Additive Manufacturing (AM), i.e. 3D printing, of metal parts is growing in popularity, but part quality can still be a limiting factor to that growth. Camera-based monitoring systems improve quality by detecting defects on-the-fly, but it often relies on incomplete handcrafted video features, or hard to interpret features derived by data-driven techniques. In this work, we propose a method based on variational autoencoders (VAEs) that produces highly informative and interpretable features for in-situ AM monitoring. Unlike handcrafted features, our technique is data-driven so that it captures all details. Unlike other data-driven methods, our technique produces features that are highly interpretable and correlate to the involved physics in the printing process. To test our technique, an object is printed with deliberate non-nominal layers and recorded at high speed with a monochrome optical camera. The video frames are then used to train a VAE as feature extractor. The VAE-computed features are then input into a classification algorithm to detect print deviations. It is shown that our proposed features outperform state-of-the-art deep learning methods, and handcrafted features, with up to 2.16% improvement, all while preserving feature interpretability, the completeness of data-driven feature extraction, and high computation speed (> 5 kHz).

In selective laser melting (SLM), the 3D-printed metal objects can deform, causing dimensional inaccuracies and potentially damaging the printer's recoater. While layerwise monitoring systems are available to detect these deformations, higher-speed detection would enable a controller to intervene to reduce this defect. In this work we propose the first such in-situ monitoring system that detects part deformations through the analysis of melt pool images. The system uses high-speed cameras with frame rates up to 20,000 frames per second in the visual and short-wave infrared spectrum to record the melt pool. An adaptive weight network is then used to predict deformation magnitudes based on the image data and contextual printing information such as part geometry and hatching patterns. The addition of this contextual information allows the model to normalize for within-layer printing variations when interpreting the image data. Our model was validated on three print ob-jects that display deformations resulting from different physical sources. Our model's predictions were compared to deformation measurements taken from X-ray CT scans and showed both a strong correlation with the ground truth, with average relative errors between 17 % and 22%, and generalized well to untrained deformation types. The prediction model is also able to operate at very high speeds (> 20 kHz) to facilitate high-speed intra-layer control loops.

Laser powder bed fusion is at the forefront of manufacturing metallic objects, particularly those with complex geometries or those produced in limited quantities. However, this 3D printing method is susceptible to several printing defects due to the complexities of using a high-power laser with ultra-fast actuation. Accurate online print defect detection is therefore in high demand, and this defect detection must maintain a low computational profile to enable low-latency process intervention. In this work, we propose a low-latency LPBF defect detection algorithm based on fusion of images from high-speed cameras in the visible and short-wave infrared (SWIR) spectrum ranges. First, we design an experiment to print an object while both imposing porosity defects on the print, and recording the laser's melt pool with the high-speed cameras. We then train variational autoencoders on images from both cameras to extract and fuse two sets of corresponding features. The melt pool recordings are then annotated with pore densities extracted from the printed object's CT scan. These annotations are then used to train and evaluate the ability of a fast neural network model to predict the occurrence of porosity from the fused features. We compare the prediction performance of our sensor fused model with models trained on image features from each camera separately. We observe that the SWIR imaging is sensitive to keyhole porosity while the visible-range optical camera is sensitive to lack-of-fusion porosity. By fusing features from both cameras, we are able to accurately predict both pore types, thus outperforming both single camera systems.

The manufacturing of metal parts via powder-bed fusion is often still facing quality issues due to microstructural porosity. Minimizing this porosity remains a priority and requires the optimization of printing process parameters. While the analysis of printed parts using X-ray computed tomography can localize and identify the pore types (e.g. keyhole or lack-of-fusion pores), these pore types can be difficult to identify across printer settings and print materials. Therefore, there is a need for a material and process agnostic approach. This work presents such an approach by considering a set of geometric pore features that do not differ considerably across print scenarios. These features are then leveraged for supervised pore type classification. The distributions of pore features were analyzed in different materials and under varying laser parameters, showing that they behave in a generic way. For classification, it is observed that they outperform other features leveraged in the state-of-the-art for pore classification in a single material, reaching up to 93.0% accuracy. Additionally, accuracies up to 90.2% for cross-material classification were observed by training on pores of one material and validating on another. These results pave the way to a general-purpose pore classification method usable across materials and process conditions.