Prediction Of Mechanical Properties Of Hot Rolled Strips Based On Support Vector Quantile Regression

As a kind of engineering structure material with high strength and better comprehensive performance,steel is widely used in modern industry.In order to meet the social demand and adapt to the market competition,modern steel enterprises must constantly improve the organizational performance of steel products.Mechanical properties,as one of the key quality indexes of steel,directly affect the use value of the product.Therefore,it is necessary to change the previous expensive and time-consuming physical test methods and complete the design of steel products with the help of advanced digital design tools.The prediction of mechanical properties of hot-rolled strip steel can be used for on-line dynamic control of product performance,optimization of steel composition and design of new products.The mechanical properties of steel are affected by many factors and there are complex interactions.Therefore,this paper will take the tensile strength of steel strip as the response variable,according to the composition parameters(chemical elements),carbon and nitrogen compounds and their process parameters obtained from the actual production line,to study how to reasonably screen the main factors affecting the prediction model of mechanical properties of steel,and establish an effective prediction model.Support Vector Machine(SVM)can solve the nonlinear problem between variables.However,Quantile Regression(QR)can well deal with data heterogeneity,is insensitive to outliers,has good "robustness",and can consider the whole distribution and other characteristics.Therefore,in this paper,SVQR model of support vector machine(SVM)and fractional number regression(QR)theory was established to explore the relationship between tensile strength of strip steel and component parameters(chemical elements),carbon nitride compounds and process parameters.First,we give the loss function and prediction model of SVQR,and the iterative least squares fitting method(IRWL)that can solve all parameters simultaneously.Secondly,we preprocess the existing data,mainly to standardize the data.Finally,according to the key influencing factors selected from the existing literature,we randomly selected 70% of the data set as the training set,trained the SVQR model,and predicted the test set composed of the remaining 30% data.The results show that,compared with the prediction methods of RR(Ridge Regression),SR(Stepwise Regression)and SVR(Support Vector Regression),the prediction error of the method we used is minimal.In addition,we added LASSO variable selection to the SVQR model to select the factors that have a greater impact on the response variables,and based on this,we established the SVQR prediction model.Compared with FS and RF variable selection methods,our error was also minimal.The validity of SVQR model is further verified.