The Study Of Combustion State Detection Method Of All-weather Torch Based On Convolutional Neural Network

During the combustion process of the flare stack,there is a problem of excessive emission of flare tail gas（flammable VOCs）,which causes major production safety accidents and seriously pollutes the atmospheric environment.Efficient combustion of flare tail gas helps to ensure safe production and reduce environmental pollution.Controlling the flow of combustion-supporting steam is the key to efficient combustion of flare tail gas.At present,most petrochemical enterprises use manual methods to regulate the flow of combustion-supporting steam to promote efficient combustion of the flare stack.Manually observe the combustion state of the torch and manually adjust the combustion-supporting steam flow.This method has problems such as poor control accuracy,low stability,untimely manual response and easy fatigue errors.In order to solve the above problems,this paper proposed an image-based non-contact flare combustion status detection method,based on convolutional neural networks to detect all-weather flare combustion status,and designed a flare high-efficiency combustion control system to intelligently regulate the combustion-supporting steam flow.When the abnormal combustion state of the flare stack is detected,increase the combustion-supporting steam flow to eliminate black smoke,thereby protecting the atmospheric environment.When the low-efficiency combustion state of the flare stack is detected,reduce the combustion-supporting steam flow to save energy.When the high-efficiency combustion state of the flare stack is detected,maintain the current flow of combustion-supporting steam to maintain the high-efficiency combustion of the flare.This paper takes the burning state of the flare stack as the research object,mainly researches the detection algorithm of the burning state based on the convolutional neural network,and designs the high-efficiency combustion control system of the flare stack based on the detection results.The specific research work is divided into the following parts:（1）The research on the qualitative detection method of flare stack combustion state based on siamese network.Aiming at the problem of difficult real-time detection or high detection cost,the non-contact detection method of flare stack burning state during daytime based on visible light images is studied,and a flare stack combustion status detection algorithm based on siamese network is proposed.By learning the similarities between the input samples,comparing the similarities to achieve qualitative detection of the flare stack burning state during the day.（2）The quantitative analysis of black smoke in abnormal combustion state of flare stack based on image segmentation.Aiming at the difficulty of the flare control system to quickly,stably and accurately regulate the combustion-supporting steam flow with the qualitative detection results of the combustion state,a quantitative detection method for the black smoke of the abnormal combustion state of the flare stack is studied,and a quantitative analysis model of flare stack black smoke based on image segmentation is proposed.The quantitative measurement problem is converted into a qualitative recognition problem.The image segmentation convolutional neural network is used to classify the black smoke pixel by pixel.The classification result is mapped into a series of continuous values through the black smoke degree evaluation function to complete the evaluation,and quantitative detection of black smoke in the abnormal combustion state of the flare stack is realized.（3）The research on the detection method of night flare stack combustion state based on migration learning.Aiming at the problem of the lack of detection methods for the burning state of the torch at night,a non-contact detection method based on infrared images is studied,and a method of night flare stack burning state detection based on migration learning is proposed.The general edge and low-level features of the structure in the computer vision task are learned through feature extraction,the weights of the convolutional layer are fixed,and the weights of the fully connected layer are retrained to solve the problem of small samples of infrared images,and then realize the detection of the burning state of the flare stack at night.（4）The design of the high-efficiency combustion control system for the flare stack.Aiming at the serious uncertainty of flare gas flow and composition,and the difficulty of precise control of process operating variables,the independent control system of high-efficiency combustion of flare is studied.The PLC controller is used as the core of the system,and the visible and infrared light high-definition cameras are used to obtain pictures of the flare stack burning scene.The image processing program is used to detect the burning state of the flare stack.The burning state is used as the feedback variable of the control system.Then combining with feedforward control,the flow of combustion-supporting steam is adjusted to achieve high-efficiency combustion of the flare stack,which is verified on the project site.