The Study Of Smoke Recognition Method Of Flare Stack Based On Convolutional Neural Network

The flare stack is a special combustion facility that guarantees the safe production of petrochemical plants and refineries.Whether it is sufficient to burn exhaust gas and toxic gas is an important factor affecting environmental pollution.In recent years,the governance of VOCs has been strengthened at home and abroad,which clearly pointed out the importance of the flare stack for the treatment of VOCs and further improved the requirements for the combustion efficiency of the flare stack.Since the conventional flare stack combustion control system is manually controlled by the worker to observe the combustion condition.Affected by the status of the staff,this method is prone to delays in observing black smoke,which leads to the lag of the smoke elimination operation.At present,with the development of industrial intelligence,replacing the manual control with intelligent control to improve the safety and reliability of the industry has become a feasible solution for improving the flare stack combustion control system.In order to solve the shortcomings of the traditional flare stack combustion control system,a convolutional neural network(CNN)model algorithm for identifying the flared flare smoke was designed.The artificial intelligence algorithm was used to monitor the combustion state of the flare stack,thus effectively avoiding the staff's reaction to black smoke.The thesis addresses the demand for the flare stack smoke recognition for the “flare stack combustion efficiency control technology development project” and the actual problem of the complex environmental background exhibited by the images collected at the experimental of the flare stack.The following aspects have been studied:1.Exploring the characteristics of CNNs to extract smoke image features,and designing a classification network for smoke images.At present,most CNN architectures for extracting image features are derived from the research of image classification networks.CNNs applied to image classification can effectively extract image features in the image that match the classification labels.This thesis analyzes the characteristics of smoke feature extraction by CNN model through visualization method,and designs the CNN to classify smoke images based on the visualization research conclusions.Comparing the CNN designed in this thesis with the latest smoke classification CNN algorithm,the algorithm has a significant reduction in network complexity and the accuracy in the corresponding smoke detection database is increased to over 99%.2.Aiming at the traditional image processing method can not effectively solve the problem of image smoke segmentation under complex backgrounds such as clear,cloudy and sky,this thesis designs two full CNN for image segmentation of flare stack smoke.The idea of segmentation model mainly focuses on the shallow features of classification model verification including a large number of smoke features,and then designs the shallow full CNN and the multi-scale feature reuse full CNN.The smoke segmentation algorithm designed in this thesis is compared with the general CNN image semantic segmentation algorithm.The effect is optimal.In addition,the complexity and execution efficiency of the algorithm model meet the industrial needs.3.Due to the fact that in actual engineering,the flare stack image storage occupies a large amount of resources,and the downsampling method is used to reduce the image storage space.To this end,the research on the super-resolution technology of the flare stack image can effectively restore the valuable smoke area in the image and enhance the visual perception and algorithm recognition ability.In this thesis,using the full smoke image patches as a training sample and by the design experience of classification CNN,a super-resolution method for the flared flare area is designed.Compared with the traditional super-resolution algorithm and the typical CNN algorithm,the proposed algorithm has higher effect in the smoke area.4.Flare stack combustion monitoring system software and black smoke monitor.In order to prove the practicability and effectiveness of the research content in practical engineering projects,two application methods for verifying the smoke segmentation algorithm based on CNN are designed.One is an interactive interface software built on the Windows system through the MFC framework.The Hikvision camera is called using the official SDK,and the smoke segmentation network model trained by the Keras is used to evaluate the black smoke area of the frame image.The second is the black smoke detector.By deploying the smoke image segmentation model to the Raspberry Pi 3B development board,the Hikvision camera is accessed through the RSTP protocol to obtain the current time image for real-time image smoke region segmentation.Both application methods can verify the practicability of the proposed algorithm and provide an effective smoke recognition function for flare stack.