Computer Vision for Quantification of Defects in Steels Printed by Binder Jetting

<p dir="ltr">Binder jet printing (BJP) is gaining potential for large-scale manufacturing of a wide range of materials at an economical rate. As printed binder jet parts are very fragile, they require post-processing to densify and increase their applicability. </p><p dir="ltr">This thesis conducts an in-depth study of various post-processing parameters related to debinding and sintering, which could potentially lead to the incorporation of several defects, including porosity, oxide inclusions, and intergranular precipitates, in 316 L stainless steel (316 L SS). These defects are detrimental to part performance, and a quantitative analysis of defects provides a quality control metric to the post-processed part. Advanced computer vision and machine learning techniques are employed to quantify defects in high-resolution optical images of the sintered and heat-treated 316 L SS, printed by binder jetting. The pipeline of this thesis starts with the background on BJP, its advantages and challenges as compared to other additive manufacturing (AM) techniques, an in-depth study of the effect of sin?tering and debinding, followed by microstructural characterization of the post-processed and heat-treated parts on which advanced computer vision techniques have been applied, to detect and quantify the defects. This thesis covers various experimental parameters and practical challenges (on a laboratory scale) encountered during post-processing BJP SS 316 L in the first three chapters, which result in the formation of multiple defects in the final part. The first three chapters are essential for understanding the fundamentals of BJP, including the effects of parameters such as sintering atmosphere, the presence of water vapor in the sintering chamber, and incomplete debinding, all of which contribute to undesirable defects in the final part. Chapter 2 focuses on understanding the optimum debinding time by estimating the thermal conductivity of the printed part. Additionally, it focuses on assessing the range of temperature over which the binder decomposes, as well as analyzing the significant products of decomposition. Intergranular precipitation of carbides, metallographic sample preparation, and collection of high-resolution microscopic images have been covered in depth in Chapter 3. Chapters 4 - 6 of the thesis deal with the use of machine learning models and the development of computer vision pipelines to detect and quantify defects like carbides, porosity & etch pits, and twin boundaries in the microstructure. Various computer vision techniques - supervised, semi-supervised, and unsupervised- have been discussed in detail in different chapters of the thesis. </p><p dir="ltr">In Chapters 4-6, the work progresses from using a supervised machine learning model that involves manual annotation of microstructural images to developing a semi?supervised and unsupervised computer vision pipeline, thereby avoiding the need for manual annotation. The developed semi-supervised computer vision pipeline gives users the flexibility to filter various features in their microstructure, and therefore has the po?tential for adaptability across different shapes, contrasts, exposures, and magnifications. The developed unsupervised computer vision technique is promising. It has potential for use in detecting and quantifying microstructural features across a diverse range of datasets, encompassing microstructures of various materials, such as Ti alloys, Ni alloys, and other steels. The contribution of the current work in this thesis is to present the AM community with a tool to detect and quantify defects, such as carbides, twin boundaries, pores, and etch pits, in their microstructural images. One could choose the technique that is most applicable to the microstructure, depending on the shape, size, and properties of the microstructural features. </p>