Applying Neural Network Technique for prediction of NOx Emission and Combustion Dynamics of an Experimental Turbulent Swirl-stabilized Combustor using flame image processing techniques

Applying Neural Network Technique for prediction of NOx Emission and Combustion Dynamics of an Experimental Turbulent Swirl-stabilized Combustor using flame image processing techniques

In the present study, direct flame images obtained from an experimental swirl stabilized combustor are used to predict the output parameters of the combustor including the level of NOx emission, amounts of noise and the level of pressure fluctuations in the combustor. For this purpose, different values of overall equivalence ratios in the range of 0.7-0.9 along with various amounts of secondary fuel injection rates between 𝑄𝑠𝑒𝑐=0.6 and 4.2 L/min are considered. Moreover, two types of secondary fuel injector are employed in this study. A method which extracts certain features from the images (i.e. physical features of the flame images) is implemented to provide proper inputs for an artificial neural network (ANN) to estimate the output parameters of the combustor. Around 70% of total image data were selected for the training process of the neural network. Meanwhile 15% of the image data is used for cross-validation process and the remaining 15% is used for evaluating the performance of the trained network. Considering the high amount of obtained correlation coefficients (even more than 0.99 for some cases) along with low calculated root mean square errors, the developed ANN demonstrated that it was able to predict the output parameters of the combustor with high accuracy.

In the present study, direct flame images obtained from an experimental swirl stabilized combustor are used to predict the output parameters of the combustor including the level of NOx emission, amounts of noise and the level of pressure fluctuations in the combustor. For this purpose, different values of overall equivalence ratios in the range of 0.7-0.9 along with various amounts of secondary fuel injection rates between 𝑄𝑠𝑒𝑐=0.6 and 4.2 L/min are considered. Moreover, two types of secondary fuel injector are employed in this study. A method which extracts certain features from the images (i.e. physical features of the flame images) is implemented to provide proper inputs for an artificial neural network (ANN) to estimate the output parameters of the combustor. Around 70% of total image data were selected for the training process of the neural network. Meanwhile 15% of the image data is used for cross-validation process and the remaining 15% is used for evaluating the performance of the trained network. Considering the high amount of obtained correlation coefficients (even more than 0.99 for some cases) along with low calculated root mean square errors, the developed ANN demonstrated that it was able to predict the output parameters of the combustor with high accuracy.