Deep learning-based detection of internal defect types and their grades in high-pressure aluminum castings

Abstract

With the increasing use of light alloy castings in automobiles, ensuring quality control is essential for safety. Xray imaging offers a practical approach to detecting internal defects in cast components. This study proposes a method to automatically and in real-time identify the location, type, and size of internal defects in aluminum parts produced via high-pressure casting. The proposed two-stage method can detect, segment, and grade defects without expensive hardware in less than a second. Using the YOLOv5 algorithm for defect detection in the first stage, a mean Average Precision (mAP) of 0.971 was achieved. In the second stage, defect grading is performed through segmentation, enabling classification in accordance with international standards without requiring additional training. The methodology provides real-time and highly accurate internal defect quality control and can be applied to different metals and standards. The dataset used in this study contains over 5,000 labelled X-ray images of aluminum cast parts, and it is made available as open access to support the NDT community and researchers.