Research On Soft Sensing Of Distillation Process Based On Data Imputation

In the distillation process,there are some important quality variables that cannot be measured online.It is meaningful for the regulation of production,but there are fewer real-time measurement methods at present.There will be a big lag in the manual sampling test and human factors will lead to some uncertainty.The soft sensing method is studied in the thesis that to solve the real-time monitoring problem of aviation kerosene quality index for the refinery distillation process.The specific research contents are as follows:The number of samples in the industrial rectification data set is spares,and it is liable to occur for the problem of data missing due to various disturbances.For both random missing and continuous missing,this thesis mainly uses a data imputation method for distillation process based on extreme learning machine.The simulation results demonstrate that it can restores the effective information carried in the original data effectively so that it provides a research foundation for soft measurement modeling under the condition of partial sample missing.This thesis proposes an improved LASSO variable selection method called EC-LASSO to solve the problem of the multiple correlations among the independent variables in the original data.EC-LASSO use the ensemble coefficients to sort the importance of the variables and then add the variables into the models by order.The best subset can be obtained according to the root mean square error criterion.The simulation results of the proposed method are verified by two sets of simulation experiments,which further demonstrates the effectiveness of the proposed method.And it also shows that this method has guiding significance in variable selection for modeling.The multiple kernel learning least squares support vector machine is combined with variable selection method to build a soft sensor model for selected variables.The new kernel function is constructed by using radial basis function and polynomial kernel function so that the advantages of this two kernel function are integrated.The simulation results illustrates that the multiple kernel learning least squares support vector machine can get a more accurate prediction results.In order to achieve the purpose of the online display of soft sensing prediction results,the WinCC software of Siemens is used to build the configuration interface,and the OPC is used to realize real-time communication between WinCC and Matla.