Robust Monitoring Of Industrial Process With Data-driven Methods

With the advent era of industrialization 4.0,the modern industrial automation systems are becoming more and more complex,informatic as well as intelligent.As the key technology to ensure product quality,process safety and the external environment-friendly production,process modeling and monitoring have become important part of modern industrial systems which should also be indispensable.In general,the process monitoring procedure can be divided into the forecasting process of production quality as well as estimation of the running status of the process equipment.Practically,the complexities of the process data entries,process variables and the system itself bring a big challenge to the robustness issue of the existing traditional data-driven based modeling and monitoring methods.Among them,data complexity refers to outliers,missing values,and big data volume;variable related complexity mainly refers to non-Gaussian,dynamics and time-varying characteristics.In addition to the complexity of data and variables,the complexity of the process system itself mainly refer to the large-scale of plant-wide systems as well as the regular requirement of maintenance and locally process update.Aiming at these issues,this dissertation proposed a systematic set of robust methods:(1)For the modeling data outlier and process multimode issue in production quality prediction,the mixture of robust principal component regression method is proposed.The Student's t distribution is utilized so that the model is more adaptive and robust against modeling outliers.Furthermore,for the multirate sampling issue of quality-related variables,the semi-supervised learning mechanism is incorporated into the robust model.The obtained semi-supervised robust model can deal with modeling data outlier as well as the multirate sampling rate issue of quality variables.(2)For the modeling issue of process data outliers,missing data and multimodality,we propose a mixture of robust probabilistic principal component analysis method.The traditional mixture of probabilistic principal component analysis is modified under the framework of t distribution.In addition,a partial learning mechanism is developed for model updating with missing data.Finally,monitoring statistics and then the Bayesian fusion method are proposed for monitoring global process with the proposed robust method.(3)Industrial processes usually behave dynamic characteristics.For this issue,two robust and dynamic modeling programs have been proposed for process fault monitoring and fault classification respectively.Firstly,we propose a robust linear dynamic system method based on t distribution.The Bayesian approach is developed for modeling procedure:latent status reasoning is presented in the Kalman filter framework to estimate the hidden state space and then parameters are deduced based on the variational Bayesian learning programs.During process monitoring,statistics are presented based on the dual Gaussian mixture construction method and can effectively deal with the construction of robust statistics for fault detection.For robust fault classifier modeling,we propose a hidden Markov model driven robust principal component analysis model,the static mixing indicator is imposed on robust probabilistic principal component analysis and then are extended by the first order Markov property for the expression of time series data.Compared with the traditional hidden Markov model,robustness has been enhanced,moreover,the proposed method can perform better on the dynamic process for fault classification.(4)Apart from the robustness against data quality,one should also consider the robust modeling of variable situations when Gaussian and non-Gaussian behaviors have been implicitly mixed.Traditional independent component analysis(ICA)focus on extracting non-Gaussianity while principal component analysis(PCA)focus on the Gaussian information extraction.Due to the absence of robust mechanism for representing information,both of them could have performance degradation for the Gaussian/non-Gaussian mixing modeling environment.This work proposes the probabilistic ICA(PICA)which can deal with the simultaneous extraction of Gaussian and non-Gaussian information.Since the residuals contain significant information,the probabilistic PCA(PPCA)is used for further analysis.The proposed two stage data extraction by PICA-PPCA should be more robust against Gaussian/non-Gaussian modeling situations.(5)For the non-robust issue of traditional methods against monitoring multimode time-varying processes,the recursive mixture factor analysis with robust mechanism is proposed.To realize the recursive updating,the recursive mixture factor analysis is induced under the general framework of recursive expectation maximization,then the Bayesian scheme is further proposed for effective and efficient component selection and model updating.For the performance degradation issue during monitoring time-varying processes,the robust monitoring and updating mechanism is proposed accordingly.Therefore,one can still conduct reasonable monitoring and updating even if some new operating modes as well as process varying factors have been occurred.In this way,the robustness against time-varying behavior has been improved.(6)For the large-scale as well as regular local updating properties of plant-wide process,the robust distributed modeling and monitoring mechanism is proposed.The robust distributed modeling and monitoring mechanism is based on unit block modeling and also with Bayesian fusion for global hierarchical monitoring.Therefore,one only needs to consider the corresponding local updating once the local unit has to been updated without affecting the monitoring process of other units.Besides,one could improve both the monitoring efficiency and the robustness of the distributed method.Based on this mechanism,a distributed Bayesian network is proposed.For modeling,we develop distributed modeling scheme for each unit and then a block model fusion algorithm is proposed.For monitoring,the missing data based reconstruction approach and Bayesian fusion method are developed and hierarchical monitoring flowchart is designed for variable level,unit block level and plant-wide level.Such hierarchical monitoring mechanism shows good visualizations for plant-wide process.The plant-wide process is also with issue of big data.Based on the robust distributed mechanism,a distributed parallel PCA(dpPCA)is proposed.The distributed and parallel computation of statistics have been proposed based on the framework of MapReduce.The combination rules have been introduced and based on which the dpPCA is constructed.For hierarchical monitoring,the Bayesian fusion and reconstruction based method have been incorporated.The proposed method solves the big data modeling issue of plant-wide process in both distributed and parallel manner.Moreover,the hierarchical monitoring scheme also inherits the robustness as well as good visualization properties for monitoring large-scale plant-wide processes.