Dvnamic Process Modeling And Monitoring Based On Canonical Correlation Analysis

Recent years have witnessed the continuous development of Made in China 2025 and In-dustry 4.0.The process industries present the trend of product variety diversification,system scale expansion and production task complexity,which puts forward more rigorous require-ments for real-time operation monitoring of process control systems.Process monitoring tech-niques have played an important role in ensuring process safety,product quality,improving economic benefits,and achieving environmental friendliness,which has become an indispens-able part in the field of process industries.In the meanwhile,with the advancement of control systems and sensing technology,large amount of process data can be collected,transmitted,and stored,which contain rich process information.However,realistic industrial processes often exhibit dynamics with noises and nonlinearity.How to deal with these complex data character-istics and effectively extract information from them to guide the operation of process industry is an urgent topic in the field of process monitoring.This paper focuses on the process dy-namics of data from process industries,aiming at the inevitable noises and natural nonlinearity in realistic industrial processes,and puts forward canonical correlation analysis-based research for dynamic process modeling and monitoring.The main contents of this paper include the following parts:(1)Aiming at the natural process dynamics of data and the inevitable noises problem in realistic situations,a Dynamic Probabilistic Canonical Correlation Analysis model is proposed,which is the dynamic extension of probabilistic CCA.The dynamic relationship between pro-cess data is constructed by multi-view modeling.As a probabilistic latent variable model,it can naturally describe the process noises in realistic process industries,which owns a more general form of process noises compared with traditional methods.The maximum likelihood estima-tion based on the expectation-maximization algorithm is used for parameter learning,in which the model is transformed to simplify the learning task.A fault detection scheme based on this model is also proposed.(2)For capturing the high-order dynamics and handling the problems of local optimum and over-fitting in probabilistic latent variable models,a Variational Bayesian Canonical Variate Analysis model is proposed.The traditional canonical variate analysis is explained through a probabilistic perspective to describe the uncertainty in process industries.Then probabilistic latent variable model is further extended to the framework of variational inference,and the model parameters are treated as random variables so as to alleviate over-fitting and to avoid local optimum by giving a priori.Moreover,high-order process dynamics can be described by data augmentation.And the corresponding fault detection scheme is proposed.In addition,a fault identification scheme based on fault relevance is proposed.The relevance between fault and process variables is measured by information criterion.By transforming classification into regression task,the smearing effect caused by data reconstruction can be avoided.(3)To deal with the nonlinear dynamic nature of industrial processes,a Kernel Dynamic Canonical Correlation Analysis model is proposed.The above mentioned approaches for dy-namic modeling are based on data augmentation.Although it can effectively extract the high-order process dynamics,the scale of model parameters inevitablely increase,resulting in the high storage cost and heavy computational load in the kernel-based model.In order to solve such problem,an improved kernel matrix approximation method is also proposed,which em-beds full-rank decomposition into the Nystrom kernel matrix approximation and essentially re-duces the storage expenses and computational complexity.Moreover,the proposed improved kernel matrix approximation scheme is genetic for kernel learning-based models.In addition,a fault identification scheme based on the above-mentioned kernel dynamic canonical correlation analysis model and improved kernel matrix approximation method is proposed,and the kernel trick is used to calculate the fault relevance for nonlinear processes.