ADAPTATION BASED ON KNOWLEDGE MODELS FOR

DIAGNOSTIC SYSTEMS USING CASE-BASE REASONING

Brigitte Chebel Morello

, Mohamed Karim Haouchine

and Noureddine Zerhouni

Institute of Automatic Control and Micro-Mechatronic Systems, 24, Rue Alain Savary, 25000 Besançon, France

Em@systec sas, 18 Rue Alain Savary, 25000 Besançon, France

Keywords: Case-based reasoning, Adaptation, Adaptation-guided retrieval, Dependency relations, Hierarchical model,

Context model, Industrial diagnostic, Diagnostic help system, Industrial diagnostic.

Abstract: The adaptation phase is a key problem in the design of Case-Based Reasoning (CBR) systems. In most

cases, adaptation methods are application-specific. Our challenge in this work is to make a general

adaptation method for the field of industrial diagnostics. This paper is a contribution to fill this gap in the

field of fault diagnostic and repair assistance of equipment. Our adaptation algorithm relies on hierarchy

descriptors, an implied context model and dependencies between problems and solutions of the source

cases. In addition, we note that the first retrieved case is not necessarily the most adaptable case, and to take

into account this report we propose in our diagnostic problem an adaptation-guided retrieval step based on a

similarity measure associated with an adaptation measure. These two measures allow selecting the most

adaptable case among the retrieved cases. The two retrieval and adaptation phases are applied on real

industrial system called SISTRE (Supervised industrial system of Transfer of pallets).

1 INTRODUCTION

The objective of this study is to build an intelligent

application based on knowledge management for

industrial diagnosis and repair in a context of

maintenance services. It is targeted maintenance

operators to aid in their daily tasks. This decision

tool is developed within the framework of the

distributed e-maintenance platform. The platform

brings a major asset to maintenance interventions

and maintenance services in general by enabling

expertise via Internet to be went directly to the user

site. Our objective is to develop a case-based

reasoning system dedicated to industrial diagnosis in

order to solve a practical problem of an industry.

CBR is the technology of experience based

system, and is an approach to problem solving by

retrieving a similar past problem from the case base

by adapt it in the new context and by learning it.This

method is well suited to the diagnosis application

because fault diagnosis is one of domain based in the

experience of the human expertise, where problems

are recurrent and can be reuse.

(Althoff, 1996) thinks the CBR is the technology

of choice to implement a knowledge based system.

Moreover, CBR is frequently proposed as a

methodology for knowledge management

application. It presents expert knowledge as past and

concrete experiences easily understandable by

human users. Our objective is to solve diagnosis

problems by reusing cases in other contexts, by

adaptation phase of CBR. This phase is complex and

is usually designed for a specific application. Some

studies into “memory-based reasoning” (Kasif,

1995) avoid this step because the wealth of the case-

base can compensate for the adaptation phase

(Stanfill, 1986). However, other authors, like us,

develop this phase to enrich the case-base. In this

context the adaptation step is the core of CBR for

better exploiting the characteristics and strength of

the CBR (Chebel-morello, 2009), (Lieber, 2007).

Furthermore, prior works on adaptation were

dedicated to a given application. Our challenge in

this paper is to define a general adaptation method

on symbolic data in the field of industrial diagnostic.

In section 3 we develop an adaptation retrieval phase

folowing by the the adaptation phase. This method is

based on the dependencies between the problem and

the solution of a solved case and exploits two

knowledge models. Three relations of dependencies

are defined and exploited to adapt a retrieved case

within an adaptation algorithm described in the same

223

Chebel Morello B., Karim Haouchine M. and Zerhouni N..

ADAPTATION BASED ON KNOWLEDGE MODELS FOR DIAGNOSTIC SYSTEMS USING CASE-BASE REASONING.

DOI: 10.5220/0003666602230229

In Proceedings of the International Conference on Knowledge Management and Information Sharing (KMIS-2011), pages 223-229

ISBN: 978-989-8425-81-2

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Section. The matching carried out at the time of the

retrieval, combined with dependency relations

between the problems and solutions, can adapt the

solution to the target problem.

This paper contains two contributions: the first

one relates to the establishment of two measures to

select the most easily adaptable among the retrieved

cases. A Retrieval Measure (RM) is combined with

an Adaptation Measure (AM) specifically defined

for diagnostic systems.

The second one is an adaptation algorithm,

dedicated to the diagnostic of industrial plants that

builds both on knowledge models and on the

dependency relations between problem and solution

descriptors.

