From: AAAI Technical Report WS-96-05. Compilation copyright © 1996, AAAI (www.aaai.org). All rights reserved. Using Description Logics for Consistency-based Diagnosis Gerd Kamp Universit~t Hamburg Vogt-KSlln-Str.30 22527 Hamburg kamp@informatik.uni-hamburg.de Abstract In this paper weargue that description logics (DL)are well suited for consistency-based diagnosis. Theyprovide a rich and expressive representation language,that is able to fulfill the representation needs. The TBoxcontains the descriptions of the different componenttypes and the modelsof correct and faulty behavior. Concrete devices are modeledas a set of assertions within the ABox.Furthermorethe powerful Tand ABoxinference services such as consistency, object classification 3 and classification can be used for simulation and diagnosis. Moreover,for certain languages (e.g..ALCY(1))) sound and complete algorithms for these inferences could be given(see e.g. [3]). Using quantitative modelsof simple mechanismsas an example domain, we show howthe basic principles of consistency-based diagnosis can be implementedusing description logics with expressive concrete domains. In addition, description logics provide the inference services necessary to organize and validate modellibraries, an aspect that is neglected in consistency-baseddiagnosis until now. Description Logics for Diagnosis 1 Holger Wache Universit~t Bremen P.O. Box 330 440 28334 Bremen wache@informatik.uni-bremen.de Consistency-baseddiagnosis1 [7; 10] is a methodfor the diagnosis of technical devices basedon descriptions of the correct and faulty behavior of components.Therefore, for representing, simulating and diagnosingthe simple bike drive train in Fig. 1 at least the following knowledgemust be represented within such a system: 1. The different types of components,e.g. wheels, gearwheels, chains, along with their attributes like force F, 2radius r and torque M. 2. The structure of assemblies, e.g the kinematic structure of the mechanismsusing kinematic pairs. 3. Physical laws like M= r x F and constraints imposed on the attributes like F > 0 and r > 0. 4. The normaland faulty behaviors (often called modelsin consistency-based diagnosis) of componentsand assemblies, e.g. the propagationof torques from one wheelto another (M1= M2)in a rotational pair. Basedon this representation diagnosis is mainlythe task of identifying the modelsof behavior that are consistent with a set of observed parametervalues. This is most often done by simulating hypothetized behaviors and comparingthe results with the observedvalues. tAlso called model-based diagnosis 2In this paperweare modeling these attributes using scalar numericvalues resulting in quantitativemodelsof behavior.Alternatively a qualitative modeling usingsymbolicvaluesas it is usedin qualitativephysicsis possible. --- \\ ,--- Figure 1: A simple bike drive train Despite these facts there are no consistency-baseddiagnostic systemsthat use description logics, at least to our knowledge. The mainreason for this is the lack of expressive concrete domainsin current DLsystems. Since we have to deal with numericparameter values, we need a description logic that is able to handle concrete domains.In order to describe physical laws and behavioral modelswe need at least a concrete domainthat is able to reason with linear systems of inequalities between polynomials4, in the TBoxas well as in the ABox. 2 Concrete Domains- State of the Art Baader and Hanschke[1; 3] have shownthat it is theoretically possible to integrate systems of nonlinear polynomi3Alsoknown as realization. 4M= r x F is a quadratic polynomialbut becomeslinear when one of the variablesis known. - 136- ais into the 1"- and ABoxreasoningservices. Theydevelopeda schemefor integrating concretedomainsinto description languagesandshowed that the decidabilityof the resulting descriptionlanguage,AgeJr(2)) is preservedfor concrete domains that fulfill certain restrictions (so called admissible concrete domains). Byusingthe decidabilityof the theoryof the elementary algebraoverthe reals [11], it caneasily be shown that systems of inequationsbetweenarbitrary polynomialsand systemsof inequalities betweenlinear polynomialsare admissibleconcrete domains[6]. Butwhatis the status of concretedomains in implementedsystems? KRSS providesa concretepart of the domain,specified as the rationalsandstrings oversomealphabetof size