Towards Accurate Root-Alarm Identification: The Causal Bayesian Network Approach

Abstract

Abnormalities in modern process industries are reported by alarms. Strong inter-connectivities within different units of a plant lead to annunciations of multiple alarms in a short period of time, which hinders a prompt operator response. A fundamental problem in such alarm floods is to identify the original source of abnormality by finding the directly affected alarm, known as the root cause. This is essentially a causal problem which has been tackled by many non-causal approaches, including transfer entropy and Bayesian network based methods, which despite their successful applications, have their own shortcomings. Causal identification requires intervention. That is, to “intervene” and manually change the distribution of a random variable A, and compare its conditional distribution with respect to another variable B before and during the intervention to examine whether A is caused by B. This acknowledged notion of causality, leads to causal Bayesian networks where unlike Bayesian networks, the edges are truly causal. Nevertheless, intervening the alarms during a fault event is implausible if not impossible, hindering the use of this rich notion of causality in alarm root cause analysis. We tackle this issue by treating abnormalities as uncertain interventions, because they, indeed, intervene the directly affected alarms and change their distributions. We then find the causal Bayesian network that best represents the alarm data. In addition to more-accurate root-cause identification, our causal Bayesian network based method has the ability to discover the root causes under multiple simultaneous abnormalities.