Automated Defect Analysis of Additively Fabricated Metallic Parts Using Deep Convolutional Neural Networks
Academic Article
Overview
Research
Identity
Additional Document Info
Other
View All
Overview
abstract
Laser powder bed fusion (LPBF)-based additive manufacturing (AM) has the flexibility in fabricating parts with complex geometries. However, using non-optimized processing parameters or using certain feedstock powders, internal defects (pores, cracks, etc.) may occur inside the parts. Having a thorough and statistical understanding of these defects can help researchers find the correlations between processing parameters/feedstock materials and possible internal defects. To establish a tool that can automatically detect defects in AM parts, in this research, X-ray CT images of Inconel 939 samples fabricated by LPBF are analyzed using U-Net architecture with different sets of hyperparameters. The hyperparameters of the network are tuned in such a way that yields maximum segmentation accuracy with reasonable computational cost. The trained network is able to segment the unbalanced classes of pores and cracks with a mean intersection over union (mIoU) value of 82% on the test set, and has reduced the characterization time from a few weeks to less than a day compared to conventional manual methods. It is shown that the major bottleneck in improving the accuracy is uncertainty in labeled data and the necessity for adopting a semi-supervised approach, which needs to be addressed first in future research.
