Page 46 - MSAM-4-3

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

Table 2. Number of “good” and “defects” powder bed images
(n=2638)
Classification Number of PB images, n
Good 1108
Defects 1530

RAM, and using Python (version 3.10, Python Software
Figure 1. Example of raw powder bed data. Early-stage defects example Foundation, United Kingdom), PyTorch (version 1.12.1,
(left); Late-stage defects example (right) Meta AI, United States), TensorFlow/Keras (version 2.13.0,
Google LLC, United States), OpenCV (version 4.4.0.40,
but also lacks real-time responsiveness. By integrating OpenCV.org, United States, Numpy (version 1.24.3,
AI-based models into the process monitoring workflow, it NumPy Developers, United States), and Matplotlib
becomes feasible to automatically detect surface defects and (version 3.7.5, Matplotlib Development Team, United
pause or stop the build when necessary. Such automated States) for model training and result visualization. All
intervention can reduce unnecessary time and material experiments were run in the PyCharm Community
loss, thereby improving overall production efficiency. Edition (version 2024.1.2, JetBrains s.r.o., Czech Republic)
To better reflect the practical requirements of quality with Anaconda (version 24.3.0, Anaconda, Inc., United
monitoring in AM, this study defines two task phases, States) to ensure dependency isolation and version control.
aligned with the hierarchical information needs in 2.2. Image classification models
industrial production – namely, binary defect detection
and spatial localization. In the first phase, the model is ResNet and EfficientNet offer multiple versions of
designed to rapidly process layer-wise image data from models, making them well-suited for transfer learning.
the EOS M290 system and determine whether the current To balance computational efficiency with performance,
layer contains potential defects, classifying them as “good” this study experimented with and selected ResNet-50
or “defects.” This phase simulates the industrial demand and EfficientNetV2B0 and implemented task-specific
for fast screening, aiming to identify anomalous layers adaptations tailored to powder bed image characteristics.
in real-time and enable timely intervention to minimize Before training, the image data was pre-processed. First,
material and time waste. ResNet-50 and EfficientNetV2B0 the three datasets were merged to expand the dataset. The
are selected for comparison due to their strong trade-off first 8 – 10 images from each print project, which appear
between accuracy and computational efficiency. ResNet overexposed due to high energy and platform reflection,
improves model depth and convergence stability through were removed, along with the last few “meaningless”
36
residual connections, while EfficientNet balances model images from the end of the print process. This resulted
size, computational cost, and accuracy using compound in a total of 2,638 powder bed images. Data were then
scaling. In the second phase, the model is further manually categorized into “good” and “defects” classes
37
required to localize and highlight defect regions with (Table 2). Finally, the dataset was split into training, test,
bounding boxes, providing spatial information to support and validation sets at a 70:15:15 ratio.
downstream inspection or corrective actions. This phase The default input size for the models was 224 × 224,
addresses the industrial need for defect traceability and while the dataset used in this experiment consisted of high-
region-specific analysis while aiming to reduce manual resolution 1280 × 1024 images. Directly inputting the original
inspection workload. The Faster region-based-CNN size would result in insufficient memory, and significantly
(R-CNN) and YOLO Version 5 (YOLOv5; Ultralytics, downsizing the images would risk losing critical features
United Kingdom) models are compared in this phase. Faster of small defect targets. Therefore, a Python script was used
R-CNN uses a two-stage approach with a region proposal to crop the “defects” images to retain only the defect region
network, offering higher precision in detecting small or and 10% of the surrounding background. To improve the
subtle defects, which benefits complex PBF-LB analysis. model’s generalization, OpenCV was then applied for data
38
In contrast, YOLOv5 performs single-stage detection with augmentation, including normalization, rotation, horizontal
faster inference and better multi-scale capability, making it flipping, scaling, and shear transformations. Finally, all
more suitable for real-time AM monitoring. 39,40 images were resized to 224 × 224 pixels.

All experiments were conducted on a laptop, equipped During transfer learning, multiple adjustments were
with Intel Core i7-12700H, 2.30GHz CPU, and 16GB of made to optimize training performance. First, ResNet-50

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