Session: Research Posters

Paper Number: 173082

Image Classification for Rotating Detonation Engine Wave Behavior: A Comparative Study of Advanced Models 

Rotating Detonation Engines (RDEs) show significant promise for enhancing the efficiency of gas turbine engines while maintaining low 𝑁𝑂𝑥 emissions. Advanced monitoring of combustion behavior within RDEs is a crucial step toward actively controlled operation in laboratory and industrial environments. Various machine learning methods have emerged from the field of computer science and artificial intelligence which have been used to advance diagnostic efficiencies of data emerging from RDEs and reduce data processing time. However, the field of computer science and artificial intelligence (CS/AI) is undergoing rapid evolution, with new models and architectures emerging at an accelerating pace. RDE monitoring techniques must take advantage of this evolution to offer computationally efficient and highly time-resolved diagnostics.

This research examines the classification of high-speed experimental RDE images using multiple models that are being compared to determine the best and fastest approach for analyzing the combustion behavior of RDEs. Successful integration of state-of-the-art data analysis methods serves as a steppingstone for further machine learning integration in RDE research and development of comprehensive real-time diagnostics.

This work advances the field of engineering by demonstrating RDE monitoring techniques using cutting edge machine learning applications. This methodology seeks to reduce the processing time needed to classify the number and direction of detonation waves, also known as the ‘wave mode,’ within drastically reduced time intervals as compared to traditional high-framerate RDE image analysis techniques. This work advances the field of computer vision by offering a practical laboratory application of cutting-edge models for image classification tasks. The integration of existing models developed by the CS/AI community with experimentally derived datasets provides insights into transfer learning and model adaptation strategies. These insights can promote further use of the practical application of these models and inform future research in real-world image classification tasks. 

The data within this study were collected from experiments using a water-cooled RDE. This RDE exists in the high-pressure combustion test facility at the Department of Energy’s (DOE) National Energy Technology Laboratory (NETL). Down-axis images are captured through a 50-mm-diameter quartz viewport at 60,000 fps by an imaging system that consists of a Nikkor 105 mm UV lens mounted to an Invisible Vision UVi 2550-10-S25 intensifier and a Photron FASTCAM SA-A high-speed digital camera. These images capture a OH* chemiluminescence emitted by detonation waves traveling around a cylindrical annulus at 5-10 kHz. Since thousands of images emerge from this image-taking process, the images are processed by using coding methods in Python then systematically filtered using an AFRL Beta filter, cropped and centered using the outer radius of the detected annulus region, and labeled according to its wave mode. 

In a previous study, convolutional neural network (CNN) architectures were trained using an initial training set containing over 100,000 images, and the current work seeks to improve this effort. This improvement expands the training data by supplementing each of the ten existing categories with 1,000-2,000 newer images that were taken during and after 2020. A comprehensive summary plot of hundreds of runs is used to identify four new classification categories with up to 12,000 images each to capture previously unrepresented combustion behaviors. Lastly, multiple modern model architectures will be trained on the training set and compared to identify which achieves the most optimal time performance while maintaining high classification accuracy. 

Using a convolutional neural network (CNN) and modern model architectures offer significant improvements over conventional RDE image processing techniques in both speed and accuracy. Model training is currently in progress, and the novel classification labels have been successfully implemented within the training set framework. While quantitative performance metrics await training completion, there is promising potential for classification speed and accuracy based on the previous work and the advancement in computer vision capabilities. 

Presenting Author: Sophia Georgieva National Energy Technology Laboratory

Presenting Author Biography: Sophia Georgieva is a recent graduate with a bachelor's in computer science from the University of Central Florida and currently serves as a summer research intern at the National Energy Technology Laboratory in Morgantown, WV. Their research focuses on applying modern image classification architectures to rotating detonation engine wave behavior.

Authors:

Sophia Georgieva National Energy Technology Laboratory
Kristyn Johnson National Energy Technology Laboratory
Justin Weber National Energy Technology Laboratory