Detection and classification of aluminum casting defects by deep learning approach

Due to their unique properties, aluminum casting parts produced by high pressure casting technique are widely used, especially in the automotive industry. However, inspection of castings is a process that requires non-destructive testing of critical components using technologies such as X-ray to examine internal defects that cannot be seen otherwise. Such a time-consuming visual inspection requires well-trained experts with the utmost care. In this study, it is aimed to detect and grade the internal defects of aluminum casting parts in accordance with an international standard and to correlate this defect information with the casting process parameters. The dataset used (Al-Cast dataset) is a unique dataset of X-ray images, all images are carefully labeled and shared open access for the nondestructive testing community. A deep learning-based object detection method is used for the detection of defects, and the detections are performed in real-time and with high accuracy (mAP value of 0.97). Using the k-means++ clustering algorithm with the information obtained from the detection, the defects are graded by determining the defect rates according to ASTM standards. In addition, the effects of these parameters on the defect rate were analyzed with a relatively small number of experiments on the defect rate output using seven different casting process parameters. The results show that especially the second phase piston speed, third phase pressure and injection fill level have a significant effect on the defect rate. The study has significant potential to be used as a whole decision support system both quality control and production phase.