A Multi-Stage Deep Learning Pipeline for Crack Detection and Segmentation in High-Speed Rail Brake Discs

The braking system of high-speed rail (HSR) is critical for passenger safety, but prolonged use of brake discs can lead to cracks, posing significant risks. Traditional manual inspection methods are time-consuming and labor-intensive. To address these challenges, we propose a multi-stage crack detection and segmentation system, a cascade mode to progressively identify cracks, leveraging deep learning techniques. The system includes automated circle detection, region cropping, and crack detection and segmentation to accurately identify cracks. Using a real-world dataset collected from HSR brake discs, our approach significantly improves data preprocessing efficiency and model performance. Furthermore, we evaluate the performance on un-seen data and compare our crack detection framework with other deep learning models. The experimental results demonstrate the system's effectiveness which is superior to other deep learning models, highlighting its potential for real-world deployment in HSR maintenance.