Research On Real-time Prediction Method Of Carbon Content At The End Of Converter Steelmaking Based On Flame Image Recognition

The end point prediction of the carbon content of molten steel is an important part of converter steelmaking.Accurate prediction will directly affect the steelmaking efficiency and help reduce energy and raw material waste.Since the carbon content of molten steel in different proportions in the molten pool can be reflected in the change of information such as the flame color and texture shape of the furnace mouth,the end point carbon content prediction method using the image feature extraction of the furnace mouth provides a new reference for traditional prediction.Flame,as a complex and changing non-structural object,has strong randomness and similarity,which brings great difficulties to feature extraction,which in turn affects the accuracy of endpoint prediction.In view of the above problems,this article will start with the flame color and texture characteristics of the furnace mouth,which are closely related to the change of the carbon content of the molten steel,study the internal structural characteristics of the furnace mouth flame,and extract the distinguishable and robust color from the complicated furnace mouth flame.Texture features can improve the accuracy of endpoint prediction,and provide a reference for the further discussion of the subsequent endpoint carbon content prediction.The main research contents of this paper are as follows:(1)From the perspective of mechanism analysis,explain the changes in the flame characteristics of the furnace mouth with different carbon content of molten steel during the converter steelmaking process;secondly,about 16000 flame images of different heats taken by industrial cameras at the end time are used as the experimental data set,Through the verification experiment proved the randomness,multi-scale and multi-directional characteristics of the furnace mouth flame.Compare the regular texture image with the flame image under the same heat,calculate the characteristic statistical values of histograms in different directions,analyze the random characteristics of the flame image of the furnace mouth according to the change of the statistical values;use the Gaussian scale pyramid to construct the multi-scale flame image Expression,combined with local extrema patterns(LEP)and gray level co-occurrence matrix(GLCM)texture feature extraction methods,and analyze the multi-scale and multi-directional characteristics of the furnace mouth flame image based on the prediction results,It provides guidance for the realization of a color texture feature extraction model with good robustness and high prediction accuracy.(2)Aiming at the random characteristic of the furnace mouth flame,which leads to the difficulty of extracting stable features and affecting the accuracy of prediction,a local texture mutual information feature extraction method is constructed and used for end point prediction.First,calculate the entropy spectrum and amplitude spectrum of the local area under the H,S,and I channels;secondly,pay attention to the complementary information of the local area,and construct the first-order derivative nonlinear mapping of the center point and the second-order derivative non-linear mapping between the neighborhood scales.Linear mapping;finally,multi-trend cross-coding of different mapping relationships is used as the texture feature of the flame image.At the same time,the third-order moment of color is used as the color statistical feature.After KNN regression model training and prediction,the carbon content error is within ±0.02% The accuracy rate of prediction reached 91.1%,and the corresponding relationship with the end point carbon content was established well.(3)Aiming at the problem that the existing image feature extraction algorithm does not make full use of the flame shape information and can only extract relatively limited flame features,this chapter treats the color and texture as a whole,and combines the randomness and multi-scale of the flame at the furnace mouth.As well as multi-directional characteristics,an improved multi-trend binary coded color texture feature representation method(IMTBCD)is constructed.First,the color channel fusion strategy is used to integrate color and texture information to obtain the color texture representation of the flame image;second,the non-uniform multi-scale expression of the color texture is constructed,taking into account the structural information of different scales;finally,IMTBCD has different flame textures from different directions The changes are encoded in multiple trends to describe more comprehensive and detailed texture difference information.The generalized regression neural network(GRNN)is used to predict the accuracy of 95.7% within a carbon content error of 0.02%,and the corresponding relationship between the flame image of the furnace mouth and the carbon content of the molten steel is established more accurately,which provides for the prediction of the end point of the converter steelmaking A certain reference.