Research Of Combustion Flame State Monitoring Technology Based On Deep Learning

Under the background of renewable energy combined to the grid and flexible transformation of thermal power units,steady combustion of boilers which are under low load condition is the key factors limiting thermal power units on reducing the minimum technical output.Therefore,the demand for the effectiveness and accuracy of flame condition monitoring and identification is greatly increased.At present,most furnaces are led by personal experience.In order to perceive the combustion state of the furnace intelligently and dig out effective combustion information deeply,the potential features of flame images are extracted by denoising convolutional autoencoding improved with Generative Adversarial Networks,the image classification is calculated by principal component analysis and fuzzy C-means clustering,the visual classification features are displayed in the form of gradient-weighted class activation map,and the combustion stability is distinguished by multi-signal fusion in this paper.The research work and conclusions of this paper mainly include:(1)Since the flame image contains the most sufficient combustion information,a potential feature extraction method based on denoising convolutional autoencoding network(DAE)was proposed.The images preprocessed pre viously are input into the network and trained along with the number of network layers and the related parameters adjusted continually.The result shows that the DAE has the ability of excellent noise reduction and robust feature extraction.(2)To solve the problem of low SSIM on DAE,an improved DAE network based on Generative Adversarial Networks(GAN)was proposed.In order to fit flame images,the network architecture,layer classes and parameters were improved,added and modified.The image feature has been extracted by DAE-GAN effectively and then divided into three classes through principal component analysis(PCA)and fuzzy C-means clustering.Comparing with various traditional feature methods,DAE-GAN shows more accurate in feature extraction.(3)The evaluation criteria of image clustering categories is studied.The image feature visualization of grade-weighted class activation map(Grad-CAM)is proposed.The result shows classification standard of flame image was distinguished by the pixel size,position and correlation of the flame color gradient change area,white area,furnace hole edge and wall area as main detection objects.This proves that the rationality and effectiveness of feature extraction by DAE-GAN and FCM clustering.At the same time,the division of detail region based on Grad-CAM also proves that deep learning is irreplaceable in the field of coal burning flame image.(4)A combustion state monitoring and identification method based on multi-signal fusion is proposed.The pressure and temperature near the burner were installed,measured,monitored and extracted for the first time.The signals processed by adaptive Kalman filter put in Long Short Term Memory Network to predict.Then the deviation of original and predicted signals are used to reconstruct the flame stability.As for combustion parameters and flame images,they were calibrated and defined empirically through the classification in advance.Finally,the decision level fusion based on Fuzzy Analytic Hierarchy Process is calculated to obtain the stability index of combustion flame state,and the state evaluation quantification is achieved.