Thesis Defense Announcement for Jackson Cornelius — 10/19/2021 at 3:00pm

October 5, 2021

Dear Faculty and Students,

You are cordially invited to my thesis defense.

Title: Estimating and Leveraging Uncertainties in Deep Learning for Remaining Useful Life Prediction in Mechanical Systems

When: Tuesday, October 19th at 3:00pm

Where: Webex
https://msstate.webex.com/msstate/j.php?MTID=mbb89b6dfa6698d38db99498dfa803a06

Candidate: Jackson Cornelius

Degree: Masters, Electrical and Computer Engineering

Committee:

Dr. John Ball
(Major Professor)

Dr. Bo Tang
(Committee Member)

Dr. Chaomin Luo
(Committee Member)

Abstract:
Remaining useful life (RUL) prediction for mechanical systems is a problem that researchers in the prognostics and health management (PHM) community have been studying for decades. Both physics-based and data-driven methods have been investigated, and in recent years, attention has been shifted to deep learning in this field. When sufficiently large and diverse datasets are available, deep neural networks can achieve state-of-the-art performance in RUL prediction tasks for a variety of systems. However, if end users are to trust the results of these predictive models, prediction uncertainty must be captured. By default, deep neural networks do not capture the uncertainty inherent in their predictions. This is an important problem to consider, especially as deep neural network-based RUL predictions models are integrated into safety-critical systems. This work addresses this problem by presenting an approach for estimating both epistemic and heteroscedastic aleatoric uncertainties that emerge in deep neural network models trained for RUL prediction. This work demonstrates that quantifying the overall impact of these uncertainties on predictions can uncover valuable insight into model performance on complex, noisy sensor data, where decisions are often made during uncertain operating conditions. First, a novel deep neural network architecture is presented that demonstrates competitive RUL prediction performance on the NASA C-MAPSS FD001 and FD003 datasets. Then, the network is adapted to estimate epistemic and heteroscedastic aleatoric uncertainties. Finally, a study is carried out to observe the effects that RUL truth data augmentation have on the perceived uncertainties in the model. Case studies on the C-MAPSS FD001 dataset show that utilizing the actual RUL truth data can yield more meaningful uncertainty estimates and more insight into the relationship between sensor data and an engine’s time-to-failure.