Metal additive manufacturing (AM) is known to produce internal defects that can impact performance. As the technology becomes more mainstream, there is a growing need to establish nondestructive inspection technologies that can assess and quantify build quality with high confidence. This article presents a complete, three-dimensional (3D) solution for automated defect recognition in AM parts using X-ray computed tomography (CT) scans. The algorithm uses a machine perception framework to automatically separate visually salient regions, that is, anomalous voxels, from the CT background. Compared with supervised approaches, the proposed concept relies solely on visual cues in 3D similar to those used by human operators in two-dimensional (2D) assuming no a priori information about defect appearance, size, and/or shape. To ingest any arbitrary part geometry, a binary mask is generated using statistical measures that separate lighter, material voxels from darker, background voxels. Therefore, no additional part or scan information, such as CAD files, STL models, or laser scan vector data, is needed. Visual saliency is established using multiscale, symmetric, and separable 3D convolution kernels. Separability of the convolution kernels is paramount when processing CT scans with potentially billions of voxels because it allows for parallel processing and thus faster execution of the convolution operation in single dimensions. Based on the CT scan resolution, kernel sizes may be adjusted to identify defects of different sizes. All adjacent anomalous voxels are subsequently merged to form defect clusters, which in turn reveals additional information regarding defect size, morphology, and orientation to the user, information that may be linked to mechanical properties, such as fatigue response. The algorithm was implemented in MATLAB™ using hardware acceleration, that is, graphics processing unit support, and tested on CT scans of AM components available at the Center for Innovative Materials Processing through Direct Digital Deposition (CIMP-3D) at Penn State's Applied Research Laboratory. Initial results show adequate processing times of just a few minutes and very low false-positive rates, especially when addressing highly salient and larger defects. All developed analytic tools can be simplified to accommodate 2D images.
Keywords: automated defect recognition; nondestructive inspection; unsupervised machine learning.
Copyright 2023, Mary Ann Liebert, Inc., publishers.