Deep learning for in-situ powder stream fault detection in directed energy deposition process

Abstract

In the directed energy deposition (DED) process, stable powder feeding through nozzles is crucial. However, abnormal powder feeding due to nozzle clogging or adhesion frequently occurs, which can cause severe defects or even lead to a process failure. Nevertheless, the abnormal powder feeding detection is challenging because the powder stream variance is subtle, and the sensing area is under a harsh environment with bright melt pool light emission and spatter ejection. This study proposes the abnormal powder feeding detection method based on a deep learning technique using melt pool images. The trained model by training dataset collected through a coaxial camera with deliberately installed clogged nozzle can discriminate between normal and abnormal states and even classify the fault nozzle among the 4 nozzles installed. The classification performance has a prediction accuracy of up to 99.2%. The result shows that the deep learning technique using melt pool images could be promising for in-situ DED process fault detection.