PSI - Issue 77

A. Polanský et al. / Procedia Structural Integrity 77 (2026) 529–536 Adam Polanský / Structural Integrity Procedia 00 (2026) 000 – 000

532

4

scratches across the entire width of the building platform in each new layer, see Fig.2 (b). Similar scratches also occur when the recoater blade catches unmelted particles or small debris. However, this usually affects a few layers and does not damage the recoater blade. Edges of parts with specific geometry, typically supports, have tendency to exhibit spots of missing powder, see Fig.2 (c). All these defects are interconnected, and several types of them maybe present in one layer. 4. Methods In the initial experiment, we trained several CNN models to evaluate the quality of the freshly spread layer in the SLM process. The input to the CNN was images of the powder bed after recoating. Subsequently, images were binary classified either as 'OK' or 'NOK' layer based on the presence of any visible defects. Examples of 'OK' and 'NOK' classes are visible in Fig. 3. Input images were captured with the built-in camera, which is part of the Powder bed system in the EOS M290 machine. This system uses one LED panel illuminating the powder bed from the side of the recoating direction, providing better defect visibility in comparison with using top lighting. The original resolution was a 1280x1024 pixel image of the building chamber as visible in Fig. 4 (a). Lately, the following operations were applied automatically by the Powder bed system: camera calibration, cropping to the region of interest, which is a 250 x 250 mm powder bed upscaling to 1000x1000 pixels, and converting to 8-bit color depth. This provides a non-deformed top view of the building platform as visible in Fig.4 (b). The image dataset consisted of 1200 'OK' layers and 1000 'NOK' layers chosen from 110 distinct print recordings of different parts, including 140,000 images. The distribution into training, validation, and testing data is visible in Table 1. For the testing set, unique print records were used to assess model accuracy. Several different types and severities of defects were chosen to identify weaknesses of trained models. Within training and validation data, care was taken to avoid similar images.

Fig. 3. (a- c) examples of CNN input image representing ‘OK’ class (d - f) examples of CNN input image representing ‘NOK’ class

Image augmentations were utilized during a training step of the models to increase model robustness.. The following augmentations were applied: random horizontal and vertical flip, brightness and contrast transformation with up to ±20% intensity, rotation up to ±1.5°, translate and scaling up to ±5%. Each of these transformations was applied during training with probability of 66%. Examples of these transformations are visible in Fig. 5.