Multimodal data fusion for high-speed in situ detection of deformation defects in selective laser sintering

Process-induced defects in the Selective Laser Sintering (SLS) of polymers continue to be common, leading to undesirable mechanical properties, discontinuation of print jobs, and generally complicating part reproducibility. Among these defects, part deformations, such as warping and local geometrical inaccuracies, remain one of the most critical. Non-homogeneous build chamber temperatures, geometry-induced thermal bottlenecks, and the overall temperature sensitivity of the SLS process make deformation defects notoriously difficult to avoid. Lately, in-situ monitoring systems have been gaining attention for their ability to detect process-induced defects by actively monitoring the process. However, existing systems often lack sufficient information and processing speed for meaningful intervention in the printing process. In this work, we advance the state-of-the-art of in-process SLS deformation detection by employing a high-speed multifaceted data-fusion approach. The proposed system combines a high-speed dynamic region-of-interest camera (2 kHz) in the visual spectrum, which observes powder bed related inconsistencies, with a short-wave infrared camera, to observe thermal-related effects. Local geometry effects are accounted for by continuously fusing the camera observations with geometrical features pre-computed from the 3D design file. Furthermore, a temporal sliding window approach ensures defect predictions are not only determined from single observations, but also from longer-evolving process conditions. A custom GPU-accelerated data-driven algorithm further ensures high-speed and accurate inference. The system was validated on two Polyamide 12 printed objects deliberately designed to produce deformation defects. Ground truth deformation targets extracted using XCT served both as training targets for the data-driven algorithm and as validation targets for a separate testing dataset. The experiments demonstrated normalized mean absolute error values of just 12% for detecting deformation magnitudes. The results show promise for the system to be leveraged for high-speed intra-layer closed-loop control of deformation defects, ultimately improving part quality and reproducibility in the SLS process.