Doctoral Dissertations

Orcid ID

https://orcid.org/0000-0002-6577-0562

Date of Award

5-2023

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Major

Industrial Engineering

Major Professor

Mingzhou Jin

Committee Members

Mingzhou Jin, Anahita Khojandi, Bradley Jared, Zhongshun Shi

Abstract

Additive manufacturing (AM) allows the production of parts and goods with many benefits over more conventional manufacturing methods. AM permits more geometrically complex designs, custom and low-volume production runs, and the flexibility to produce a wide variety of parts on a single machine with reduced pre-production cost and time requirements. However, it can be difficult to determine the condition, or health, of an AM machine since complex designs can increase the variability of part quality. With fewer parts produced, destructive testing is less desirable and statistical methods of tracking part quality may be less informative. Combined with the relatively more complex nature of AM machines, qualifying AM machines and monitoring their health to perform maintenance or repairs is a challenging task.

We first present a case study that demonstrates the difficulty of monitoring the qualification of an AM machine. We then discuss some unique challenges AM presents when calibrating and taking measurements of laser power, and we demonstrate the relative insufficiency of this method in tracking the qualification status of an AM machine and the quality of the parts produced.

Next, we present a framework that reverses the directionality of monitoring AM machine health. Rather than monitoring machine subsystems and intermediate metrics reflective of part quality, we instead directly monitor part quality through a combination of witness builds and witness parts that provide observational data to define the health status of a machine. Witness builds provide more accurate data separated from the noisy influence of parts and parameter settings, while witness artifacts provide more timely data but with less accuracy.

Finally, machine health is modeled as a partially observed Markov decision process using the witness parts framework to maximize the long-term expected value per build. We show the superiority of this model by comparison to two less complex models, one that uses no use no witness parts and another that uses only witness builds. A case study shows the benefits of implementing the model, and a sensitivity analysis is performed to show relevant insights and considerations.

Files over 3MB may be slow to open. For best results, right-click and select "save as..."

Share

COinS