Decision-making systems are known as the main pillar of industrial alarm systems, and they can directly effect on system’s performance. It is evident that because of hidden attributes in the measurements such as correlation and nonlinearity, thresholding systems faced wrong separation defining by Missed Alarm Rate (MAR) and False Alarm Rate (FAR). This study introduced a novel extended adaptive thresholding based on mean-change point detection algorithm and shows that it is more efficient than other existing thresholding algorithm in the literature. Number hypothetical and industrial examples are given to delineate the capabilities and limitation of proposed method and prove its effectiveness in an industrial alarm system.
A Novel Extended Adaptive Thresholding for Industrial Alarm Systems
1. A Novel Extended AdaptiveThresholding
for Industrial Alarm Systems
Mahdi Bahar-Gogani
Koorosh Aslansefat and Mahdi Aliyari Shoorehdeli
2. Topics
Introduction
Definitions and Basic Concepts
Thresholding
• Simple Threshold
• Adaptive Threshold
Extended Adaptive Threshold (EAT)
EAT with Delay Timer
Mean-Change Point Detection
Results (4 different examples)
2
3. Introduction
3
in large industrial systems, there are thousands of sensors in
different areas to monitor physical or environmental
conditions of the plant during operation.
Whenever a process variable exceeds a certain threshold, an alarm is raised (in auditory or visual
form) to indicate equipment malfunction, process deviation, or any other abnormality in the plant
Monitoring
4. Introduction
4
Example:
in the nuclear power plant accident occurred atThree Mile Island in 1979, which is the worst nuclear accident in the
US history, operators were faced with redundant information, much of it irrelevant and illusory during the accident
An alarm system generates and processes alarms to present
abnormal behaviors of systems to the operator.
these systems are vital assets for process safety and
efficient operation of modern industrial plants.
Alarm
Management
operators often face many more alarms than they can handle immediately,
mainly due to the excessively large number of nuisance alarms
5. Alarm Management
5
Purpose of Alarm
Management
The main purpose of alarm management is the
reduction of these three parameters.
Definition:
consists of methods, tools, standards (such as ISA-18.2 and EEMUA-191), and
activities that improve system performance by improving the effectiveness of
alarm systems.
The performance of an alarm management system usually can be specified by
three indices, namely, the false alarm rate (FAR), missed alarm rate (MAR),
and averaged alarm delay (AAD).
6. Alarm Management
The other purposes of Alarm Management :
Elimination of multiple alarms from the same cause
Root cause identification of alarms
Prioritizing and grouping the alarms
Tuning the alarm limits and delays
6
7. Alarm Management : Some Available Functions
Filters: IIR filters, averaging filter, median filter
Delay Timer
Deadband
Threshold
7
8. Definitions
8
In practice, because of uncertainty and noisy situation, in normal conditions,
sometimes an alarm is raised while no fault occurs in the system.
in abnormal conditions, maybe some alarms are not recognized by an alarm system
False Alarm
Missed Alarm
15. Extended Adaptive Threshold
15
1 1
1 1
( 1) (1 ) ( ),
( 1) (1 ) ( )
n n n
n n n
k k
m m k m k
( ) ( ) ( )n n n n
T k m k k
2 2
2 2
( 1) (1 ) ( ),
( 1) (1 ) ( )
Abn Abn Abn
Abn Abn Abn
k k
m m k m k
( ) ( ) ( )Abn Abn Abn Abn
T k m k k
we consider two adaptive threshold
window length = 25
momentum factor= 0.5
17. Tuning Factors
17
It is clear that proposed method could be reduced FAR and MAR
simultaneously by increasing Alfa parameter in considered interval.
excessive increasing Alfa could be ruined MAR and FAR.The reason is that due to
Excessive large Alfa maybe the maximum ( minimum) value of abnormal (normal)
data is lower (bigger ) than the threshold. So alarm systems could not detect
abnormal (normal) condition after normal (abnormal) signal occurred.
18. Tuning Factors
18
Genetic Algorithm-based Optimization
arg min, , , ,N Abn N Abn
Jw w
1 2 3
, ,N Abn
FAR MAR AAD
J w
RFAR RMAR RAAD
19. Non-Parametric Performance Assessment
19
Find FAR & MAR in EAT
𝐹𝐴𝑅 =
0
𝑥1
𝑓𝐴𝑁 𝑥 𝐹𝑄 𝑥 𝑑𝑥
𝑀𝐴𝑅 =
0
𝑥2
𝑓𝑃 𝑥 𝐹𝐴𝑏𝑛 𝑥 𝑑𝑥
Validation: using Monte Carlo
Proposed
Solution
Monte Carlo (1e07 Iteration)Performance
Ind./Method VarianceMean
0.0312630.76e-040.031264MAR
0.0312630.76e-040.031264FAR
24. Hypothetical Example 3
24
𝑥( 𝑡)~𝑁(0,0.5) 𝑡 ≤ 500
𝑥( 𝑡)~ (1.5,0.6) 500 ≤ 𝑡 < 1300
𝑥( 𝑡)~𝑁(0,0.5) 1300 ≤ 𝑡 < 2000
𝑥( 𝑡)~ (1.5,0.6) 2000 ≤ 𝑡 < 2900
𝑥( 𝑡)~𝑁(0,0.5) 2900 ≤ 𝑡 < 3700
Non-parametric
intermittent
without label
Usually, operator record normal data and designer could be defined each part
of data as a supervisor, so the normal behavior of a system is known.
sometimes there is no information about labeling. In this case, data must be
separated as an unsupervised system.
