Page 21 - ESAM-1-1

Engineering Science in
Additive Manufacturing ML in MAM monitoring and control through images

sensing methods for measuring the thermal-fluid fields respective areas during the printing process are closely
in the melt pool, Lu et al. proposed a physics-based linked to the state of the melt pool. For instance, Repossini
129
compressive sensing algorithm to reconstruct the complete et al. segmented splash images within the laser heating
130
internal temperature and fluid fields of the melt pool based zone during the LPBF process, revealing an unstable
on the surface temperature measurements captured by melt pool alongside intense spatter spatial distribution.
the IR camera. Besides, researchers have combined melt In addition, Yin et al. discovered a strong correlation
131
pool morphology and thermography data for a more between spatter area, quantity, and laser energy density.
comprehensive understanding of melt pool evolution. It was also found that the formation of plumes results
Smoqi et al., for instance, captured the thermal images of from material evaporation and the heating of surrounding
72
the melt pool during the printing process and cut them to gases. Similarly, Ye et al. captured plume and spatter
132
uniform size. Then, they conducted the separation of the images, utilizing DBN to successfully recognize melt pool
melt pool body and ejecta using image processing. Based states. Images depicting the melt pool typically exhibit a
on this, four features, such as melt pool length, mean ejecta heightened dynamic range and intricate details, furnishing
spread, mean ejecta temperature, and distribution of melt an extensive reservoir of information for ML algorithms.
pool temperature, are further extracted to represent the By harnessing ML algorithms, thermodynamic evaluations
physical information of objects. Then, KNN was utilized to can be conducted by leveraging melt pool image data.
deal with these features and assign corresponding labels. Besides, through real-time analysis of melt pool images,
Furthermore, they also evaluated the severity of the defects. ML facilitates operators in monitoring the melt pool’s
On these results, they managed to explore the relationship condition, pre-empting potential issues, and promptly
between input energy density and defects. The framework tweaking manufacturing parameters, thereby streamlining
is shown in Figure 16. and enhancing the MAM process.
In delving further into melt pool characteristics,
researchers have noted that the captured images often 3.2.3. Internal defects
contain impurities, primarily identified as spatters and The analysis of acoustic and spectral image data plays a
plumes. Through the application of CNN methods, it crucial role in detecting internal and surface defects in
was observed that the occurrence of spatters and their the MAM processes, including pores, cracks, and foreign

Figure 15. Melt pool morphology recognition based on YOLO algorithms. Reproduced with permission from Elsevier. Copyright © 2023 The Author(s).
50
Distributed under a CC-BY 4.0 license (https://creativecommons.org/licenses/by/4.0/).

Figure 16. The framework of k-nearest neighbors-based defect prediction

Volume 1 Issue 1 (2025) 15 doi: 10.36922/esam.8548