Inference Methods for Detecting the Root Cause of Alarm Floods in Causal Models

Abstract

Driven by the oil and chemical industry and amplified by the digitization and automation of the industry, the issue of alarm management has been gaining more and more importance. In highly automated and complex industrial systems, on the one hand, a large number of messages and alarms arise and, on the other hand fewer and fewer employees must be able to handle them. This amount of alarms is called alarm flood and it is a huge safety risk in facilities such as refineries. Therefore, it is necessary to reduce these alarm floods, thus reducing downtime, supporting the operator and preventing catastrophes. A novel approach to reducing alarm floods is concerned with learning the causal relationships between the alarms. The learned interrelations of the alarms are represented by a causal model. Based on these causal model, a root cause analysis is carried out to find out the cause of an alarm flood. This makes it possible to dramatically reduce the number of alarms and messages by displaying only the potential root causes. Therefore, we validate the approach of identifying the root cause of an alarm flood by a given causal model. The three most common inference methods are investigated and their suitability for practical application is evaluated on two demonstrators from SmartFactoryOWL.