Fuzzy Systems in Use for Human Reliability Analysis

1
Fuzzy Systems in Use for Human Reliability
Analysis

• Myrto Konstandinidou
• Zoe Nivolianitou
• Nikolaos Markatos
• Christos Kyranoudis

Loss Prevention Prague, June 2004
2
Outline

• Introduction
• The Fuzzy Logic as a modeling tool
• Methods for Human Reliability Analysis
• The CREAM methodology
• Development of the Fuzzy Classification System
• Results
• Conclusions

3
Introduction

• HRA is a critical element for PRA
• Most important concerns
• - the subjectivity of the methods
• - the uncertainty of data
• - the complexity of the human factor per se
• Fuzzy logic theory has had many relevant
  applications in the last years

4
Fuzzy Logic as a modeling tool (1)

• Fuzzy logic (FL) is a very useful tool for
  modeling
• - complex systems
• - qualitative, inexact or uncertain information
  
• FL resembles the way humans make inference and
  take decisions
• FL accommodates ambiguities of real world human
  language and logic

5
Fuzzy Logic as a modeling tool (2)

• Applications
• - Automatic control
• - Data classification
• - Decision analysis
• - Computer Vision
• - Expert systems

The most used fuzzy inference method Mamdanis
method (1975)
6
Fuzzy Logic as a modeling tool (3)

• Definitions
• FL allows an object to be a member of more that
  one sets and to partially belong to them.
• - Fuzzy set
• - Degree of membership
• - Partial membership

7
Fuzzy Logic as a modeling tool (4)

• The 3 steps of a FL system
• Fuzzification the process of decomposing input
  variables to fuzzy sets
• Fuzzy Inference a method to interpret the values
  of the input vectors
• Defuzzification the process of weighting and
  averaging the outputs

Fuzzification
Defuzzification
Crisp Output
Crisp Input
Inference
8
Methods of Human Reliability Analysis

• Fundamental Limitations
• Insufficient data
• Methodological limitations
• Uncertainty
• Most important methods developed for HRA
• THERP
• CREAM
• ATHEANA

9
CREAM Methodology (1)

• The choice of CREAM was made because
• It is well structured and precise
• It fits better in the general structure of FL
• It presents a consistent error classification
  system
• This system integrates individual, technological
  and organizational factors

10
CREAM Methodology (2)

• Control Modes
• Scrambled
• Opportunistic
• Tactical
• Strategic

Definition of Common Performance Conditions
(CPCs) to be used in FL model
11
Development of a Fuzzy Classifier (1)

• Experience
• - Accident analysis
• - Risk assessment
• - Human reliability

• Data
• - Diagrams of CREAM
• - MARS Database
• - Incidents and accidents from the
  Greek Petrochemical Industry

12
Development of a Fuzzy Classifier (2)

• The Development of the Fuzzy Classification
  System for Human Reliability Analysis

13
Development of a Fuzzy Classifier (3)

• STEP 1 Selection of the input parameters

14
Development of a Fuzzy Classifier (4)

• STEP 2 Development of the Fuzzy sets
• Each input is given a number based on its quality
• 0 (worst case) - 100 (best case)
• Time of day from 000 (midnight) to 2400
• Output scale 0.510-5 - 1.0100

15
Development of a Fuzzy Classifier (5)
16
Development of a Fuzzy Classifier (6)

• Output fuzzy sets
• Probability of a human erroneous action

17
Development of a Fuzzy Classifier (7)
Input variable
18
Development of a Fuzzy Classifier (8)
Action Failure Probability
1
0
-5.30E00
-4.30E00
-3.30E00
-2.30E00
-1.30E00
-3.00E-01
Strategic
Probability interval
Output
Tactical
Opportunistic
Scrambled
19
Development of a Fuzzy Classifier (9)

• STEP 3 Development of the fuzzy rules
• Based on CREAM basic diagram
• Simple linguistic terms
• Logical AND operation

20
CREAM basic diagram
21
Development of a Fuzzy Classifier (10)

• Fuzzy model operations

22
Scenarios

• Five independent scenarios characterizing 5
  different industrial contexts
• Scenario 2 represents a best case scenario
• Scenario 4 represents a worst case scenario
• Scenarios 4 and 5 have slight differences in the
  values of input parameters

23
Results of test runs
Probability interval
Control Mode
Scenario
1.010-3ltplt1.010-1
1
Tactical
0.510-5ltplt1.010-2
2 (Best case)
Strategic
1.010-2ltplt0.5100
3
Opportunistic
1.010-1ltplt1.0100
4 (Worst case)
Scrambled
1.010-1ltplt1.0100
5
Scrambled
24
Comments on the results

• All FL model results in accordance with CREAM
• Best case scenario very low action failure
  probability
• Worst case scenario very high action
  failure probability
• Small differences in input have impact to output
• The results can be used directly in PSA methods
  (event trees, fault trees, etc.)

25
Conclusions (1)

• FL system to estimate the probability of human
  erroneous action has been developed
• Based on CREAM methodology
• 9 input variables
• 1 output parameter

26
Conclusions (2)

• Test runs for 5 different scenarios
• Very satisfactory results
• Main difference between FL model and CREAM
  probabilities estimation are exact numbers
• The results can and will be used in other PSA
  methods

27
Further goals

• Model calibration with data from the Greek
  Petrochemical Industry
• Addition of other CPCs or PSFs
• Expansion to other fields of the chemical
  industry
• Application in other fields of technology
• (e.g aviation technology, maritime
  transports, etc)

28
Acknowledgments

• The Financial support of the EU Commission
  through project PRISM GTC1-2000-28030 to this
  research is kindly acknowledged