Page 16 - ESAM-1-1

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

Table 2. In situ signals and monitoring objects in MAM
Types Monitoring object Sensor equipment Pros & cons Ref.
LPBF DED
Vision Melt pool morphology, deposited CCD, CMOS, ICI, line Easily doable and intuitive display, but 46,71,72 49-51
layer geometry, spatter, plume, crack, camera, and 3D camera susceptible to lighting conditions
roughness, and distortion system
Thermal Temperature distribution, spatter, plume, Pyrometer, IR camera, Continuous monitoring and obvious anomaly 72-78 53,79,80
inclusions, pores, and crack near-IR camera, detection, but limited to in-depth material
hyperspectral line analysis
X-ray Melt pool morphology, pores, cracks, lack X-ray machine High penetration depth and detailed internal 81-86 87-89
of fusion, and inclusions structure, but requires specialized equipment
Acoustic/ Pores, phase transformation, inclusions, Acoustic emission High sensitivity to material properties, but 90-94 65,95-97
Spectral crack, and material composition sensor, microphone, and susceptible to environment noise
spectrometer
Abbreviations: CCD: Charge-coupled device; CMOS: Complementary metal-oxide-semiconductor; ICI: Image sensor interface; IR: Infrared;
DED: Direct energy deposition; LPBF: Laser powder bed fusion; MAM: Metal additive manufacturing.

features and defects or quality attributes. Because these model maintain long-term dependencies by deciding how
linkages are intricate, non-linear, and poorly understood, much of the prior hidden state should be carried over to the
ML modeling serves as an efficient method to help build present state. The reset gate then determines how much of
and analyze these intricate relationships. ML techniques the prior hidden state should be disregarded, enabling the
excel at uncovering patterns in data that may not be model to concentrate on fresh data when needed, as shown
apparent through traditional analysis methods, making in Figure 8B. For instance, if the melt pool exhibits unusual
them valuable tools for identifying and understanding the dynamics (e.g., sudden changes in size or temperature), the
underlying factors influencing the quality of manufactured GRU model can flag this as a potential defect. In addition,
components. According to the features of the training data, GRUs can predict future states of the process, enabling
the ML models are typically divided into four categories: proactive control measures. However, supervised learning
supervised, semi-supervised, unsupervised, and RL, as faces challenges such as high data annotation costs, data
seen in Figure 7. imbalance, and image quality variability, which can be
addressed through techniques such as semi-supervised
3.1.1. Supervised learning learning, data augmentation, and image preprocessing.
Common supervised learning algorithms applied for in
situ monitoring in MAM include multilayer perceptron, 3.1.2. Semi-supervised learning
support vector machine (SVM), classification and Semi-supervised learning provides an effective solution to
regression trees, and k-nearest neighbors (KNN), aiming to challenges such as limited labeled data and high annotation
reduce the difference between forecasts and ground truth costs by leveraging a sizable pool of unlabeled data as well
values. Besides, long short-term memory, gate recurrent as a small amount of labeled data. Various techniques
104
unit (GRU), and transformer models come with built-in are utilized to enhance model performance, including
attention mechanisms that can learn important regions in self-training and consistency regularization methods
images or sequence information, helping the model focus like Mean Teacher and Fix Match. In addition, generative
on key parts and enhance performance. 100,101 Specifically, models, such as generative adversarial networks (GANs)
the KNN aims to compare unlabeled data with and variational autoencoders, along with graph-based
102
labeled data existing in the dataset and then extracts the methods, are employed to generate pseudo-labels, improve
classification label of the data (nearest neighbor) with the robustness, and effectively capture relationships within the
closest characteristics in the sample, shown in Figure 8A. data. Specifically, GANs are a potent class of generative
In this way, the numerous images can be classified, models that can be effectively applied for image-based
predicting the defects. As for GRU, it can be used to analyze in situ monitoring of MAM. GANs are made up of two
105
a sequence of thermal images captured during the printing neural networks – a discriminator and a generator – that
process. The gating mechanisms enable it to excel at are trained concurrently and competitively, as shown in
103
capturing critical temporal dependencies, allowing it to Figure 9. The generator creates synthetic images (e.g., melt
retain important information over time while discarding pools, defects) from random noise, while the discriminator
irrelevant data. In particular, the update gate helps the classifies labeled data (e.g., defect types) and distinguishes

Volume 1 Issue 1 (2025) 10 doi: 10.36922/esam.8548