Degree Type

Thesis

Date of Award

2020

Degree Name

Master of Science

Department

Mechanical Engineering

Major

Mechanical Engineering; Human Computer Interaction

First Advisor

Eliot Winer

Abstract

As sensing and data collection capabilities have dramatically increased in recent years, many areas from medicine to entertainment to engineering have to rethink how products are designed, delivered and maintained. In engineering fields data is everywhere and its use as a decision aid, in the constant stream of tradeoff decisions, is critical to delivering more robust products and services accurately and efficiently. Thus, the need to develop intelligent methods to analyze and visualize large datasets, to enable human understanding, is critical. One method that has been proven effective in this endeavor is the self-organizing map (SOM). However, SOMs require substantial computational resources and time to train, making them impractical for large datasets or datasets that may be added to over time. If this issue could be overcome, this approach could be widely adopted. This thesis studies the concept of using a subset of data to represent the characteristics of a full data set via a SOM. The correlation of a subset and full dataset SOM was studied on two different test cases. The percent difference of node weights was used to compare map representations between the partial and full datasets. A node alignment process was designed and implemented to enable a more accurate comparison of two SOMs. The methodology was evaluated on two test cases. A hundred comparisons of node weights from subset and full datasets maps were completed per test case. Results showed that pairing node weights by row and column designation did not accurately compare two different SOMs. The alignment process was then performed on ten samples of map comparisons per test case. Results of the aligned nodes provided a much more accurate comparison of SOMs from partial and full datasets.

The results of this study show that with a good representative subset of data very similar nodal weights can be reached through map training compared to using the full dataset. This allows a trained SOM to be available as a decision aid in a fraction of the training time compared to using the full dataset.

DOI

https://doi.org/10.31274/etd-20200902-9

Copyright Owner

Holly Deann Baiotto

Language

en

File Format

application/pdf

File Size

50 pages

Share

COinS