at least 2. Werestrict our analysis to the numericdomain. Terms withinthe concretedomain.Instead, they all rely on the respectiveoperatorsof the abstract domain,resulting in obvious performancehits. Since KRSS and CLASSIC only provideandin the abstract domain,it is impossibleto definethe missingcomparisonoperator < in these systems. Theonly wayto circumventthe abovelimitations in some wayis via constructs(satisfies, test-h/c) that allowto include arbitrary Lisp functions into conceptdescriptions5. Obviously this constructs cannotbe used at all within the TBox reasoningservices. Thus, the handling of concrete domain expressionsas a blackbox mayresult in incompleteor wrong classifications. In Section7 wewill see that classification basedon concretedomainexpressionsis indeedan interesting TBoxinference. W.r.t. the ABoxreasoning they have the majordisadvantageof being lisp functionsand not constraints. All this showsthat current implementations of description logics do not provide the mechanisms weneed in order to fulfill the representationneedsdescribedin Section1. aft I:++;/ (o. ~) rmulae (a < #) e o e ® ® ® (~_<#) (,~ >#) Operator -9~o (~v ¢) ^~ ~e~°n~er I Van®) O ® ® O @ @ O ® ¯ explicitly ontained C) @ via the abstractdomain implicitlyexpressible ® ® Table1: Expressiveness:LinearSentences Table1 summarizes the systems’capabilities to handlesystemsof (in)equations betweenpolynomials. Thefirst andforemostobservationis the completelack of anyfunction symbols.Therefore,current terminologicalsystemscan at best handlearbitrary comparisonsbetweenattributes or attributes and numbers.Only LOOM defines a relation +, but it cannotbe usedwithinconceptdefinitions. Thereforeit is impossibleto define eventhe simplestequations betweenunivariatelinear polynomials like z = y + 1 or d = 2. r. Notto mentionlinear multivariatepolynomials like u = 2. (l + w) or nonlinear polynomialslike M= F. KRSS and CLASSIC show another notable limitation. Unlike the other systemsthey provideonly > and< as comparison operators,andonetermof the equationis restricted to be a constant. This makesit evenimpossibleto define a square as a speeiai rectanglewith I = w.Onehas to use the abstract equality predicate equalin order to describethis. Furthermore,no systemimplements logical operatorsor quantifiers 3 Expressive concrete domains Thereare two possibilities for implementing the concrete domainextension of a terminological system: providing a 6genericinterface to algorithmsrealizing a concretedomain or hardwiringthe integration of specific algorithms7. A genericinterface makesit easy to providemultipleconcrete domains,resulting in a modularsystemarchitecture. Thus onecan select the properalgorithmsneededto fulfill the requirementsof the specific application, andcouplethemwith the descriptionlogic throughthe genericconcretedomaininterface. For example,a large numberof algorithmsfor solvingsystemsof algebraic (in)equationsare andhavebeendeveloped. Furthermore,a couple of this algorithmshave beenincorporated into CLP(R)-systems (e.g. [5]). Using a CLP(R)systemas the concrete domainsolver, weparticipate from the progressmadein this area throughsimplyexchanging the CLP(7~)-system at the interface. 3.1 Concrete predicates Thus, we implementedCTL[6], a system based on TAXON andthe description logic A£g~(7))whichrealizes a generic concrete domaininterface. CTLimplementsthe KRSS syntax, whichwe had to be extended in the following way (for details see [6]): Predicateterms P describe concrete predicates. Theyare either a predicate namePNor a list ((namer)(x~ x,~)(exprv)) consisting of a dom ain identifier (namer)and a list of variables zl... zn. The expression (exprv)actually definesthe concretepredicate. (define5 KRSS, LOOM,and CLASSICprovide such constructs. ~Thisis mainlythe taskof decidingff a finite conjunction of concretedomain expressions is satisfiable. 7The built-in concrete domains in CLASSIC, LOOMand KRIS are more or less examples for this hardwiredintegration. - 137- constraint PNP) assigns a namePNto a predicate term P. Further (constrain RI ... R,P) is an additional concept term operator with the followingsemantics:(constrain nt ... R,,P) z -- {a 13bl,...,bn((a, bl) E n~ A... (a, bn) E RznA (bl,..., bn) e pZ)}. Since CTLis basedon the plain tableaux calculus using the rules describedin [3] only the respectiveconcretedomain roles hadbe implemented appropriatelyin order to realize the generic concretedomaininterface. Since/~is a list of the admissiblevalues for eachparameter, the computed restrictions can thenbe deliveredtogether with the generatedmodel. 