Increasing industry acceptance of powder bed metal Additive Manufacturing requires improved real-time detection and classification of anomalies. Many of these anomalies, such as recoater blade impacts, binder deposition issues, spatter generation, and some porosities, are surface-visible at each layer of the building process. In this work, the authors present a novel Convolutional Neural Network architecture for pixel-wise localization (semantic segmentation) of layer-wise powder bed imaging data. Key advantages of the algorithm include its ability to return segmentation results at the native resolution of the imaging sensor, seamlessly transfer learned knowledge between different Additive Manufacturing machines, and provide real-time performance. The algorithm is demonstrated on six different machines spanning three technologies: laser fusion, binder jetting, and electron beam fusion. Finally, the performance of the algorithm is shown to be superior to that of previous algorithms presented by the authors with respect to localization, accuracy, computation time, and generalizability.

粉末床金属增材制造的行业认可度不断提高，要求改进实时检测和异常分类。这些异常中的许多异常，例如重涂机刀片的撞击，粘合剂的沉积问题，飞溅的产生以及一些孔隙，在构建过程的每一层都是表面可见的。在这项工作中，作者提出了一种新颖的卷积神经网络体系结构，用于逐层粉末床成像数据的逐像素定位（语义分割）。该算法的主要优势包括能够以成像传感器的原始分辨率返回分割结果，在不同的增材制造机器之间无缝地传递所学知识以及提供实时性能的能力。在六种不同的机器上演示了该算法，这些机器涵盖了三种技术：激光融合，粘合剂喷射和电子束融合。最后，在定位，准确性，计算时间和通用性方面，该算法的性能显示出优于作者提出的先前算法。