Docstoc

jha

Document Sample
jha Powered By Docstoc
					A Web-based Intelligent Hybrid
  System for Fault Diagnosis

          Gunjan Jha
       Research Student
 Nanyang Technological University
           Singapore
          Presentation Overview

•   Traditional Hotline Service Support
•   Related Work & Techniques
•   The WebService System
•   Customer Service Database
•   The Hybrid Approach
•   Summary & Conclusion


3/23/99               AAAI, SSS-'99       2
          Traditional Hotline Service
                   Support
• Customers located worldwide make long
  distance calls to the service centre
• The service engineer provides an advice to the
  customer by referring to the Customer Service
  Database (Knowledge Base)
• The service engineer may need to pay an
  onsite visit if the advice does not work


3/23/99              AAAI, SSS-'99                 3
                                          Problem


                                       Telephone line
                    Service                               Customer
                                          Advice
                   Engineer


              Search           Get
              system          result


                       Hot-line                         Customer
                       advisory                          service
                        system                          database




          Traditional Hotline Service Support

3/23/99                                 AAAI, SSS-'99                4
    Disadvantages of the traditional
      customer support process

• Expensive overseas telephone calls
• Expensive onsite trips by service engineers
• The need to train and maintain experienced
  service engineer
• Dependence on the service engineers and the
  customer service database


3/23/99            AAAI, SSS-'99                5
  Online Customer Service Support
• BBS (Bulletin Board System)
• ARWeb, Cognitive E-Mail, Target WebLink and
  ClearExpress WebSupport [Muller 96]
• Muller, N.J., 1996. Expanding the Help Desk through
  the World Wide Web. Information Systems
  Management, 13(3): 37-44.
• Compaq, NEC [Chang 1996]
• K. H. Chang, et al., 1996. A Self-Improving Helpdesk
  Service System Using Case-Based Reasoning
  Techniques. Computers in Industry, 30(2): 113-25.

3/23/99                AAAI, SSS-'99                     6
          The WebService System
                                             Inte lligent Fault
                                            Diagnosis System


                                                 Mainte nance
                                                  Program
                                                                                       Customer
                                                                                        Serv ice
                                                                                       Database

                   Serv ice Engineer


                                                Intelligent Fault
                                               Diagnosis Engine                           CLIPS
                                                                                        Rule-base




                                                                                     Neural Netw ork
                                                  Web Serve r                           Indexing
                                                                                       Database



                                                                                        Databases



                                                       Internet




            User                                                                                       User
                       Web Brow s er   Web Brow s er          Web Brow s er   Web Brow s er




3/23/99                                         AAAI, SSS-'99                                                 7
          Customer Service Record Database

• A fault record consists of
  – fault condition
  – checkpoints
• Example

Fault Condition: CASSETTE DETECTION ERROR.
Checkpoints: (1) IS THE CASSETTE 'SITTING' PROPERLY.
             (2) ENSURE THAT THE TAPE GUIDE IS PROPERLY SET.
             (3) CONFIRM THE OPTICAL MODULE. (T.G. PG. 10).



3/23/99                     AAAI, SSS-'99                      8
          Intelligent Fault Diagnosis
                  Techniques
•   Case based reasoning (most popular)
•   Artificial neural network (Learning Systems)
•   Rule based reasoning (for Quasi-static systems)
•   Miscellaneous techniques
          Fuzzy logic, Genetic algorithms, Decision trees
          and Statistical techniques
• Hybrid techniques

3/23/99                    AAAI, SSS-'99                    9
          The Hybrid Approach
• Based on hybrid CBR-ANN-RBR approach
• Integrate Neural Network into the CBR cycle for
  indexing, retrieval and learning
• Use Rule-based reasoning for Case-Reuse and
  assistance in carrying out the diagnosis
• Major tasks:
   – Knowledge Acquisition, Retrieval, Reuse, Revise
     and Retain


3/23/99                AAAI, SSS-'99                   10
                          Fault Diagnosis Process

                                                              Reuse of
                                                           serv ice records
                                                                                      user
     User                            retrieved                           service    feedback
                                  fault-condition                        records
             Problem
             scenario
            description

     Pre-                    Neural                       LVQ3        Rule-
                                                                                                Rev ise w ith
 processing of              netw ork                      neural     based
                                                                                               user feedback
   user input               retriev al                   netw ork    engine




    Customer
                           Know ledge                                update
     Service                                                                         revised
    Database               acquisition                              databases
                                                                                   information
                                                          Retain by updating
                                                              databases



3/23/99                                             AAAI, SSS-'99                                               11
               Knowledge Acquisition

                                                      Customer
                                                       Serv ice
                                                      Database

               Neural Network Training                                  Rule Generation


                                                                   Generation of
                                 Pre-processing of
          NN Training                                             control rules and
                                  f ault-conditions
                                                                  checkpoint rules




          NN Indexing
                                                                     Rule-base
           Database




3/23/99                                  AAAI, SSS-'99                                    12
               Forming a Weight Vector
                          Fault-condition                                    List of Keywords

     CUTTER & ANVIL CANNOT ENGAGE IN AFTER 1ST PCB.
                                                                        Index     Keyword
                                  Pre-processing
                                                                         1        ANVIL
                                                                         2        BREAK
                                                                         3        CUTTER
               Keywords Extracted             Index          key word    4        CAMERA
                                                              index      5        FAULTY
               CUTTER                            3
                                                                         6        GUIDE
               ANVIL                             1
                                                                         7        NC
               NOT_ENGAGE                        8
                                                                         8        NOT_ENGAGE
               PCB                               9
                                                                         9        PCB
                                                                        10        SHAKY
                                       Weight Vector

