Masters Theses

Orcid ID

http://orcid.org/000359309

Date of Award

8-2020

Degree Type

Thesis

Degree Name

Master of Science

Major

Mechanical Engineering

Major Professor

Subhadeep Chakraborty

Committee Members

Suresh Babu, Hairong Qi

Abstract

The purpose of this paper is to report recent results demonstrating feasibility of active monitoring and fault probability estimation in the Selective Laser Melting (SLM) process in a Renishaw AM250 machine, through analysis of layer-by-layer surface profile data of Fe3Si powder. The data was collected in-situ during the metal additive manufacturing of a Heat Exchanger section, comprised of a series of conformal channels. Specifically, a shallow artificial neural net (ANN) was trained with high-resolution powder bed surface height data from a laser profilometer and then linked to post-print CT scans which provided the truth-data labelling of each site as faulty or nominal. Various measures of accuracy and performance demonstrate excellent performance of the ANN, suggesting that the ANN is capable of discovering strong correlations between surface roughness characteristics and the presence and size of faults. These results were generated by grouping the profile data using post scan CT data, which would not be available in-situ. As such, further work was performed to apply the NN and use insights gained from development of the NN to identify faults in-situ using only the available in-situ data. First an application of the NN was tried on un-preprocessed data but failed to reach satisfactory levels of accuracy. Next, a deeper understanding of the internal process was developed by systematically studying fault sights, their roughness values, and interaction with the NN. This data and insight guided future steps. It was found that faults tended to have higher peaks and lower valleys in close proximity than nominal regions and were being correctly classified as such by the NN. As such, the next step involved developing an algorithm to a priori determine what to consider as a grouped region using the presence of extreme profile data height values. This is referred to as “min-max” stitching later on in the report. The final step involved iterating on this algorithm and revisiting the raw data to try and uncover any other noticeable trends.

Comments

Portions of this document are under review for publication

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

Share

COinS