Page 25 - ESAM-1-1

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

enrich existing MAM datasets through the creation of the issues of low robustness and repeatability in MAM.
synthetic images that closely emulate real-world scenarios. This highlights the importance of advancing both
This data augmentation strategy elevates the quality of areas concurrently to drive the automation of MAM
training data, furnishing ML models with a broader technologies. These advancements have substantial
spectrum of examples to glean insights from, particularly research implications and are fundamental in overcoming
in scenarios where genuine data is scarce. Moreover, GAI the industry-specific limitations that presently hinder the
plays a pivotal role in anomaly detection within MAM broad adoption of MAM technologies.
processes by constructing a baseline model of normal This work offers a thorough review of earlier investigations
images. Any deviations from this established norm can into in situ monitoring and process control in the context
be swiftly identified as potential anomalies, facilitating
the early detection of defects, irregularities, or deviations of MAM. The examination delves into the characteristics of
from standard manufacturing practices. In cases where diverse in situ imaging systems utilized in MAM processes.
datasets are imbalanced, GAI can generate synthetic data Furthermore, it clarifies the applications of ML techniques
for underrepresented classes, effectively harmonizing the in feature extraction and correlation establishment based on
dataset and amplifying model efficacy. This methodology a range of image types. In addition, the evolution of process
194
ensures that ML models are honed using a more inclusive control mechanisms within MAM is meticulously reviewed.
and diversified dataset, culminating in heightened accuracy Finally, the review also discusses challenges and outlines
and reliability in anomaly detection. Furthermore, GAI can future directions concerning the utilization of ML in image-
mimic process variability by generating images that mirror based MAM process monitoring and control.
diverse process parameters or material characteristics. By Acknowledgments
integrating synthetic data into existing datasets, models
can undergo comprehensive training, resulting in superior None.
model performance and generalization. In addition, by
integrating GAI-generated data with transfer learning, Funding
models can be fine-tuned for specific MAM processes or This work was supported in part by the University Grants
materials. This hybrid approach facilitates the seamless Committee of Hong Kong (No. RMGS24EG04) and
adaptation of pre-trained models to novel MAM domains, HKUST-Bright Dream Robotics Joint Research Institute
harnessing the combined strengths of synthetic and (No. OKT24EG05).
authentic data to enhance model performance significantly.
Conflict of interest
5.2.3. Hybrid approaches combining physics-informed
modeling and ML Yanglong Lu is an Editorial Board Member of this journal but
was not in any way involved in the editorial and peer-review
Hybrid approaches that integrate physics-based modeling process conducted for this paper, directly or indirectly.
with ML techniques offer a robust framework for MAM Separately, other authors declared that they have no known
process monitoring. By combining the strengths of both
paradigms, these approaches can provide more accurate competing financial interests or personal relationships that
predictions and better control over the manufacturing could have influenced the work reported in this paper.
process. Future research should focus on developing Author contributions
standardized datasets, lightweight algorithms, and hybrid
methodologies to address computational overheads and Conceptualization: All authors
ensure data quality. Visualization: Jian Wang
Writing – original draft: Jian Wang
Nevertheless, formidable hurdles such as data quality Writing – review & editing: Yanglong Lu, Xin Zhang
assurance, computational overheads, and the imperative
of robust training datasets demand immediate attention. Ethics approval and consent to participate
Future research endeavors should pivot towards
the cultivation of standardized datasets, lightweight Not applicable.
algorithms, and hybrid methodologies that fuse physics- Consent for publication
based modeling with ML paradigms.
Not applicable.
6. Conclusions

Monitoring and controlling processes are vital elements Availability of data
within closed-loop systems that are crucial for tackling Not applicable.

Volume 1 Issue 1 (2025) 19 doi: 10.36922/esam.8548