Page 22 - ESAM-1-1

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

bodies. ML-based methods have significantly enhanced 4. Process control in MAM
the identification of anomalies within the captured images.
As a pivotal facet of MAM automation, process control is
For instance, Li et al. introduced a multi-scale currently at the forefront of opportunities and challenges,
64
spatially interactive fusion CNN to demonstrate the propelled by advancements in sensing technologies and
efficacy of acoustic feature representation in reflecting data processing capabilities. Various control strategies,
defect information during the LPBF process. Similarly, such as proportional integral derivative control, sliding
Rahman et al. confirmed the reliability of acoustic mode control, predictive control, and adaptive learning
94
images as a monitoring tool through the application of control, have been employed to enable process feedback
the K-means clustering method. Furthermore, Hossain control in MAM. In recent years, the discourse around
et al. leveraged CNNs to process wavelet images for leveraging ML algorithms for process control in MAM has
15
the prediction of potential defects such as cracks and gained significant traction. With their ability to perceive,
keyhole pores. In a different approach, Jayasinghe et al. learn, and evolve on their own, ML techniques are well-
90
employed an auto-regressive time-series model to analyze positioned to be the cornerstone of intelligent control
photodiode images in the LPBF process, successfully systems in MAM. These methods are instrumental in
154
identifying porosity within printed build layers. Acoustic monitoring and regulating a multitude of parameters and
and spectral images also stand out as rich reservoirs of variables throughout the manufacturing process to achieve
data for ML models. These images boast diverse frequency desired outcomes.
ranges and intricate patterns, offering insights into
material properties. Internal defects manifest as abnormal 4.1. Melt pool control
fluctuations within these images, thereby providing Melt pool control stands as a critical component of MAM,
valuable cues for ML algorithms to detect and analyze focusing on the management and optimization of melt
anomalies effectively. pool behavior during the printing process. Achieving the
In the domain of in situ monitoring for MAM, ML required part qualities, such as surface finish, mechanical
stands as a cornerstone. Through the application of strength, and dimensional precision, requires careful
ML algorithms, real-time sensor data can be effectively control of the melt pool.
analyzed, enabling the monitoring and regulation of the For instance, Rezaeifar et al. developed a control
155
manufacturing process. Table 3 provides an overview system to manage melt pool width, reducing surface
of common ML approaches for in situ monitoring in roughness in the LPBF process. Devesse et al. introduced
156
MAM. The integration of ML in MAM encompasses a a CNN-driven control system to adjust the melt pool
wide array of functionalities, including the prediction size by modulating laser power. In DED, continuous
of melt pool characteristics, defect detection, process heat accumulation often results in irregular melt pool
optimization, and quality control. By establishing ML shapes, impacting part properties. Researchers devised
models, it becomes possible to forecast and analyze melt an adaptable controller with layer-dependent gains
pool behaviors, improve the precision of defect detection, to maintain melt pool width in real-time, enhancing
optimize process parameters for enhanced manufacturing microstructure uniformity by adjusting laser power during
efficiency, and enable real-time monitoring and feedback the DED process. Gibson et al. further enhanced
158
157
mechanisms. control by modulating print speed and deposition rate per
layer to manage melt pool size and process stability.
Table 3. ML applications for in situ monitoring in MAM Accurate temperature control of the melt pool and
surrounding area is crucial for managing the solidification
ML architecture LPBF DED rate, reducing residual stress, and ensuring part quality.
CNN 64,91,133-138 49,50,126,38,139 Bernauer et al. utilized a CNN-based approach to correlate
159
SVM 116,140-142 143,144 melt pool temperature with weld bead geometry, adjusting
KNN 72,145,146 96,147,148 printing parameters for stable microstructural properties.
160
Tree algorithms 75,149 143,150,151 Smoqi et al. controlled melt pool temperature through
Physics-informed 72,113,126,129 152,153 laser power adjustment for uniform microstructure with
DBN 46,132 \ reduced porosity.
Abbreviations: CNN: Convolutional neural networks; DBN: Deep 4.2. Deposited layer
belief networks; DED: Direct energy deposition; KNN: K-nearest
neighbors; LPBF: Laser powder bed fusion; MAM: Metal additive Ensuring the quality and uniformity of the powder bed is
manufacturing; ML: Machine learning. paramount for achieving stable melt pool formation and

Volume 1 Issue 1 (2025) 16 doi: 10.36922/esam.8548