Detection of Combustion Instabilities Cased by Premixed Flames Using Convolutional Autoencoder

In this paper, using a simple combustion device consisting of a premixed burner, a rectangular cylinder, and a visualization window, the pressure fluctuation level and fame images were measured when the flame position and operating conditions were varied.By applying a Convolutional Autoencoder (CAE) to the acquired images and extracting features, a Combustion Instability Index (ΔCAE TI_err ) was defined that can quantify flame conditions.By organizing the correlation between the proposed index and combustion oscillation levels, we evaluated the possibility of detecting signs of increased combustion oscillation.The results showed that the proposed Combustion Instability Index and the combustion oscillation level were highly correlated.The mechanism that causes the increase in combustion oscillation level was discussed by evaluating the effect of operating conditions on flame distribution using Grad-CAM data analysis, which enable visualization of the index on a two-dimensional plane.