Copyright © 1996, AAAI

Development of Self-Maintenance Photocopiers

Y. Shimornura a ml K. Ogawa a,nd S . TanigawaMina Industrial Co., Ltd .

Tamatsukuri 1-2-28, Chuo-ku, Osaka, Japan .Email : {simoinura, ogawa, tanigawa}'u,mita.cojl>

Abstract

The traditional reliability design methods are im-perfect, because the designed systems aim at lessfaults but once a fault happens they may hard-fail. To solve this problem, we propose the con-cepts of self-maintenance machine (SMM) thatcan maintain its functions flexibly even thoughfaults happen . In order to achieve capabilities ofdiagnosing and repair planning, a model basedapproach that employs qualitative physics is pro-posed . Regarding repair executing capability,control type repair strategy is proposed . A pro-totype of the self-maintenance machine is devel-oped and it succeeded to maintain its functionsas far as its structure does not change . Thisprototype, however, revealed the following prob-lems that may arise when applied to a commercialproduct as embedded software . The problems in-clude ; (1)performance of the reasoning system,(2)the system size, (3)effects of environmentalchanges, and (4)roughness of qualitative repairoperations . In order to solve these problems, wepropose a new reasoning method based on vir-tual cases and fuzzy qualitative values . This is amethodology of knowledge compilation that givesbetter reasoning performance and can cleat withreal world applications like the self-maintenancemachine . By using this method, we finally de-veloped a commercial photocopier that has self-maintenability and is more robust against faults .The commercial version has been supplied world-wide as a product of Mita Industrial Co., Ltd .,since April, 1994 .

IntroductionWhile maintenance tasks are hardly automated andneed laborious work, it is becoming more crucial thatmachines do not break or, even if they do, can berepaired quickly because of customers' requirements,costs, and so on . The disasters of Three Mile Islandand many aircraft accidents reveal that traditional re-liability design method is imperfect, because these sys-tems are designed for less faults but nevertheless once

Y. Umeda a.nd T . TomiyarnaGraduate School of Engineering, the University of Tokyo

Hougo 7-3-1, Bunkyo-ku, Tokyo, Japan.Email: {umeda, tomiyama}(lzzz .pe.u-tokyo .ac jp

a fault happens they may hard-fail . The same prob-lem exists in the maintenance of photocopiers . To re-duce such a problem, we should develop soft-fail (asopposed to hard-fail) machines that can still maintaintheir functions even if faults happen .

Although some researchers proposed Al techniquesfor supporting fault diagnosis (e.g . (de IileerW'illiams 1987), (Hamilton 1988), (Milne Sr, Trave-Massuyes 1993)) or technical service (Bell el al. 1994),there was few research about automation of diagno-sis and repair as an embedded software . However,the photocopier industry has been pursuing such tech-niques, because photocopiers essentially need contin-uous maintenance and, therefore, maintenance costsincluding employing service engineers and supplyingparts are huge .

In this paper, we first introduce the concept of self-maintenance machines (SMM) (Umeda et al. 1994)and its reasoning techniques . Second, we discuss prob-lems in applying these techniques to a commercialmachine as an embedded software and propose threenew techniques called virtual cases, imitational faultmethod, and job elements as solutions to these prob-lems. Finally, we illustrate our experience of devel-oping a commercial self-maintenance photocopier todemonstrate the feasibility and industrial relevance ofthe self-maintenance machine as a design strategy .

Task DescriptionThe ultimate objective of this project is to automatethe maintenance of photocopiers . However, it is infea-sible or at least. very difficult to do so completely withthe current technology-, for example, consider automa-tion of parts change or cleaning . Instead, we here takean approach to maintain functionality of a machine fora while, which we call functional maintenance, and towait for a scheduled visit of service man. This ap-proach is useful enough for photocopiers, because theservice engineer will visit customers on a regular ba-sis and will maintain the machine in a steady state .For exan-tple, suppose a photocopier malfunctions on

