Page 18 - ESAM-1-1

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

Unlike supervised learning, unsupervised learning is deep Q-networks (DQN), which handles high-
107
identifies hidden patterns, structures, or anomalies in the dimensional state spaces by fusing Q-learning and deep
data by clustering, dimensionality reduction, or generative neural networks. In DQN, a CNN estimates the expected
modeling. One effective unsupervised learning technique cumulative reward for performing a particular action
is the DBN, a hierarchical generative model composed in a particular condition, as determined by the Q-value
106
of multiple layers of stochastic latent variables. DBNs function. The CNN processes raw image data (e.g., melt
are built using stacked restricted Boltzmann machines pool or layer images) to extract features, which are then
(RBMs), where each RBM layer learns to capture used to predict Q-values for all possible actions. While
increasingly abstract features from the input data, as shown the experience is recorded in a replay buffer, the agent
in Figure 10. In the context of MAM monitoring, a DBN investigates the environment during training by acting and
processes raw image data (e.g., melt pool or layer images) observing rewards. The temporal difference error between
through its layers, with each layer extracting higher-level the target and predicted Q-values is minimized to train the
features such as edges, textures, or patterns associated with DQN. In MAM monitoring, the DQN agent can learn to
defects or process anomalies. During the unsupervised adjust process parameters in real time to minimize defects
pre-training phase, the DBN learns to reconstruct the input or optimize melt pool stability based on the observed
data by minimizing the energy function of each RBM layer, image data. For example, if the agent detects irregularities
allowing it to simulate the data’s underlying distribution. in the melt pool, it can dynamically adjust the laser power
Once pre-trained, the DBN can be fine-tuned using a or scan speed to correct the issue. It is suitable for real-
small amount of labeled data for specific tasks like defect time MAM process control, where actions are machine-
detection or process state classification. For example, the setting adjustments based on the agent’s interactions with
learned features can be used to identify anomalies in melt its environment.
pool behavior or layer deposition by detecting deviations Traditional ML methods in MAM face challenges,
from the normal data distribution. such as the need for extensive data and a lack of physical
interpretability. Physics-informed ML (PIML) techniques
3.1.4. RL
have been developed to solve these issues, leveraging both
RL trains algorithms through a reward-and-penalty physical principles and statistical correlations between
mechanism for sequential decision-making. In this data to enhance interpretability, reduce spuriousness,
framework, an RL agent observes the state of the and incorporate domain-specific knowledge into ML
manufacturing process (e.g., melt pool images or layer- algorithms. Some available PIML algorithms include but
wise deposition patterns) and takes actions (e.g., adjusting are not limited to physics-informed neural network,
108
laser power, scan speed, or other process parameters) to physics-informed fully convolutional networks,
109
maximize a reward signal, which reflects the quality of physics-based compressive sensing, 110-112 and physics-
the manufactured part or the efficiency of the process. constrained dictionary learning. 113-115 By incorporating
As shown in Figure 11, one prominent RL algorithm physical principles into the ML models, the reliability and

Figure 10. The framework of deep belief networks.

Volume 1 Issue 1 (2025) 12 doi: 10.36922/esam.8548