Detecting significant alarms using outlier detection algorithms

Although alarms in plants are designed to notify any anomaly or faults in order to prevent accidents or to improve process, it is very difficult for the operators to identify meaningful alarms, since there are large volumes of false and nuisance alarms. Outlier detection algorithms are used to identify anomaly in data, and thus they can be used to suggest abnormal alarms. In this research, we analysed real world alarm data collected from an iron processing company and constructed data features for algorithmic outlier detection. With the data we identified outlier alarms and compared their run length with nonoutlier alarms. Our results demonstrated that outlier alarms detected by the algorithms have significantly difference patterns in average run length compared to the normal alarms.