4 Representation As wehaveseen in Section I, wemustbe able to describe the different componenttypes of a mechanism as well the kinematicstructure. In kinematicsthis is normallydonewith links andkinematicpairs [9]. In order to describethe drive 3.2 Generating models and computing parameter train of Fig.l, it is sufficient to definerotational links and restrictions tensionlinks as specializationsof generallinks. ThetermiDescription logic systemsbased on a tableaux methodexnologyin Fig.2a defines a link as somethingthat carries a plicitly generatea modelin the courseof the consistency test. force: llnk.foree. Rotationallinks (rotational-link)are links Normally only the result of the consistencytest is presented that in additionhaveattributes for a radius(rot.radius) and to the user andthe constructedmodelis moreor less thrown torque(rot.torque).link.forceandrot.torqueare not negative, away.In technical domainsand especially in our applicarot.radiusis strictly positive, andthe torqueis the productof tion (but also e.g. in configuration)the calculatedmodel radius andappliedforce. Finally wedefine a number of addithe mainobject of interest. Thereforewerealized a function tional links suchas wheel,vhalnetc. withoutfurther specify(show-model) within CTLthat returns the modelMfor a con- ing them. sistent ABox. In order to describe the structure of a mechanism, kineThis schemacan be extendedto concretedomainsby using matic pairs are used. Aklnematle-palrdescribesthe connecquantifier eliminationtechniquesfromcomputeralgebras. A rionbetween twolinks palr.llnkl andpalr.llnk2. Depending on subtaskof the consistencytest for an ABox,is testing whether the degreesof freedomof the relative motionof the links and a conjunction of concretepredicatesis satisfiable, i.e. check- the type of connectiondifferent types of pairs can be idening whetherthe sentenceS = 3xl ...qxk(pl A ... A Pn) is tified. In the terminologyshownin Fig.2bwerestrict us to true. Tarskisbasic idea for decidingthe theory of the elethe descriptionof pairs between tworotational-links(rot-pair) mentaryalgebra wasto transforman arbitrary sentenceinto andpairs between onerotatlonal-linkandonetenslon-llnk(rotan equivalentquantifier free oneS [11]. Thevalidity of this tension-pair). Notethat weare able to describerot-tensionsentenceis then easy to check. Overthe years quantifier pair via an or construct. In addition to the component types eliminationtechniqueshavebeenvasty improved,moreeffiandthe structure of the device,descriptionsof the correct as cient generalalgorithmshavebeendevised(e.g. Cylindrical wellas the differentfaulty behaviormodelsare neededfor the Algebraic Decomposition(CAD)[2]) and further improved consistency-based diagnosisas well as the simulationof the (e.g. PCAD [4]) and specialized algorithmsfor subsets device. Theterminologyin Fig.2c describesthe correct bethe theory have been found and realized. MostCLP(7~)- havior of rotational-palmandrot-tension-palm.Whereasthe systemsmakeuse of the Fourier-Motzkin algorithmfor quan- torqueis propagated in ok-rot-pair,ok-rot-tension-pairproparifler eliminationoverlinear systemsof inequalities(see e.g. gatesthe force. [8]). The terminologyin Fig.2d showssomeexemplaryfaulty Besidescheckingthe validity of a sentence, quantifier behaviorsof a rotational pair. Apair is slipping (slippingelimination as a generic methodcan be used to computethe rot-pair) if bothtorquesare strictly positiveanddifferent. restrictions imposedon the parametersof the modelM: pair is broken(broken-rot-pair)if onetorqueis strictly positive, the other zero. Aseconddevelopermightdistinguish 1. Let X= z~,..., z,~ a list of parametersin M. betweenstrong andweakslipping pairs (strong-slipping-rot2. Forzi EX pair resp. weak-slipping-rot-pair) as it is depictedin Fig.2e. (a) Eliminateall quantiflersbut z~ fromS. I.e. transNotethat in our languageit is possibleto describethe weak formS into a sentence Si = 3xiRi, Ri quantifier slipping pair as the negationof a strong one. Thisallowsfor free. a simpledescriptionof a weakslipping pair, andreducesthe sourcesof possiblefaults. It furthereasesthe modification of S~in disjunctivenormalform,i.e. trans(b) Transform the knowledge base, e.g. a changeof the limit betweenstrong formSi into S~ = dnf(S~). and weakslipping. (c) Collectthe matrixR~of g~. 