Shimomura 225

Figure is Fundamental Architecture of SWIM

Friday evening . The service engineer will not (or can-not) come until Monday morning, but some impor-tant documents must be copied immediately . The self-maintenance photocopier will maintain its function tomake copies by some means, but it might work sloweror the copy quality might be slightly worse until theservice engineer shows up . The self-maintenance pho-tocopier, therefore, can reduce sudden, unscheduledvisits of service engineers, which will greatly reducethe maintenance costs, while minimizing damages tothe customers' satisfaction .One of the features of our approach is that such a

self-maintenance photocopier can be developed withexisting mechatronics technologies ; namely, a photo-copier is already equipped with sensors, actuators, anda CPL? even for a small to medium-sized analoguephotocopier . This suggests that we can build a self-maintenance photocopier just with an additional rea-soning system that maintains its functions .

The Framework of SMMWe define a self-maintenance machine (SMM) as a ma-chine that can maintain its functions for a while eventhough faults happen . SMM requires the following fivecapabilities .

Monitoring capability" Fault judging capability" Diagnosing capability" Repair planning capability" Repair executing capability

In order to achieve these capabilities, we developedan architecture for SMNI depicted in Figure 1 . In thisarchitecture, the monitoring capability and the repairexecuting capability are executed by sensors and ac-tuators, respectively . The fault judging capability, di-agnosing capability, and repair planning capability areobtained through the computer .

One of the key issues of SM 11 is how- to actuallyexecute repairs . Here, we will introduce two types ofrepairing strategies, i .e ., the control type and t hr: func-tional redundancy type ." Control Type

226 QR-96

Fault History-"---- Fault Data Base

Intelligent CAD

Some repairs of S\I:NI can be accomplished by ad-justing parameters without, changing or reorganiz-ing the structure . This type of repair is achievedby controlling actuators to recover part or whole ofthe requested functions, when, e.g ., deterioration ofapart happens . In some cases, part of the requestedfunctions might be lost or still degraded, but crucialfunctions will be maintained; this is called functionaltrade-off.

" Functional Redundancy TypeIf a machine has a mechanism that repairs itself byreconfiguring its structure, the range of repairingwould be enlarged . For achieving this mechanism,we have proposed a new repair strategy, called func-tional redundancy, which uses potential functions ofexisting parts in a slightly different way from theoriginal design (Urneda et al. 1994) . For example,in case of emergency that the engine is broken, acar with a manual transmission can run for a whilewith a starting inotor of which original function is"to start the engine ." This functional redundancyresults from the fact that the starting motor canperform its potential function "to generate drivingforce" instead of "to start the engine" by changingthe power flow of the car .In the following sections, we mainly discuss the con-

trol type SMM .

PrototypingWe developed a prototype that executes the controltype repair in order to examine the feasibility of theself-maintenance photocopier (Lmeda et al. 1994) .Figure 2 depicts the prototype in which a photocopier(the object machine) is connected to an MS-DOS'based personal computer . The personal computer col-lects sensory data from the photocopier and outputscontrol data to it . The personal computer is also con-nected to a Macintosh` computer . The personal com-puter converts quantitative sensory data to qualitativedata and sends them to the Macintosh . Figure :3 showsan example of a quantity space that traps quantitativevalues to qualitative values and landmarks . The Nlac-intosh reasons about repair plans from the qualitativesensory data with the method described in Section .This reasoning system is implernented on Sntalltalk-80'3 .Reasoning Technique for the Prototype . Thereasoning systein of the prototype is base on a model-based reasoning technique that employs qualitativephysics (Weld &, de Kleer 1989), because it can solve

' NIS-DOS

is

a

registered

trademark

of

MicrosoftCorporation .

Z ylacintosh is a registered trademark of Apple Com-puter . Inc .

'Srnalltalk-80 is a registered tradernark of ParcPlaceSystems . inc .

Figure 2 : The Prototype of a Self-Maintenance Photo-copier

DrumVs

0 Normal(0,Normal) , (Normal,inf+)

0 1 .61-1 .77 (V)Figure 3 : An Example of Quantity Space

the following problems that heuristics basedsystems encounter .

Reusability of knowledge

Problems of knowledge acquisition

expert

Because knowledge dependent on the object machineis localized and isolated in the model, the systemcan deal with different object machines by changingobject models .

