Decentralized Multi Model Process Monitoring Based On Data-Driven Methods

In recent years,with the continuous development of industrial automation and the improvement of science and technology,the modern large industries are accompanied by large scale and complicated process.How to ensure the safety of industrial processes,reduce production costs,and ensure product quality has become an urgent problem for modern industry.In this context,process monitoring research has become particularly important.With the development of computer technology,process data information is more and more abundant.On this basis,data-driven process monitoring methods have received extensive attention,and multivariate statistical process monitoring has become a research hotspot in the field of process monitoring.Although some research results of MSPM have been reported,there are still some problems that need to be discussed and addressed for the actual industrial process.Based on the existing research results,this paper proposes some new process monitoring methods for some problems existing in traditional fault detection methods.The specific contents are as follows:(1)Considering the complexity of modern industrial process data,there are different autocorrelation between measurement variables and traditional process monitoring methods don't recognize this.Based on dynamic principal component analysis(DPCA)and independent component analysis(ICA),this paper proposes fault detection using mutual information based variable-weighted dynamic PCA method(MI-WDPCA)and fault detection using mutual information based variable-weighted ICA method(MI-WICA).These methods use the decentralized modeling technique,and establish the corresponding PCA(ICA)model for each measured variable according to its correlation.In order to facilitate the judgment of whether there is a fault,Bayesian inference is used to detect multiple PCA(ICA)models.The indicators are integrated into a final probabilistic indicator.The experimental comparison shows that the method achieves better monitoring results than the traditional process monitoring method.(2)For modern complex industrial processes,variables are usually nonlinearly related,while traditional PCA and ICA monitoring algorithms are still linear essentially.In order to monitor the nonlinear process effectively.This paper proposes a multi-model based KPCA fault detection method(MMKPCA)based on kernel principal component analysis(KPCA).This method selected different kernel functions to establish fault detection models respectively.The monitoring statistics obtained by different KPCA models are combined into a final probabilistic index by Bayesian inference.This method achieves better monitoring results than traditional nonlinear process monitoring methods demonstrated by experiment.Finally,based on the summary of the work of this paper,the future researches in of the process monitoring are discussed.