A neural network for identification and classification of systematic internal flaws in laser powder bed fusion

Quality control of mechanical components is crucial to ensure their expected performance and prevent their failure. For components manufactured additively, quality control performed in-process is particularly interesting, as the sequential deposition and remelting of layers represent a possibility to mitigate existing flaws. The first step towards closed-loop control is to ensure that the monitoring setup and the data analytics approach can flag and discriminate flaws. This study aims to assess the potential of a layerwise monitoring system associated with a supervised machine learning approach to identify and classify internal flaws in laser powder bed fusion of Hastelloy X. For that, systematically generated internal flaws were mapped ex-situ in 72 distinct process conditions. The outputs of the near-infrared long-exposure acquisition system were labeled according to the ex-situ characterization and used to train a fully convolutional neural network. The network was then used to classify previously unseen monitoring images into three classes, according to the predominant flaw type expected, lack of fusion, keyhole porosity, or residual porosity. Accuracy, precision and recall over 96% are obtained, indicating that the monitoring system combined with this supervised machine learning approach successfully identifies and classifies internal flaws