wheels and axles, we commonly use phased-array ultrasonic testing. However, the disadvantage is that ultrasonic inspection methods rely too much on the intuition of skilled workers and as the workload increases, a large amount of data is not used effectively, which can easily lead to safety hazards. To deal with these issues, an efficient detection method emerges as the times require. we collected ultrasound-based B-scan defect data for wheels and axles, by expert manual annotation to establish a database of various types of defects in wheels and axles of existing trains. By using the improved YOLO-v5-based algorithm for training validation and testing, improving the feature extraction layer and adding a small target detection layer for difficult defects. Finally, by adding an attention mechanism to improve the training accuracy and using active learning strategies for data enhancement to make it more applicable to ultrasound images, the experiments significantly improved detection efficiency and stability, with a high defect detection rate and a significantly decreased false alarm rate. The algorithm has good performance with laboratory data. The algorithm has good performance in laboratory data and can meet the application requirements in the actual wheel and axle inspection data, we tested more than 3000 different pictures which are all from the real data collected by ultrasonic testing, with the defect detection alarms reaching 100%, detection speed reaching real-time detection, and false alarms being controlled to within 2%. More importantly, with the self-upgraded of algorithm and new data collection, the detection efficiency will improve gradually.