Inspection of sandblasting defect in investment castings by deep convolutional neural network

Investment castings often have surface impurities, and pieces of shell moulds can remain on the after sandblasting. Identification defects involves time-consuming manual inspections in working environments high noise poor air quality. To reduce labour costs increase health safety employees, automated optical inspection (AOI) combined with a deep learning framework based convolutional neural networks (CNNs) was applied for detection sandblasting defects. Four classic CNN models, including AlexNet, VGG-16, GoogLeNet, ResNet-34, were training predictive classification. A comprehensive comparison reveals that GoogLeNet v1 could accurately determine whether there Among four AlexNet VGG-16 most accurate, prediction accuracy 99.53% 99.07% qualifying products both 100% defective products. v4 ResNet-34 did not perform as expected defect prediction. The reasoning behind performance is attributed to restrictedness investment casting dataset use models residual architectures. Finally, direct technique AOI structure fast flexible computational interface demonstrated.