Labeling of Data
25. 25
Main Problem :
How can we find abnormal data ????
FCM
change point
detection
Solution
Paper: Performance Assessment and Design for Univariate Alarm Systems Based on
FAR, MAR, and AAD
26. 26
For {𝑥(𝑖)}𝑖=1
𝑇
, find one mean change point
𝑈𝑖,𝑇 = 𝑉𝑖,𝑇 𝑖 = 1
𝑈𝑖.𝑇 = 𝑈𝑖−1.𝑇 + 𝑉𝑖.𝑇 𝑖 = 2.3. … . 𝑇
𝑉𝑖.𝑇 =
𝑗=1
𝑇
𝑠𝑔𝑛 𝑥 𝑖 − 𝑥(𝑗)
Find the time instant 𝑖 𝑚𝑎𝑥 maximizing 𝑈𝑖.𝑇
𝑃 = 2𝑒𝑥𝑝
−6 𝑚𝑎𝑥
1≤𝑖≤𝑇
𝑈𝑖.𝑇
2
𝑇2 + 𝑇3
that 𝑥(𝑖 𝑚𝑎𝑥) is the change point
of {𝑥(𝑖)}𝑖=1
𝑇∝= 0.01
∝< 𝑝∝> 𝑝
End
Divide {𝑥(𝑖)}𝑖=1
𝑇
into two subsections:
{𝑥1(𝑖)}𝑖=1
𝑖 𝑚𝑎𝑥
And {𝑥2(𝑖)}𝑖 𝑚𝑎𝑥
𝑇
according to
𝑖 𝑚𝑎𝑥
Algorithm
find all the change points
27. 27
data must be divided into normal and
abnormal parts
Change points are known
Consider slope of 𝑈𝑖,𝑇 in each intervals
positive sign of slope of 𝑈𝑖,𝑇means this part consisted of normal
data and negative one means this part contain abnormal data.
, ,
( ) ( )i T e i T s
e s
U D U D
m
D D
28. Hypothetical Example 3
28
Normal/AbnormalSign of slopeIntervals
NormalNegative[1,499]
AbnormalPositive[500,1300]
NormalNegative[1301,1999]
AbnormalPositive[2000,2899]
NormalNegative[2900,3700]
0 500 1000 1500 2000 2500 3000 3500 4000
-8
-6
-4
-2
0
2
4
6
8
10
x 10
5 Mean-Change Point Detection
Time (Sec.)
Value
U
29. Hypothetical Example 3
29
0 500 1000 1500 2000 2500 3000 3500
-2
-1
0
1
2
3
4
5
6
Example3
Extended Adaptive Threshold, N
=5 ,Abn
=3
Time (Sec.)
x(t),Measurement
Normal
Abnormal
Extended Adaptive Threshold
0 500 1000 1500 2000 2500 3000 3500
0
0.5
1
Alarm Signal with 3 delays & Simple Threshold
Time (Sec.)
x(t),Measurement
0 500 1000 1500 2000 2500 3000 3500
0
0.5
1
Alarm Signal with 3 delays & Adaptive Threshold
Time (Sec.)
x(t),Measurement 0 500 1000 1500 2000 2500 3000 3500
0
0.5
1
Example 3: Alarm Signal in with 3 delays & Extended Adaptive Threshold
x(t),Measurement
0 500 1000 1500 2000 2500 3000 3500
-3
-2
-1
0
1
2
3
4
5
6
7
Example 3: Test Data by applying DEAT
Time (Sec.)
x(t),Measurement
Normal
Abnormal
EAT from train data
:Optimal Parameters
window length = 38 ,𝛼 𝑁
= 5.32 ,𝛼 𝐴𝑏𝑛 = 3.08
30. Example 4: V94-2 Gas Turbine Measurement
108 sensors
Vibration sensor
blue color of the curve belongs to the abnormal condition (before systems overhaul) and green
color of the curve appertain to the normal condition (after systems overhaul)
30
EAT
(α = 79.38)
EAT
(α = 52.71)
Simple Threshold
(with deadband)
AT
Simple
Threshold
Performance Ind
0.0161480.0900550.3646540.3075690.460713MAR
0.0671450. 0151310.2289100.1395270.192783FAR
31. Capacities of the EAT
The first conspicuous potential of the proposed method is to reduce MAR and
FAR simultaneously.
The mean change detection algorithm was modified and simplified in this
study with less computational complexity.
With the combination of both EAT and N-sample delay timer, the nuisance
alarms are reduced. The use n-sample delay timer enables us to implement
extended adaptive threshold on online measurements.
In this study, a brief solution for performance assessment of EAT has been
proposed.
A multi-objective Genetic Algorithm was applied for designing appropriate
extended adaptive threshold.
31
32. Limitations of the EAT
In this study, it is assumed that the abnormal measurements have a unique
pattern, especially in mean value. However, in some industrial examples and
after malfunction occurrence, there are different types of faults effected on the
abnormal measurements. As a future work, if instead of the proposed
algorithm, the other methodology with capabilities of fault diagnosis to be
replaced.
The combination of EAT and n-sample delay timer cause an inevitable delay
on the alarm system.
In this study, it is assumed that type of abrupt fault (rising or falling) is
known. It can be solved with more complex computation in change point
detection algorithm.
32