Robustness of the reasoning systemsThe system can deal with unknown cases thatare not described explicitly in the heuristics basedknowledge by applying generic physical principles tothe model.

Since models can be prepared from design informa-tion of the object machine and the knowledge aboutphysical principles can be acquired systematically, itis easier to acquire the model based knowledge thanthe heuristics based knowledge .

Knowledge Representation . The knowledge forthe reasoning system consists of a model of the objectmachine (which we call a parameter model) and knowl-edge about physical principles that describes fault phe-nomena that might occur on machines in general .Figure 5 depicts an example of the parameter model

of the prototype (Figure 4) that describes physicalcharacteristics of the machine with parameters andtheir mathematical relations . In this model, functionparameters are used for treasuring functions . For in-stance, Os denotes the density of the output image onthe paper . Control parameters signify parameters con-trollable through actuators . An example of this is IIICthat can be controlled by the halogen lamp controller .

Ero0er HIC 'Phc~it~n~eter.XHalogen Lamp

.Main Charger

tvtaif,. Charge Controll erVnC

CL117,p--1

Figure 4 : The Structure of the Photocopier

Original Paper

Exposure Unit

P ParameterP !'unction Parameter

control Parameter©A"nsing Parameter

ACB A=B'( A<C A=B-C

A<c A=BBC

A `C A=B/C

Figure 5 : Parameter Model of the Prototype

Sensing parameters are monitored by sensors . For ex-ample, X should be monitored by the photometer.

In this prototype, we have selected two parametersas the function parameters ; namely, Os, the density ofthe output image on the paper and Sp at the separationunit indicating whether or not the output paper comesout from the machine . Since faults might change thestructure and/or attributes of the object machine, wedescribe fault phenomena based on the idea of processof Qualitative Process Theory (Forbus 19134) . In ourapproach, every phenomenon is associated with condi-tions under which this phenomenon occurs and effectsit causes (see Figure 6 explained below) .

Figure 6 :nothena

D

< E

Phenomenon Plcondition : value(F) > Limitleffect: increase(C)

Phenomenon P2condition : value(F) < Limit2effect: break(B = D - E)

Example of the Parameter Model and Phe-

Shimomura 227

Algorithm of the R,easonitug System . The rea-soning system of the prototype executes fault judg-ntent, fault diagnosis and repair planning with the al-gorithm explained below by using an example param-eter model depicted in Figure 6 .1 . Fault JudgmentFirst, the system identifies fault symptoms from thesensory data .

. Fault Diagnosis1 . Search for Fault Cause CandidatesThe qualitative reasoning system searches for allphenomena (called fault causes) of which influencesare related to the symptom directly or indirectly . InFigure 6 in which functional change is the increaseof A, phenomena P1 and P2 are selected .

2 . Fault SimulationNext, the reasoning system carries out qualitativesimulation about the behavior of the object machinefor each fault cause candidate based on the QSI.M al-gorithm (Kuipers 1986) . This simulation determinesa faulty state of the object machine (fault Triodel) foreach fault cause . The fault models provide the repairplanning step with information about the structureand the state of the faulty object machine.

3 . Identify Fault CandidateThe reasoning system selects the fault model whichmatches the most with the sensory data .

. Repair PlanningThe system reasons out a repair plan from the diag-nosed fault model . Here, based on the ideas of func-tional maintenance, we aitn at recovering the lost func-tions (i .e . fault symptoms) rather than fixing the faultsthemselves (i .e . structural or attributive changes) thatare objectives of the traditional repairs . This proce-dure consists of the following three steps :1 . Generate Repair Candidates

After the user selects a fault symptom to be recov-ered, the system derives desired changes of the con-trol parameters from the faulty parameter model. InFigure 6, repair objective is to decrease the value offunction parameter A and the control parameters areC and E . WVe can reason out that decrease of C' andincrease of F, are the candidates of repair operation .

2 . Repair SimulationThe reasoning system qualitatively simulates the be-havior of the object machine when each reasonedrepair operation is executed . As a result, the rea-soning system examines whether or not a selectedrepair method can achieve the repair objective andwhether or not a selected repair method will causeside effects to the object machine .