The proposed approach will be applied

throughout this paper on a Supervised Industrial

System for pallets TRansfEr (SISTRE) (Rasovska,

2007). It represents a flexible production system and

it is composed of five robotized working stations

which are served by a transfer system of pallets

organized into double rings (internal and external).

Each station is equipped with pneumatic actuators

(pushers, pullers and indexers) and electric actuators

(stopper) as well as a certain number of inductive

sensors (proximity sensors). An inductive read/write

module allows to identify and locate each pallet and

to provide information relative to required operation

in a concrete station. The moving of the pallets is

ensured by friction on belts which are involved by

electric motors. Each pallet has a magnetic label that

is used like embarked memory. This memory can be

read in each working station thanks to magnetic

read/write modules (Balogh) and allows the

memorisation of the product assembly sequence.

These labels thus enable to track the pallet path

through the system. The feasibility of our approach

will be studied in Section 6 through 125 generic

cases.

2 CASE REPRESENTATION

2.1 The Diagnostic Case

The case base reflects the experiment by the link of

dysfunctional mode of component and the cause of

this fault, and the action of repair.

Indeed, this representation is based on the

standard definition of diagnostics which is the

following one (Maintenance terminology, 2001):

“they are the actions carried out for the detection of

breakdown, its localization and the identification of

the cause”. We exploit these three parts in the

formalization of the case, which, moreover, relies on

the knowledge models of the equipment to be

diagnosed. Thus, we have the localization and the

functional part in the problem space of case, the

detection part of the failure class and the

identification of the failing component in the

solution space of case.

A case is composed by Problem and Solution

part: Case= (ds1, ds2, dsi …., Ds1, Ds2, Ds3, Ds4).

The problem part is composed by two kinds of

descriptors: (i) the “localization” descriptors are

linked at a conceptual graph (see Figure2) where the

node is the value of the descriptor determined in this

part, and the solution is the failed zone. The failed

zones are composed of the components potentially

failed.

(ii) The “supervisor” descriptors are defined by

three attributes relating to the component value, its

state and its functional mode (Figure 3):

= (

tate

value

A functional mode is an operating mode of a

component of the equipment. Abnormal operating

mode: corresponds to a system malfunction that is to

say there was a failure.

The solution part is composed by four

descriptors the first one Ds

, is relative at the class of

the fault component, the second Ds

is dedicated at:

the cause of failure, the next one describes Ds

: the

Repair action and the last one Ds

define the zone of

the failure.

Table 1: Generic structure of the case.

Example:

Let a specific component in equipment a puller. This

component can have two state linked at it position:

[front; back] and can have two functional modes

[normal, abnormal].

The descriptor associated at the puller in a diagnosis

case can write:

Table 2: Case example.

2.2 Models Associated with the Case

Moreover, the knowledge representation is based on

two models associated with the case-base, namely:

KMIS 2011 - International Conference on Knowledge Management and Information Sharing

224

the context model and the components taxonomy

model.

2.2.1 The Components Taxonomy Models

The equipment analysis determines sets of their

components. Every group of components is

regrouped by functional classes, and constitutes a

components' hierarchy which is common to the

source problem descriptors “ds” and source solution

“Ds”. The SISTRE hierarchical model of

components is described in Figure 1.

A case will have a formalization object and will

define a hierarchy of descriptors containing both

problem and solution descriptors (Haouchine, 2008).

Equipment (component)

Actuator

Magnetic sensor

Conveyor

Pneumatic

actuator

Electric

actuator

Balogh

Stopper

Brace

Belt

Motor

Pallet

Balogh0

Balogh1

Puller

Pusher

Instance

Class

Indexer

…

Detector

Speed

transmission

Figure 1: A part of the SISTRE’s components hierarchy.

The descriptors of the localization part are exploited

by a context model in two phases of the CBR. In the

elaboration phase, the user is asked a dynamic tree

of questions. Adaptation phase will select the correct

element to be substituted in the adaptation

algorithm.

In our study, the adaptation cost is quantified by

a measure called AM which taking account a crucial

descriptor related to the diagnostic: the functional

mode.

2.2.2 Context Model

The context model is a contextual graph allowing

the localization of components comprising a failure

and selection of concerned components compared to

the set. Therefore, the context model enables to

inform the “localization” descriptors in order to

determine the failure zone and the components

potentially failing. The course of a pallet will be

followed. Using a contextual graph, as shown on

Figure 2, components likely to be failing will be

localised.

Figure 2: A part of contextual graph of the equipment.

