Session: 16-01-04: Mechanical Performance III

Paper Number: 173271

Integrated Physics - Constrained Dictionary Learning for Quantitative Porosity Estimation in Laser Powder Bed Fusion Process 

Abstract

High porosity is a key factor that hinders the application of laser powder bed fusion (LPBF) in manufacturing high reflective metals. The existing methods for pore monitoring mainly focus on identifying individual pore types, which provides limited guidance for improving printing settings. Besides, many research works ignore the significant storage and transmission burden caused by collecting a large amount of necessary data for training classifiers. To address these problems, this research proposes an integrated physics-constrained dictionary learning (IPCDL) method to achieve simultaneous porosity estimation, compression, and super-resolution of metallographic images of copper manufactured by LPBF. Specifically, a label-consistent dictionary learning strategy is utilized to force the generated sparse vector to be discriminative among different classes and similar in the same class. The physical constraint of the low-resolution imaging strategy is considered during dictionary optimization to enable the reconstruction of high-resolution images from low-resolution ones. Based on this, the high-resolution camera can be replaced with a low-resolution camera, which has a lower cost and produces images that are easier to transmit, thereby yielding improved data acquisition efficiency. Moreover, a residual-based graph sample and aggregate (GraphSAGE) algorithm is integrated with dictionary learning to achieve improved classification performance on porosity with unclear between-class boundaries. The shallow features contained by optical images and hierarchical features represented by sparse vectors are concatenated to build the input graphs for GraphSAGE. Experimental results show that IPCDL achieves over 91% accuracy in classifying high-resolution images at a compression ratio of 40. Moreover, even with a compression ratio of 4.9, the high-resolution images can still be reconstructed from low-resolution ones which has 87.75% fewer pixels compared to high-resolution images. The classification accuracy on low-resolution images is over 89% in five-fold cross-validation experiments. The classification accuracy, generalization ability, and robustness against imbalanced data of IPCDL significantly outperform Unet, Resnet, and traditional dictionary learning methods. The influence of printing parameters on porosity distribution is also studied based on the estimated porosity results. It is found that adopting a relatively high laser power and low scanning speed contributes to reducing the generation of lack of fusion pores. However, if the laser power exceeds 400 W, the scanning speed should be lower than 500 mm/s to minimize the keyhole pores. Higher laser power and lower scanning speed are required when printing boundary regions compared to central region of the parts. Additionally, using a hatch space close to the laser beam diameter and a bidirectional scanning strategy with small rotation angles is beneficial for improving copper density.

 

Keywords: Laser powder bed fusion, Porosity detection, Dictionary learning, Physics constraint, Compression and super-resolution 

Presenting Author: Longye Pan The Hong Kong University of Science and Technology

Presenting Author Biography: Miss Longye Pan received her bachelor's and master's degrees in Mechanical Engineering from Shandong University, China. She is currently a PhD candidate at The Hong Kong University of Science and Technology. Her research interests focus on using physics-constrained dictionary learning for additive manufacturing process monitoring.

Authors:

Longye Pan The Hong Kong University of Science and Technology
Yanglong Lu The Hong Kong University of Science and Technology