Page 45 - MSAM-4-3

Materials Science in Additive Manufacturing AI-driven defect detection in metal AM

and metallographic images in industrial manufacturing. assessment. The workflow includes layer-wise high-
The experimental results demonstrated that a learning resolution image acquisition, data pre-processing and
rate of 0.01, Adam optimizer, and Inception-ResNet-v2 annotation, and model optimization through transfer
network achieved the best classification accuracy for learning and model evaluation. Results demonstrate that
31
30
image classification tasks in AM. Khan et al. used the proposed object detection approach enables accurate
optical tomography data, monitored layer by layer, and and efficient defect localization, significantly reducing
combined random forest algorithms to detect porosity manual effort and improving build success rates, thus laying
and lack of fusion defects. Although an accuracy of the foundation for smarter real-time monitoring in AM.
99.98% was achieved, the risk of overfitting existed due
to the small training dataset of only 100 images. Their 2. Methods
study further proposed that more efficient detection 2.1. Data collection and model selection
could be achieved with CNNs and larger-scale datasets.
31
Abhilash and Ahmed used the ResNet50 CNN model to The experiment collected three sets of images from
32
classify AM component surface conditions to determine EOSTATE PowderBed (EOS Gmbh, Germany), which
polishing conditions and optimized surface quality while consists of a high-resolution (2D bitmap data resolution
eliminating defects, achieving a prediction accuracy of of 300 – 400 µm) visual wavelength camera on the EOS
96% and significantly reducing manual inspection and M290 metal machine. The EOSTATE PowderBed captured
material waste. Lee et al. proposed a local detection layer images of the printing builds, including defects
33
method based on 3D-CNN to perform defect classification (Table 1). The sets contained 1383, 287, and 1016 images
and local volume fraction prediction, utilizing data from taken layer-by-layer before and after powder bed melting.
the melt pool monitoring system and micro-computed The feedstock used for test samples was 316L stainless
tomography. This method could classify porosities caused steel. The powder size distribution was Dv(25) = 22 µm,
by the lack of fusion and locked pores in the PBF-LB Dv(50) = 37 µm, and Dv(75) = 58 µm, measured using
35
process. Wen et al. combined simple CNN, YOLOv4, laser diffraction. The recoating speed was set at 80 mm/s.
34
33
and Detectron2 to perform classification, object detection, AM defects differ from typical object detection targets,
and segmentation tasks on scanning electron microscopy as they do not have uniform or consistent contours.
images of AM parts, achieving excellent accuracy and Instead, they are abstract, mixed, and evolve gradually
defect detection capabilities in static and dynamic videos. throughout the printing process. For example, Figure 1
These studies demonstrate the significant potential presents grooves formed by debris dragging across the
of CNNs for defect detection in AM, warranting further surface, scattered powder, and pits or voids. In the early
validation of the large-scale image dataset constructed in stages, slight surface irregularities or powder residues
this work. Previous approaches have primarily focused on caused by recoater vibrations may have low contrast against
binary classification – determining whether defects exist – the background. However, if these defects are not detected
without providing spatial localization, which increases the and addressed in time, they can escalate into more severe
burden on manual inspection. The demand for agile, real- issues. The model is expected to identify these subtle defect
time monitoring also exposes the limitations of pixel-level patterns early.
segmentation in practical applications. This study improves In practical industrial settings, the printing of a single
detection accuracy based on classification models and AM part can take up to dozens of hours, depending on its
further incorporates advanced object detection techniques size and complexity. On a typical industrial PBF-LB system,
to enable both identification and localization of defects in such as the EOS M290, printing one layer alone generally
metal 3D printing, enhancing model interpretability and takes between 10 and 60 s. Since, it is difficult to predict
applicability. All models are trained and systematically when and where a defect may occur during the process,
compared on a unified, large-scale dataset to ensure operators are often required to monitor the build frequently.
generalizability and provide a comprehensive performance This manual monitoring is not only time-consuming

Table 1. Printing parameters

Set Laser power (W) Scan speed (mm/s) Hatch distance (mm) Layer thickness (µm) Volumetric energy density (J/mm )
3
1 80 – 250 780 – 3120 0.09 40 8.89, 17.81, 26.71, 35.61
2 215 1083 0.09 40 55.15
3 195 1083 0.09 20 100.03

Volume 4 Issue 3 (2025) 3 doi: 10.36922/MSAM025150022