Submitted Papers

Manuscripts currently under review.

Zohoury, Kian, Cornelius, Jackson, Petrich, Jan, Cole, Vernon J, and Reutzel, Edward W. “3D Convolutional Neural Networks for Robust In-Situ Defect Segmentation in Powder Bed Fusion Additive Manufacturing.” Journal of Computing and Information Science in Engineering (2024). American Society of Mechanical Engineers.

Abstract: Defect detection is critical in additive manufacturing (AM), as it reduces the proliferation of structurally-unsound components. While X-ray computed tomography (XCT) is costly and only available post-build, in-situ sensors enable machine learning to draw real-time inferences that help steer repairs and process optimization. However, annotating their signatures for supervised learning is challenging due to lack of human interpretability. Automated labeling and homographical registration techniques aim to solve this issue, but fail at establishing one-to-one mappings between in-situ build plate coordinates and their truth labels originating from the XCT ex-situ coordinate frame. Therefore, models that fail to compensate for noisy data pairs result in reduced performance in defect segmentation. In this paper, we propose a 3D convolutional neural network (3D CNN) encoder-decoder architecture to segment layer-wise defects in 2D, at the native in-situ sensor resolution. By considering a volume of adjoining build layers, 3D CNNs can learn spatiotemporal features associated with multi-layer defects and mitigate registration errors, which manifest as coordinate offsets that weaken data pair correlations. Compared against 2D variants, 3D in-situ networks demonstrated robust defect segmentation and reduced the gap between ex-situ networks that ingest perfectly-registered data. Moreover, this approach addresses natural class imbalances in AM processes, where defects (i.e. porosities) account for less than 0.1% of total volume.