Application of computer vision and deep learning for flame monitoring and combustion anomaly detection

This paper is devoted to study combustion of pulverized coal in a 5 MW thermal furnace with tangential scroll supply of coal-air suspension and cylindrical reaction chamber. Deep learning approaches were used to monitor the combustion of a coal flame in a furnace and to determine combustion anomalies from flame images. We have developed a deep neural network autoencoder, which is a combination of convolutional layers, fully-connected layers and upsampling layers. The autoencoder was trained to reconstruct the combustion regimes that corresponded to high values of the coefficient of excess air. The trained autoencoder was then used to identify abnormal burning regimes with a lower excess air ratio, in which there is an increase in the amount of unburned coal dust. The best classification model had AUC ROC quality metric value on a test image sample AUC ROC = 0.8. The average precision of the model was 77% and the average recall was 66%. The metrics obtained is limited by the quality of the image labeling due to poorly controlled experimental conditions. The paper concluded that the use of upsampling with convolutional layers show themselves better than the deconvolutional layers, and the combination of convolutional layers with fully-connected layers constitutes the optimal architecture of the model.