Advanced Methods for Alarm Monitoring and Alarm Flood Analysis Based on Industrial Data

Abstract

Alarm systems are critical assets in process industries to ensure process safety and efficiency. However, the number of alarms has grown exponentially in the past decades due to easy alarm configuration in computerized systems and lack of proper alarm rationalization. As a consequence, the presence of nuisance alarms and alarm floods severely compromise the performance of alarm systems: nuisance alarms could distract operators with incorrect indications of abnormalities; while alarm floods lead to increased operational risks as operators are overwhelmed by massive alarms and thus may overlook critical alarms. Motivated by this, this thesis focuses on the development of data-driven methods for alarm monitoring and alarm flood analysis. Three research topics are considered. First, to address nuisance alarms during the start-up operations of industrial equipment, a new alarm monitoring method is proposed, which comprises an offline design stage to capture equipment dynamics during start-ups and an online algorithm for alarm monitoring. Second, a systematic pattern matching method is proposed to capture similar alarm floods across different processes, where the alarms are associated with the same fault types, but configured with different tag names. The obtained results could facilitate root cause analysis and lead to generalized solutions. Last, a pattern mining method is proposed to extract compact alarm sequential patterns from Alarm & Event (A&E) logs with the incorporation of time stamps, tolerance of alarm order switchings, and distillation of compact results. Therefore, the proposed method is capable of avoiding influences of order ambiguities and also minimizing the redundancy of extracted patterns. The effectiveness and practicality of the proposed methods are demonstrated by case studies using alarm data from a large-scale industrial facility. Based on the proposed methods, equipment start-ups are effectively monitored while nuisance alarms are significantly reduced; notable alarm sequential patterns are discovered from historical alarm data, which could facilitate alarm suppression, root cause analysis, and decision support for operators.