Page 15 - ESAM-1-1

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

imaging system is required to capture and analyze the deposited materials. Spectral imaging is valuable for
emitted spectral signatures. The setup incorporates a material composition analysis and process monitoring.
light source to uniformly illuminate the sample, ensuring Integrating different imaging methods allows for a more
consistent spectral data acquisition. In the monitoring comprehensive understanding of the printing process,
process, it is necessary to keep accurate positioning within improving quality control and overall efficiency in MAM.
the spectrometer’s field of view. As the sample is illuminated, A summary of the monitoring objects targeted by various
spectral data are collected across the designated wavelength imaging methods in MAM is provided in Table 2.
range. Then, the collected data is processed using specialized
software to create spectral images offering insights into the 3. ML for image-based in situ monitoring
material’s chemical composition, structural properties, In monitoring MAM processes, extensive image data
and any other occurring changes. For instance, Lough et can be produced that offers valuable insights into the
68
al. utilized OES to record spectral information of light in fabrication process. Leveraging ML techniques allows for
LPBF printing. They examined the relative intensities and the extraction of meaningful information and features from
chemistry of excited species and methodically explored the these images. This capability improves MAM processes by
relationships between process parameters and spectrum analyzing image data in real time to oversee the printing
features. Similarly, in the resistance DED method, Liu process and identify anomalies, defects, or deviations from
et al. investigated the connection between spectral features the desired outcomes.
69
and deposition stability. Their study revealed a significant
increase in productivity when applying hot-wire voltage. 3.1. ML models
OES, as an in situ technique, demonstrates substantial In situ monitoring serves the purpose of collecting process-
potential in characterizing internal defects in MAM. relevant data simultaneously during fabrication, enabling
The Fourier transformation of spectral signals in LPBF the determination of the component’s state throughout
and the use of tomography detection to measure pores production. ML methods have emerged as a potent tool for
were explained by Montazeri et al., and 90% prediction analyzing image-based data to detect defects and optimize
70
accuracy was attained by training an ML model for pore the manufacturing process in real time. This is due to ML’s
prediction using Fourier transform coefficients as input. ability to uncover concealed patterns in multimodal and
The experimental setup and procedure for spectral- high-dimensional data. When integrating ML algorithms
based in situ monitoring of MAM necessitate specialized with captured image data for feature extraction and
equipment like a spectrometer and appropriate lighting process optimization, addressing labeling errors, such as
sources. Although the technical demands for spectral false or missing labels, is paramount as they significantly
imaging are more advanced compared to some imaging impact the model’s accuracy and reliability. High-quality
techniques, the setup is generally manageable in a laboratory feature extraction and selection from image-based data
setting with the requisite equipment and expertise. are essential to ensure the selection of all relevant features
98,99
Reproducibility hinges on factors such as spectrometer without redundancy. The integration process typically
calibration, consistency in sample preparation, and precise commences with preprocessing the captured image data
data processing. Challenges may emerge in optimizing to enhance its quality and prepare it for feature extraction.
spectral data acquisition, deciphering intricate spectral This may involve tasks like noise reduction, normalization,
information, and ensuring image reliability. and resizing to maintain dataset consistency. Following
preprocessing, ML algorithms are applied for feature
It is evident from the references provided that substantial extraction. CNNs are commonly employed for this task
advancements have been made in MAM’s in situ sensing due to their proficiency in learning hierarchical features
tools and techniques. Numerous imaging techniques from images. CNNs can automatically extract features at
have been employed to mine crucial information for various levels of abstraction, capturing pertinent patterns
monitoring and enhancing the quality of the printing within the image data. Feature selection constitutes
process. In summary, visual imaging is commonly used for another pivotal step in the integration process. This step
extracting morphological features like surface defects and involves selecting the most informative features from the
dimension deviations. By tracking the thermal behavior extracted set while discarding redundant or irrelevant
of components both locally and in bulk, thermal imaging ones. Techniques like principal component analysis
makes it possible to extract geometric characteristics (PCA) or recursive feature elimination (RFE) can be
from temperature distributions. X-ray imaging can detect employed to fine-tune the feature selection process and
internal defects and visualize melt pool dynamics. Acoustic enhance the model’s performance. The subsequent step
imaging can be used to track the interior conditions of involves establishing a correlation between the extracted

Volume 1 Issue 1 (2025) 9 doi: 10.36922/esam.8548