Page 43 - ESAM-1-4

Engineering Science in
Additive Manufacturing Machine learning for biomedical metal AM

rather to enable the model to learn and memorize the data process is deviating from normal conditions, or when ML
patterns characteristic of qualified states through extensive models predict impending defects, the system instantly
normal process data. Once real-time sensor signals deviate generates commands to dynamically adjust process
significantly from this learned normal pattern, the system parameters (e.g., laser power and scan speed), thus steering
automatically triggers an anomaly alert. Models like the process back on track and preventing defects. Within
150
autoencoders learn intrinsic patterns and distributions such a closed-loop framework, ML plays an indispensable
from large amounts of normal data, enabling them to role: it enables accurate, real-time diagnosis, and prediction
identify anomalous signals deviating from these patterns. of process states and quality indicators by automatically
146
Abranovic et al. employed a convolutional long short- extracting features from multi-sensor data and establishing
147
term memory autoencoder (ConvLSTM Autoencoder). high-dimensional, non-linear mappings. This capability
151
Trained unsupervised using only video sequences of to create a quantitative, intelligent process-quality mapping
molten pools under normal operating conditions, the provides the foundational insight necessary for precise
model quantified process stability through its ability to parameter adjustments.
predict future frames. This approach successfully achieved Effective parameter tuning is central to realizing such
online detection and localization of four typical defects closed-loop control, and a variety of practical approaches
without requiring any annotated defect data. have been developed. Gerdes et al. predesigned and
123
Therefore, ML-based real-time diagnosis lays the core implemented a programmed feedforward control strategy
foundation for achieving closed-loop control, transitioning involving an automatic increase in print speed at the
152
from sensing to decision-making and execution. midpoint of the print path. Armstrong et al. employed 3D
148
The ultimate goal of an intelligent monitoring system laser scanning between layers to detect geometric errors,
is not merely to identify defects but also to dynamically thereby adjusting subsequent layer paths and parameters.
intervene in the process parameters by integrating ML’s Both of these methods are effective ways to improve the
decision-making capabilities with the actuators of the geometric accuracy and consistency of printed structures
AM equipment, thereby proactively suppressing defect during the AM process.
formation. This facilitates a fundamental shift in the The above-mentioned control models of parameter
quality control paradigm from post-inspection toward tuning establish the practical foundation for closed-loop
online prediction and adaptive manufacturing. control. To achieve higher level autonomous optimization,
strategies such as model predictive control (MPC) are
4.4. Closed-loop control and adaptive widely adopted. This approach utilizes dynamic models to
manufacturing
predict future behavior and optimize parameter adjustment
The ultimate objective of quality monitoring is to integrate decisions. A core challenge in this field is obtaining accurate
the sensing and decision-making capabilities of ML with and efficient dynamic system models, with the technical
the actuators of AM to create an intelligent closed-loop approach evolving from traditional methods toward data-
control system. The core mechanism operates as follows driven techniques. Early research relied primarily on system
149
(Figure 15): when online sensor data indicates that the identification-based linear MPC. For example, Liu et al.
153

Figure 15. Schematic diagram of the intelligent closed-loop control system. Image created by the authors.
Abbreviation: ML: Machine learning.

Volume 1 Issue 4 (2025) 21 doi: 10.36922/ESAM025440031