Page 40 - ESAM-1-4

Engineering Science in
Additive Manufacturing Machine learning for biomedical metal AM

inconsistencies in powder materials, and resulting defects extremely high spatio-temporal resolution, including
(such as porosity, lack of fusion, and cracks) remain major brightness, size, geometry, and fluctuations (Figure 13B).
obstacles to its widespread industrial adoption. Traditional This visual information provides an intuitive basis for
methods heavily rely on post-printing destructive testing assessing molten pool stability. 130-133 The molten pool’s
or CT, which are not only costly and inefficient but, more dimensions, circularity, and fluctuations reflect energy
critically, incapable of intervening in or repairing defects input and melting conditions, serving as crucial precursors
that arise during the manufacturing process. 122 to forming instability and defect generation. Yang et al.
134
At present, ML technology is driving a fundamental employed a high-speed camera sensing method to capture
paradigm shift in AM quality control. The core lies in dynamic image sequences of the molten pool, spatter,
integrating ML to fully leverage the vast data generated and other phenomena. The key features of the molten
during manufacturing. This shifts quality control from a pool under ultrasonic disturbance were extracted and
passive, post-inspection approach to a proactive, in-process reconstructed by integrating a fully connected layer with a
135
prevention, and intervention model. It transitions from a convolutional autoencoder approach. Mi et al. similarly
passive, offline, sampling-based system to an active, online employ the cameras to capture sequential images during
intelligent monitoring system. This reduces reliance on the DED process. Further utilization of deep CNNs
human expertise, enabling real-time quality monitoring, enabled precise segmentation of the molten pool contour
early defect alerts, and even autonomous process and minute spatter particles, achieving an accuracy rate
optimization. The shift aims to make every biomedical of 94.71%. Despite being susceptible to interference from
123
metal printing process transparent, controllable, and intense light, visual sensing has become a vital tool for
reliable, thereby ensuring that products possess high detecting surface defects and geometric deviations due to
reliability and batch-to-batch consistency. 124 its advantages of being information-rich and highly real-
time.
4.2. Sensing technology for AM monitoring
Acoustic and spectral sensing: Microphones or AE
In the field of AM, the integrated application of multiple sensors can collect acoustic signals and stress waves
monitoring methods enables comprehensive perception of generated during printing by plasma, spatter, material
process states. Different sensor types capture information phase transitions, and even microcracks, thereby
from distinct dimensions of the physical process, providing sensitively indicating the formation of internal defects 136-138
the basis for training and inference of ML models. 125 (Figure 13C). Rahman et al. employed AE sensing
139
Thermal imaging sensing: Typically, thermal cameras combined with the K-means clustering algorithm to
operate on line-scan or push-broom principle, where a achieve continuous in situ monitoring of multi-layer
single imaged line is optically dispersed onto a 2D sensor deposition states during the WAAM process, verifying the
to simultaneously record one spatial and one spectral consistency of acoustic signals in identifying process states
dimension, with full-field temperature mapping achieved under varying material and process conditions. Spectral
through subsequent scanning of this line across the target sensors capture spectral information with high sensitivity
140
(Figure 13A). 123,126-128 Thermal imaging cameras provide and rich content. For instance, Montazeri et al. employed
direct records of process thermal history by monitoring a multispectral optical emission sensor, combined with
temperature field distributions across the molten pool Fourier transform imaging and maximum likelihood
and its heat affected zone. Temperature history and estimation, to achieve real-time monitoring of pore defects
gradients serve as critical indicators for material phase during the LB-PBF process.
transformations, residual stresses, and defect formation. Different sensors have distinct strengths (Table 2); for
Abnormal cooling rates or localized overheating can lead instance, acoustic and spectral sensors are more sensitive
to undesirable phase changes and are also closely linked to internal defects such as porosity and cracks; thermal
to defects such as porosity and hot cracks. Liu et al. sensors precisely reflect energy input and thermal history;
129
employed infrared thermal imaging cameras to capture while visual sensors inherently excel at capturing surface
real-time thermal image sequences during large-scale AM defects and geometric deviations. Thus, multi-sensor data
processes. By integrating a CNN long short-term memory fusion has become an inevitable trend. Fusing multi-sensor
model, they achieved temporal prediction of future data overcomes the limitations of single data sources by
temperature distributions, enabling early identification integrating heterogeneous data into high-dimensional
and early warning of abnormal temperatures. feature vectors. This provides ML models with a more
Visual sensing: High speed cameras can capture comprehensive and complementary information view,
the morphological dynamics of the molten pool with enabling more precise defect diagnosis. Gaikwad et al.
142
141
Volume 1 Issue 4 (2025) 18 doi: 10.36922/ESAM025440031