Industrial Process Monitoring Using Structured Models

With the rapid development of science and technology,the modern process industry is be-coming more sophisticated and complicated,and the pollutant emission standards set by the gov-ernment are becoming stricter.Process monitoring,as an important means to maintain production safety in the process industry,build product quality assurance system,and enhance social and environmental benefits,has been successfully applied in recent years.For process monitoring,the data-driven method represented by multivariate statistical analysis has attracted widespread attention from researchers due to its simple concept and easy implementation.Although many forms of research results have been achieved,the multivariate statistical analysis still has certain shortcomings in fault isolation and diagnosis.The main reason is that it does not consider the structured information describing the process operation states from the perspective of structure,including the variable correlation,causal characteristics,etc.Thus,the ever-increasing data scale makes it difficult to circumvent the so-called "smearing effect"(abnormal information is prop-agated between faulty and normal variables)for a pure data-driven method.In order to mine the hidden structured information in the data as much as possible,this thesis conducts the research of process monitoring methods using structured models.Based on multivariate statistical analysis,series of graphical models,probability linear discriminant analysis,and so on,the relevant fault detection and diagnosis methods are developed with feature variable selection as the key tech-nique,combined with prior knowledge of process characteristics and system principle.During the design of an approach,the research work of process modeling,fault information extraction,and data visualization is performed.The details can be summarized as follows.1.A model of SJSPCA is proposed to address the "smearing" problem usually encountered by PCA in fault isolation and diagnosis.The model fully mines the correlation of process data using a structured regularization term consisting of a l2,1 norm and a Laplacian regularization term,realizing the structured expression of abnormal process information.The l2,1 norm has an excellent characteristics of consistent selection of feature variables.The graphical Laplacian constraint can mine the potential graphical structure relationship among the data.Both of them act on the loading matrix of PCA,which realizes the structured selection of process variables,and directly generates a prerequisite for the structured expression of abnormal process information.In the stage of fault isolation,the shrinking step deletes the useless information in the loading matrix by row,leaving some non-zero elements to sufficiently characterize the fault information;in the step of isolation,the data recovery matrix of PCA can show the faulty/abnormal information in the corresponding rows belonging to the principal or residual subspace of loading matrix,so that the problematic variables can be identified at a glance2.For the visualization of the relationship between variables in a time-varying process,a SSGL was constructed by combining a sparse Gaussian graphical model,a structured regulariza-tion term,and a data moving window scheme.The structured regularization term in this model includes two l2,1 norms.The first norm regularization is used to obtain the overall change trend of the time-varying process,the other term captures the abnormal events with a sudden in the pro-cess,and both of them work together to mine the graphical structural relationship hidden in the process data.In a unified graphical model-based process monitoring framework,fault detection is performed by testing the equivalence between two graphical model structures.The results of fault isolation can be achieved by observing the changes of a series of test graphical structures.Alterna-tively,more stable and precise results can be obtained by solving the reconstruction optimization problem for an abnormal graphical structure.3.An I-PLDA model is proposed to deal with the problem of large within-class variance in the data,which is problematic for the original model.The proposed model introduces a corre-sponding within-class loading matrix for each class of data,which embeds the category informa-tion of the process samples into the constructed model through an inverse within-class covariance matrix.This improvement presets the model foundation for the mode identification using the cri-terion of maximal cosine similarity and also enhances the performance of fault detection using the conducted statistics.For fault isolation in the same probabilistic framework,the Laplacian priors are adopted for the reconstruction of probabilistic generation model for a faulty sample,produc-ing a structured expression of abnormal information in the changing parts of the corresponding within-class loading matrix by implementing sparse probabilistic inference algorithm.Applica-tions show that the faulty variables can be well localized by sparse probabilistic reconstruction,which effectively circumvents the "smearing" effect encountered by other traditional methods4.A MRPLDA model is proposed to deal with the problem of the industrial process dataset,including outliers,non-Guassianness,and large within-class variance.The construction of this model depends on two key steps:1)Predetermining the prior distributions for the within-class latent variable and data noise to adjust the sensitivity of statistics like data mean and covariance,so that the model can avoid the interference of outliers or abnormal data points in the training dataset;2)Constructing a mixture probabilistic model containing a limited number of RPLDAs to handle with the other problems mentioned above in the framework of mixture probabilistic distributions.Based on the constructed model,two sets of state and class latent variables are introduced to form a mixture probabilistic model classifier,which can be used to identify the process faults.Thus,the task of fault detection and isolation in process monitoring can be structured as the classification behavior of the proposed fault classifier.Meanwhile,a state inference strategy,consisting of probability approximation,evidence inference,and voting based decision,is also proposed.This strategy further enhances the classification ability of the constructed fault classifier.