Multimode Industrial Process Fault Diagnosis Method Based On Bayesian Independent Component Analysis

With rapid development of modern industries,industrial systems become complicated and large-scale.Fault diagnosis technology shows its great importance in ensuring process safety and improving production quality.Since modern industrial computer control system can acquire and store huge amounts of real-time process operation data,multivariate statistical data analysis based fault diagnosis method has shown its significant theoretical and practical value.Aiming at the multimode industrial process monitoring problem,this paper analyzes the multimode and non-Gaussian characteristics of process data and studies multimode process fault diagnosis method based on Bayesian independent component analysis(BICA).Furthermore the proposed methods are verified by simulation analysis.Firstly,aiming at multimode industrial processes with non-Gaussian features,the paper proposes a fault detection method based on Bayesian independent component analysis(BICA)algorithm.In this method,Bayesian inference and ICA are combined to establish a probability mixture model for multimode data.The ICA model parameters are obtained by the iterative optimization algorithm and the mode of each observation is determined by Bayesian inference simultaneously.Then considering process data of auto-correlation,matrix dynamic augmentation is applied to BICA method and a dynamic BICA(DBICA)method is proposed to monitor the multimode non-Gaussian dynamic process.Lastly case studies on one continuous stirring tank reactor(CSTR)simulation system and the Tennessee Eastman(TE)benchmark process are used to demonstrate that the proposed method is more effective than ICA and DICA methods.Then,in order to solve the problem that traditional variable contribution plot can not indicate the information transmission relationship among fault variables,this paper proposes a multimode process fault recognition method based on information transmission contribution plot.In the proposed method,the contributions of BICA ensemble monitoring statistics are calculated.Furthermore,the nearest neighbor transfer entropy is used to describe the transitivity among variables and mine cause-and-effect relationship of fault variables,which helps to determines the fault source variable and fault propagation process.Finally,simulations on a numerical example and the CSTR system are performed to show the effectiveness of the proposed approach.Finally,considering that overall statistical modeling submerges the local variables information,a novel multimode process fault diagnosis method based on variable partitioning BICA(VPBICA)is proposed.This method applies the similarity analysis to divide the monitored variables into mode-related sub-group and mode-unrelated sub-group.Then BICA algorithm is implemented for mode-related variables,and ICA algorithm is implemented for mode-unrelated variables.Two statistical sub-models are developed to monitor process changes.The simulation results on the CSTR system show that the method is more effective than the ICA and BICA methods,and helps to identify fault variables.