Study On Combustion Stability Of Hierarchical Flame Image

The combustion state of utility boiler directly affects the thermal efficiency and equipment life of boiler,so it is necessary to monitor the stability of boiler combustion accurately and effectively in real time to ensure the safe and economic operation of boiler.The flame image of furnace is the most direct representation of the combustion stability of boiler.The traditional stability discrimination method uses classifier to learn the characteristics of flame image extracted manually,and then to determine the combustion stability.There are many problems in this method,such as complex operation,poor anti-interference ability,low recognition accuracy and low generalization performance.In recent years,deep learning,transfer learning and other methods have achieved great success in the field of image recognition.Among them,transfer learning can transfer the knowledge learned in one field to other fields,which greatly improves the recognition effect,and is usually better than the effect of deep learning.At present,there are few methods of deep learning and migration learning to judge the stability of flame image,especially the research of combustion stability based on migration learning is basically blank.Therefore,the combination of migration learning to achieve combustion stability has a great prospect.According to the operation characteristics of 660 MW ultra supercritical coal-fired unit,this paper collects and establishes a fine new flame image data set,including the flame image of each layer of burner of the unit boiler in different periods and working conditions,covering the full states of stable combustion,unstable combustion,start and stop of the boiler,fire extinguishing and critical fire extinguishing.Three kinds of classic deep convolution neural networks are selected,and the truncated singular value decomposition(TSVD)algorithm is introduced to optimize the convolution layer.Training the feature extractor and classifier of transfer learning to construct the deep transfer neural network,and finally realized the combustion stability discrimination based on the model parameter transfer learning method.In addition,the relationship between the transfer learning performance and the number of layers of neural network is studied and analyzed,as well as the specific stability judgment results of each layer of burner.This paper collects and establishes a fine new level combustion image data set,which includes the flame images of 3 levels burners in different periods and different working conditions,including the images of stable or unstable combustion,start and stop of the boiler,fire-fighting.Three classic CNN are selected,and the truncated singular value decomposition(TSVD)algorithm is introduced to optimize the convolution layer.Training the feature extractor and classifier of migration learning to construct the deep migration neural network,and finally realized the combustion stability discrimination combining the deep learning and migration learning,and compared with the results of combustion stability discrimination only using the deep learning mode,studied the relationship between the migration learning performance and the number of neural network layers.The experimental results show that the improvement of network convolution based on TSVD algorithm is effective,while the deep transfer learning achieves an ideal result in the discrimination of combustion stability,and the training time is greatly reduced,and the overall accuracy of Res101 and Dense201 is over 90%.The prediction accuracy of Dense201 model is 94.3%,which is better than the combustion stability judgment results of deep learning.