3 . Select Repair OperationThe reasoning system selects a repair operation thatcan repair or improve the objective function and hasthe least number of side effects .

228 QR-96

Examples. Vigure i illustrates an example opera-tion of the prototype . lit this example, die_ symptom"Paper Os is high" and the fault cause `'deteriorationof the halogen lamp" are identified correctly by the sys-tetn . Since the reasoning system outputs only qualita-tive repair operations, the repair executing algorithmincludes a heuristic strategy that changes parametervalues in a stepwise manner . lit this repair procedure,the system first tried to increase the voltage ofthe lampcontroller . However, the controller reaches its controllimit without. recovering the function because of de-terioration of the lamp . Then, the system executedanother repair operation in which the main chargercontroller decreased the retain charger voltage and itsucceeded in recovering the function .

This example demonstrates both the feasibility andthe flexibility of SNIRI . The latter is achieved, becauseit can generate repair plans adaptively according to thestate of the object machine based on the model basedapproach .

Problem DescriptionBased on the techniques described in above section,we have developed a commercial product of self-maintenance photocopier . However, some problemsturned out to apply the techniques developed for theprototype system to the commercial machine . In thissection, we analyze these problems and propose tech-niques that solve the problems.

Since the reasoning system of the prototype reasonsout, all possible faults based on physical principles andexecutes fault simulation for each diagnosis task, thesystem has the following features :1 . It can diagnose unexpected faults of the machine in-

cluding causally related multiple faults (but limitedto single cause) .

2 . It can diagnose faults that change the physical struc-ture of the machine .

However, to use the system as embedded software ofthe commercial products, the following problems canbe identified :1 . The size of the system and the reasoning speed are

unaccept able .Since the prototype system is implemented onSmalltalk-80, the size of the system is large as em-bedded software for commercial products and thereasoning speed is significantly slow .

2 . Changes of the environment lead to mistakes in di-agnosis .In the prototype system, a qualitative space map-ping is used for converting quantitative sensory datainto qualitative values . However in the real world, itis not easy to obtain the trapping statically becauseof imprecision of the sensors and changes of the en-vironntent around the machine . lucorrect, mappingresults in mistakes lit the fault diagnosis .

I suggest RM (MainCharger VnCLow)7:20:18 pmI suggest RM (MainCharger VnCLow)7:20:23 pm

HalogenLamp HIC UpHalogenLamp HIC UpHalogenLamp HIC Up is LimitedNowMainCharger VnC DownMainCharger VnC DownMainCharger VnC DownMainCharger VnC Down

3 . Roughness of the qualitative repair operation .Since the prototype system is based on the qual-itative reasoning, the system can deal with faultsflexibly. On the other hand, the qualitative repre-sentation of states of the machine is too rough tojudge whether an repair operation is correct or not .For solving these problems, we have developed the

following new techniques that will be discussed below .Virtual case based reasoning for problem 1 .

" The imitation al fault method for problem 2 ." Division of the quantity space and the job elements

for problem 3 .

Application Description

Virtual CaseTo solve the problems about. the reasoning speed andsystem size, we adopt the single fault assumption, andcompile all possible states of the machine includingfaulty states as compiled knowledge . For this purpose,we have developed a tool called a virtual case com-piler. A virtual case is a set of qualitative values ofthe parameters of the parameter model and representsa state of the object machine . The virtual case com-piler generates all possible virtual cases through qual-itative simulation of the prototype reasoning systemover all faults and all operations of actuators . One ofthe advantages of this method is that the compiler usesthe qualitative simulator without any modifications . Italso means that the qualitative simulator can be sepa-rated from the diagnostic engine . As a result, we canreduce the size of embedded software and the reason-ing speed increases, because this makes the reasoningalgorithm much simpler .

C"3r jrrt-t-."

TransUnlt VtC

HalogenLamp HIC lLirntl

MainCharger VnC

TransUnit Ds

Drum Vs

Drum X

Figure 7 : Screen flardcopr of the Reasoning System

Paper Os

Separation Sp

0

N Limit0s

0

N LimitSp

Table 1 : The Virtual Cases

Table 1 shows examples of the virtual cases that arecompiled from the parameter model shown in Figure5 . In Table 1, a virtual case "a"' that can be identifiedby the sensory values " Ds=+", "Vs=+",and "X=--"corresponds to a set of qualitative states that includesfaultv states .

