A Web-based Intelligent Hybrid System for Fault Diagnosis Gunjan Jha Research Student Nanyang Technological University Singapore.

A Web-based Intelligent Hybrid System for Fault Diagnosis

Gunjan Jha

Research Student

Nanyang Technological University

Singapore

3/23/99 AAAI, SSS-'99 2

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 3

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 4

C ustom erservice

database

Searchsystem

G etresu lt

C ustom erServiceEng ineer

Hot-lineadvisorysystem

Telephone line

Advice

Prob lem

Traditional Hotline Service Support

3/23/99 AAAI, SSS-'99 5

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 6

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 7

The WebService System

Internet

Web BrowserWeb BrowserWeb Browser Web BrowserUser

Intelligent FaultDiagnosis System

Service Engineer

Intelligent FaultDiagnosis Engine

Web Server

CLIPSRule-base

Neural NetworkIndexing

Database

MaintenanceProgram

Databases

User

CustomerService

Database

3/23/99 AAAI, SSS-'99 8

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 9

Intelligent Fault Diagnosis Techniques

• Case based reasoning (most popular)

• Artificial neural network (Learning Systems)

• Rule based reasoning (for Quasi-static systems)

• Miscellaneous techniquesFuzzy logic, Genetic algorithms, Decision trees

and Statistical techniques

• Hybrid techniques

3/23/99 AAAI, SSS-'99 10

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 11

Fault Diagnosis Process

LVQ3neural

network

Revise with user feedback

Knowledgeacquisition

Reuse of service records

Retain by updatingdatabases

CustomerService

Database

Neuralnetworkretrieval

servicerecords

revisedinformation

Problemscenario

description

User retrievedfault-condition

userfeedback

updatedatabases

Pre-processing of

user input

Rule-basedengine

3/23/99 AAAI, SSS-'99 12

Knowledge Acquisition

CustomerService

Database

NN Training

NN IndexingDatabase

Rule GenerationNeural Network Training

Pre-processing offault-conditions

Generation ofcontrol rules andcheckpoint rules

Rule-base

3/23/99 AAAI, SSS-'99 13

Forming a Weight Vector

Keywords Extracted Index

CUTTER 3ANVIL 1NOT_ENGAGE 8PCB 9

Index Keyword

1 ANVIL 2 BREAK 3 CUTTER 4 CAMERA 5 FAULTY 6 GUIDE 7 NC 8 NOT_ENGAGE 9 PCB 10 SHAKY

CUTTER & ANVIL CANNOT ENGAGE IN AFTER 1ST PCB.

Pre-processing

List of Keywords

.75 00 0 000

Fault-condition

keywordindex

Weight Vector

.75.75.75

3/23/99 AAAI, SSS-'99 14

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 15

Unknown Keyword Synonym Index

STANDSTILL STAY 12

THE CARRIER DID NOT TRANSFER THE PCB DURING LOADING.

Keywords Extracted Index

CARRIER 2DETECT 3LOAD 5NOT_TRANSFER 7PCB 8SENSOR 11

PCB

PCB DETECTOR SENSOR

STANDSTILL

0 00 01111111 000

Input Vector

keywordindex

User Input

Index K eyword

1 ANVIL 2 CARRIER 3 DETECT 4 DRIVE 5 LOAD 6 NOT_INSERT 7 NOT_TRANSFER 8 PCB 9 PUMP10 RAIL11 SENSOR12 STAY13 TRANSFORMER

Pre-processing

keyword index

User

List of keywords

Feedback onSynonym

3/23/99 AAAI, SSS-'99 16

3/23/99 AAAI, SSS-'99 17

3/23/99 AAAI, SSS-'99 18

3/23/99 AAAI, SSS-'99 19

Reuse of Checkpoint Solution

CustomerService

Database(Unique Set)

Rule-basedInference

Engine

CheckpointsolutionRetrieved

fault-condition

User

Feedback

Rule-base(Checkpoint

Rules)

3/23/99 AAAI, SSS-'99 20

User's feedback

Checkpointpriority

modification

SuccessfulCorrect fault-condition

and checkpointFailure

Checkpoint solutions fail

CustomerService

DatabaseWeights update

Update input vector

Update keyword list

Update fault-conditions to

keyword-list index

NN indexingdatabase update

Rule-basegeneration

Rule-baseNN

IndexingDatabase

Fault diagnosisthrough serviceengineer's help

service report

Pre-processigand training

Update servicerecord

Maintenanceprogram

Serviceengineer

Rule-base update

servicerecord

3/23/99 AAAI, SSS-'99 21

Case Retain or Maintenance Module

3/23/99 AAAI, SSS-'99 22

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 23

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 24

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 25

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 26

Performance comparison of three methods to update checkpoint priorities.

0

1000

2000

3000

4000

5000

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

Checkpoint priority

Fre

qu

ency

co

un

t

method1

method2

method3