COMPUTER VISION-BASED AUTOMATED DETECTION OF CASTING DEFECTS IN ШХ15 BEARING RINGS

This study presents a computer vision-based system for automated detection and classification of casting defects in ШХ15 (AISI 52100 / 100Cr6) bearing rings produced from secondary metallic materials. A dataset of 2,400 surface images was collected from bearing ring specimens manufactured by centrifugal and gravity casting methods at different recycled material ratios (20–80%). Images were captured using an industrial CCD camera at 5 megapixel resolution under controlled illumination. Four defect categories were defined: porosity, hot cracks, shrinkage cavities, and non-metallic inclusions. Three convolutional neural network (CNN) architectures — custom lightweight CNN, ResNet-50, and VGG-16 — were trained and evaluated using transfer learning. The fine-tuned ResNet-50 model achieved the best overall performance with accuracy of 96.8%, precision of 95.4%, recall of 96.1%, and F1-score of 95.7%. Porosity was detected with the highest accuracy (98.3%), while non-metallic inclusions presented the greatest classification challenge (93.2%). The developed system enables real-time inspection at 12 images per second, providing a practical alternative to manual visual inspection in bearing ring quality control.