Improved Principal Component Analysis And Multi-state Bayesian Network Integrated Method For Chemical Process Diagnosis

Most of the raw materials and intermediate products in chemical industry are corrosive,toxic,flammable and explosive.In addition,with chemical industry becoming highly integrated and large-scale,the correlation among process variables becomes more and more complicated.Sometimes a small local fault can cause a series of severe consequences.Therefore,it is vital for chemical industry to detect the fault in time and accurately diagnose the root cause,propagation path and change trend of the fault.At present,the availability of instrument system provides a large amount of raw material and the development of data and information science provides advanced tools for data mining.Therefore,data-driven methods have gradually become the hottest research direction in the field of fault diagnosis.In this paper,a hybrid fault diagnosis method based on improved principal component analysis(IPCA)and multi state Bayesian network(MBN)is proposed.In the fault detection stage,the improved principal component analysis method is used to detect and identify fault variables.In the diagnosis stage,the multi-state Bayesian network trained by process data is used to receive the fault variables identified in the first stage,and then the network is updated to identify the propagation path,root cause of the fault,and the change trend of the key process variables in the form of probability.Johnson transformation is introduced into principal component analysis(PCA)to transform the process data to a normal distribution,which can alleviate the strict requirements of the traditional PCA on the normal distribution of process data.Moreover,Spearman rank correlation coefficient matrix is used to replace the covariance matrix of the traditional principal component analysis to improve the feature extraction of nonlinear process data.These two improvements greatly enhance the ability of fault detection.Multi state Bayesian network divides variable nodes into five states: high fault,high normal,normal,low normal and low fault.This is advantageous than the normal and fault binary classification of Bayesian network in the current fault diagnosis literature,and can diagnose the propagation path,root cause and changing trend of fault more specifically.In addition,in the process of building MBN,the state partition of nodes is conducted by nonparametric kernel density estimation method,then the conditional probability of nodes is obtained by maximum likelihood estimation method,and some causal connections of nodes which are difficult to judge by process knowledge are determined by transfer entropy calculation.This method was first explored and verified on a simple depropanizer distillation process,and then applied to the fault diagnosis benchmark TE process.Compared with the traditional principal component analysis,kernel principal component analysis and dynamic Bayesian network,the application results show that this method has the advantages of high fault detection rate,low misdiagnosis rate and low calculation load.In addition,it can not only diagnose the root cause and propagation path of the fault,but also accurately diagnose the changing trend of process variables.