A Dynamic Mode Decomposition Based Deep Learning Technique for Prognostics

Download (1.17 MB)
posted on 01.12.2020, 00:00 by Khaled Mohammad A Akkad
Remaining useful life is one of the key indicators for mechanical equipment health and condition-based maintenance requirements. The field of prognostics and health management heavily relies on remaining useful life estimation. Due to the magnitude of monetary costs associated with failure of mechanical equipment, remaining useful life prediction has become one of the pillars of the prognostics and health management field. The availability of industrial big data enabled remarkable research efforts in prognostics. Much of this effort is directed at the development and improvement of deep learning based prognostic techniques by means of creating hybrid models, hyperparameter optimization, and other approaches. This dissertation aims to improve the remaining useful estimation capabilities of deep learning prognostic techniques by integrating a physics based approach into deep learning schemes. This physics based approach is the Koopman operator which produces infinitely linear representations of nonlinear systems with known equations. These infinite representations need to be approximated via a data driven approach for the purposes of obtaining health indicators useful for predicting the remaining useful of industrial machines and equipment. Dynamic mode decomposition is a data driven approach for approximating the modes of the Koopman operator. In this dissertation, dynamic mode decomposition is incorporated into a variety of deep learning prognostic schemes to enhance the performance of the remaining useful estimation. Two industrial applications are utilized to validate the proposed approach. The first application is the NASA spiral bevel gear vibration data. The second application is the NASA commercial modular aero-propulsion system simulated vibration data of turbofan engines. The proposed approach demonstrates an increase in accuracy of remaining useful estimation in both applications and across all datasets therein when the dynamic mode decomposition is in incorporated into the deep learning prognostic schemes.



He, David


He, David


Mechanical and Industrial Engineering Department

Degree Grantor

University of Illinois at Chicago

Degree Level


Degree name

PhD, Doctor of Philosophy

Committee Member

Li, Lin Hu, Mengqi Brown, Michael Bechhoefer, Eric Qu, Yongzhi

Submitted date

December 2020

Thesis type