Imitational Fault Method

N: normal value+ : luger ntun normal.:smaller than normal

A qualitative space mapping is needed for convertingquantitative sensory data into qualitative values . how-ever, due to errors and instability of the sensors andchanges of the environment around the machine, it. i sdifficult to obtain this mapping of which accuracy is fa-tal to the correctness of the fault diagnosis . In order tosolve this problem, we have developed two techniques ;namely, the fu:~zification of qualitative value ( Umeda,Tomiyarrra, k Yoshikawa 1992) and the invitationalfault method. The fuzzification of qualitative value is amethod to relate numerical sensory data to qualitativevalues by means of fuzzy theory (Zadeh 1965) .The imitational fault method calibrates the mapping

between sensory values and fuzzy qualitative values dy-

Shimomura 229

cases Sensor Parameters Qualitative States Fault SymptotttDs Vs x r

® I

b + + Nt r

5 1.0Parameter Ds

Parameter Vs

Parameter X1 .0

Figure 8 : Fuzzy Qualitative Space Map

Division of the Quantity Space

230 QR-96

ti 0.0 1

1/

`

0 of

'~

N11:

0.0 12.2 3.1

1 .7 40

2 .6 3.2Normal

Faulty Normal

Faulty Normal

Faulty

narnically by checking u1) the fuzzy Inenlbership func-tions through imitationally caused faults every tilllea fault repair operation is executed . 'Tile innitationalfault method is executed in tile following steps :1 . After a repair operation, tile system causes faults in-

tentionally by controlling ail actuator until a synlp-tom occurs . The actuator that Inay change valuesof all the sensors is appropriate for this . Currently,the designer selects the actuator to cause inutationalfaults .

2 . The system revises the fuzzy qualitative space snap-ping for all sensory parameters of each sylnptoln todistinguish faulty and normal sensory data.

Figure 8 shows examples of the fuzzy qualitativespace mappings generated by this method when asymptom "the increase of parameter Os" is irnitation-ally caused by the decrease of the halogen lamp con-troller that is an actuator for Hl (see Figure 5) . Forexample, the mapping for tile paratrleter Ds in Figure8 signifies that the symptom definitely appears anddisappears when its sensory value reads snore than 3 .1and less than 2.2, respectively, but it is not obvious toconclude that the symptom appears when the value isbetween 2 .2 and 3 .1 .

As described later, virtual cases are ordered accord-ing to their matching ratios to the machine's state asthe result of diagnosis . This order of cases identifiestile most probable candidate of fault and divides thequantityspace of the object machine into some areas inthe following way. Here, tile quantity space is definedas a vector space of all sensory parameters of whichdirnension is the rlnrnber of the sensory parameters inthe rrtachine . The order of cases can be consideredequivalent to the order of distances to the cases of thecurrent machine's state .

Figure 9 depicts an example of this divided spacethat consists of two sensory para.nneters P 1 and P2 and,therefore, that has four cases, viz ., a(N, +), b(+, +),c(N, N), and d(-1-, N), if only the positive regions ofthe parameters are considered . These four cases canbe ordered according to their matching ratios . For III-

stance, with a certain set of the sensor values, case 1> isthe most probable and case d is tile second most . 'litsorder, bd, corresponds to an area which is a division ofthis quantity space . Ne Inay consider a certain number

P2

P2

a,b,c,d : casesorder

Job Elements

N

Ndiagnosis range : 4

diagnosis range : 1Figure 9 : Vxanlple of Qualitative Space Division

of cases in this ordering . This number is called rliag-nosis range ail(], thus, determines the nunnber of tiledivision of tile space . For example, Figures 9 (A) and(B) show the spaces of which diagnosis ranges are fourand one, respectively . The number of the divided areasin the quantity space affects repair accuracy, we shoulddecide this nulllber experimentally with respect to thediagnosis range . The diagnosis range should be largeenough to execute the repair operations precisely, butit should not be too large, because the divided quan-tity space becomes too precise to be detected correctlyby the sensors . Currently, the diagnosis range of thecommercial product is experimentally determined .The Inaxinwm number of divided areas of the quan-

tity space, N)nax, is given by the following equation,where n denotes the number of sensors .

