PSI - Issue 83

Miloslav Kepka et al. / Procedia Structural Integrity 83 (2026) 138–145

144

3.3 Impact of powder bed defects on a part's integrity The presence of these defects in the powder bed negatively impacts the mechanical properties of the final component. Without in-situ monitoring, it is nearly impossible to detect these internal flaws through post-build visual inspection. Therefore, developing monitoring systems is essential for providing "insight" into the part's integrity without relying solely on costly non-destructive testing. Preliminary experiments within the RTIdigi project have confirmed the presence of internal 'Lack of Fusion' (LOF) defects in printed 18Ni300 maraging steel, specifically in regions where 'wave' anomalies persisted over several consecutive layers. This area will be further explored in subsequent studies.

Fig. 9. Lack of fusion defect due to the presence of wave defects inside the 18Ni300 printed part.

3.4 Monitoring possibilities Advanced monitoring solutions include infrared thermography or optical tomography. These systems monitor process stability via thermal signatures; however, they require expert operation, often lack automated defect detection, and are characterized by high implementation costs. A more cost-effective alternative is the use of standard CMOS cameras. These capture "After Recoating" (AR) and "After Exposure" (AE) images of the build area. While more affordable, the manual analysis of thousands of layers remains inefficient and user-unfriendly without advanced automation. The RTIdigi project focuses on developing a robust, user-friendly automated system. In the initial phase, an algorithm utilizing CNN for binary classification of AR images was developed. The system automatically categorizes layers as either "OK" or "NOK". The model was trained on a dataset of 1,000 OK and 1,000 NOK layer images. Using the ConvNext-Tiny architecture, we achieved a peak accuracy of 96.8% on an independent test set. The limitation of this binary classification lies in the lack of spatial awareness regarding part position and the inability to differentiate among specific defect types.

Fig. 10. Showcase of OK and NOK layers.