A Study On Soft Measurement Modeling Based On Nonlinear Partial Least Squares

Soft measurement modeling based on data should solve such problems: limited data, highly correlated variables, data with noise, high nonlinearity and time-varying. This thesis focus on the soft measurement modeling based on nonlinear Partial Least Squares (PLS), which can solve these problems.PLS can eliminate the correlation between variables and harmful noise, and is applicable to process modeling, especially in the cases of too many variables compared to limited samples. In this thesis, the mechanism of selection of PLS's latent variables is explained using singular value decomposition (SVD), and the performance of PLS model is analyzed. The application of PLS model is introduced as well.In this thesis Neural Network/Partial Least Squares (NNPLS) is used for nonlinear soft measurement modeling. A novel regularised hybrid NN training algorithm is used to increase training speed, prediction precision, and improve model's generalization capability. Using such method to build MFR soft measurement model, the off-line result is better than general NN model. A kind of sensitivity analysis method for NNPLS model is also proposed, which can also be used for variable selection and model order determination.To accelerate model on-line modification, a recursive PLS/ Neural Network (RPLSNN) algorithm is also developed. The model structure of RPLSNN combines a linear function with Gaussian units, and recursive Partial Least Squares (RPLS) is used to modify linear parameters. An adaptive scheme is also proposed to add and delete Gaussian units, thus the Gaussian units' response can track the changing of input data distribution. For determining the Gasussian centers, an improved k-means clustering algorithm is developed to distribute centers uniformly over input space and exclude redundant centers automatically. A real data set for MFR soft measurement is used to demonstrate the effectiveness of the algorithm.