Alarm Management

Published on May 2016 | Categories: Types, Instruction manuals | Downloads: 33 | Comments: 0 | Views: 293
of 11
Download PDF   Embed   Report

Alarm Management

Comments

Content

Seion
A Statistical Method for Alarm System Optimisation

White Paper
By
Dr. Tim Butters
Data Assimilation &
Numerical Analysis Specialist

[email protected]
www.sabisu.co

Seion Alarm Management

White Paper

Contents
1 Introduction

2

2 Challenge

2

3 Solution
3.1 Alarm Clusters . . . . . . . .
3.1.1 Dependent Alarms . .
3.2 Cluster Linkage . . . . . . . .
3.3 Visualisation . . . . . . . . . .
3.3.1 Adjacency Matrix . . .
3.3.2 Force Directed Graph .

3
3
4
5
5
5
6

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

4 Conclusions

.
.
.
.
.
.

6

List of Figures
1
2
3
4
5

Cluster Identification Method . . . . . . . . . . . . .
Name Ordered Adjacency Matrix of Alarm Data . . .
Cluster ID Ordered Adjacency Matrix of Alarm Data
Frequency Ordered Adjacency Matrix of Alarm Data
Force Directed Graph of Alarm Data . . . . . . . . .

© 2014 Sabisu Ltd.

1

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

. 4
. 7
. 8
. 9
. 10

V1.0

Seion Alarm Management

1

White Paper

Introduction

Alarm systems are an important part of any industrial installation. They
provide a vital safety role in alerting operators to unexpected behaviour, and
in performing forced shut-downs if a severe danger is detected. It is therefore
vital that a plant’s alarm system performs in the most optimal way possible,
as failure to do so not only poses a great risk to safety, but also profitability.
There are several metrics that can be used to quantify the performance
of an alarm system. One of the most important is “operator load”. This
is the number of alarms each operator has to address in a given time-frame
(typically 10 minutes). Operator load not only gives a good indication of how
well the alarm system is configured, but also how viable it is that all alarms
are being acknowledged and actioned in the appropriate manner. If a system
is poorly configured the operator load in the control room can be so large
that it is impossible for the operators to properly register each fault. This
can force them to manually prioritise the alarms and attempt to decipher
the root cause of the problem, which can be extremely difficult. In the past
poor alarm system configuration has led to large-scale industrial accidents
such as the explosion at the Texas City refinery, and the nuclear incident at
three mile island in the USA.
Seion is an alarm management decision support tool that can be used to
optimise the performance of industrial alarm systems. It utilises a novel statistical method of analysing alarm system performance developed by Sabisu,
which can identify clusters of alarms that often act together. This provides
information that can be used to detect redundant alarms that provide little
to no extra information. These alarms can then be analysed and, if appropriate, removed from the system. This process reduces the operator load, and
makes it easier for them to identify the physical cause of any alarm incident.
With all changes being enacted through your companies change management
system.

2

Challenge

As the cost of sensors has fallen it has become commonplace to install a large
number across a plant, ensuring that every physical component is monitored.
Although this may be necessary for some aspects of operation, it is not usually
necessary to configure an alarm for each one, a practice that has also become
© 2014 Sabisu Ltd.

2

V1.0

Seion Alarm Management

White Paper

common. Many of these sensors could be reporting on physically linked
systems, causing several alarms to sound for each single ‘event’.
Although there are extant solutions available for the configuration of
alarm management systems, many of them are complex in nature. They
can require a computational model of the plant, and permission to intercept
alarms in real-time, hiding alarms the automatic system has deemed unnecessary. Although these solutions can be effective, they are often costly and
require a large amount of configuration time.

3

Solution

Sabisu have developed an alarm management configuration and optimisation
tool that utilises historical data already available from the plant. This is a
decision support system that requires very little set-up, and ensures that you
remain in complete control of your alarm system.

3.1

Alarm Clusters

The main objective of Seion is to identify alarms that act in clusters. These
are alarms that are likely to be linked in some physical way (e.g. temperature
conduction between sensors) and therefore indicates a level of redundancy
that is likely to increase the operator’s alarm load.
To identify these clusters the alarm log for the system is analysed, and
alarms that occur within a specified time window of one-another a statistically significant amount of the time are marked as a linked pair. Each linked
pair consists of a principle alarm (P ) and an incident alarm (I), with the
principle alarm implying the incident alarm: P → I. It may be found that
one principle alarm has several corresponding incident alarms, with many of
these forming their own linked sets as shown in figure 1.
Clusters of this nature indicate potential redundancy in the system, as
the statistical threshold will only identify alarm clusters if the alarms forming
them occur together a large percentage of the time. This suggests that it is
rare to see the principle alarm without its corresponding incident alarm(s),
and therefore the incident alarm is adding very little extra information. It
is, however, needlessly increasing the load on the operator. The statistical
thresholds used for these calculations guarantee not only that each incident
alarm is often seen with its principle alarm when compared to the number of
© 2014 Sabisu Ltd.