N77Lax=2" x, P'., x 11((ti-V)+Er~,k)

(a>3)

(1)r-3 k-3

'Table 2 show's the result of area division of a quan-tity space consisting of three sensors, and the spaceis divided into 96 areas using the rnaxinluln diagnosisrange of eight . 111 'Table 2, symbols such as a and bdenote virtual cases a.nd each area corresponds to anorder of these virtual cases according to their Ina.tchingratios .

Each repair operation through controlling all actua-tor means a transition of areas ill tile divided quan-tity space . However, there are possible ail(] impossibletransitions from one area to neighboring areas becauseof physical characteristics of the object nnachine . 11-though these transition rules for the cotnluercial prod-uct are experimentally obtained, future work shouldaim at deriving these rules automatically front tile pa-rameter model . We call this rule a. job element thatrepresents the relation between all operation of all ac-tuator and area transition . 'Table 3 show's examples ofjob elements . The job elements are described for eacharea ill tile divided quantity space .

cwx.~"xu:nw:aiw.arxa."tu_al case comppilatio

E~~

virtual casesmaximum divided quantity spacejob element

~i sethn g dia gnosis rang ,

deformation of the quantity space~and job element

"

Table 2 : Area Dividing of a Quantity Space of Three Sensors

machine ~tio

"no symptom occurs

Reasoning Algorithm of the CommercialPhotocopier

Figure 10 : The Reasoning Algorithm of the Commer-cial Photocopier

In this section, we explain the reasoning algorithm ofthe commercial photocopier . Figure 10 represents thereasoning algorithm .Models and Knowledge .

During the developmentprocess, first the following knowledge for each kind ofmachine must be prepared.

1 . The parameter model set. up manually and virtualcases compiled from the parameter model with thevirtual case compiler .

2 . The lnaximurnly divided quantity space and job el-ements .

Then, the following data must be determined .

3 . The diagnosis range .

Table 3 : The Job Elements

4 . The divided quantity space and the job elementsmodified with respect to the diagnosis range .

5 . The fuzzy qualitative space map for each sensoryparameter obtained through the invitational faultmethod .

Diagnosis . 'The commercial version ofself-maintenance photocopier diagnoses itself based onvirtual cases by searching for a virtual case that hasthe highest matching ratio with the sensory data of theobject machine . The matching ratio is calculated bycomparing the fuzzy qualitative values of sensors withthe qualitative values in each virtual case as given inequations (2) and (3) .

C = I -

C(p1) 2 + C(p2) 2 +

" " + C(pra)2/n

G(pi)Gm(gi)-Gs(gi) (i=1,2,n)

whereC: the matching ratiois the number of a sensorpi : a sensory parameterqi : a qualitative valueGin : a fuzzy value of qi in a virtual caseGs: a sensed fuzzy value of qi

Table 4 lists ordered virtual cases as an example ofthe fault diagnosis for the following fuzzy qualitativedata when the symptom "increase of parameter Os"occurs . In this data, the value of parameter Ds, forinstance, indicates that the fuzzy grade of `+' (largerthan normal) is 0 .6 and the fuzzy grade of `N' (normal)is 0.4 (see Figure 8) .

Area Name Order of Cases Area Name Order of CasesA-1 abdcefhg A-49 eafblldgcA-2 abdceflig A-50 eaflibdgcu u II u u

A-47 dhcagebf A-95 llgedfcabA-48 dhcgaebf L A-9(i llgefdcab

Job Element Area : A-34Operation Next Areas Operation Next Areas

H1 up A-36 Ill down A-33, A-28, A-27V11 up A-28, A-33, A-27 Vn down A-36Vb up A-36 Vb down A-33

Ds(+ : 0 .6, N : 0 .4) (4)V s(+ : 0 .6, N :0 .4) (5)A(- : 0 .8, N : 0 .2) (6 )