An example of a context model concerning the

descriptor “Ds

” is shown on Figure 3.

Figure 3: A context model of “Ds

” descriptor.

The context allows the localization of components

problems and the selection of the right descriptors

compared to all others. Therefore, these present

components

constitute the context in which the

failing component is identified. A dependency

relation is associated with these components.

2.3 The Diagnostic Case Base

We take inventory of 125 cases in SISTRE case base

the case problem part is composed of seven

descriptors. The first two descriptors define the

localization of the failure. This localization is

determined by “ds

: zone”, “ds

: palette site”.

Let us consider the example of case S1 in the

Table 3. This case represents a problem on the “D1

detector”. The localization part determines that there

is a failure on the entry of the principal ring. Then,

the supervisor part provides the components state

implied in this place. The S1 stopper is in “top”

position which has a “normal” functional mode. The

balogh0 has value “1”, which means that it must

enter the working area so that it can be treated by a

robot. Finally, the D1 sensor does not detect the

presence of the pallet which is in “abnormal” mode.

The solution part is made up of a class descriptor

of failing component, of a descriptor identifying the

failing component, of the repair action and of the

failure zone.

ADAPTATION BASED ON KNOWLEDGE MODELS FOR DIAGNOSTIC SYSTEMS USING CASE-BASE

REASONING

225

Table 3: A part of the SISTRE case base.

3 RETRIEVAL PHASE

There are two categories of retrieval phase: the first

to be called “simple retrieval" and the second

“combination retrieval/adaptation”. Our study is

focused on the second type. We note that the most

similar case is not always the best candidate for

adaptation (Smyth, 1995) Consequently we propose

an adaptation guided retrieval method applied at the

industrial diagnostic based on two measures the first

one of similarity the second one of adaptation.

3.1 Retrieval Measure

We propose four local similarity measures are

exploited: - For the value of

value

, which belongs

to the hierarchical model of descriptors, φvalue is

developed.

value

then φvalue = 1 (“dt” for target

descriptor)

value

≠

value

then

 φ

value

= 0.8 if the value are in the identical level

value

=val1 and

value

=val2

 φ

value

= 0.6 there is one level of differences

value

=val1 and

value

=val4

 etc…(see Fig. 5).

Figure 4: Example of descriptor hierarchy.

- For the descriptor value

state

and for the

functional, φ

state

and φ

is calculated in the same

way.

 If

state

then φstate = 1 and If

then φ

= 1

 If

state

≠

state

then φstate = 0 and If

then φ

= 0

The similarity metric depends on the formalization

of the case. Note that all the descriptors are not all

inquire. The similarity measure will reflect the

presence of descriptors in the case.

To take into account presence and/or absence of

information in descriptors, a local similarity

presence

is developed.

 φ

presence

= 1 if component information is

present in

ds and dt descriptors



presence

= 0, if not

The global similarity measure (1) is obtained by

aggregation of these functions on the whole set of

KMIS 2011 - International Conference on Knowledge Management and Information Sharing

226

descriptors. From this measure, a set of cases can be

selected.

∑

×××

esence

presence

state

value

ϕϕϕϕ

(1)

Where m: represent the number of problem

descriptors.

The retrieval phase associates the RM measure

with a kNN algorithm in order to choose the set of

the most similar cases to the target case. In order to

select the most adaptable retrieved source case, we

have introduced a measure called “adaptation

measure" which will emphasize the functional mode

values of descriptors.

3.2 Adaptation Measure

The Adaptation Measure “AM” (2) takes into

account the source cases descriptors which are

different from case target and will be only linked to

the class and to the functional mode compared to the

solution descriptors. The adaptation measure is

conditioned by the functional mode value. Indeed, a

strong weight is affected to the dysfunctional mode

related to the failure.

∑

⋅

value

state

Class

).(

ϕϕϕ

λϕ

(2)

λi is the associated weight according to the

functional mode.

 If FM = normal Æ λi = 20

 If FM = normal/abnormal Æ λi = 21

 If FM = abnormal Æ λi = 22

A weight is associated to the functional mode

because this last is considered as being important in

the determination of the failing component. The

number of different descriptors is determined by the

denominator in the equation (2).

The retrieved source case having the greatest

adaptation measure value among the retrieval source

cases will be the candidate chosen for the adaptation

step.

4 ADAPTATION PHASE

We propose an adaptation algorithm based on the

context model, on the dependency relations between

various problem and solution descriptors and on the

descriptors hierarchical model.

