Causal Alarm pattern analysis by the Integration of Technical Information from engineering documents
Pattern Recognition to Identify Alarm Floods in Industry
Alarm Management Systems (AMS) are a crucial part of modern automated Production Systems (aPS), due to their important role in assuring the safe operation of the system. Alarm messages are used to inform the system operator about abnormal situations. However, many existing industrial AMS suffer from poor design and performance, e.g. abnormal events often result in various further alarms that must be handled by the operators in a short time. Furthermore, causal dependencies between alarm sources increase the number of alarms presented to the operator simultaneously. This situation is referred to as an alarm flood. In this case, the operator might not be able to deal with all the alarms appropriately. This may result in an incorrect decision by the operator and a hazardous situation. Therefore, an efficient and reliable AMS is mandatory to avoid operator overload as well as to increase the safety by fast and accurate root cause detection. In order to improve AMS and to overcome alarm flood problems, several approaches exist to identify sequences of occurring alarms automatically. These approaches are essentially based on sequence pattern detection in historical alarm data. Due to their limitations, these approaches can miss some important alarm patterns or find invalid patterns. Furthermore, they will fail to detect the patterns in the case of new abnormal situations and unknown alarms during runtime, which have not been seen yet in the historical data. In order to overcome these challenges, the applicants intend to improve AMS methods by considering additional information sources like plant structure data from engineering documents and process data from the process control system, in combination with the historical alarm data. The integration of these additional, automatically extracted information with historical alarm data analysis will be used for root cause analysis and detection of causal alarm patterns. Furthermore, in the case of new alarms during runtime, methods will be investigated which update the former causal patterns by including unknown alarms in proper patterns. Finally, the improved algorithms should support the operator during runtime of the plant by a reduction of alarm floods and better identification of root causes of alarms.