Research On Prediction Method Of FeO Content In Sinter Based On Deep Learning

The iron and steel industry is of great significance and is the material basis for the development of the national economy.Due to the lack of high-quality iron ore resources in our country,sinter is the main raw material for iron and steel production.The FeO content of sinter is an important factor determining the physical and chemical properties of sinter.In order to improve the quality of sinter production,it is necessary to monitor the FeO content of sinter for feedback control.Due to the long process flow,the sampling test does not meet the real-time requirements,and because the sintering process is complex and changeable,it is difficult to predict through mechanism modeling.Data-driven deep learning does not require complex artificial feature design based on mechanisms,and has powerful representation learning capabilities due to multi-layer nonlinear changes.Therefore,this paper analyzes the sintering process mechanism and characteristics,so as to select the production parameters related to the FeO content of the sinter and infrared sectional images of the sintering machine tail,and uses the deep learning model to extract the image features and predict the FeO content in sinter.The main work is as follows:(1)For the sintering production parameter data with poor quality and complex relationship,this paper first cleans the missing values and outliers by statistical analysis method,and then calculates the Pearson correlation coefficient and distance correlation coefficient between the data to remove redundant and irrelevant features,and finally uses standardization and decision tree-based discretization to obtain continuous and discrete features of production parameters,respectively.(2)For the sectional image data of the sintering machine tail,it is difficult to obtain the corresponding FeO content.In view of the problem that the model is easy to overfit due to insufficient training samples,this paper establishes a convolutional neural network Dense Net with high accuracy and few parameters,and then adopts transfer learning and semi-supervised learning method Fix Match to learn effective model parameters,and finally uses the rest of the Dense Net except for the classification layer to extract image features.(3)In order to achieve accurate and stable FeO content prediction,this paper concatenates parameter features and image features as input,and proposes Shortcut Connection Neural Factorization Machine as the prediction model by adding attention mechanism and shortcut connection to Neural Factorization Machine.The optimal hyperparameters of the prediction model are obtained through a large number of controlled experiments and cross-validation.In summary,the proposed method for predicting FeO content in sinter achieves 1.139% mean relative error on the test set.The results show that the method is feasible and effective,and can accurately predict the FeO content in sinter,thereby providing accurate guidance information for optimizing the sintering quality.There are 29 figures,4 tables,74 references...