Data-driven detection of flame shape transitions in a two-stage lean premixed aeronautical combustor
Abstract
This study presents two detection strategies of flame shape transitions in a two-stage lean premixed aeronautical combustor: a supervised microphone-based method and an unsupervised manifold-learning framework. Pressure measurements and high-speed flame images from 13 experiments are analyzed to characterize the transition between lifted and attached flames. The microphone-based strategy identifies transitions through characteristic drops and valleys in the plenum pressure signals. Conversely, the unsupervised approach employs Uniform Manifold Approximation and Projection (UMAP) to embed the flame images into a three-dimensional low-dimensional manifold (LDM). The LDM reveals two dense stable attractors and a low-density transitional trajectory, enabling a clear separation of the flame states. The comparison with a classical supervised correlation-based approach shows that the manifold- and microphone-based frameworks detect transitions with mean lags of −2.3 ms and +25 ms, respectively, indicating that the LDM better captures incipient structural changes in the flame dynamics. The manifold representation thus provides a physics-consistent encoding of transitional dynamics and early precursors to flame attachment. A real-time implementation of the manifold-based approach is also introduced and relies on the definition of a stability envelope in the latent space. This enables online detection of the transition to a new stable flame state within a few milliseconds relative to the correlation-based reference, including early warnings of up to 2 ms and lags of up to 3.1 ms. These results demonstrate the potential of the unsupervised manifold-based method for the monitoring and early detection of flame shape transitions, contributing to the development of robust control strategies in staged aeronautical combustors. Novelty and significance statement: Flame shapes play a key role in the stability and performance of staged lean premixed combustors, directly affecting efficiency and emissions. However, their real-time detection and characterization remain challenging due to the nonlinear, high-dimensional nature of combustion dynamics. This work introduces novel data-driven strategies for detecting flame shape transitions. An unsupervised strategy applies, for the first time, the Uniform Manifold Approximation and Projection (UMAP) algorithm to high-speed combustion imaging to reveal low-dimensional manifolds that separate stable and transitional flame states. This representation enables robust identification of structural flame changes and supports real-time detection with millisecond-scale precision by defining a stability envelope in the latent space. The strong agreement between the newly proposed strategy and a classical correlation-based method demonstrates the potential of low-dimensional data-driven representations for monitoring combustion dynamics and improving instability detection in advanced low-emission combustors.
Related
- Source-Driven Transitions in Axisymmetric Euler Flow
- Sharp minimax risks and phase transitions in sparse submatrix detection
- Model-Driven Deep Learning for Distributed Detection with Low-Bit Quantization
- A machine-vision framework for electric vehicle charging hole localization based on contour-driven detection and geometric verification
- Statistical-Algorithmic Phase Transitions in Stochastic L2-Boosting
BibTeX
@article{mateo2026datadriven,
title = {{Data-driven detection of flame shape transitions in a two-stage lean premixed aeronautical combustor}},
author = {Emilio Mateo and Antoine Renaud and Bérengère Podvin and Salvatore Iavarone},
journal = {Proceedings of the Combustion Institute},
year = {2026},
month = sep,
doi = {10.1016/j.proci.2026.106427},
url = {https://hal.science/hal-05766015v1},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Data-driven detection of flame shape transitions in a two-stage lean premixed aeronautical combustor
AU - Emilio Mateo
AU - Antoine Renaud
AU - Bérengère Podvin
AU - Salvatore Iavarone
PY - 2026
DA - 2026/09/11
JO - Proceedings of the Combustion Institute
DO - 10.1016/j.proci.2026.106427
UR - https://hal.science/hal-05766015v1
AB - This study presents two detection strategies of flame shape transitions in a two-stage lean premixed aeronautical combustor: a supervised microphone-based method and an unsupervised manifold-learning framework. Pressure measurements and high-speed flame images from 13 experiments are analyzed to characterize the transition between lifted and attached flames. The microphone-based strategy identifies transitions through characteristic drops and valleys in the plenum pressure signals. Conversely, the unsupervised approach employs Uniform Manifold Approximation and Projection (UMAP) to embed the flame images into a three-dimensional low-dimensional manifold (LDM). The LDM reveals two dense stable attractors and a low-density transitional trajectory, enabling a clear separation of the flame states. The comparison with a classical supervised correlation-based approach shows that the manifold- and microphone-based frameworks detect transitions with mean lags of −2.3 ms and +25 ms, respectively, indicating that the LDM better captures incipient structural changes in the flame dynamics. The manifold representation thus provides a physics-consistent encoding of transitional dynamics and early precursors to flame attachment. A real-time implementation of the manifold-based approach is also introduced and relies on the definition of a stability envelope in the latent space. This enables online detection of the transition to a new stable flame state within a few milliseconds relative to the correlation-based reference, including early warnings of up to 2 ms and lags of up to 3.1 ms. These results demonstrate the potential of the unsupervised manifold-based method for the monitoring and early detection of flame shape transitions, contributing to the development of robust control strategies in staged aeronautical combustors. Novelty and significance statement: Flame shapes play a key role in the stability and performance of staged lean premixed combustors, directly affecting efficiency and emissions. However, their real-time detection and characterization remain challenging due to the nonlinear, high-dimensional nature of combustion dynamics. This work introduces novel data-driven strategies for detecting flame shape transitions. An unsupervised strategy applies, for the first time, the Uniform Manifold Approximation and Projection (UMAP) algorithm to high-speed combustion imaging to reveal low-dimensional manifolds that separate stable and transitional flame states. This representation enables robust identification of structural flame changes and supports real-time detection with millisecond-scale precision by defining a stability envelope in the latent space. The strong agreement between the newly proposed strategy and a classical correlation-based method demonstrates the potential of low-dimensional data-driven representations for monitoring combustion dynamics and improving instability detection in advanced low-emission combustors.
SN - 1540-7489
SN - 1540-7489
ER -