Research On The Construction Of Soft-sensing Model Of Hydrocarbon Flame Equivalence Ratio Based On Machine Vision And Machine Learning

Energy shortage and environmental pollution are the hot issues of social concern and scie ntific research.In recent years,the rapid development of industry promotes the economy to a new level,but also makes the environmental problems increasingly prominent.As the main for m of energy supply in china,burning fossil fuels plays an important role in improving energy utilization and reducing pollutant emissions.Among them,the combustion efficiency of gas fu el is greatly affected by the change of equivalence ratio to a certain extent.The measurement and control of equivalence ratio can not only effectively improve the fuel utilization and comb ustion efficiency,but also help to reduce the emissions of NO?x,unburned hydrocarbon and p articulate matter,so as to achieve efficient and clean combustion.Studies have shown that the equivalence ratio is closely related to the chemiluminescence of free radicals in the combustion process of hydrocarbon flame.The equivalence ratio can be obtained indirectly by measuring the concentration ratio of C₂^*toCH^*.The chemiluminescence of flame free radical is a phenomenon of the emission spectrum when the excited group has e nergy level transition in the combustion process.Its luminescence color is the characterization form of the emission spectrum of free radical and contains a lot of information that can reflec t the combustion state.Therefore,through the relationship between flame color emission spectr um free radical equivalence ratio,a soft sensing model of equivalence ratio based on flame co lor can be established to reflect the flame combustion state.The traditional equivalence ratio detection method based on color modeling usually uses a single color variable to fit the functional relationship between combustion equivalence ratio an d color variables,however,due to the different effects of the scattering of the optical system and the spectral response of the image sensor,The soft sensor model constructed only by a si ngle variable often has large measurement error.Therefore,this paper proposes to establish a multi input/single output color equivalence ratio model by using machine learning method,w hich takes the multi-color features of flame collected by machine vision as the model input.Firstly,this paper takes methane flame as an example to collect the image data of methan e premixed flame under different working conditions,and preprocess the image,including abno rmal data elimination(fuzzy flame image)and image filtering,the digital flame colour distribu tion(DFCD)technology can separate the yellow red flame from the blue flame,and finally o btain the target area of the flame image,namely the region of interest(ROI).Secondly,the ratio of each color component in different color space models of flame is d esigned as feature,and the first-order color moment,second-order color moment,third-order co lor moment,maximum value,minimum value and standard deviation of different features are e xtracted.The feature selection is realized based on the combination of wrapper and filter,and finally the high-quality features reflecting the combustion state of flame are obtained.Finally,on the basis of feature engineering,the soft sensing model of methane premixed f lame equivalence ratio is established based on BP neural network,elm algorithm and SVM ma chine learning algorithm.The characteristics of different algorithms are compared and analyzed.The improved SVM algorithm is applied to propane premixed flame equivalence ratio soft se nsing modeling to verify the feasibility of the proposed method sex.