Modular Bayesian Network Based Industrial Alarm Root Cause Analysis

Timely and accurate alarm root cause analysis is one of the effective methods to relieve the problem of "alarm overloading" for the process industries.The key links that affect the performance of alarm root cause analysis are establishing a precise and concise causality network for alarm variables and developing a reasonable traceability analysis mechanism.The overall optimality is strictly requested for the process industrial system.Furthermore,the correlation between the alarm variables is complex,and variables have a strong self-correlation,the state of the variable is not only influenced by its related variables,but also influenced by its own historical state.The transfer entropy can accurately measure the strength of causal relationship between variables by overcoming the difficulties of strong nonlinear and serious self-correlation.However,the causality network based on the transfer entropy only will loses its overall optimality,since transfer entropy can only be calculated in pairs.The Bayesian Network(BN)structure can represent the causal relationship between variables,however,the Bayesian Information Criterion(BIC)and Bayesian Dirichlet(BDe)scoring functions currently used in the process industry area have poor applicability.To solve the above problem,we proposed a family transfer entropy(FTE)based on the transfer entropy calculation method and the concept of "family" in Bayesian Network structure,which is used to globally evaluate the strength of causality between variables and their parents.In order to evaluate the overall performance of the network structure and avoid overfitting of Bayesian Network structure learning,we further introduced the penalty item and Family Transfer Entropy Tests(FTET)was proposed as the final scoring function.Combining the proposed FTET,a complete learning process for the process industrial Bayesian network structure learning was developed.Beyond that,based on simple process knowledge and data analysis methods,the requirements and rules for sub-module partitioning of large-scale systems,as well as sub-modules fusion and optimization methods are proposed to guarantee the efficiency and accuracy of learning.The Bayesian Network structure learning experiments of stochastic process and TE process show that FTET is more adaptable than BIC and BDe currently used in the process industry.The learned network structure has higher degree of accuracy and simplicity.To achieve quantitative alarm root cause analysis,an alarm traceability analysis method based on Bayesian Network reasoning was proposed.This method maximizes the average posterior probability of alarm status,and outputs the analysis result in the form of probability.The proposed method was applied to the alarm traceability analysis of TE process.Experiment result indicates that this method can correctly locate the root alarm variable.