Research On Process Monitoring Method Based On Subspace Separation And Priori Knowledge

The monitoring method based on data driven has been paid more and more attention in the field of control,and many scholars have made remarkable achievements in the study field.For the nonlinear,nongaussianity and multi-stage features of batch processes,this paper presents a multi-phase batch process monitoring method based on subspace separation.Because there being a small number of labeled samples with containing latent information to have a certain guiding role for fault detection,this paper presents a process monitoring method based on prior knowledge,and the method can improve the generalization ability,comprehensive performance and accuracy of process monitoring model.Specific studies are as follows:(1)In order to solve the problems that exist in the multi-stage batch process monitoring,this paper presents a multi-phase batch process monitoring method based on subspace separation.Firstly we do preprocess the batch processes data,and then we can detect the variable change of the phase transition process by analyzing the relationship between phases,then we do decompose the subspaces for each phase,so subspaces of different changes can be extracted,finally we establish a comprehensive monitoring system to monitor the various stages of change by KICA.Through the experiment we can see the process monitoring effect of the whole monitoring system is very good.(2)To explore the latent information of labeled data,this paper puts forward process monitoring based on the prior knowledge.The detection method uses the local regularization to make the optimal solution ideal,but local regularization can also cause inadequate learn problems,so we can also use global regular to revise our monitoring system.Experiments show that the process monitoring method based on prior knowledge not only is feasible,and its process monitoring precision is very high.Through the experiment,we can also know the the process monitoring effect of the experiment depends largely on of proportion of the labeled sample and model parameters.