Alarm Pattern Recognition in Continuous Process Control Systems using Data Mining

An alarm management system with the Human Machine Interface in a process control system is used to alert an operator of any abnormal situation, so that corrective action can be taken to ensure safety and productivity of the plant and quality of the product. An alarm system reporting many alarms even during the normal state of the plant is due to chattering alarms, duplicated alarms, intermittent equipment problems and certain alarms configured in the system which may not have any importance. In such a situation the operator may miss certain critical alarms leading to undesirable outcomes. So, to have an optimum alarm system, the unwanted alarms have to be identified and eliminated. In this paper, we propose an offline method to identify repetitive, frequent sequences or patterns using PrefixSpan and Bi-Directional Extension algorithms. With the identified sequences or patterns, plant operation experts can improve the effectiveness of the alarm system through alarm rationalization so that this will help the operator in making the plant more safe, reliable and productive. The main objectives of this work are the following: (i) to use a definitive method to represent alarm data in an alarm log which is Temporal data as Itemsets without a need for complex mathematical, statistical or visual methods; (ii) to use data mining algorithms for identifying Frequent sequences which can be implemented on a normal computing resource such as Personal computer; (iii) to apply the method to the complete alarm data available no matter how big they are; (iv) to study and establish that the chosen method is possible to be applied to larger sized datasets.