Restricted (Penn State Only)
Gobert, Christian J
Graduate Program:
Mechanical Engineering
Master of Science
Document Type:
Master Thesis
Date of Defense:
March 22, 2017
Committee Members:
  • Asok Ray, Thesis Advisor
  • Shashi Phoha, Committee Member
  • Edward Reutzel, Committee Member
  • Guha Manogharan, Committee Member
  • Jan Petrich , Committee Member
  • Additive Manufacturing
  • Machine Learning
  • Powder Bed Fusion
  • Flaw Detection
  • Image Processing
Additive manufacturing (AM) has garnered appeal as having the potential to provide cost-savings for high-end low-volume parts, to enable cost-effective consolidated and complex part designs for enhanced performance, and to make supply-chain management operations more efficient. Certifying adequate part quality remains a barrier to implementation, and process monitoring and part qualification during metallic AM processes may serve as a crucial enabler for the widespread adoption of metallic AM in industrial, military and commercial settings. Post build evaluation techniques, such as x-ray computerized tomography (CT), can identify discontinuities in AM parts, however post-build inspection is costly, and embedded defects may be difficult or impossible to repair. The layerwise construction of metallic powder bed fusion AM(PBFAM) parts presents the opportunity to repair defects during the build process, and so the ability to detect flaws in situ could enable cost-effective in-process re-melting and correction of detected flaws in AM parts. High resolution imaging of build layers may provide a cost-effective means to realize in situ flaw detection, compared to melt pool monitoring and post-build evaluation techniques, to achieve in-process certification. This research investigates and explores a discontinuity detection scheme implemented in a metallic PBFAM system using supervised machine learning on sensor data collected in situ during each layer of the AM build process. A PBFAM system was instrumented with a high resolution digital single lens reflex (DSLR) camera that captured multiple images of the entire build platform following each powder recoating and laser fusion step of each layer in the build. Ground truth labels defining part condition, e.g. anomalous or nominal, were extracted from post-build high resolution CT scans of a single AM part utilizing advanced image processing tools. The anomalous features detected within the CT scan were manually verified to correspond to discontinuities within the component, and were depicted in a 3D representation to provide insight into the development of part flaws during the build process. An affine transformation was generated to map the CT scan data and 3D location of discontinuities into the DSLR layerwise image data using reference points in the AM part, enabling the definition of anomalous and nominal ground truth labels of discretized locations obtained from CT scan data to be properly located within the DSLR layerwise images. Based on knowledge of the AM process and the accuracy of the generated affine transformation, multi-dimensional features were extracted from labeled anomalous and nominal DSLR image locations. Supervised machine learning was employed to study the severability of the labeled anomalous and nominal features directly within the domain of the in situ layerwise images. Linear support vector machine (SVM) models were used to create two discontinuity detection strategies, which were then validated through four-fold cross-validation. Both detection strategies were (i) evaluated with several performance metrics and assessed with respect to discontinuity size, (ii) validated by identifying similar discriminatory separation between cross-validated models, and (iii) evaluated with the presence of sensor noise added to the DSLR layerwise images.