An Adaptive Graph-Based Approach for Similarity Analysis and Early Classification of Alarm Floods

The display of a large number of alarms on human-machine interface screens of industrial control systems in a short period is referred to as an alarm flood. This can overwhelm human operators, and prevent them from making proper decisions and actions in a timely manner. This paper proposes an adaptive graph-based approach for modelling alarm floods, which can successfully accommodate gradual changes in the display of alarms within their sequences. Based on this approach, methods are introduced for similarity analysis as well as supervised and unsupervised early classification of alarm floods, offering online prediction of the types of incoming alarm floods in real-time. This can lead to identifying common root-cause abnormalities to assist the human operators. The performance of the proposed approach and methods is evaluated using the Tennessee Eastman process.