     .75   0    .75   0      0     0      0      .75   .75   0




3/23/99                                       AAAI, SSS-'99                                     13
          Checkpoint Rule for a Fault-
                  condition
    (defrule MAIN::chkpt_rule-7
          (phase accept.fault.cond)
          (fault.cond AVF_CHK007)
          =>
          (assert (check-seq AVF_CHK007-1 AVF_CHK007-2 AVF_CHK007-3
           AVF_CHK007-4 AVF_CHK007-5))
          (assert (help-seq AVF_CHK007-1.GIF AVF_CHK007-2.GIF
           AVF_CHK007-3.GIF AVF_CHK007-4.GIF AVF_CHK007-5.GIF))
    )




3/23/99                         AAAI, SSS-'99                         14
                                                                User                Feedback on
                          User Input                                                  Synonym


               THE CARRIER DID NOT TRANSFER THE                        Unknown Keyword     Synonym      Index
                      PCB DURING LOADING.
                                                                       STANDSTILL          STAY          12

                    PCB                STANDSTILL
                                                                                            keyw ord index
                     PCB DETECTOR SENSOR                           List of keyw ords

                                                                              Index      Keyword
                                Pre-processing
                                                                              1          ANVIL
                                                                              2          CARRIER
                                                                              3          DETECT
          Keywords Extracted               Index
                                                                              4          DRIVE
                                                           keyw ord
                                                                              5          LOAD
          CARRIER                           2               index             6          NOT_INSERT
          DETECT                            3
                                                                              7          NOT_TRANSFER
          LOAD                              5
                                                                              8          PCB
          NOT_TRANSFER                      7
                                                                              9          PUMP
          PCB                               8
                                                                              10         RAIL
          SENSOR                            11
                                                                              11         SENSOR
                                                                              12         STAY
                                                                              13         TRANSFORMER


                                       Input Vector

           0    1     1    0   1       0     1     1   0    0      1     0    1     0



3/23/99                                                AAAI, SSS-'99                                            15
3/23/99   AAAI, SSS-'99   16
3/23/99   AAAI, SSS-'99   17
3/23/99   AAAI, SSS-'99   18
          Reuse of Checkpoint Solution
                           Customer
                                                          Rule-base
                            Serv ice
                                                         (Checkpoint
                           Database
                                                           Rules)
                          (Unique Set)




                                           Rule-based    Checkpoint
              Retrieved                                   solution
                                            Inf erence
            fault-condition                  Engine

                                                                       User
                                                            Feedback




3/23/99                                  AAAI, SSS-'99                        19
                                               Use r's fe e dback

                                 Successful
                            Correct fault-condition                      Failure
                               and checkpoint                     Checkpoint solutions fail
       NN indexing
                                           Rule-base update
     database update


                                            Checkpoint             serv ice
          Update key word list                priority             record            Fault diagnosis
                                            modif ication                            through serv ice
                                                                                     engineer's help            Serv ice
                                                                                                                engineer
            Update f ault-                                                                    serv ice report
                                             Rule-base
            conditions to
                                             generation
          key word-list index
                                                                                     Update serv ice            Maintenance
                                                                                        record                   program

          Update input v ector


                                          NN                          Customer
                                                                                      Pre-processig
                                       Indexing       Rule-base        Serv ice
           Weights update                                                              and training
                                       Database                       Database




3/23/99                                                   AAAI, SSS-'99                                                       20
          Case Retain or Maintenance Module
3/23/99                 AAAI, SSS-'99         21
                   Summary
• The research has successfully demonstrated the
  effectiveness of hybrid CBR-ANN-RBR approach for
  the fault diagnosis problem
• Performance analysis have proved the approach to be
  much accurate and efficient than the traditional CBR
  techniques (Nearest Neighbor)
• Future work focuses on incorporating genetic
  algorithm and data mining techniques for better
  accuracy and efficiency

3/23/99                AAAI, SSS-'99                 22
          Performance Analysis
• Performance compared with traditional CBR
  systems using kNN technique
• Retrieval Accuracy: test data from customer
  service database (size ~ 15000)
     – ANN: 93.2%
     – kNN1: 76.7% (Fuzzy Trigram)
     – kNN2: 81.4% (Euclidean distance based matching)


3/23/99                 AAAI, SSS-'99                23
   Performance Analysis (…continued)

• Retrieval Accuracy: test data from the user input (size
  = 50)
   – ANN: 88%
   – kNN1: 78% (Fuzzy Trigram)
   – kNN2: 72% (Euclidean distance based matching)
• Average Retrieval Speed (test size ~ 15000)
   – ANN: 1.9s
   – kNN1: 12.3s
   – kNN2: 9.6s
3/23/99                 AAAI, SSS-'99                   24
          3 possible methods to Update
            Checkpoint-Rule Priority
• Method 1: No need to change the priorities of
  checkpoints.
• Method 2: Assign priority “1” to the checkpoint
  that solves the problem and decrease the priorities
  of the checkpoints ahead of it by “1”.
• Method 3: Swap the priority of the checkpoint that
  solves the problem with the one just ahead of it.


3/23/99               AAAI, SSS-'99                25
   Performance comparison of three
 methods to update checkpoint priorities.
                            5000


                            4000
          Frequency count




                                                                                             method1
                            3000
                                                                                             method2
                                                                                             method3
                            2000


                            1000


                              0
                                   1   2   3   4   5     6    7   8    9     10   11   12   13   14   15
                                                       Checkpoint priority




3/23/99                                                  AAAI, SSS-'99                                     26

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:0
posted:2/21/2012
language:
pages:26