A training-free machine learning approach for 3D powder bed reconstruction and defect detection in PBF-LB/M

Abstract We introduce a novel in situ quality monitoring approach for the Laser Powder Bed Fusion (PBF-LB/M) process that operates independently of prior experimental data and can be integrated into existing systems. We employ a cost-effective, high-resolution CMOS camera coupled with a bright and darkfield lighting arrangement to acquire data layer-by-layer. Using computer vision techniques, specifically the GrabCut algorithm, we automate the inspection and the virtual reconstruction of fused part geometries. Our results, compared with nominal slice data and computed tomography scan data, indicate a robust accuracy in quantifying geometric deviations (mean deviation of 15.60 $$\upmu$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mi>μ</mml:mi> </mml:math> m and 15.20 $$\upmu$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mi>μ</mml:mi> </mml:math> m respectively, Jaccard scores of 0.99). The proposed training-free framework, unlike traditional machine learning methods, requires no labeled datasets or prior training, offering a cost-effective and adaptable solution. This eliminates the dependency on material, scan strategy, or part geometry, which typically hinders the scalability of conventional approaches. Additionally, our method facilitates automatic defect detection such as recoater strikes, recoater hopping, and powder shortages, with balanced accuracy scores of up to 0.87. These advantages highlight the framework’s potential as a practical tool for in situ process monitoring and quality assurance in PBF-LB/M systems.