smart alarm monitoring - university of albertaslshah/files/nserc_irc2007... · 2009. 6. 9. · 4...

Smart Alarm Monitoring

Iman Izadi, Research FellowDepartment of Chemical and Materials Engineering

University of Alberta

2

NSERC-Matrikon-Suncor-iCORE IRC Seminar; 3 December 2007

Outline

MotivationFalse alarms, causes and remediesModel base performance monitoringMultivariate vs. univariate Error processingCase studiesConcluding remarks

3


Why Monitoring?

Improving efficiency and reliabilityReducing unplanned downtimeAvoiding production lossPreventing equipment damage

Environment protectionSafety

MOREPROFIT

4


History

Traditionally, a limited number of variables are monitored.If a variables exceeds a certain limit, an alarm is raised.After an alarm is raised, the operator must understand and acknowledge the alarm, detect the event, and react accordingly.SPC (Statistical Process Control) and SQC (Statistical Quality Control) charts are used for performance monitoring (e.g., Shewhart chart).For overall performance of the process, a few “quality variables” or “key performance indicators (KPI)” are monitored.

5


Old Control Room

Historically, alarms were expensive and difficult to implementEach alarm had to be hardwiredLimited space in control panelsResulting in careful design and review before implementation

6


Modern Control Room

Today, hardware and software advances has made it easy to add alarms at minimal costLarge increase in the quantity of alarmsReducing the quality and efficiency of alarms

7


Why False Alarms?

Thousands of alarms configured and continuously addedBad design, bad tuningDifferent operating statesProcesses change over time Equipments change and degradeSeasonal changesNoise

8


Standards and Regulations

1999 EEMUA 191 v1World-wide standard

2007 EEMUA 191 v22008 ISA S18.02 - Management of Alarm Systems for the Process Industries2008 ANSI Adoption of ISA

9


EEMUA Benchmarks

EEMUA Oil & Gas PetroChem Power Other

Average Alarmsper hour <6 36 54 48 30

Average StandingAlarms 9 50 100 65 35

Peak Alarms/hour(10 Minutes) 60 1320 1080 2100 1080

Distribution %(Low/Med/High) 80/15/5 25/40/35 25/40/35 25/40/35 25/40/35

10


How to Reduce False Alarms

Can be addressed in different levels in the plant (design, operation, management)Improve the design

Multivariate instead of univariateMore accurate models Better configuration (tuning)Processing of data

Alarm Management

11


Multivariate vs. Univariate

Pressure

0 20 40 60 80 100

92

94

96

98

100

102

104

Pre

ssur

e

Pressure versus Temperature

Temperature

0 20 40 60 80 10041

42

43

44

45

46

47

48

4941 42 43 44 45 46 47 48 49

92

94

96

98

100

102

104

12


Why Multivariate Analysis

Hundreds to thousands of variables are measuredVariables are highly correlated with one another

It is not easy to understand the correlation between variablesStatistical rank of data is very lowRank is independent of the number of variables measuredThe behavior of the process can be expressed by a few independent variables (sources of variation in the in the process)

Low signal-to-noise ratio Little information in any one variable

13


Model-based Performance Monitoring

The actual behavior of the process is compared with the behavior predicted by the model.(perfect model) and (no abnormalities)

(error is white noise)The prediction error is processed.The result is compared with a threshold.

Actual Process

Process Model

+_

errorprocessing alarm

14


Modelling

First principles modelData-driven modeling

Dynamical modeling (system identification)Statistical modeling (PCA, PLS, CCA, …)

A new set of independent variables are created (from a combination of the original variables) that capture the variability of the process

Neural networks…

15


Error Processing

After the error is generated, it should be processed to facilitate

Threshold designNoise reductiondetection delay and false alarm rate trade-offfalse alarm rate and missed alarm rate trade-off

16


Reducing False Alarm Rate

False alarm rate can be reduced using simple processing techniques

FilteringReduces noiseRemoves spikesModifies distribution

Time delayDead-band

Reduces alarm chattering

17


Filtering

Filters are applied to variables to:reduce noiseremove bad or unwanted dataextract featuresmodify statistical distributionsseparate frequency components

Common filtersMoving averageMoving varianceMoving rangeEWMAFinite Impulse Response (FIR)

18


Filter: Moving Average

0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

19



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

20



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

21



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

22



-2 0 2 40

0.1

0.2

0.3

0.4

false alarm = 102missed alarm = 157

23



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

24



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

0 500 1000 1500

-1

0

1

2

3

25



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

0 500 1000 1500

-1

0

1

2

3

-1 0 1 20

0.5

1

1.5

26



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

0 500 1000 1500

-1

0

1

2

3

-1 0 1 20

0.5

1

1.5

27



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

0 500 1000 1500

-1

0

1

2

3

-1 0 1 20

0.5

1

1.5

28



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

-1 0 1 20

0.5

1

1.5

false alarm = 39, missed alarm = 1

29



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

0 500 1000 1500

-1

0

1

2

3

-1 0 1 20

0.5

1

1.5

30



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

0 500 1000 1500

-1

0

1

2

3

-1 0 1 20

0.5

1

1.5

31



0 500 1000 1500

-2

0

2

4

6

-2 0 2 40

0.1

0.2

0.3

0.4

-1 0 1 20

0.5

1

1.5

32


Filter: Moving Variance

0 500 1000 1500-10

-5

0

5

10

33



0 500 1000 1500-10

-5

0

5

10

-5 0 50

0.1

0.2

0.3

0.4

34



0 500 1000 1500-10

-5

0

5

10

-5 0 50

0.1

0.2

0.3

0.4

35



0 500 1000 1500-10

-5

0

5

10

-5 0 50

0.1

0.2

0.3

0.4

36



0 500 1000 1500-10

-5

0

5

10

-5 0 50

0.1

0.2

0.3

0.4

0 500 1000 1500

0

5

10

15

20

37



0 500 1000 1500-10

-5

0

5

10

-5 0 50

0.1

0.2

0.3

0.4

0 500 1000 1500

0

5

10

15

20

0 5 10 150

0.5

1

38



0 500 1000 1500-10

-5

0

5

10

-5 0 50

0.1

0.2

0.3

0.4

0 500 1000 1500

0

5

10

15

20

0 5 10 150

0.5

1

39



0 500 1000 1500-10

-5

0

5

10

-5 0 50

0.1

0.2

0.3

0.4

0 5 10 150

0.5

1

40


Time Delay

0 500 1000 1500

-2

0

2

4

6

41


Time Delay

42


Alarm Chattering - Deadband

43



44



45



46



47



48


Case Study 1

Performance Monitoring of a Pump

49


Case Study 2

Performance Monitoring of Electric Motors

50


Concluding Remarks

Alarm management and engineering is critical in today’s industries.Multivariate and model-based methods offer better monitoring performance than univariate and signal-based methods.False alarms can be avoided by processing the prediction error and applying simple techniques.The major trade-off in alarm design is between detection delay and false alarm rate.

51


Acknowledgements

Dr. Dave Shook from MatrikonDr. Sirish Shah and Dr. Tongwen ChenCPC Group MembersNSERC-Matrikon-Suncor-iCORE for financial support

smart alarm monitoring - university of albertaslshah/files/nserc_irc2007... · 2009. 6. 9. · 4...

Documents