If the solution class of the best chosen source case is

similar to the problem class then the algorithm uses

the hierarchical model. If the class is different, then

the algorithm uses the contextual model to localise a

set of potential failure component and then uses the

hierarchical model.

4.1 Dependency Relations (DR)

The influence of a descriptor problem “ds” on the

solution descriptors “Ds” is expressed by a

dependency relation. A dependency relation is a

triplet (dsi, Dsj, DRij). DRij gives us the type of

relationship between the problem and the solution to

a given case. Three relation types are defined: DRij

⊂

(No relation, Low, High).

 DRij = High: there is a high dependency

relation between dsi and Dsj descriptors.

Indeed, dsi descriptor is strongly relevant

compared to Dsj descriptor.

 DRij = Low: there is a low dependency

relation, i.e., the descriptors are connected

thanks to the context which will be

characterized by a contextual model.

 DRij = No relation: there is independence

between dsi and Dsj.

These dependency relations will be exploited in the

adaptation algorithm.

4.2 Adaptation Algorithm

The algorithm (algorithm 1) relies on the context

model, the descriptors hierarchical model and the

dependency relations. This algorithm adapts

descriptor by descriptor. The substitution's

adaptation, by generalization and by specialization

will be taken into account in the algorithm.

Three possible scenarios are treated differently by

the algorithm

 DR = high and same class between problem

and solution descriptors.

 DR = high and different class between problem

and solution descriptors

 DR = Low

This algorithm deals with the adaptation of one

descriptor at a time. It is conditioned by the solution

A problem descriptor ds

is strongly relevant compared

to a solution descriptor Ds

when the value of ds

descriptor is crucial in the determination of Ds

value. The

change of ds

value is directly reflected on Ds

value.

ADAPTATION BASED ON KNOWLEDGE MODELS FOR DIAGNOSTIC SYSTEMS USING CASE-BASE

REASONING

227

descriptor class found at retrieval step. After the

retrieval phase which makes it possible to select a

retrieved case (

ret

) thanks to both RM and

AM measures, the adaptation phase engages. The

initialization step creates a list of couples having a

relation either high or

low. According to the nature

of the relation, the treatment differs. Consequently

the second step will depend on the DR values and

the classes of the descriptors

Algorithm 1: Adaptation Algorithm.

1. If by browsing through the list, a value of “DR

= high” is found then the couple is selected

and class of “

ret

” and “

ret

”is compared. If

they have the same parent class, the influence

of this substitution will be considered in

“

ret

” to assign this new value to

ret

.,(on

the contrary, algorithm look at the context list

descriptors) and selects the “dti” descriptor

which belongs to the same parent class as

“

ret

”.We note that target descriptor “dt*”.

Then, the value of reminds will be determined

and which will be thereafter to be affected in

“Dtj”.

2. If in the list there is only the DR =low .the

algorithm selects the parent class of

ret

descriptor. Then, it identifies the dti descriptor

Retrieval descriptors problem.

Retrieval descriptors solution.

belonging to the same parent class as

ret

which will change status (dti Æ dt*). After

that, the relationship dt* will influence the

transformation of the

ret

solution which will

be affected thereafter to “Dtj”.

3. Finally, when all DR values are equal to “no

relation” then there is no adaptation.

5 VALIDATION & CONCLUSION

In this section we present two experiments (i) first

one concerns the need of adaptation phase in our

system, (ii) and second one studies the performance

of the adaptation algorithm. We used a leave-one-

out cross-validation method for the first two parts to

assess SISTRE's ability to accurately adapt retrieved

cases for a case base containing 125 cases

 The need of adaptation:

Accuracy rate of diagnosis system with and

without adaptation phases are compared. The

results show that the proposed method with

adaptation selects the cases which are the best

adaptable ones by obtaining 88% of accuracy.

If the adaptation algorithm is powerful one can

get a good performance concerning the CBR

system applied to a limited number of cases.

However, we find bad results without the

adaptation, only 58.1% accuracy rate with the

retrieval step. These results show that in our

system this adaptation phase is essential to have

a good result.

 Performance of adaptation algorithm: This

experiment is designed to study the accuracy of

the help diagnosis system; overall accuracy,

and more precisely the accuracy of only

retrieval cases.

Table 4: Results of the adaptation.

We note that the accuracy is 88% reflecting that

110 cases were adapted correctly to the set of 125

cases. This accuracy is computed using “Ds

” as the

component responsible for failure. The obtained

accuracy rate for which the adaptation measure is

well chosen and the adaptation algorithm is treated.

