Survey of Machine Learning Based RDE Diagnostics: Evaluation and Comparison for Broad Application in Experimental Facilities

Abstract

Laboratory rotating detonation engine (RDE) operations have become increasingly reliable in recent years, giving strong indication that industrial applications are becoming 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, various machine learning methods have been developed to advance the efficiency of RDE diagnostic techniques from conventional post-processing efforts to lab-deployed real-time methods. This work evaluates and compares various convolutional neural networks (CNNs) trained in previous NETL studies according to metrics effecting diagnostic feasibility, external applicability, and performance. Each CNN surveyed, including image classification, object detection, and time series classification, is used to develop total RDE diagnostics and evaluated alongside conventional techniques with respect to real-time capabilities. Real-time capable diagnostics are deployed and evaluated in the laboratory environment using an altered experimental setup, which is outlined herein for possible adaptations in external experimental facilities. Of particular interest, image-based CNNs are applied to externally provided images to approximate dataset restrictions. Conventional methods and object detection are found to offer diagnostic feedback rates of 0.017 and 9.50 Hz currently limited to post-processing, respectively. Image classification using high-speed chemiluminescence images, and time series classification using high-speed flame ionization and pressure measurements, achieve classification speeds enabling real-time diagnostic capabilities, averaging lab-deployed diagnostic feedback rates of 4 and 5 Hz, respectively. Object detection, while currently limited to post-processing usage, achieves the most refined diagnostic time-step resolution of 20 µsec. Image and time series classification require the additional correlation of sensor data, extending their time-step resolutions to 80 msec. Comparisons show that no single diagnostic approach outperforms its competitors across all metrics. Instead, each diagnostic is uniquely suited for a set of primary objectives and constraints. This finding justifies the need for a machine learning portfolio containing a host of networks selected and modified to address specific needs throughout the RDE research community.