Industrial Application of Deep Neural Network for Aluminum Casting Defect Detection in Case of Unbalanced Dataset

Abstract

We have developed a deep neural network for casting defect detection. The approach is original because it assumes the use of data related to the casting manufacturing process, i.e. measurement signals from the casting machine, rather than data describing the finished casting, e.g. images. The defects are related to the production of car engine heads made of silumin. In the current research we focused on the detection of defects related to the leakage of the casting. The data came from production plant in Poland. The dataset was unbalanced. It included nearly 38,500 ob-servations, of which only 4% described a leak event. The work resulted in a deep network consisting of 22 layers. We assessed the classification accuracy using a ROC curve, an AUC index and a confusion matrix. The AUC value was 0.97 and 0.949 for the learning and testing dataset, respectively. The model allowed for an ex-post analysis of the casting process. The analysis was based on Shapley values. This makes it possible not only to detect the occur-rence of a defect but also to give potential reasons for the appearance of a casting leak.