Page 90 - MSAM-3-4

Materials Science in Additive Manufacturing Super-resolution method for L-PBF

SR. The entire reconstruction process consists of three (1) A chain network based on the attention mechanism
28
parts based on convolutional neural networks (CNN) is designed for extracting high-frequency information
and has shown better reconstruction results compared to from melt pool images. It considers the characteristics
traditional methods. DL models can autonomously learn of simple background and important boundary
the high-dimensional mapping relationship between information in the melt pool image, preserves low-
LR and SR images, thereby achieving better SR effects. level features containing important spatial information
Consequently, more and more achievements are being through residual-in-residual (RIR) structure, and
made in this field, such as Fast SRCNN (FSRCNN), Very reassigns weights through the attention mechanism.
29
31
30
Deep SR network, Enhanced Deep SR network (EDSR), (2) Furthermore, unlike the block training of the
and residual channel attention network (RCAN). Many traditional SR algorithms, the melt pool image is fully
32
excellent network structures suitable for SR have also been input into the network for training. Therefore, a U-Net
proposed, such as the U-shaped model Dual Regression branch is proposed to enable the network to learn
Network based on U-Net and the SR Generative the overall morphology semantic information of the
33
34
Adversarial Network (SRGAN) based on the concept of melt pool. Finally, a dual-path melt pool SR network
35
generative adversarial networks. (MPSR-Net) is proposed.
Applying SR to melt pool monitoring is attractive (3) The SR evaluation metrics of the network reached the
because it can improve the quality of monitoring PSNR of 36.13 and the SSIM of 0.926. In addition,
signals at a non-hardware level. However, there is little a strategy is proposed to use the accuracy of the
research concerning this field in L-PBF. Because melt reconstructed features of the melt pool as a new
pool monitoring is an atypical environment. Sun et al. evaluation. The average Mean Absolute Percentage
36
proposed an improved SRGAN network for surface Error (MAPE) of the final melt pool contour
defect detection in AM parts. Similarly, Zhang et al. reconstruction is only 4.52%, and the Intersection
21
proposed a layer-by-layer surface SR method based on over Union (IoU) indicator is as high as 0.939.
U-Net for the L-PBF process. These proposed methods can 2. Methods
demonstrate the application value of SR in AM. However,
they mainly focus on part-scale surface monitoring, as 2.1. Experimental platform
such equipment is easier to deploy. At present, only a few The experimental platform used for capturing melt pool
related works have been made in melt pool monitoring. images consists of two parts: an L-PBF equipment, and a
37
Zhu et al. proposed an in situ AM system that integrates high-speed camera system, as shown in Figure 1.
the SR model, which can achieve SR reconstruction of melt
pool images. However, the proposed SR model is mainly The high-speed camera model is the FASTCAM Mini
applied to offline monitoring and has not been specifically AX200 type 200K-M-16GB with 20000 FPS. A filter of
improved for the characteristics of the melt pool. The 700 – 1000 nm was used to ensure that only the radiated
current SR method used for melt pool monitoring lacks light containing melt pool information is collected.
targeted reconstruction of important information such as Experiments were conducted at four different laser powers:
melt pool edges or variability in melt pool dynamics, and 90 W, 110 W, 130 W, and 170 W, with a constant laser
the dataset used is usually collected by non-professional scanning speed of 100 mm/s. The powder material used
high-speed cameras which make the melt pool susceptible was 316L stainless steel powder with a diameter ranging
to interference from metal powder. Meanwhile, melt from 10 μm to 45 μm. The original images were enlarged
pool dynamics is the main target of monitoring, but by a factor of two and cropped to obtain the region of
existing research only uses general indicators such as peak interest (ROI) of the melt pool.
signal-to-noise ratio (PSNR) and Structural Similarity 2.2. Melt pool image analysis and preprocessing
Index Measure (SSIM) to evaluate model performance,
which leads to the need for further confirmation of the The characteristics of melt pool images are reflected in the
reliability of the model in reconstructing the melt pool following aspects:
features. Melt pool images differ significantly from natural (1) Dynamic variability. The formation mechanism of the
images, and classical networks may have varying effects melt pool involves interactions between the laser and
and applicability in different scenarios. Proposing the metal powder, as well as physical phenomena such as
appropriate SR method for melt pool images requires a heat conduction, fluid dynamics, and metallurgical
comprehensive consideration of factors such as the melting reactions. The shape, size, brightness, and other
process and image features. The contribution of this study features of the melt pool will dynamically change
can be summarized as: during the building process.

Volume 3 Issue 4 (2024) 3 doi: 10.36922/msam.5585