Research On Carbon-hydrogen Flame Equivalence Ratio Model Based On Machine Learning And Data Fusio

As one of the key research issues in the field of combustion science,energy conservation,and emission reduction are the challenges that need to be addressed to achieve the“double carbon”goal for China’s future sustainable development.Currently,more than eighty percent of China’s energy supply comes from the combustion of fossil fuels.Meanwhile,despite the rapid development of clean energy,it is difficult to replace fossil fuels as the main source of supply in the short term.Natural gas is a relatively clean fossil fuel whose main products during combustion are carbon dioxide,water,etc.Methane（CH₄）,the main component of gas,is one of the basic hydrocarbons.And among the variables that can be measured,the equivalence ratio（Φ）is one of the important parameters that affect its combustion state and pollutant emissions.Combustion is defined as a violent chemical reaction process with the release of light and heat.During the reaction process,the excited radical could release the optical signal at specific wavelengths when it came back the ground level.It was demonstrated that the chemiluminescence intensity ratio of CH^*and C₂^*can be used to indicateΦ.It was further found that,in the RGB model,the varying trend of average intensities of blue（B）and green（G）channels of flame images resembles that of CH^*and C₂^*chenil-uminescence intensities.Therefore,in the field of combustion diagnostics,B/G tends to be used to indicateΦas the colour marker of CH^*/C₂^*.However,flame colour is the integration of the chemiluminescence emitted within the visible wavelengths,which has limited spectral resolution.In addition,there are some difference between the colour response of flame obtained from different colour cameras due to their different optical characteristics.This has limited the practical appli-cation of colour-basedΦmeasurement.In this thesis,the conventional chemiluminescence-based and colour-basedΦmeasurement were improved using machine learning algorithms,and the errors caused by these two measurement methods were analyzed.Then,a data fusion-based color image simulation method was proposed to investigate the effect of imaging process on colour-basedΦmeasurement,and various algorithms were proposed to minimize these effects.Finally,considering the limitations of digital colour cameras that the spectral resolution is relatively low,a machine learning-based flame chemiluminescence spectrum reconstruction method was proposed,which attempts to estimate the chemiluminescence emission intensity from digital colour images.Firstly,on the basis of the conventional chemiluminescence-based flameΦmeasurement,the correlation between flame chemiluminescence andΦwas quantitatively analyzed using Pearson correlation coefficient,and the chemilumine-scence emission,which has a relatively high linear response to the variation ofΦwas determined as the input.Based on it,various machine learning algorithms were used to construct the improved chemiluminescence-based flameΦmeasurement model.Secondly,according to the results of the above quantitative analysis of the correlation between chemiluminescence emission intensity andΦat different wave-lengths that several chemiluminescence within the visible wavelengths shows a high linear response to the variation ofΦ.It was also found that several chemiluminescence within R band has a relatively high linear response to the varyingΦ.Based on the investigation,this thesis proposed to use the average intensity of R channel to construct the new colour-based equivalence ratio indicator.Then,various machine learning algorithms were used to establish the improved colour-basedΦmeasurement model using these colour features.Then,in order to quantitatively investigate the effect of the imaging process on colour-based flameΦmeasurement,a data fusion-based colour image simulation method was proposed by integrating the hyperspectral images with the colour camera spectral sensitivity functions.Based on these simulated images,the effects of imaging precess were analyzed and various algorithms were used to minimize these effects.Finally,considering the limitations of digital colour cameras that the spectral resol-ution is low,this thesis proposed a flame chemiluminescence spectrum reconstruction method using machine learning algorithms.Based on the reflectance reconstruction method,we obtain the optimal colour features by investigating the correlation between flame chemiluminescence and flame colour under varyingΦconditions,and various machine learning algorithms were used to construct the flame spectrum reconstruction model to estimate the chemiluminescence emission intensity at different wavelengths.