Study On The Prediction Model Of FeO Content In Sinter Based On Infrared Image Of Machine Tail

The intellectualization of the manufacturing process is the main development direction of China’s steel industry at present.Sintering is an important process in the steel production process,and its product quality index plays a key role in the stable operation of blast furnace,energy saving and carbon reduction,and cost reduction and efficiency.The Fe O content in sinter is one of the comprehensive indicators,and the real-time detection of Fe O content is of practical significance to realize intelligent and optimal control of the sintering process,improve the stability of Fe O content in sinter,and reduce the process energy consumption of the whole ironmaking process.In this study,firstly,by studying the infrared image sequence at the machine tail,a method to select the best cross-sectional image within a single discharge cycle by calculating the average temperature change rate between frames in a specific temperature zone is proposed.Then,combining the sintering process theory with digital image processing technology,12 image feature parameters related to the sintering process state are extracted,such as the area-rate of molten zone,the area-rate of high temperature zone,etc.On this basis,the gradient boosting decision tree(GBDT)model with image feature parameters as input,the GBDT model with sintering state parameters and image feature parameters as input,and the convolutional neural network model(VGG-16,Res Net-18)with cross-sectional images as input are developed,respectively.For the characteristics of the cross-sectional image,Res Net-18 is improved in terms of convolution kernel and downsampling block.By training and testing the model with the data collected from the field,the prediction results show that the performance of the single-input GBDT model can be improved extensively by the introduction of the sintering state parameter.Prediction models using cross-sectional images as input are superior to those using image feature parameters and state parameters as input,in which the Res Net-18 model has a better fitting effect.The Fe O content prediction model of sinter built with the improved Res Net-18 has a hit rate of 91.2% and an RMSE below 0.3%.A system for prediction of Fe O content of sintered ore was developed using Visual C#.The preview of the video stream at the machine tail,monitoring of key image feature parameters and Fe O content prediction results are realized through modular programming to provide the necessary analysis results of the sintering process for the field operators.