LAUSR

Abstract Causal relations between sources of industrial alarms can result in alarm floods, leading to a large number of simultaneously occurring alarms. Hence, various approaches exist to detect such alarm floods in historical alarm data with the purpose of operator support, root cause analysis and predictive maintenance. However, such approaches often suffer from randomly occurring, non-related alarms, resulting in invalid data patterns. Furthermore, the high amount of different alarm messages limits the approaches’ capabilities due to high computational costs. To overcome both problems, this paper introduces a graph-based approach to automatically split historical alarm data into groups of statistically depending alarms. The resulting groups, based on the conditional probability between alarms, can be extracted automatically, showing promising results for further data-mining approaches analyzing the groups’ dynamics individually. The developed method is evaluated based on historical alarm data recorded from a real industrial manufacturing plant.