The Woodruff School

SUMMARY

Metal powder-bed fusion is an additive manufacturing (AM) process which enables fabrication of functional metal parts with near-net-shape geometries. The drawback to the metal powder-bed AM processes is lack of precision and the high chances of defect formation. This work addresses the efficacy and development of in-situ quality inspection for powder-bed using high-resolution visual images. For the first time, an imaging setup and the required machine vision (MV) algorithms are developed, implemented, and evaluated to inspect the part cross sectional geometry and to provide an assessment of porosity, in metal powder-bed AM, by direct visualization of porosity. An imaging setup is developed to produce images in-situ from each layer that visualize the fused objects in the layer of powder and the surface quality in terms of fusion and formed defects. Appropriate image processing algorithms are designed and implemented for detection of fused geometric objects and estimation of the geometric parameters. Geometric objects are detected with a boundary point-to-point (root-mean-square) error of 81 microns. Image processing algorithms are developed to detect porosity directly from the camera images of the layers. Information about location, shape, and size of defects can be inferred from them. In addition to detection, a high-level, intelligent approach is proposed and implemented that provides a qualitative assessment of porosity by identifying defective regions. For this purpose, a statistical Bayesian framework is developed and trained based on specific features extracted by means of pattern matching. The Bayesian network is designed to maximize the Figure of Merit and leads to precision of 89% and negative predictive value of 83%. In addition to offering an efficient MV-based inspection system, this work also provides an infrastructure for developing more precise and confident imaging and detection systems for powder-bed AM for visible-light camera images. It also offers a foundation for developing MV systems for powder-bed AM that use other sources of 2D measurements such as height maps, microscopic images, and stereo imaging.

SUBJECT:	Ph.D. Dissertation Defense

BY:	Masoumeh Aminzadeh

TIME:	Thursday, September 29, 2016, 1:00 p.m.

PLACE:	Love Building, 109

TITLE:	DEVELOPMENT OF A MACHINE VISION SYSTEM FOR IN-SITU QUALITY INSPECTION IN POWDER-BED ADDITIVE MANUFACTURING

COMMITTEE:	Dr. Thomas Kurfess, Chair (ME) Dr. David Rosen (ME) Dr. Richard Cowan (MI) Dr. Chen Zhou (ISYE) Dr. H. Jerry Qi (ME)

SUMMARY Metal powder-bed fusion is an additive manufacturing (AM) process which enables fabrication of functional metal parts with near-net-shape geometries. The drawback to the metal powder-bed AM processes is lack of precision and the high chances of defect formation. This work addresses the efficacy and development of in-situ quality inspection for powder-bed using high-resolution visual images. For the first time, an imaging setup and the required machine vision (MV) algorithms are developed, implemented, and evaluated to inspect the part cross sectional geometry and to provide an assessment of porosity, in metal powder-bed AM, by direct visualization of porosity. An imaging setup is developed to produce images in-situ from each layer that visualize the fused objects in the layer of powder and the surface quality in terms of fusion and formed defects. Appropriate image processing algorithms are designed and implemented for detection of fused geometric objects and estimation of the geometric parameters. Geometric objects are detected with a boundary point-to-point (root-mean-square) error of 81 microns. Image processing algorithms are developed to detect porosity directly from the camera images of the layers. Information about location, shape, and size of defects can be inferred from them. In addition to detection, a high-level, intelligent approach is proposed and implemented that provides a qualitative assessment of porosity by identifying defective regions. For this purpose, a statistical Bayesian framework is developed and trained based on specific features extracted by means of pattern matching. The Bayesian network is designed to maximize the Figure of Merit and leads to precision of 89% and negative predictive value of 83%. In addition to offering an efficient MV-based inspection system, this work also provides an infrastructure for developing more precise and confident imaging and detection systems for powder-bed AM for visible-light camera images. It also offers a foundation for developing MV systems for powder-bed AM that use other sources of 2D measurements such as height maps, microscopic images, and stereo imaging.