Deep learning-based multisensor fusion for in situ defect prediction in additive manufacturing

While metal additive manufacturing has enabled rapid manufacturing of complex part geometries, non-destructive part evaluation remains cost prohibitive for many applications. Progress has been achieved with in situ process monitoring, but machine learning methods for predicting detailed part quality from process data have not been widely adopted. This paper presents a supervised deep learning approach for defect prediction in laser powder bed fusion (LPBF). The supervised deep learning approach leverages multiple in situ sensors which capture images and video of the build process in multiple spectral bands including visible, near-infrared (NIR), and long wave infrared (LWIR). Defect labels are obtained by computed tomography (CT) scans of finished parts and subsequent volumetric pore analysis. Sensor fusion techniques developed to align multi-sensor data and CT flaw labels to the build coordinate system are presented in detail. Multiple neural network architectures are demonstrated, with emphasis on feature level fusion to support processing multiple sensor inputs in native resolution. Defect prediction accuracy of 97.86% on independent test set is demonstrated.