The safety and operational efficiency of railway infrastructural heavily relies on the condition of railway wheelsets and effective defect detection. Traditional inspection methods such as visual inspection and ultrasonic testing show limitations in terms of precision, operational efficiency and scalability. In this study, a novel approach is presented for real-time wheelset defect detection using state-of-the-art deep learning model: Neural Compact Refined Architecture-You Look Only Once model (NCRA-YOLO). A large dataset consists of 7487 high-resolution annotated images of different wheelset. The dataset used for training and testing utilized advanced data augmentation techniques to enhance performance. The model demonstrated superior accuracy of 100% in fracture and 83% in spot class. This model is the most effective solution with only 1.3 million parameters and 3.8 GFLOPS that proves its effectiveness in real time defect detection. The model achieved a precision of 94.7% and a mean average precision at IoU thresholds of 50% to 95% (mAP50:95) of 96.1% during cross validation, making it as the most effective model for generalizing to unseen data.

Muhammad Zakir Shaikh; Sahil Jatoi; Enrique Nava Baro; Bushra Abro; Agata Manolova; Bhawani Shankar Chowdhry

2025 28th International Symposium on Wireless Personal Multimedia Communications (WPMC)

https://ieeexplore.ieee.org/abstract/document/11351276

Railway Safety; Wheelset Defect Detection; Deep Learning; NCRA-YOLO; Computer Vision