Abstract
Internal defects, e.g., lack of fusion and porosity, are major quality concerns in Laser Powder Bed Fusion (L-PBF). In post-process part inspection, X-ray Computed Tomography (XCT) is used to scan the part to reveal the defective regions inside. 2-dimensional XCT images are obtained showing the part’s cross-section at different heights (layers). Segmenting the defects from raw XCT images is necessary to locate the defected regions, evaluate the part’s quality, and enable root cause analysis. This study proposes two methods for defect segmentation, one is based on deep learning (DL) and the other based on classic machine learning (ML), and compares them with statistical image thresholding approaches (i.e., K-means, Bernsen’s, Otsu’s Thresholding). A discussion about the method-level difference among these methods is provided, revealing the merits of the proposed DL and ML methods in fast defect segmentation and transfer learning across printing conditions. A Case study is done by applying the DL, ML, and statistical image thresholding methods on real XCT images of L-PBF specimens. The defect segmentation accuracy and efficiency of the proposed DL and ML methods are evaluated. A guideline is developed for automatic defect segmentation from XCT images of L-PBF-ed parts by combining statistical image thresholding and DL/ML methods.
