Properties of chemical species distribution in hypersonic boundary layers at high enthalpies
Abstract
The research objective is to develop a machine learning (ML) model capable of diagnosing the state of a micro-meso scale combustor in real-time using chemiluminescence spectra acquired from multiple locations within the combustor, despite a limited amount of data. Although previous studies have utilized flame images, these methods required complete optical access to the combustor and a significantly large amount of labeled data for successful model training, making them impractical. Furthermore, identifying all unstable modes of a combustor and collecting signals under those states are challenging. Therefore, we specifically propose to use chemiluminescence spectra acquired from multiple locations within the combustor and develop two new ML approaches to address the issues related to limited data and the absence of labeled data for abnormal combustor states. The first ML approach, designed to detect anomalies without training data on abnormal combustor states, is based on the observation that the decoder in an encoder-decoder model, trained solely on normal operation signals, does not perform well on abnormal signals. The second ML approach, aimed at successfully obtaining a predictive model with limited available data, leverages the observation that signals vary gradually with changes in the combustor s operating conditions. This gradual variation can be exploited as constraints to further guide model training. Considering the challenges of data collection, which is not only time-consuming but often dangerous, our research output will facilitate the application of ML-based diagnosis methods to real-world problems, ultimately requiring significantly less data. Importantly, although the approach will be demonstrated for micro-meso scale combustors, it can be applied to other disciplines that involve image-like inputs.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 05, 2025
- Source ID
- FA86552417001
Entities
People
- Mario Di Renzo
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Salento