Causal Inference in Industrial Alarm Data by Timely Clustered Alarms and Transfer Entropy

In large-scale industrial plants, alarm management system (AMS) has a critical role in safety and efficiency of the plant. High degree of connectivity in large-scale plants results in high degree of dependencies between the generated alarms, and thus in any abnormal condition, a huge number of alarms are presented to the operator. This phenomenon is known as alarm flood, which might lead to a hazardous situation if the operator cannot handle them. Therefore, an efficient alarm analysis system is required to assist the operator by detecting the sequence of alarms and the root-cause analysis between them. In this paper, a data-driven method using the alarm log file is proposed to detect the causal sequence of the alarms. In this method, an efficient alarm clustering based on time distance between the alarms is proposed to keep the timely close alarms in one cluster. This clustering approach can help to preserve the neighboring alarms in one cluster. By similarity analysis between the detected clusters, the similar clusters can form a category of alarms. Each category and the clusters inside them are further analyzed for root-cause detection by means of transfer entropy. Finally, the proposed method is evaluated with an industrial alarm data.