An Autoencoder-Driven Clustering Framework for Preventing and Suppressing Flashbacks
C. Karaca (TU Delft - Aerospace Engineering)
I. Langella – Mentor (TU Delft - Flight Performance and Propulsion)
Nguyen Anh Khoa Doan – Graduation committee member (TU Delft - Aerodynamics)
M. Floris – Mentor
Steven J. Hulshoff – Graduation committee member (TU Delft - Aerodynamics)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
This thesis addresses thermoacoustic instabilities and flashback in hydrogen combustion. A simplified model of Ansaldo Energia's GT36 reheat combustor was simulated at 20 bar using Large Eddy Simulation (LES), revealing unsteady flame dynamics driven by strong pressure oscillations. To detect flashback precursors, LES-derived time series were sampled at the combustor wall after an analysis identified suitable monitoring locations, representing a step toward practical sensor placement.
Fourteen thermodynamic, velocity, and species mass fraction signals were reduced via autoencoders with 2–4 latent variables; the three-latent representation emerged as optimal, isolating transition sharpness and mid-frequency modes. Clustering of this space with a modularity-based algorithm consistently identified precursors with maximum lead times of ~42 μs and virtually no false positives. In one case, a flashback was successfully predicted and suppressed. Robustness analyses confirmed generalization across locations and noise levels, demonstrating that wall-based latent clustering advances predictive flashback control toward real-world deployment.