Semester

Spring

Date of Graduation

2022

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Mechanical and Aerospace Engineering

Committee Chair

Andrew Nix

Committee Co-Chair

Donald Adjeroh

Committee Member

Donald Adjeroh

Committee Member

V’yacheslav Akkerman

Committee Member

Cosmin Dumitrescu

Committee Member

Donald Ferguson

Abstract

In recent years, the possibilities of higher thermodynamic efficiency and power output have led to increasing interest in the field of pressure gain combustion (PGC). Currently, a majority of PGC research is concerned with rotating detonation engines (RDEs), devices which may theoretically achieve pressure gain across the combustor. Within the RDE, detonation waves propagate continuously around a cylindrical annulus, consuming fresh fuel mixtures supplied from the base of the RDE annulus. Through constant-volume heat addition, pressure gain combustion devices theoretically achieve lower entropy generation compared to Brayton cycle combustors. RDEs are being studied for future implementation in gas turbines, where they would offer efficiency gains in both propulsion and power generation turbines. Much diagnostic work has been done to investigate the detonative behaviors within RDEs, including point measurements, optical diagnostics, thrust stands and other methods. However, to date, these analysis methods have been limited in either diagnostic sophistication or to post-processing due to the computationally expensive treatment of large data volumes. This is a result of the substantial data acquisition rates needed to study behavior on the incredibly short time scale of detonation interactions and propagation. As laboratory RDE operations become more reliable, industrial applications become more plausible. Real-time monitoring of combustion behavior within the RDE is a crucial step towards actively controlled RDE operation in the laboratory environment and eventual turbine integration. For these reasons, this study seeks to advance the efficiency of RDE diagnostic techniques from conventional post-processing efforts to lab-deployed real-time methods, achieving highly efficient detonation characterization through the application of convolutional neural networks (CNNs) to experimental RDE data. This goal is accomplished through the training of various CNNs, being image classification, object detection, and time series classification. Specifically, image classification aims to classify the number and direction of waves using a single image; object detection detects and classifies each detonation wave according to location and direction within individual images; and time series classification determines wave number and direction using a short window of sensor data. Each of these network outputs are used to develop unique RDE diagnostics, which are evaluated alongside conventional techniques with respect to real-time capabilities. Those real-time capable diagnostics are deployed and evaluated in the laboratory environment using an altered experimental setup via a live data acquisition environment. Completion of the research tasks results in overarching diagnostic capability developments of conventional methods, image classification, object detection, and timeseries classification applied to experimental RDE data. Each diagnostic is employed with varying strengths with respect to feasibility, long-term application, and performance, all of which are surveyed and compared extensively. Conventional methods, specifically detonation surface matrices, and object detection are found to offer diagnostic feedback rates of 0.017 and 9.50 Hz limited to post-processing, respectively. Image classification using high-speed chemiluminescence images, and timeseries classification using high-speed flame ionization and pressure measurements, achieve classification speeds enabling real-time diagnostic capabilities, averaging diagnostic feedback rates of 4 and 5 Hz when deployed in the laboratory environment, respectively. Among the CNN-based methods, object detection, while limited to post-processing usage, achieves the most refined diagnostic time-step resolution of 20 µsec compared to real-time-capable image and timeseries classification, which require the additional correlation of a sensor data window, extending their time-step resolutions to 80 msec. Through the application of machine learning to RDE data, methods and results presented offer beneficial advancement of diagnostic techniques from post-processing to real-time speeds. These methods are uniquely developed for various RDE data types commonly used in the PGC community and are successfully deployed in an altered laboratory environment. Feedback rates reported are expected to be satisfactory to the future development of an RDE active-control framework. This portfolio of diagnostics will bring valuable insight and direction throughout RDE technological maturation as a collective early application of machine learning to PGC technology.

Share

COinS