Page 24 - ESAM-1-1

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

crucial information within images. The fine-tuning and model’s generalization proficiency. Effective strategies
192
optimization of models are similarly constrained. encompass enhancing data diversity, refining feature
188
Furthermore, a limited dataset can introduce data skewness extraction processes, mitigating sample deviations,
and imbalance, hindering the model’s ability to effectively anticipating environmental shifts, and conscientiously
capture the intricacies of real-world scenarios. Ultimately, selecting suitable algorithms. These proactive measures
the accuracy of monitoring and control systems may suffer are instrumental in enhancing the model’s generalizability,
due to insufficient data, thereby compromising the system’s thereby facilitating its performance across varied
ability to predict and manage key parameters in MAM applications and environments.
processes. This, in turn, can impact product quality and These challenges underscore the complexity and
189
production efficiency. importance of leveraging ML in image-based MAM
5.1.3. Rapid and accurate response for real-time process monitoring and control. Overcoming these
monitoring and control hurdles while harnessing the potential of ML is crucial
for enhancing process efficiency, ensuring quality, and
Real-time monitoring and controlling data transmission managing operations in MAM processes.
pose pivotal challenges in modern technological landscapes.
The challenges associated with data transmission 5.2. Prospects
encompass bandwidth constraints and latency limitations, In the realm of image-based MAM process monitoring
impacting the real-time functionality of monitoring and control, ML techniques hold significant promise for
systems. Moreover, ensuring data integrity stands out as future advancements. Here are a few potential avenues.
a critical concern, given the potential risks of data loss or
corruption during transmission, which could compromise 5.2.1. Generalized models based on transfer learning
the system’s accuracy. In addition, the processing of The integration of transfer learning into the domain
190
extensive image data presents its own set of challenges, as of MAM offers a multitude of benefits, ranging from
these data typically demand substantial storage space and streamlined model training and enhanced adaptability to
bandwidth, requiring efficient management to minimize improved performance and specialized feature extraction.
transmission delays and system strain. Besides, the This approach not only minimizes the need for extensive
computational complexity of advanced ML algorithms can annotation efforts and computational resources but also
introduce significant latency in decision-making processes, facilitates continuous learning and scalability across diverse
potentially compromising the timeliness of control actions. processes and materials. By leveraging pre-trained models,
The dynamic nature of MAM processes requires ML models transfer learning jumpstarts the learning process for MAM
to adapt quickly to changing conditions, necessitating applications, significantly expediting model development.
robust online learning capabilities that are challenging to A key advantage of transfer learning lies in its ability to
implement effectively. These challenges are compounded harness pre-existing knowledge of general image features,
by the need for precise synchronization between data yielding heightened performance in defect detection,
acquisition, processing, and control actuation, creating process monitoring, and quality control tasks within the
a complex temporal coordination problem that must be MAM domain. The continuous learning capabilities
193
addressed for successful implementation. of transfer learning models empower them to adapt to
evolving MAM processes and materials by periodically
5.1.4. Generalizability of ML model retraining on new data, ensuring their relevance and
The generalization capability of an ML model stands as efficacy in dynamic manufacturing environments. In
a critical challenge. Various factors contribute to this conclusion, transfer learning-based generalized models
challenge, including data diversity, feature extraction present a wealth of advantages for ML applications in
intricacies, sample deviations, environmental fluctuations, image-based MAM process monitoring and control. Their
and algorithm selection. The absence of data diversity adaptability, efficiency, performance enhancements, and
191
can render a model unable to effectively extrapolate to scalability make them a promising approach for bolstering
novel scenarios, while challenges in feature extraction defect detection, process optimization, and quality control
may hinder its capacity to accurately differentiate among in MAM applications.
various feature types. Moreover, sample deviations and
environmental fluctuations can detrimentally impact the 5.2.2. Generative AI for data generation
model’s generalizability, impeding its adaptability in real- Generative AI (GAI) offers a set of opportunities to propel
world settings. Crucially, the selection of an appropriate advancements in image-based MAM process monitoring
ML algorithm plays a pivotal role in determining the and control. One of its key assets lies in the capacity to

Volume 1 Issue 1 (2025) 18 doi: 10.36922/esam.8548