Page 42 - ESAM-1-4

Engineering Science in
Additive Manufacturing Machine learning for biomedical metal AM

fused heterogeneous sensor data from a co-axial pyrometer state to be automatically classified. These models can learn
and a high-speed video camera and developed a sequential defect features directly from image data, enabling the
decision analysis neural network model to achieve accurate current state to be automatically classified. For instance,
prediction and assessment of the geometric quality of Scime et al. addressed powder bed anomalies during
144
single tracks in LB-PBF. the LB-PBF process by constructing a multi-scale CNN
using an enhanced AlexNet architecture to automatically
4.3. Application of ML methods in real-time defect analyze grayscale images of the powder bed after powder
detection spreading. This model autonomously learned and achieved

In the AM process, ML provides a powerful technical high-precision recognition and classification of six types of
means for achieving in situ process monitoring and quality powder bed anomalies, achieving an overall classification
assessment by analyzing real-time multi-source sensor accuracy of 97% while pinpointing defects with pixel-level
145
data. Based on the characteristics of data and challenges precision. Similarly, in the L-DED process, Chen et al.
encountered in actual production, two primary technical developed a multi-sensor fusion digital twin framework
approaches have emerged: precise identification and based on supervised learning strategies. This model enables
localization of known defects, and acute detection of real-time identification and localization of known defect
unknown anomalies when labeled samples are scarce. types such as cracks and critical-pore voids, subsequently
generating a virtual quality map registered with the part’s
Supervised learning methods play a crucial role when
common defect types in manufacturing processes, such as 3D volume, achieving a defect classification accuracy of
96%. It significantly outperforms single-sensor approaches,
spatter, spheroidization, and poor powder distribution, are thereby demonstrating the effectiveness and robustness of
well understood and sufficient annotated data have been supervised methods in multimodal fusion scenarios.
accumulated. For in situ quality monitoring of LB-PBF,
Knaak et al. developed a technique that marries high However, a more prevalent challenge in industrial
143
dynamic range optical imaging with CNN (Figure 14). settings lies in the diverse and unpredictable nature of
This synergy provides high-spatial-resolution capabilities anomalies, making it extremely costly or even impractical to
and facilitates layer-by-layer prediction of surface obtain large quantities of labeled defect samples. Faced with
roughness, thereby enabling real-time quality assessment this challenge, unsupervised or semi-supervised anomaly
and process optimization. These models can learn defect detection methods demonstrate unique value. Their core
features directly from image data, enabling the current principle is not to directly identify specific defects, but

Figure 14. Framework for layer-wise monitoring and optimization of laser powder bed fusion processes. 143
Abbreviations: CNN: Convolutional neural network; RF: Random forest; RL: Reinforcement learning; LPBF: Laser powder bed fusion.

Volume 1 Issue 4 (2025) 20 doi: 10.36922/ESAM025440031