Ph.D. Defense
Rahul Rameshbabu
(Advisor: Prof. Dimitri N. Mavris)
Enabling Rapid, Online Damage Diagnosis of Rotating Machines Through Uncertainty-Aware Surrogate Modeling and Bayesian Inference
Thursday, April 24
2:00 p.m.
Weber SST II, CoVE
and
Abstract
Rotating machinery plays a critical role in the aerospace industry, supporting systems from aircraft propulsion to manufacturing equipment. Faults in such machinery, if left unaddressed, can lead to costly repairs or even catastrophic failure. Recent initiatives—such as Industry 4.0, digital engineering, and the Internet of Things (IoT)—have accelerated the development of sensing and computing technologies that enable high-frequency condition monitoring of vibration signals. By analyzing vibration response data at specific frequencies, parameters associated with common faults (e.g., imbalance, misalignment, cracks, bearing defects) can be inferred. These inferred parameters can then inform maintenance and control strategies aimed at preventing failure in safety-critical systems, minimizing downtime, and maximizing overall performance.
Model-based methods for damage diagnosis leverage vibration measurements and physics-based models of rotating machinery to solve an inverse problem to infer fault parameters of interest. Bayesian model calibration is a promising approach to solve the inverse problem while accounting for uncertainties arising from sensor noise and model discrepancies. However, to make Bayesian calibration feasible for real-time, online diagnosis, surrogate models are employed to approximate the potentially expensive rotordynamics simulations. While this surrogate-accelerated approach enables rapid inference, several technical challenges arise in its implementation for effective online damage diagnosis.
The first challenge is the need for a scalable, probabilistic surrogate modeling method for rotordynamics simulations. To address this, the thesis proposes a deep ensemble neural network approach for constructing surrogate models. The second challenge involves accurately quantifying and propagating surrogate model uncertainty through Bayesian inference during damage diagnosis. This is addressed by leveraging techniques from the uncertainty quantification (UQ) literature for regression models. The final challenge is the development of a Bayesian inference method suitable for rapid, online use. To this end, the thesis proposes an Extended Kalman Filter–based method that continuously assimilates vibration data to enable real-time fault diagnosis. Each proposed method is validated individually using a simulated case study involving imbalance diagnosis of a six-stage compressor. Finally, the complete framework is demonstrated on a physical test rig, showcasing its effectiveness in diagnosing imbalance faults in real hardware.
Committee
- Prof. Dimitri Mavris – School of Aerospace Engineering (advisor)
- Prof. Graeme Kennedy – School of Aerospace Engineering
- Prof. Elizabeth Cherry – School of Computational Science and Engineering
- Dr. Olivia Pinon Fischer – School of Aerospace Engineering
- Mr. Andrew Dugenske – Georgia Tech Manufacturing Institute