Session: 05-04: Combustion Measurements 1

Submission Number: 179134

Optimal Sensor Placement for Flow Field Classification in an H2 Combustor: Foundations and Experimental Results 

The latest technical requirements, standards, and policies in terms of efficiency, environment, safety, operation flexibility and operation costs imply the need for better combustion monitoring in gas turbines. In the experimental research project KI-PIRINHA (Key enabling technologIes on Performance, EfficIency and Resilience by Artificial INtelligence for Hydrogen Applications), a health monitoring of H2 combustion flame stability in recirculating combustors is implemented and experimentally evaluated.

Deviation from desired flow characteristics may yield instable flame containment (“flashback”), potentially damaging inlet nozzles, structural integrity of the combustor chamber or sensors installed. Suboptimal combustion conditions decrease performance and efficiency and result in increased exhaust gas emissions. To detect unintended combustion regimes, the combustion chamber is equipped with multiple opto-acoustic sensors whose position is optimized for the abnormal regime detection task.

A reduced order model is created based on transient CFD simulations and Laser Doppler Velocimetry (LDV) of the flow field within a prototype combustion chamber. This knowledge about the dynamical properties of the flow field enables selection of a set of CFD simulation model parameters during development and locating sensors at the most informative location in combustion experiments.

We apply a flow regime classification algorithm of experimental velocity data for the process of selecting optimally informing sensor locations: We use modal decomposition of simulated transient flow fields for tailoring an optimal basis by Singular Value Decomposition (SVD). We truncate the set of modes to maintain the most informative ones and that best support sensor placement (Sparse Sensing). Further dimensionality reduction from the feature space to a discriminating space (linear discriminant analysis) allows for classification. In the low dimensional discriminator space, we train a nearest centroid classifier for anomaly detection.

The task of flow regime anomaly detection is interpreted as a classification task, where measurement data is assigned to previously learned classes. Here, each class represents the flow regime of a CFD result that is determined by one distinct set of parameters. These parameters may be a distinct setting for inlet conditions leading experimentally determined to yield stable and instable combustion regimes.

We present the mathematical foundations and demonstrate how optimal sensor placement using prior knowledge (CFD data) yields higher information gain than a conventional approach based on accessibility. We compare our results with those achieved with random sensor placement.

Finally, we show measurement of combustion experiments within the prototype combustion chamber and results for the flow regime classification as the anomaly detection.The exhibition of results for the specific use case proves the suitability of the method and the detailed foundations allow the community for easy adaption to various user specific applications.

Presenting Author: Arno Fallast University of Applied Sciences FH JOANNEUM GmbH, Institute of Aviation

Presenting Author Biography: A. Fallast is a senior researcher and lecturer at the Institute of Aviation at the University of Applied Sciences, Graz where he founded and now heads the “Intelligent Systems Lab” research group. He is conducting his PhD research at the Technical University Graz, Institute of Machine Learning and Neural Computation, researching AI methods for anomaly detection in dynamic systems.
He received his bachelor’s degree and his master’s degree in aviation engineering from the FH JOANNEUM in Graz, Austria. For 13 years he has been working as a researcher and lecturer, and his research interests include:
- System identification of unmanned vehicles and dynamical systems.
- Anomaly detection for dynamic systems.
- AI based flight control.
- Flight simulation and trajectory optimization for UAM vehicles.

Authors:

Arno Fallast University of Applied Sciences FH JOANNEUM GmbH, Institute of Aviation
Lukas Andracher University of Applied Sciences FH JOANNEUM GmbH, Institute of Aviation
Samuel Lesak University of Applied Sciences FH JOANNEUM GmbH, Institute of Aviation
Fynn Thilker University of Applied Sciences FH JOANNEUM GmbH, Institute of Aviation
Andrea Hofer Combustion Bay One e.U., Advanced Combustion Management
Fabrice Giuliani Combustion Bay One e.U., Advanced combustion management