Diagnostics in Additive Manufacturing Using Image-Based Machine Learning

Abstract

Additive Manufacturing (AM) is a vital capability in the aerospace industry. Blue Origin manufactures a substantial share of engine parts via metal AM. To meet growing customer demand, the company must dramatically increase engine throughput and, thus, 3D prints. Blue Origin has identified non-destructive testing (NDT) – particularly, Computed Tomography (CT) scanning – as an unsustainable bottleneck to expanding AM capacity. Not only is this process expensive, but, critically, there are not enough aerospace-grade CT machines in the world to support projected throughput. Without process change, meeting customer demand will soon become impossible. Yet, these scans provide important quality control, and any reduction in NDT must be accompanied by assurances of engine part integrity. This thesis introduces a diagnostic system that safely alleviates the bottleneck, and further yields insights that end-stage NDT alone cannot provide. The proposal is a machine learning system that evaluates the manufacturing process itself, examining layer-by-layer photographs captured during printing. It is predicated on two hypotheses: (1) These images, considered together, provide a synthetic 3D illustration of the build process; and (2) Machines can be taught to assess these process signatures dependably. The resulting system provides rich diagnostics. It achieves near-perfect anomaly recognition – 100% when using conservative defect thresholds. Operationally, the system can (at minimum) safely enable a 37-54% reduction in NDT, translating to millions of dollars in annual cost savings. In practice, this reduction will likely be higher. The system further enables early process intervention and a more data-driven approach to manufacturing intelligence. This work turns what began as an unsustainable bottleneck into an opportunity for enhanced quality control, process intelligence, and long-term manufacturing resilience.