Study And Application Of Prediction Technology For Mechanical Properties Of Hot-rolled Strip Based On Data-driven

Recently,China's iron and steel industry is upgrading in an omni-oriented way in the direction of "intelligence,digitalization and informatization".In the long-term production process,iron and steel enterprises have accumulated rich data,which comprehensively reflects the internal relationship between the various links in the steel production process,and has great meaning.It has become the key to upgrade and transformate of current iron and steel enterprises to use industrial data and artificial intelligence technology to establish the corresponding relationship between composition,process and mechanical properties of hot-rolled products and realize the online prediction of mechanical properties in the process of iron and steel production.Under the background of "Made in China 2025" and industrial key technology upgrading,this research is carried out based on industrial data processing,develops high-precision mechanical property prediction model for different sample data by selecting appropriate modeling and optimization algorithm,and finally realizes the industrial application of mechanical property prediction technology for hot-rolled strip.The main contents of this research are as follows:(1)Aiming at the problems of outliers,data redundancy,unbalanced data distribution and over-fitting existing in the modeling of industrial big data,combined the theory of mathematical statistics with the theory of rolling process and machine learning,the data processing methods,such as hierarchical clustering,outliers elimination,data normalization and data equalization,were proposed.Considering the dimension disaster caused by the direct introduction of parameter modeling,the correlation analysis method based on Forward Selection was used to select the important parameters.The results show that the sample data is more stable and shows reasonable regularity after processing.In addition,the selected parameters can effectively simplify the model structure on the premise of ensuring the accuracy of the model.(2)Aiming at the problems of over-fitting and low efficiency in the traditional feedforward neural network,the Bayesian regularization method and particle swarm optimization(PSO)were introduced to improve the traditional feedforward neural network.Based on the improved neural network,the prediction model of mechanical properties of multi-brand C-Mn steel was established.Meanwhile,the mean influence value method and control variable method were used to study the regularity of the model.The results show that the optimized Bayesian regularized neural network has good generalization ability,and the model reflects the laws of physical metallurgy.(3)Based on small dataset of Ti microalloy high strength steel,the Hotelling T2 statistic and upper control limit were calculated to identify and remove the abnormal points in the sample data.The thermodynamic model of titanium carbonitride precipitation was introduced,and the mechanical property prediction model based on small sample Ti high strength steel was established by combining principal component analysis and support vector machine.The results show that the mechanical properties prediction model based on SVM algorithm has high generalization ability compared with the traditional BP neural network.(4)Based on industrial big data processing technology,correlation analysis technology and high-precision mechanical properties prediction technology,a data-driven prediction software for mechanical properties of hot-rolled strip was developed.Based on 1780 hot rolling production line,this software was used to predict the mechanical property of typical steels(SS400?Gr50-11?C610L and C510L)with high precision.