Date Approved

6-5-2017

Embargo Period

6-5-2017

Document Type

Thesis

Degree Name

MS Computer Science

Department

Computer Science

College

College of Science & Mathematics

First Advisor

Baliga, Ganesh R.

Second Advisor

Lobo, Andrea F.

Third Advisor

Tinkham, Nancy L.

Subject(s)

Reliability (Engineering); Neural networks (Computer science)

Disciplines

Computer Sciences | Operations Research, Systems Engineering and Industrial Engineering

Abstract

In the field of engineering, it is important to understand different engineering systems and components, not only in how they currently perform, but also how their performance degrades over time. This extends to the field of prognostics, which attempts to predict the future of a system or component based on its past and present states. A common problem in this field is the estimation of remaining useful life, or how long a system or component functionality will last. The well-known datasets for this problem are the PHM and C-MAPSS datasets. These datasets contain simulated sensor data for different turbofan engines generated over time, and have been used to study estimations of remaining useful life.

In this thesis, we study estimations of remaining useful life using different methods of data analytics, preprocessing, post-processing, different target functions used for training models, and their combinations. We compared their performance primarily using scores from the 2008 Prognostics and Health Management Competition. Our basic feedforward neural network outperforms previously competitors and other modern methods. Our results also gave us a ranking between the top 10 and 15 based on a 2014 benchmark. We have improved on the results of previously, published methods, primarily focusing on the Prognostics Health Management competition score of our results for better comparisons.

Share

COinS