Soft-staged Modeling Method Of Batch Processes Based On Kernel Function

Batch process plays an important role in the fields of industrial production. The existence of non-stationary, nonlinear, multi-stage characteristics and some process variables that cannot be measured directly reduces application of advanced control methods and hinders automation level and production efficiency of batch process. So the study of soft sensor technique and the soft sensor modeling for batch process is very important. As an important feature of batch processes, multi-stage is of great influence on soft sensor modeling. Although the duration of transition process is very short, it has obvious difference from that of other period of time, and the division of stage can improve the modeling of multistage soft sensor model and enhance the accuracy of model. Aiming at this issue, this paper developed a soft-transition modeling method based on kernel function.Focusing on the transition between the multi-stage batch process, we proposes a new sliding window weighted MPCA (SWMPCA) method based on the analysis of multi-way principal component analysis (MPCA). This method improves the impact of uncertainty factors such as influence of process noise and multiphase transition process on the division of stage. Studying the soft sensor modeling theory and method, we propose a multi-stage batch process method of least squares support vector machine (LS-SVM) with parameters for the inseparable relationship between the transition process and the operation process. This method can improve the accuracy of global single model precision. Finally, the simulation software of penicillin fermentation process is used to validate the proposed method. The results show that the division of stage method based on SWMPCA can correctly divide the process into different stages, the online monitoring model constructed according to the multi-stage can reduce misdiagnosis rate an emissive judgment rate compared existing method, and the soft sensor modeling constructed by LS-SVM method with weights can improve the accuracy.