Shimomura 231

symptom occurs

collectutg sensory data

mapping on fuzzyqualitative space map

calculating matching ratiowZ

nupi~t 1t1011

of tha actuator

creating fuzzyqualitative space map

fuzzy qualitative iial,~'.*space rr,7

divided quantity space of3 sensors

symptom ; '?occurs

Figure 11 : The Algorithm of Fault Diagnosis

Figure 12 : The Concept of Repair with Job Elements

Fault diagnosis, thus, identifies an area in the quan-tity space of the object machine . Figure 11 shows thealgorithm of the fault diagnosis based on the combina-tion of virtual cases and the irnitational fault method .

Repair . After fault diagnosis, the system executesrepair planning and a repair operation with the jobelements as follows :

1 . The system searches for a path from the current areato the goal area by combining job elements . It issimilar to techniques which have been developed insoftware analysis in respect of using for all executionpaths that some code might take (Srinlani &; Malatya1992) .

2 . It manipulates the actuators in order to changethe state of the object machine along the path,continuously comparing the current state with thesupposed-to-be states described in the `Next Areas'of the current job element

3 . If the current state of the machine is different fromthe supposed-to-be state, the system stops the op-eration immediately and restores the state to theinitial one . In such a case, the diagnosis range ofthe machine might have changed by, for example,

232 QR-96

vJatdt diagnosis

searching a path

apply the operationdescribed in the ith clcuic :u

Figure 13 : The Algorithm of the Fault Repairing

Fu,

Virtual Case Base

Job Element BaseCreator ::. . .

.;:

-}-',Inference Engine~~_

deterioration of sensors .

Object Machine

Figure 14 : The Structure of the Commercial System

Figures 12 and 13 show the concept of repair withjob elements and the repairing algorithm, respectively .

Executing the Invitational Fault Method .

Afterexecuting repair, the system modifies the fuzzy qualita-tive space trap of each sensory parameter by executingthe invitational fault method .

Application Use and Payoff

fault L)Lcurs*

We have developed a commercial photocopiers basedon the techniques described so far . Figure 14 and 15depicts its architecture and exterior . Figure 16 is ascreen hardcopy of the virtual case compiler used forthe development . Figure 17 compares a faulty outputimage and a repaired one . The inference system of thecommercial product was developed on the C and as-sembler languages, and takes at most 10 minutes for adiagnosis and a repair including revision of the fuzzyqualitative space trapping . The system size is about280 Kbytes including 16 virtual cases and its diagnosisrange is 2 . These results are acceptable as an embed-ded system . Figure 18 shows typical results of someexperimental diagnoses that demonstrates the diagno-sis against fuzziness of the sensory information .

Figure 15 : The Exterior of the Commercial Machine

Table 4 : Example of Fault Diagnosis

Figure 16 : An Screen Hardcopy of the Virtual CaseCompiler

While the knowledge compilation techniques basedon the virtual case greatly improves the reasoningspeed and the system size, the system has less flexi-bility against faults with structural changes comparedwith simulation based systems .The commercial version photocopier has been sup-

plied to the market world wide since April, 1994 . Al-though it is too early to evaluate payoff of the prod-uct, we predict the decreases of down-time and mainte-nance costs paying off the cost of the development . Atleast, service engineers, particularly in the US, preferthis self-maintenance photocopier to traditional ones,because it can reduce their work while maintain thelevel of service .

Maintenance

Conclusion

Faulty Image

Repaired ImageFigure 17 : The Result of Fault Repairing

Figure 18 : The Result of the Fault Diagnosis

Application Development andDeployment

The commercial system was developed by 10 design-ers over two years following the development of theprototype system in the "self-maintenance machine"project at the University of Tokyo in 1992 . The designteam includes mechanical, electrical, and software en-gineers . The designers spent the most of the time ofdevelopment on experiments for the sensors and soft-ware development and they also have started to deploythese techniques to other products .

We have not yet found out any necessity of mainte-nance of the knowledge or the reasoning algorithm ofthe system .

