Study Of Flame Image Recognition And Combustion Diagnosis Method Based On Data Driven

Combustion is an essential way of fossil fuel utilization,widely applied in energy and power,metallurgy and chemical,aerospace and other fields.The on-line and accurate diagnosis of combustion state is of great significance to prevent potential safety risks and improve economic performance.The combustion diagnosis method based on flame imaging has the advantages of non-invasiveness,anti-interference and high sensitivity,presenting broad application prospects.This thesis aims to establish a data-driven method for flame image recognition and combustion state diagnosis,and the flame imaging technique,digital image processing technique and artificial intelligence technique are combined to identify and predict the combustion condition category,combustion pollutant emission and combustion stability.Experiments are performed on a 4.2MW rotating-cup heavy-oil combustion system.Firstly,the advance of combustion state diagnostic techniques is summarized,the direct and indirect combustion diagnosis methods are introduced,and the combustion diagnosis techniques based on the flame image are emphatically elaborated.Then,the basic theory and design of the data-driven method are presented in details,and the basic knowledge involved in flame image feature extraction and feature analysis is summarized.Fuerther the performance evaluation index of the prediction model is described,which provides technical support for data-driven flame image recognition and combustion state diagnosis.A semi-supervised learning model based on denoising autoencoder,generative adversarial network and Gaussian process classifier is proposed to identify the combustion conditions of the heavy-oil combustion system.Through the combustion test,the flame images of the 15 combustion conditions created at three operating loads are collected to establish a"combustion conditions-flame images"dataset.The flame image features are extracted by the denoising autoencoder with a deep structure,and sent to the Gaussian process classifier to identify the combustion conditions.Experimental results show that the denoising autoencoder improves its feature learning ability by using the adversarial training mechanism,and accurately extracts the robust features of low-quality flame images,thus solving the problem of image noise interference.Compared to the traditional neural network models such as autoencoder and support vector machine,the proposed semi-supervised learning model has higher recognition performance,with a recognition accuracy of 97.5%.After fine-tuning the Gaussian process classifier,the semi-supervised learning model can achieve the recognition range expansion,with a recognition accuracy of 95.19%for new conditions,showing strong generalization ability.Aiming at the prediction of pollutants（NO_x and CO₂）emissions from the heavy-oil combustion system,an ensemble learning model consisting of stack denoising autoencoder,artificial neural network,support vector regression,least square support vector machine,extreme value learning machine and Gaussian process regression is proposed.The results show that the stack denoising autoencoder can effectively extract image features for pollutant emission prediction,and its performance is better than classical feature extraction methods such as principal component analysis and stack autoencoder.The ensemble learning model takes advantage of four single prediction engines（including artificial neural network,support vector regression,least square support vector machine and extreme value learning machine）,thereby reducing the risk of model structure design and hyperparameter selection and significantly improving prediction performance,with the NO_x prediction accuracy of R²=0.97 and the CO₂prediction accuracy of R²=0.96.In addition to high-precision point prediction,the ensemble learning model can also generate the confidence interval information,which is used to quantify the prediction uncertainty and increase the reliability of the prediction results.Finally,a combustion stability monitoring method based on stack sparse autoencoder is proposed.A classification model and a regression model used for qualitative and quantitative prediction of combustion stability are established,and these two models are experimentally verified on the ethylene diffusion flame.The results show that the image features extracted by the stack sparse autoencoder,after being processed by cluster analysis and statistical analysis,can generate qualitative and quantitative stability labels of a single image,thereby forming an image label production method.The classification model can construct the nonlinear mapping between image features and qualitative labels（0 or 1）,used to identify the stability category of a single flame image.The regression model can construct the nonlinear mapping between image features and quantitative labels（0 to 1）,used to estimate the stability degree of a single flame image.In order to verify the applicability of the combustion stability monitoring method in different combustion devices,it is applied to the heavy-oil combustion system.Further the results show that the established classification model and regression model can accurately monitor the combustion stability under different combustion loads,and the best operating conditions of the heavy-oil combustion system are recommended based on obtained results.In summary,the data-driven flame image recognition method proposed in this paper can realize the recognition of combustion conditions,the prediction of combustion pollutant emission concentration,and the diagnosis of combustion stability.It has superior performance in robustness,generalization ability and response speed,providing important support and guarantees for online combustion optimization and adjustment.