(646e) Learning Based Automated Identification of Nuisance and Correlated Alarms

Alarms systems are paramount for safe, efficient and reliable operation of modern industries. Ideally, alarms are suitably configured to provide an indication to process operators as soon as deviation from normal plant behavior is detected. The process operator can then take suitable remedial actions to ensure optimal plant behavior.

In practice, incorrectly configured alarm systems lead to problems including detection delays, alarm chattering, nuisance alarms and alarm flooding [1]. Due to the relative ease and low cost of measurement of process variables and collection of data, more alarms can be easily configured. However, this leads to the issue of generation of nuisance alarms which are alarms that require no response from the operator. Usually, an underlying fault in the plant propagates across several different units, leading to multiple alarms generated from different sources. In abnormal situations hundreds of alarms can arise in quick succession leading to operator overload. It is estimated that the US industry alone loses billions of dollars every year due to incorrect handling of alarms [2].

In this work, we present a method for the automatic identification of nuisance alarms and correlated alarms based on historically labelled alarm data. The method is designed to operate with an incoming stream of data – newly available labelled data can be immediately used to update the machine learning based models. The method can be easily deployed over the cloud and scaled up to process alarm data at a very high frequency. The method is applied to network monitoring alarm systems spanning several sites. Results indicate the extremely high accuracy of our method for identification of nuisance and correlated alarms and demonstrate its utility in reduction of alarm flooding.

[1] B. R. Hollifield and E. Habibi, Alarm management: seven effective methods for optimum performance, ISA, 2007.

[2] W. Hu and A. W. Al-Dabbagh and T. Chen and S. L. Shah, Process Discovery of Operator Actions in Response to Univariate Alarms, IFAC-PapersOnLine, 49, 1026—1031, 2016.