In this paper, we have proposed a design strategy forthe self-maintenance machine and its prototype . We

Shimomura 233

Virtual Case Matching RatioOrder Fault Ds Vs X CDs ~~~~'' C

1 IIT +1 .0 +1 .0 -1 .0 0.4 OA 0.2 0.652 Vn control +1 .0 +1 .0 N 1 .0 0 .4 0 .4 0.8 0 .433 Vb control +1 .0 N1 .0 N1 .0 0 .4 0 .6 0.8 0 .384 Vt control N 1 .0 N 1.0 N 1 .0 0.6 0 .6 0.8 0 .33u u u u u

QuantitativeSensory Data

NormalE- a-Ds Vs X Ds Vs X Ds Vs X

Qualitative Fuzzy (+0.1,+0.9,-1 .0) (+1 .0,+0.9,-0.1) (+0.8,+0.6,-0.2)

Value Non Fuzzy (N,N,) (+,N,N) (N,N,N)

Diagnosis Fuzzy ° halogen lamp10main charge main charge

Results Non Fuzzy" transfer developing bias transfer

Pressure Control Motors

Paper Speed Sensor

" Making repair planning inore flexible by allowingusers to select. repair goals based on the functionaltrade-of(Umeda et al. 1994) .

" Developing a machine that can treat faults withstructural changes based on the concepts of func-tional redundancy (Vineda et al . 1994).

234 QR-96

Figure 19 : The Paper Handling Unit

then proposed a method to reduce the systern size andto improve the performance of reasoning, combinedwith a method to treat sensory data with fuzzy quali-tative reasoning and imitational fault method . We alsoproposed a method to execute repair operations moreaccurately. The development of a commercial versionof the self-maintenance photocopier demonstrated itsfeasibility .

Future issues include :

We have also started to develop new system in whichwe aim to repair faults about paper handling in pho-tocopiers based on sane techniques described in thispaper . In this system, we will treat with faults aboutpaper handling, e .g ., paper jamming, askew feeding,etc, by controlling some mechanical actuators built in

Figure 20 : The Exterior of New Experimental Machine

independent intelligent control units (see Figure 19-20) .

ReferencesBell, D . ; Bobrow, D. ; Falkenhainer, B . ; Frocnherz,M . ; Saraswat, V . ; and Shirley, M. 1994 . Rapper :The copier modeling project . In Nishida, `f ., ed ., Pro-ceedings of 81h. Inter-national Vf"orkshop on QualitativeReasoning about Physical Systems, 1-12 .de Kleer, .1 ., and Williams, B . C . 1987 . Diagnosingmultiple faults . Artificial Intelligence 32(l):97--130 .Forbus, K. 1984 . Qualitative process theory. ArtificialIntelligence 24(3) :85-168 .Hamilton, T. P . 1988 . Helix : A helicopter diagnosticsystern based on qualitative physics . Artificial Intel-ligence in Engineering 3(3):141-150 .Kuipers, B . 1986 . Qualitative simulation . ArtificialIntelligence 29(3) :289--338 .Milne, R. ., and Trave-Massuyes, L . 1993 . Real-timemodel based diagnosis of gas turbines . In Proceedingsof IJc41-9.g Il'orkshop - Engineering Problems forQualitative Reasoning -, 49--57 .Srimani, P ., and Malaiya, Y. 1992 . Steps to practicalreliability measurement . IEEE Sopu;are 9(4)_10 -- 12 .Umeda, Y . ; Shirnomura, Y. ; Tomiyama, T. ; andYoshikawa., H . 1994 . Using functional maintenanceto improve fault tolerance . IEEE Expert 2(3):25-31 .Umeda, Y. ; Tomiyama, T. ; and Yoshikawa, H . 1992 .A design methodology for a self-maintenance ma-chine . In First International Conference on Intelli-gent Systems Engineering, number 360 in ConferencePublication, 179984 . The Institution of ElectricalEngineers (IEE) .Weld, D., acid de Kleer, J ., eds . 1989 . Readings inQualitatire Reasoning about Physical Systems. SanMateo, (' :1 : :Morgan-Kaufniann .Zadeh, L . A . 1965 . Fuzzy sets_ Information (.'ontrol8:338 353 .