Mode decomposition and convolutional neural network analysis of thermoacoustic instabilities in a Rijke tube

Thermoacoustic instabilities are prevalent in a variety of combustion systems including gas turbine combustors and liquid-fueled rocket motors. The limited ability to predict when and how the onset of significant thermo-acoustic instabilities will occur in these systems limits the ability implement active control approaches. The lack of predictive capability and sensing of the onset of instability typically requires the inclusion of passive damping in the combustor design. One possible avenue towards new paradigms in combustor design and optimization is the active control of combustion instabilities. In order to assess the capabilities for detection of combustion instabilities, proper orthogonal decomposition (POD) and a convolutional neural network (CNN) approach are being examined for classification of important dynamics. In this study, high-speed image sequences are collected for a turbulent premixed, bluff-body-stabilized flame in a Rijke tube. The level of thermoacoustic instability is varied through positioning of the flame in the tube. A deep convolutional neural network (CNN) classifies flame images via automatic visual feature. The CNN model is trained on sequential image frames extracted from hi-speed flame videos and demonstrates high accuracy for different flame characteristics from the perspective of instability. The performance of CNN and POD in identifying unstable and stable behavior is evaluated.