Process Monitoring And Quality Prediction Of Basic Oxygen Furnace Steelmaking Process Based On Functional Data Analysis

The quality of molten steel in the basic oxygen furnace(BOF)steelmaking process greatly affects the properties of steel products.Stable production process and accurate quality prediction are very important to improve the quality of molten steel.With the development of sensor networks,cloud platforms,and 5G communication,massive production data can be obtained to provide sufficient information for the online control of the BOF steelmaking process.However,due to the complex metallurgical reaction mechanism,experiential process parameter control,and large between-batch variations in the production data,the accuracy of the existing control models needs to be improved for abnormal production state monitoring,ending time forecast,and dynamic prediction of molten steel’s quality throughout the steelmaking process.Therefore,this paper is going to be based on the functional data analysis theory and study the process monitoring and quality prediction of the BOF steelmaking process.The main research contents are as follows.(1)Aiming at the problem that the large between-batch variations in the production data conceal the features of abnormal production state and decrease the monitoring accuracy,the Mahalanobis distance-based functional derivative support vector data description method(MD-FDSVDD)is proposed and applied to the online monitoring of BOF steelmaking process.First,functional data analysis is used to smooth the discrete sequences of functional variables and extract their derivative functions,which is in order to remove the between-batch variations among the production data and enhance the features associated with anomalies.Then,the derivative functions are corrected based on Mahalanobis distance to address the strong linear correlation challenge.Finally,the transformed derivative functions are used to estimate the control limit for process monitoring.The production data of BOF steelmaking process validate that the proposed MD-FDSVDD method achieves 80.75%and 82.14%detection rates of splashing anomaly and drying anomaly respectively for the abnormal production process,where the advance alarming time of splashing anomaly and drying anomaly is 30.5s and 25.1s,respectively.The MD-FDSVDD method has better monitoring accuracy than the existing functional monitoring methods.(2)Aiming at the problems of short time margin of ending time forecast and low prediction accuracy on abnormal batches,the functional data analysis-phase plane method(FDA-PP)is proposed and applied to the ending time forecast of BOF steelmaking process.First,the fitting functions and the first-order derivative functions of the total proportion of CO and CO2 in off-gas are extracted based on the functional data analysis and used to construct the functional phase plane.The phase trajectory characterizes the velocity and acceleration of metallurgical reaction in the molten pool,which is out of the viewpoint of metallurgical dynamics.Then,the late blowing period is recognized by estimating a boundary for the stable state of phase trajectory,so providing the starting point for ending time prediction.Finally,the phase trajectory in the late blowing period is predicted and used to further forecast the ending time.The production data of BOF steelmaking process validate that the proposed FDA-PP method achieves 19.31%and 17.01%mean relative prediction error(MRPE)of ending time on splashing batches and drying batches,respectively,which is more accurate than the classical phase space method,the support vector machine method(SVM),and the BP network method.(3)Aiming at the problems that long production cycle and complex production data features increase the difficulties of dynamic prediction of molten steel’s carbon content and temperature throughout the entire steelmaking process,the functional kernel partial least squares method(FKPLS)is proposed and applied to the dynamic prediction of carbon content and temperature of the molten steel during the entire process.First,mixed functional data of scalar and time series is smoothed based on functional data analysis,by which the three-way data matrix of batches × variables× time is transformed into a two-way matrix of batches × functions of variables.Then,the kernel function is introduced to map the independent variable functions to a high-dimensional feature space,and the functional kernel principal components are extracted with the constraint that maximizes the covariance of the functional kernel principal components between the independent variables and the dependent variables.The functional regression model is established using the extracted principal components.Finally,the regression coefficient function is estimated through basis function expansion.The production data of BOF steelmaking process validate that,for the medium phosphorus steel,the proposed FKPLS method achieves less than 15.31%and less than 0.66%MRPE of molten steel’s carbon content and temperature respectively during the entire production process.The prediction accuracy is better than the existing functional PLS(FPLS)method,Gaussian process functional regression(GPFR)method,tensor PLS(TPLS)method,multiway KPLS(MKPLS)method,and multiway PLS(MPLS)method.(4)Aiming at the problems that large variations of process parameters as well as complex dephosphorization and rephosphorization principles make the phosphorus content prediction has larger statistical uncertainty,the functional relevance vector machine method(FRVM)is proposed and applied to the prediction of molten steel’s phosphorus content and its probabilistic confidence interval.First,the discrete sampling sequences of production data are smoothed to continuous functions based on functional data analysis.Next,through the basis function expansion,the regression between independent variable functions and dependent variable function is transformed into the regression between basis function coefficients.Then,based on Bayesian learning,the probability distribution of the regression weight of basis function coefficients is estimated.Finally,the probability distributions of the basis function coefficients of the dependent variable are predicted,then the dependent variable function and its probabilistic confidence interval can be solved.The production data of BOF steelmaking process validate that,for the medium phosphorus steel,the proposed FRVM method achieves less than 18.09%MRPE of molten steel’s phosphorus content during the entire production process,and can provide a confidence interval of phosphorus content with 89.34%prediction interval coverage probability(PICP)under the confidence degree 0.1.Compared with the existing GPFR method,FPCA-RVM method,and multiway RVM(MRVM)method,the prediction results obtained by the proposed FRVM method have lower prediction error and statistical uncertainty.