Geometry-Agnostic CNN for Porosity Detection in LPBF using Low-Cost NIR and Optical Imaging with XCT Ground Truth

A major barrier to commercializing laser-powder bed fusion (L-PBF) additive manufacturing is the lack of reliable, geometry-agnostic, defect detection methods. Most current solutions or systems are expensive and geometry-dependent which makes them unable to generalize across part shapes, materials or systems. This research explores a novel, geometry agnostic, low- cost approach to in-situ monitoring and porosity detection using near-infrared (NIR) imaging and optical imaging for stainless steel 361L material. The NIR and optical modalities were aligned with post build X-ray computed tomography (XCT) data for ground truth labeling. To generate porosity in this experiment, a spatter generator was used to create seeded porosity in printed metal cylinders. These cylinders were then XCT scanned to characterize localized pores. Layer-wise NIR and optical images were captured using low-cost, off-axis sensors. Semantic image segmentation was applied to the XCT data to identify connected pore structures in each layer. Composited XCT layers were then aligned with composite NIR and optical layers. A convolutional neural network (CNN) was trained to classify each NIR layer as containing a major pore or not, using XCT-labeled slices for supervision. To achieve geometry-agnostic inference, a secondary sub-sampling approach was introduced, wherein local image regions were classified independently and aggregated to infer overall layer porosity classification. The approach was trained on cylindrical geometries and validated on a cuboid from a separate build with no spatter generators, to further showcase generalization.

revised: