分层旋流火焰特征提取与燃烧不稳定预测

Combustion instabilities are unstable phenomena arising from the coupling of heat release rate oscillations and pressure fluctuations. The chemiluminescence image is one of the most accessible data types in experimental combustion research. This paper adopted three feature extraction methods：hand-crafted，unsupervised neural network(USN)，and sound pressure level supervision neural network(PSN). The features were extracted from two data sets：transient image and time-averaged image. Based on the extracted features，their significance and physical characterization were analyzed，and their sound pressure level prediction accuracy and transfer ability were tested. The results show that the proposed PSN in this paper can achieve feature extraction and instabilities prediction with excellent robustness，accuracy and transfer ability based on the reproduction of the physical meaning of flame.