Routine assessment of alarm system is necessary to make sure its performance is not degrading. Alarm historian is the best and easily available source to evaluate the performance of the alarm system against standard benchmark statistics. Alarm data is usually perceived as long strings of garbled text stored in a data base. Our research has produced some novel tools (has been integrated in our alarm management platform) that help control engineers to evaluate the performance of the alarm system with ease.

Utility of these tools is presented here using real industrial alarm data.
Shown below is the High Density Alarm Plot (HDAP)

Color of each bar in the plot indicates the number of alarms in a 10 minute window (horizontal axis) on the corresponding tag (vertical axis). A week’s worth of alarm data for an industrial alarm system is shown here. The tag names are arranged in such a way that the first tag has the maximum number of alarms over the given time period and as we go down, the number of alarm occurrences decrease. This unique representation is very instrumental in identifying chattering alarms, related alarms, redundant alarms and other discernible patterns.

Alarm Similarity Color Map (ASCM) for the same dataset is shown below

ASCM is very useful in detecting similarities in alarm occurrences. Related and redundant alarms appear well clustered in this plot. This plot also clusters alarms that contribute to plant instability.

Run Length Distribution (RLD) is a very useful tool in detecting and analyzing chattering alarm tags. Delay timers can then be properly designed. The RLD for a certain tag is shown below

Alarm flood analysis is an advanced and challenging issue of alarm management our group is studying. An incident when the number of alarms annunciated is more like to exceed operator’s response capability is referred as an alarm flood. A very general way to define an alarm flood is based on the alarm rate per 10 minutes. An alarm rate greater than 10 per 10 minutes is commonly called an alarm flood. Below a plot of alarm rate per 10 minutes versus time is shown where alarm floods can be detected when the plot crosses an alarm rate of 10 per 10minutes.

It has been seen that many of these alarm floods are very similar to each other in terms of the patterns of alarm annunciation. Our research is to analyze these patterns and cluster similar alarm floods in to groups. Similarities in alarm patterns can be interpreted as similarities in fault propagations within a process. This study will aid to develop the methods to prevent such groups alarm floods which are regular in time and due to a same set of interrelated process variables.

Different pattern analysis techniques were applied to investigate pairwise similarity and were clustered using an unsupervised clustering algorithm in our study. In following figure, a case study on 39 alarm floods from real industrial data is shown where the similar alarm floods were clustered together and indicated by the clusters of darker pixels.

A further investigation on these groups of alarm floods were also carried out to validate the similarity in alarm patterns. Following few of such alarm flood sequences are shown to illustrate the similarity in patterns of alarm annunciations.

Similarity investigation on alarm floods categories different sets of alarms which are raised in a sequence. A rationalization or a causality analysis of these alarm sets will help to prevent such large sequence of alarms in future and help to reduce the number alarm floods in process industries.