Description
Additive manufacturing (AM) has expanded the ability to increase part complexity, enabled rapid prototyping, and extended end user product customization. However, challenges to part quality in AM, such as porosity remain. This thesis investigates defect detection, the first step in increasing AM part quality. During this investigation, it was determined that image-based detection would present an inexpensive and readily available solution. However, image-based detection requires time-consuming classification. It was also determined that porosity is a critical defect. Therefore, automated image-based porosity detection for Powder Bed AM (PBAM) was the focus of this study. For image classification to be performed, a set of images were labeled. Each of these images contains a set of metadata including bounding boxes within the image and the label of the region. Bounding boxes were also cropped out of the images and saved as separate images. Features were then manually extracted from these regions and applied to supervised machine learning algorithms. However, it was determined that these features are weak classifiers and additional features may be required. The best performing supervised machine learning (SML) algorithm was the Boost Tree at 74% accuracy with all features utilized. Automated feature extraction was then assessed using transfer learning on a ResNet50 convolutional neural network (CNN). The CNN was applied to two region generated algorithms and was found to provide a consistent precision of 70% for a 50% recall and 75% recall with a 50% intersection of union (IOU). Faster R-CNN, an algorithm that both proposes regions and classifies them, was trained, and tested with the dataset and the ResNet50 network. It was found that the R-CNN resulted in 81% recall and 75% precision at a 50% IOU. The culmination of these efforts provides a foundation for in-depth ML studies on porosity detection for PBAM. The current model can be used to assist labeling new images and increase the database size. Additional, SMLs can be considered and benchmarked against the results herein. Overall, the work presented establishes a baseline for future work and the ability to iteratively develop models for porosity detection.
