Multimode Soft Sensor Modeling Based On Higher Order Statistics Mode Identification

The industrial process is limited by the expensive measuring instruments or the harsh measuring environment,and it is difficult to measure the key quality indicators in real time and accurately.The establishment of a function mapping relationship between auxiliary variables and dominant variables to realize the soft sensor modeling method for predicting unknown dominant variables is an effective solution to the above problems.The soft sensor modeling method requires that the modeling data and real-time data satisfy the independent and identical distribution assumption,and the normal process data for modeling comes from a single stable production condition.In the actual industrial process,the changes in working conditions make the production process exhibit multi-working conditions and multi-modal characteristics.Under multiple operating conditions,real-time data cannot select modals with similar data distribution for modeling and prediction,which will cause deterioration of the performance of the soft sensor model.Reasonably divide the modals from the production process of multiple operating conditions and select corresponding modal modeling An effective means to improve the prediction effect of modeling.To this end,based on modal identification,this paper studies a class of multi-case soft sensor modeling based on high-order characteristic modal identification.The main research contents of the full text are as follows:(1)Aiming at the multi-modal process including transitional modes,in order to better reveal the running state and distribution change law of the multi-modal process and improve the subsequent modeling accuracy,a maximum mean discrepancy(MMD)is proposed The multimodal process transition modal identification method introduces a sliding window to cut the data,uses the maximum mean difference to measure the local data distribution difference,and performs modal identification on offline multimodal data.(2)Aiming at the problem of multi-model soft sensor model misalignment under multi-modal transition process,based on the effectiveness of relative entropy in measuring process data distribution during process monitoring,a relative entropy indicator is introduced in the sliding window to measure the relative Entropy is used to identify stable and transitional modes,and after the support vector regression model is established for historical modes,the "switching" mode is used to predict the output.(3)In view of the misalignment of the soft sensor model prediction caused by the inconsistency between the modeling data and the real-time data,two transfer learning methods of geodesic flow kernel and subspace alignment are introduced to adapt the subspace after the dimensionality reduction of the principal component analysis Then,the working condition data after the domain adaptation is projected into the partial least squares latent space,and finally the latent variables are modeled by support vector regression.