Parsimoniously Modeling For Soft Sensor Based On Feature Subspace Regression

Nowadays, there are many correlated secondary variables of soft sensor for complex objects, which make poor performance, low generalization and bad stability. Aiming at the problem, after the discussion of some feature extraction and regression methodology several parsimoniously modeling methodology are presented based on Feature Subspace Regression, such as PCA+NN, PCA+SVR, KPCA+NN, KPCA+SVR, KPLS (KCCA+LS), etc. Meanwhile, the methodology can solve the modeling problem of soft sensor with seldom prior knowledge.It firstly uses feature extraction method, as PCA (Principal Component Analysis) CCA (Canonical Correlation Analysis) or KPCA (Kernel PCA) or KCCA (Kernel CCA) to extract the input matrix of second variable, obtaining the linear/nonlinear uncorrelated sample data. On the basis, the SVR (Support Vector Regression) with SRM (Structure Risk Minimization) or NN (Neural Network) or LS (Least Squares) with ERM (Empirical Risk Minimization) are utilized to realize the structure identification of information fusion part. In this way, performance, stability and generalization of soft sensor are improved.Furthermore, the accuracy and the complexity of the modeled plant are estimated by AIC (Akaike Information Criterion). The smaller AIC value is, the better modeling structure is. In order to verify the effectivity, three types of soft sensor with linear dynamic plant, nonlinear static plant and nonlinear dynamic plant are studied with different parsimonious methods. Simulation results show that KPLS is good for its high accuracy, PCA+SVR method is most effective when solving soft sensor modeling with linear plant, and to nonlinear plant soft sensing problem, KPCA+SVR method is more excellent.On the basis, a new type of sleeping fidget soft sensor is presented and also studied as the modeling problem. Compared with some parsimonious modeling methods, PCA+NN is evaluated best with AIC criterion of high accuracy, good generalization and low complexity to this problem.