Classification of Friction Cracks in Railway Brake Discs Using Convolutional Neural Networks

Brake discs are critical components of a train’s pneumatic braking system and require monthly preventive maintenance, including cleaning, measurement, and visual inspection. Visual inspections, guided by standards such as UNE-EN 14535-3 and the Faiveley Transport technical manual, involve assessing cracks classified by size and type (penetrating, incipient, superficial). This study presents an automated solution for crack detection and classification using thermal imaging and artificial intelligence. A low-cost Flir Lepton 3.1R thermal camera was used to identify cracks based on size and location. The proposed method integrates the YOLOv8m model with a U-Net (ResNet50 backbone) and a CNN for improved detection and segmentation. Compared to previous approaches using YOLOv4 or Mask-RCNN, this combination provides higher precision in identifying thermal cracks on railway brake discs. The process begins with thermal image acquisition and data augmentation through noise injection and geometric transformations, producing a robust dataset for training the neural networks. Experimental results demonstrate the feasibility of applying AI and thermal imaging for predictive maintenance, achieving a 97% detection accuracy. Furthermore, the study confirms that using higher-resolution thermal datasets enhances model performance, making this approach a cost-effective and reliable tool for automated brake disc inspection.