For this subset of cases, the accuracy rate is 91.66%.

KMIS 2011 - International Conference on Knowledge Management and Information Sharing

228

Within the study framework on technical

diagnostic and repair assistance system, an

adaptation-guided retrieval method has been

proposed. Our previous studies have enabled us to

formalize the case of a supervised industrial system

of pallets transfer (SISTRE). This formalization is

adapted to our method. In fact, we set up a

formalization of object of the cases, associated to the

descriptors hierarchical model. This model is

common to problem and solution descriptors of the

case-base cases and a model relating to the

application context. All steps depend on the cases

formalization and the associated knowledge models.

This modelling has influenced the proposed

similarity measure as well as the adaptation

measure. The latter is directly related to the

functional mode of the supervised components (an

attribute specific to the descriptor). The retrieval

phase is related to the adaptation phase using the

conjunction of similarity and adaptation measures.

This conjunction makes it possible to select among

the retrieved cases the most adaptable. The

adaptation phase will exploit the dependency

relations between the problem and the solution.

We are proved the feasibility of this diagnostic

help system. To build it in any type of industrial

equipment, two knowledge model need to be

elaborate.

To avoid the cost of the development of

knowledge models, we are currently working to use

these algorithms with models (functional events and

components models) developed in web-maintenance

platform. This model is defined in the domain

ontology of maintenance, in the context of

Semantic-maintenance and life cycle (SMAC)

Project.

ACKNOWLEDGEMENTS

This work was carried out and funded in the

framework of SMAC project (Semantic-

maintenance and life cycle), supported by European

program Interreg IV between France and

Switzerland

REFERENCES

Althoff K. D., Bartsch-Spörl B., 1996, Decision support

for case based application, Wirtschaftsinformatik,

ISSN 0937-6429, 1996, vol. 38, no1 pp. 8-16

Bridge, D., Ferguson, A., 2002, An expressive query

language for product recommender systems. Artificial

Intelligence Review 18(3-4), 269-307 (2002)

Chebel-Morello B., Haouchine M.-K., Zerhouni N., 2009.

A methodology to conceive a case based system of

industrial diagnosis. In World Congress of

Engineering Asset Management, WCEAM'09, Greece

Cordier A., 2008: Interactive and Opportunistic

Knowledge Acquisition in Case-Based Reasoning,

PhD thesis, Laboratoire d'InfoRmatique en Images et

Systèmes d'information, University of Lyon I,

November.

Haouchine, M. K., Chebel-Morello, B., Zerhouni, N.,

2008: Adaptation-Guided Retrieval for a Diagnostic

and Repair Help System Dedicated to a Pallets

Transfer. In 3rd European Workshop on Case-Based

Reasoning and Context-Awareness. 9th European

Conference on Case-Based Reasoning, ECCBR 2008,

Trier, Germany

Haouchine, M. K.: 2009 Rememoration guide par

l’adaptation et maintenance des systèmes de diagnostic

industriel par l’approche du raisonnement partir de

cas. PhD thesis, Automatic en Micro-Mecatrnoic

Department, Franche-Comt University

Kasif, S., Salzberg, S., Waltz, D., Rachlin, J., Aha, D.,

1995: Towards a Framework for Memory-Based R.

NECI Technical Report

Lieber, J., 2007: Application of the Revision Theory to

Adaptation in Case-Based Reasoning: the

Conservative Adaptation. In 7th International

Conference on Case-Based Reasoning - ICCBR'07,

4626, pp. 239-253.

Maintenance terminology. 2001 European standard, NF

EN 13306.

Rasovska, I., Chebel-Morello, B., Zerhouni, N., 2007: A

Case Elaboration M for a Diagnostic and Repair Help.

FLAIRS.

Smyth, B., Keane, M. T., 1998: Adaptation-guided

retrieval: Questioning the similarity assumption in

reasoning. Artificial Intelligence 102(2), 249-293.

Smyth, B., Keane, M. T 1995: Experiments on

Adaptation-Guided Retrieval in Case-Based Design.

In: Veloso, M., Aamodt, A.; (eds.): Proceedings of the

1st International Conference on Case-Based

Reasoning. LNAI, Vol. 1010, Springer, Berlin 313-

324.

Stanfill, C., Waltz, D., 1986: Towards memory-based

reasoning. Communications of the Association for

Computing Machinery, volume 29 pp. 1213-1228.

ADAPTATION BASED ON KNOWLEDGE MODELS FOR DIAGNOSTIC SYSTEMS USING CASE-BASE

REASONING

229