Analysis of Defect Associated with Powder Bed Fusion with Deep Learning and Explainable AI

Abstract

Research into the detection, classification, and prediction of internal defects using surface morphology data of parts created via powder bed fusion-type additive manufacturing has become a hot topic in the previous decade thanks to the development of deep learning. However, there is no other evidence to evaluate the model other than accuracy and metrics. In this paper, a novel data set is compiled from various literature and other sources to evaluate the black box model using explainable artificial intelligence (XAI). The data set contains three major powder bed fusion defects: gas porosity, lack of fusion, and balling. The anomaly was initially found using convolutional neural networks (CNN) and transfer learning. Based on test data, a model comparison was performed to determine the best accuracy and an F1 score. VGG16 has outperformed all other models in terms of accuracy, with an F1 score of 98.6 percent. Further, the model has been compared with the existing state-of-the-art model for classification in the domain of powder bed fusion defects. Finally, VGG16 was employed to interpret and explain the test data set. The LIME explanations revealed that the feature predicted by the model was present in conjunction with the fault. As a result, we are confident that the proposed model with XAI would considerably improve the fairness and trustworthiness of the output result in the powder bed fusion field. This can also aid in the automation of additive manufacturing in the realm of Industry 4.0.