Latent Variable Regression Modeling And Applications For Industrial Process Data

It is quite important to quality control and operating performance in the process industry that the key performance indices such as product quality variables are measure timely.The research topic of this thesis is focused on regression modeling and soft sensor application for prediction of key performance indices in the process industry.Principal component regression,Partial least squares and independent component regression are three of the most popularly latent variable modeling methods that have been used for regression modeling in the past years.In order to address drawbacks of those latent variable models,we have carried out some improvements on soft sensor applications of those methods,the main contributions of this thesis are summarized as:Due to the shortcoming of principal component regression in nonlinear data modeling,a linear subspace based method is proposed.Since the principal components extracted by the principal component analysis method are orthogonal to each other,the diversity of linear subspaces which are constructed through those principal component directions can be well guaranteed.A variable contribution index is defined for variable selection in each linear subspace.For online prediction of a new data sample,a Bayesian probabilistic combination strategy is developed for results integration from different linear subspaces.Based on two industrial examples,online soft sensing performance can be improved by the proposed method.In order to provide a probabilistic viewpoint for the latent variable model,the traditional partial least squares model is extended to the probabilistic form.The Expectation-Maximization algorithm is introduce for parameter learning of the developed model.For complex modeling,the basic probabilistic PLS model is extended to the mixture form.Besides,a new semi-supervised form of the probabilistic PLS model is also developed for training those datasets which include both labeled and unlabeled samples.With the incorporation of additional unlabeled data samples,regression performance of the probabilistic PLS model is improved.Compared to other two latent variable models,the structure of the independent component regression model is not quite stable,due to its initialization step.In order to improve its robustness,a two-level independent component regression model is developed and applied for soft sensing.By incorporating Bayesian combination strategy,the prediction results generated from different subspaces through the directions of independent components are effectively integrated together.Based on simulation case studies on two benchmark spectral datasets,the prediction performance of the proposed method is improved,compared to existing methods.Ensemble learning is introduced for performance enhancement of those three different types of latent variable models.Through the bagging strategy upon three different model structures,various sub-models are developed for the same regression problem,the performances of which are complementary to each other.The least squares algorithm is used for parameter optimization of the ensemble learning strategy.Based on the simulation case study on an industrial process,the prediction performance can be improved the ensemble learning method.