Case Based Reasoning fault diagnosis for cars

1
Case Based Reasoningfault diagnosis for cars

• Sven Burmester, Andreas Goebels

2
Content

• Introduction
• Modeling
• Similarity
• CBR Cycle
• Retrieve
• Reuse
• Revise
• Retain
• Conclusion

3
CBR-Cycle
NewCase
LearnedCase
Retain
Retrieve
Database
RetrievedCase
Tested/repairedCase
General Knowledge
Reuse
Revise
SolvedCase
4
Problem Tasks
CBR System
synthetical tasks
analytical tasks
classification
rating
prognosis
configuration
planning
Diagnosis
design
decisionsupport
5
Example

• Problem Long distance light failure in cars
• Attributes Case I Case II
• --------------------------------------------------
  ---------
• car type Golf II Audi A6
• construction year 1993 1995
• batterie voltage 13.6 V 12.5 V
• lighter glas OK broken
• Solution fuse defect lamp defect
• current Problem brake light failure
• Attributes new case
• -------------------------------------------------
• car type Audi 80
• construction year 1989
• batterie voltage 12.6 V
• lighter glas OK
• Solution ?

6
Modeling / Case Representation

• Attribute value representation
• Object oriented representation
• Graph representation
• First order logic representation
• Hierarchical case representation

7
Modeling Attribute Value Representation

• Attributes A1, , An
• Valuations T1, , Tn ? i unknown ? Ti
• Values a1 ? T1, , an ? Tn
• Case C ? T1 x x Tn x Solutions
• Set S C1, , Cn of cases

a1 an s
8
Modeling Attribute Value Representation

• Attributes

9
Modeling Object oriented Representation
Car Failures
car body
car engine
car illumination
spark plug

rust spots

front illumination
back illumination
left front lamp
right front lamp
10
Similarity

• Distance Measure dist S2 -gt ?
• x ? S dist(x, x) 0
• x, y ? S dist(x, y) dist(y, x)
• Example Hamming Distance
• Transformation
• bijective function f ? -gt 0, 1
• f(dist(x, y))1 dist(x, y) (dist(x,
  y)1)-1sim(x, y)
• f(0) 1
• Similarity Measure sim S2 -gt 0,1
• x ? S sim(x, x) 1
• x, y ? S sim(x, y) sim(y, x)

11
CBR-Cycle
NewCase
LearnedCase
Retain
Retrieve
Database
RetrievedCase
Tested/repairedCase
General Knowledge
Reuse
Revise
SolvedCase
12
Similarity Example

• Problem Similarity between C1 / C2 and the new
  Case Cnew

C1
C2
Cnew

• Similarity of ordered symbols Difference ?
  Distance
• dist(1993, 1989) 1993 1989 4
• sim(1993, 1989) 0.79
• dist(1995, 1989) 1995 1989 6
• ? sim(1995, 1989) 0.55

13
Example

• Similarity of unordered symbols
• Similarity table sim(x, y) sx, y
• reflexive function gt symmetric table
• unknown attributes
• optimistic / pessimistic / expected value
  strategy

14
Example
C1
Cnew

• Weight each Valuation w( )

• sim(C1, Cnew) 0.2 0.50 0.1 0.79
• 0.3 0.91 0.4 1.00 0.82
• sim(C2, Cnew) 0.2 0.80 0.1 0.55
• 0.3 0.99 0.4 0.00 0.44

15
Case Base

• Problems
• find enough (reasonable) example cases
• preprocess the cases correct / efficient
• find errors / contradictions (consistency of
  Case Base)

16
Retrieval

• Problem Searching a huge Knowledge Base is time
  consuming
• sequencial retrieval O(n)
• ? Find efficient Retrieval methods
• two-stage retrieval
• retrieval with kd-trees
• retrieval with networks
• retrieval with Fish and Shrink
• Retrieval tasks
• most similar case
• k most similar cases (ordered / unordered)
• all cases with at least simmin similarity

17
CBR-Cycle
NewCase
LearnedCase
Retain
Retrieve
Database
RetrievedCase
Tested/repairedCase
General Knowledge
Reuse
Revise
SolvedCase
18
Reuse
Solution adaption
Diagnosis situation
no adaption
transformation based
generativ


substitutionaladaption
structuraladaption
19
CBR-Zyklus
NewCase
LearnedCase
Retain
Retrieve
Database
RetrievedCase
Tested/repairedCase
General Knowledge
Reuse
Revise
SolvedCase
20
Revise

• Could the problem be solved successful?
• Typical questions
• contains the case useful new knowledge?
• occurs the problem frequently?
• is it possible to store the knowledge in the case
  base?

21
CBR-Zyklus
NewCase
LearnedCase
Retain
Retrieve
Database
RetrievedCase
Tested/repairedCase
General Knowledge
Reuse
Revise
SolvedCase
22
Retain

• The learning component of CBR
• Case Base
• add new case
• and / or reorganize Case Base

23
Conclusion

• Hard to define a similarity function
• Hard to organize the Case Base
• You can use existing solutions
• You can find solutions for new problems
• Incomplete problem descriptions possible
• Low request time
• Requires sparse maintenance effort
• Requires sparse general knowledge
• Improves over time
• High user acceptance

24
Literature

• R. Bergmann, M. Richter Wissensverarbeitung für
  Electronic Commerce, 2001
• O. Niggemann, B. SteinWissensbasierte Systeme,
  2000
• T. M. Mitchell
• Machine Learning, 1997

25
This is the End

• Thank you for your attention