3

V1.0

Seion Alarm Management

White Paper

Figure 1: Diagram showing basic cluster identification. Alarms A, B and C
continually sound together. The coloured brackets show the search window
for the corresponding alarm (the window for alarm C is not shown). Alarm
A is the principle alarm for a cluster containing B and C as incident alarms,
and alarm B is the principle alarm for a cluster containing alarm C as the
incident alarm. There are no alarms within the search window of C, therefore
it is not the principle alarm for any cluster.
occurrences of the principle alarm, but also that the linked pair are seen in
the alarm log relatively often. This ensures that a phenomenon seen only a
couple of times will not be included, as the statistical confidence is not high
enough.
3.1.1

Dependent Alarms

Dependent alarms are a slightly different concept from alarm clusters, in
which the statistical threshold governing the significance of a linked pair
with respect to the occurrences of the principle alarm is not necessarily met.
However, the incident alarm is never seen by itself, only ever following its
principle alarm. This could identify redundancy, as with the cluster analysis, or it could represent a distinct separate event. Dependent alarms are
automatically identified by Seion, but should be treated differently when considering the best course of action as the alarms identified as constituting a
dependent set may provide extra information compared to the set’s principle
alarm alone.

© 2014 Sabisu Ltd.

4

V1.0

Seion Alarm Management

3.2

White Paper

Cluster Linkage

As can be seen in figure 1, clusters are identified by searching within a set time
window of each alarm. If an alarm continually sounds after a given principle
alarm, but the time between the two is longer than the search window no
link will be made between them. Increasing the time window is one way to
find such links, however, caution must be used as there is a higher chance of
forming erroneous links between alarms that are not actually linked (this is
especially true if there are a number of ‘noisy’ alarms in the system).
To combat this, Seion constructs an approximation to the transition matrix of the system. This allows the identification of links between alarm
clusters by calculating the n − th power of the transition matrix, where n is
the number of time windows in which to search (i.e. the 2nd power searches
for links within 2 search windows, the 1st power returns the originally identified clusters).
This cluster linkage method provides a possible way to identify the clusters that should be tackled first, as if one cluster implies another then removing it will further reduce the alarm load through the reduction in activation
of the second cluster.

3.3

Visualisation

As well as providing a list view of the cluster analysis, Seion also incorporates
interactive visualisations to provide easy interpretation of the results. These
visualisations are in the form of a dynamically sortable adjacency matrix,
and a force directed graph.
3.3.1

Adjacency Matrix

This visualisation shows which alarms are linked in clusters by grouping
them together into block rectangles on a grid. The clusters are coloured to
aid identification, although some alarms that appear in more than one cluster
will be black. The connectivity of the alarm is reflected in its opacity, with
highly connected alarms appearing darker than less connected alarms.
The adjacency matrices for some example alarm data are shown in figures
2 - 4. Figure 2 shows the matrix sorted by alarm name. This ordering makes
it simple to locate an alarm of interest and quickly identify the key alarms
it is linked with. Figure 3 shows the matrix ordered by cluster ID, which
© 2014 Sabisu Ltd.

5

V1.0

Seion Alarm Management

White Paper

attempts to group each cluster together so that their relative sizes can be
assessed. Figure 4 shows the adjacency matrix sorted by frequency, which
is a measure of the ‘connectivity’ of an alarm. For an individual alarm, the
more alarms linked with it, and the more times it appears in the alarm log,
the higher the frequency. This provides a simple method to rank each alarm
by its effect on the alarm load, with the worst offending alarms in the top
left of the matrix.
3.3.2

Force Directed Graph

A force directed graph (FDG) of the alarm clusters from the example data
is shown in figure 5. The FDG shows clusters by ‘linking’ each constituent
alarm with a line, with the arrow on the line indicating which alarm implies
the other. As with the adjacency matrix, the clusters are coloured to aid
identification.
This ability to identify principle and incident alarms allows the quick
identification of key alarms to target when implementing suppression rules.
If an alarm has many arrows leading from it to other alarms it is a highly
connected principle alarm. If an alarm has many arrows leading to it, it is
mainly seen as an incident alarm. To remove a cluster, suppression rules
should be triggered by highly connected principle alarms, suppressing the
resultant incident alarms.

4

Conclusions

Seion provides a statistically driven solution to the problem of alarm management optimisation. This statistical nature allows historical alarm data to
be utilised to find clusters of alarms that indicate redundancy in the system.
This redundancy leads to an increased operator load, therefore identifying
and dispersing these clusters increases both the safety of the plant, and also
profitability through the reduction in forced shut-downs. The use of statistical analysis makes this solution highly portable, leading to very short set-up
times and a high level of cost effectiveness.
By identifying clusters of alarms, and providing tools that allow those
clusters to be ranked based on their effect on the alarm load, it is easy to
target the areas of the system that will have the largest positive impact.

© 2014 Sabisu Ltd.

6

V1.0

Seion Alarm Management

White Paper

Figure 2: Adjacency matrix of example alarm data sorted by alarm name.
This ordering makes it easy to find specific alarms and the alarms they are
linked with.

© 2014 Sabisu Ltd.

7

V1.0

Seion Alarm Management

White Paper

Figure 3: Adjacency matrix of example alarm data sorted by alarm cluster
ID. This ordering makes it easy to compare the alarm cluster sizes.

© 2014 Sabisu Ltd.

8

V1.0

Seion Alarm Management

White Paper

Figure 4: Adjacency matrix of example alarm data sorted by alarm frequency. This ordering makes it easy to identify the most problematic alarms
with respect to alarm load. The highest frequency alarms are in the top left
of the matrix, and the lowest frequency in the bottom right.

© 2014 Sabisu Ltd.

9

V1.0

Seion Alarm Management

White Paper

Figure 5: Force directed graph of example alarm data. Arrows indicate
which principle alarms imply which incident alarms.

© 2014 Sabisu Ltd.

10

V1.0

Sponsor Documents

Or use your account on DocShare.tips

Hide

Forgot your password?

Or register your new account on DocShare.tips

Hide

Lost your password? Please enter your email address. You will receive a link to create a new password.

Back to log-in

Close