Georgia Institute of Technology
NOVEL ARCHITECTURE AND MACHINE LEARNING MODELS FOR SYSTEM PROGNOSTIC AND REMAINING USEFUL LIFE PREDICTION WITH UNCERTAINTY QUANTIFICATION
Abstract
dc:description.abstractUnexpected failures in engineering systems or equipment often lead to significant disruptions and losses. A key element for maintenance planning and system safety analysis is prognostic and remaining useful life (RUL) prediction. Accuracy in prognostic and RUL prediction is important to sustain equipment reliability, reduce total maintenance costs, and prevent unexpected failures. This thesis addresses two prevalent challenges in data-driven prognostic models related to prediction accuracy and interoperability. This thesis first proposes a novel computational architecture for accurate prognostic and RUL prediction with uncertainty quantification. This computational architecture integrates signal preprocessing, Deep Learning (DL), and nonstationary Gaussian process regression (NSGPR) to address the first challenge regarding prediction accuracy. Based on this architecture, this thesis develops three models for aircraft engine, rotating machinery, and lithium-ion battery prognostic applications for performance testing and benchmarking against other best-in-class alternatives. The models are tailored for different problem setups and the characteristics of the input data. The first model is tested using the engine prognostic dataset (C-MAPSS), the second model using the PHM12 bearing vibration dataset, and the third model using the NASA Ames battery prognostic dataset. The results in each case show significantly higher accuracy and tighter uncertainty bounds compared with other best-in-class models in the literature. The nature of the uncertainty quantification provided in RUL prediction is investigated and compared with the more prevalent dropout method in DL. One advantage of the proposed method is its capacity to properly capture heteroscedastic uncertainty (aleatoric) with the NSGPR, whereas the dropout method is more generally confined and better suited for homoscedastic uncertainty. Both methods capture equally well epistemic uncertainty. Second, the other major challenge in data-driven prognostic models is the need to improve interpretability and provide realistic and trustworthy predictions. Although data-driven and ML-based models have significant advantages for prediction accuracy, especially for high dimensional problems, one of their major drawbacks is the insufficient interpretability of the predictions. To demonstrate how this challenge can be addressed, this thesis develops a physics-informed dynamic deep autoencoder (PIDDA) by integrating physics equations into a novel DL model for prognostic and degradation prediction. PIDDA includes three elements: an autoencoder, a physics-informed model training; and a physics-based prediction adjustment. The PIDDA model is tested and benchmarked using the battery prognostic dataset. The computational experiments demonstrate that PIDDA (1) provides significantly higher prediction accuracy; (2) requires less prior data for its predictions; (3) produces more informative and interpretable predictions than alternative models. To analyze the effectiveness of using physics equations, an ablation study of PIDDA is conducted in this thesis.
Degree
thesis:*- Level thesis:degree_level
- Doctoral
- Department dc:contributor.department
- Aerospace Engineering
- Grantor dc:publisher
- Georgia Institute of Technology
- Year dc:date.issued
- 2022
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Xu, Zhaoyi
- Advisor dc:contributor.advisor
-
- Saleh, Joseph H.
- Committee members dc:contributor.committeemember
-
- Mavris, Dimitri N.
- Feron, Eric
- Theodorou, Evangelos
- Orlando, Thomas
Subjects
dc:subject × 4Rights
- Language dc:language.iso
- en_US
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- http://hdl.handle.net/1853/66546