| Abstract | Defect detection in laser powder bed fusion (LPBF) parts is a critical step for in their quality control. Ensuring the integrity of these parts is essential for a broader adoption of this manufacturing process in highly standardized industries such as aerospace. With many challenges to overcome, there is currently no standardized image analysis and segmentation process for the defect analysis of LPBF parts. This process is often manual and operator-dependent, which limits the repeatability and the reproducibility of the analytical methods applied, raising questions about the validity of the analysis. The pore segmentation step is critical for porosity analysis since the pore size and morphology metrics are calculated directly from the results of the segmentation process. In this work, Ti6Al4V specimens with purposely induced and controlled porosity were printed, scanned 5 times on two CT scan systems by two different operators, and then reconstructed as 3D volumes. The porosity in these specimens was analyzed using manual and Otsu thresholding and a convolutional neural network (CNN) deep learning segmentation algorithm. Then, a variance component estimation realized over 75 porosity analyses indicated that, independently of the operator and the CT scan system used, the CNN provided the best repeatability and reproducibility in the LPBF specimens of this study. Finally, a multimodal correlative study using higher resolution laser confocal microscopy observations was used for a multi-scale pore-to-pore comparison and as a reliability assessment of the segmentation algorithms. The validity of the CNN-based pore segmentation was thus assessed through improved repeatability, reproducibility, and reliability. |
|---|
