A high-speed multi-modal in-situ monitoring system for targeting multiple defect types in selective laser sintering

Selective laser sintering (SLS) of polymers still faces quality issues due to process-induced defects. Microstructural porosity and part deformations remain among the most critical defects, causing undesirable mechanical properties, print discontinuations, and complicating part reproducibility. These problems limit the practical applicability of SLS and necessitate the use of post-process quality inspections, increasing costs. In-situ monitoring systems, in which sensors record and analyze the printing process, are part of the solution. However, the current state-of-the-art methods primarily focus on post- print defect detection. These methods lack the high-speed requirements to enable meaningful intra-layer control and often target only single defect types. While this helps reduce post-process quality inspection, it fails to address the core challenge: preventing defects. In this work, we address these limitations by implementing a high-speed in-situ monitoring system, operating at up to 2 kHz, that detects both porosity and deformation defects, thereby enabling real-time corrective measures. Our approach uses dynamic region-of-interest high-speed cameras in both the visual and short-wave infrared spectrum, combined with a GPU optimized image processing pipeline, to facilitate the high throughput requirements. The processing pipeline is based on machine learning methods that process the image data using general-purpose feature encoders, enhanced by local printing context and temporal information, and can readily be extended to additional defect types. To validate its effectiveness, six cylindrical parts were printed in Polyamide 12 and recorded by the proposed system. The prints were designed to deliberately produce different types of porosity and deformation defects. X-ray CT analysis then provided ground truth defect values to train and test the signal processing pipeline. The experiments demonstrated normalized mean absolute error values of 14% for predicting porosity and 11% for predicting deformations while operating at the required speeds. These results show promise for the system to be effectively used for high-speed, intra-layer closed-loop control, significantly improving the SLS process.