Page 44 - MSAM-4-3

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

porosity, and impurity-induced defects, among others. a systematic AM database for ML-assisted AM research.
4-6
In traditional AM processes, product quality is ensured Only a few of the six existing AM research databases contain
by several hours of manual monitoring of the printing image datasets for quality monitoring, and annotated
process, leading to low productivity and increased costs. datasets are almost non-existent. In 2024, Liu et al. also
22
23
7,8
Therefore, improving monitoring methods and ensuring identified the scarcity of image-based datasets in the AM
product quality to enhance reliability has become a key field, thereby warranting more open datasets to facilitate
research focus. 9,10 quality evaluation and defect detection in AM processes. 23
In recent years, with the continuous development and Researchers have proposed various methods to address
maturation of artificial intelligence (AI), machine learning data imbalance and scarcity. 23,24 In the study by Westphal
(ML) has been widely discussed and applied in the field and Seitz, they combined transfer learning with pre-
25
of quality control in AM. 1,11-15 Kadam et al. employed trained VGG16 and Xception models to classify small
16
a series of pre-trained models for feature extraction datasets, with results indicating that VGG16 performed
combined with various ML algorithms to achieve fault best across multiple performance metrics, validating the
detection in the fused deposition modeling (FDM) effectiveness of CNN-based defect detection for non-
process, demonstrating the effectiveness of ML algorithms destructive quality assurance. Szymanik et al. proposed
26
in anomaly detection. In 2024, Herzog et al. reviewed the an innovative approach that combines enhanced signal
1
application of various ML methods and in situ monitoring analyses to improve the accuracy and effectiveness of defect
technologies in metal AM defect detection. They surveyed detection in materials with low thermal conductivity,
50 independent studies published since 2017, most based enabling precise detection of defects in thermal imaging
on computer vision algorithms to classify defective images. and automatic identification of their type and size. Ansari
Among image classification-based detection algorithms, et al. designed artificial implant defects that simulate
27
over 75% used supervised learning methods that detect various shapes and sizes of pores and combined a CNN
anomalous data by labeling the data as “normal” and one model with X-ray tomography data to successfully detect
or a few “anomalies.” These methods reported accurate defects as small as 0.2 mm with up to 97% accuracy,
favorable accuracy rates of approximately 75 – 95%. 1 demonstrating the potential of CNN models in PBF pore
28
Given the effectiveness of AI-based methods, many detection. Kozhay et al. developed a CNN algorithm
researchers aim to build AI-based automated quality for detecting and classifying defects in FDM-printed
7
monitoring systems. 17,18 Khan et al. developed a real-time images, achieving 90% accuracy. However, the system’s
defect detection system based on a convolutional neural performance outside specific layouts was limited due to
network (CNN) model to improve the automation of fused insufficient data, highlighting the importance of defect
28
filament fabrication and reduce production losses and data diversity for algorithm optimization.
labor costs. Although the model achieved an accuracy of In previous research, CNN models have been widely
84% in identifying geometric anomalies, it struggled to recognized as the most suitable AI model for defect image
detect vertical plane defects, and the insufficient dataset led classification and detection tasks in the AM field. 12,29 Fu
to inconsistencies in the model. Conversely, Cannizzaro et al. systematically summarized ML algorithms for defect
7
29
et al. introduced a real-time defect monitoring and detection in PBF-LB processes, noting that supervised
19
detection system for metal powder bed fusion (PBF) learning is the most commonly used ML method. It
that classifies five different types of powder bed defects performs classification and regression by learning the
and monitors the profile of each printed layer. They also relationship between input and output in labeled datasets,
explored using generative adversarial networks to generate making it widely applied in PBF-LB systems for defect
synthetic images of powder bed defects to address the detection and classification. CNNs extract features
19
scarce labeled data for training and testing ML models. through convolution layers, reduce dimensions through
Tamir et al. proposed the integration of digital twins and pooling layers, and output through fully connected
11
parallel systems into AM for real-time process monitoring layers, showcasing excellent image processing and pattern
to optimize intelligent manufacturing processes through recognition capabilities. CNNs are the most popular
the development of virtual models. deep learning algorithms for image recognition, image
The diversity and size of the dataset directly impact the classification, and object detection, making them well-
performance of ML models, and high-quality datasets are suited for defect detection in Laser-Based AM processes. 29
30
crucial for improving prediction and detection accuracy. As early as 2019, Han et al. proposed a deep CNN-
Nonetheless, researchers often face challenges in finding based AI technique to effectively monitor surface quality
suitable target datasets. 20,21 Zhang et al. emphasized creating and detect defects by analyzing metal fracture micrographs
22

Volume 4 Issue 3 (2025) 2 doi: 10.36922/MSAM025150022