Data Driven Fault Detection of Premixer Centerbody Degradation in a Swirl Combustor

Raghul Manosh Kumar, Benjamin Peters, Benjamin Emerson, Kamran Paynabar, Nagi Gebraeel, Timothy Lieuwen

Modern gas turbines are equipped with multiple sensors, with the intent of monitoring the working and components of the turbine. Formally referred to as condition monitoring, this calls for the need to be able to quickly analyze large volumes of data to determine if the gas turbine is operating under optimal conditions, and if not, how far from optimal it is. One common method of monitoring the gas turbine health, is via acoustic pressure transducers mounted close to the annular or can combustors. Such combustor flowfields frequently employ a swirl configuration, to which a centrebody is a crucial part. These combustor centrebodies are exposed to the high flame temperatures and are hence prone to ablation. It is a desirable outcome of condition monitoring, to predict this centrebody ablation as it occurs, to avoid unnecessary shutdown and restart of the gas turbine.

This paper aims to study centrebody ablation effects on flame dynamics and subsequent combustor operation in a laboratory scale, unwrapped annular triple nozzle, swirl stabilized atmospheric pressure combustor, where the fault is simulated using 4 different centrebody lengths, ranging from optimal to worst case operating conditions. Simultaneous high speed CH\* chemiluminescence and high fidelity acoustic pressure measurements at various locations on the combustor are obtained. Multiple repeats at each fault level provide ample samples for a statistical study involving testing and training routines. One goal of the paper is to look at the effect of decreasing centrebody lengths on flame dynamics in this configuration. The paper also aims to develop an analytics framework that leverages high-dimensional data for diagnosing combustor centrebody faults.

This framework consists of data curation, feature engineering, and diagnosis. Data curation ensures data quality by detecting outliers and accounting for missing or corrupted data. Informative features are selected through a feature engineering process consisting of three primary tasks: feature extraction, dimensionality reduction, and data fusion. Feature extraction leverages both data-driven and physics-based models to construct robust features from the acoustic and image stream data.

Through statistical regularization, a Machine Learning tool designed to avoid model overfitting, optimal sensor locations are identified. Finally, dimensionality reduction methods decompose the large condition monitoring data structures into smaller, low rank structures containing the most informative aspects of the raw data. Thus, these structures are now ideal for condition monitoring tasks. Finally, a robust mapping from the space of low-rank structures to the length of the centrebody is devised to infer the state-of-health of the centrebody.


Data Driven Fault Detection of Premixer Centerbody Degradation in a Swirl Combustor

Division / Sponsor:
International Gas Turbine Institute (IGTI)

Paper Proceedings: https://asmedigitalcollection.asme.org/GT/GT2020/volume/84058

Turbo Expo - Virtual Conference

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.