3. Returnthe list R = R1,... Rn of collectedmatricesR~. Sthi$is calledvariableelimination or projection in the field of CLP(~) 5 Simulation Simulationis basedon the followingprocedure: - 138- Figure 2: TheKrss code 1. Descriptionof the device:Instantiating,relating the instancesto primitiveconceptsandbuildingup the device structureby relating the instances. 2. Determination of the modelsof behaviorfor the different deviceparts. 3. Assertionof someparameterrestrictions. 4. Checkingthe consistency of the ABox.Thevalues and restrictions of additionalparametersare calculatedduring the model-generatingprocess. Furthermore,additional instances maybe generated. 5. Presentation of the generatedmodeland parameterrestrictions. Eventuallybacktracking in order to calculate another model. It is clear that the assertion of behaviormodelsandparameterrestrictions canbe freely intermixed.In addition,the consistencytest andtherefore the modelgenerationcan be triggered after every step. In the followingwewill give a brief overviewhowsimulationworks, using the drive train depictedin Fig.1. First the structure of the drive train has to be described:Instanceshaveto be instantiated, primitive conceptsfor this instances mustbe given andthe links and the kinematicpairs mustbe related (Fig.2f). In a secondstep (Fig.2g)wegive the valuesof the radii the crankarm-1, chainrlng-1,sprocket-1 andrear-wheel-1 and the force that is appliedto the crankarm as starting 9. values Wethen state that all components exposetheir normalbehavior. Finally,wegivea valuefor the force appliedto orankarm1. Nowthe values for nearly all missingparametervalues can be calculated(see Figure3). W.r.t. the bottom-bracketspindle-1 and the roar-axle-1 0 could be excludedfromthe admissiblerangesof the respectiveforces(since the force appliedto the erankarm is strictly positive). It is not surprising 9The values are given in Meter and Newton that nofurther restriction is possible,giventhe fact that no radius is givenfor these links. Neverthelessthe propagation of the respectivetorquesto their adjoininglinks is possible dueto the lawof torquethat holdsfor intact rotation-pairs. Figure3: Thegeneratedmodelandparameterrestrictions Arbitraryother parameterrestrictions as well as an arbitrary selectionof behaviorsare possible,but dueto the lack of space wecan give no further examples. 6 Diagnosis Diagnosisis supportedin the followingway: 1. Descriptionof the structure of the device. 2. Assertion of the observed parameter values and (re)realization of the affected components. This gives an upperboundfor the behaviorsthat are consistentwith the observedvalues. Weak realization in additiongives a lowerboundfor the . possible modelsof behaviorand therefore hints which parametervalues shouldbe determinednext. This values are thanasserted, the components (re)realizedetc. - 139- in CTL,wecan guaranteethat all missingandaccidentalsubWeillustrate this method,usinga single rotational-pairas an example.First Fig.2h showshowto describe a rotational sumption relations in the modellibrary are detected. pair within the ABox.Fig.2i then showshowdifferent sets 8 Summary and outlook of parametervalueslead to different diagnoses.Whilein the first variant identic torquevalueslead to ok-rot-pair, giving Descriptionlogics with expressiveconcretedomainsare well a zero value for the torque of onelink and a non-zerovalue suited for consistency-baseddiagnosis and simulation of for the for force of the other link suffice to recognizeit as technical devices. Additionallythey providethe inferences a broken-rot-pair.Thisis possiblebecausethe radius is rethat are neededin order to organizeandmaintainlarge model stricted to be strictly positive. Usingthe physicallawit can libraries. Actualworkfocuseson interfacing to a computer be concluded that the torqueof the secondlink is strictly posi- algebrasystemfor quantifier eliminationover quadraticsentive whichleads to the recognitionas a broken-rot-pair. In the tences [12]. Further workconcentrateson concrete domains third variant the pair is realizedas a weak-slipping-rot-pair,over other base types, e.g. using CLP(FD) systemsfor dedueto thecalculated ratio of thetorques. scribing qualitative models.Finally, weexplore other applicationareas suchas configurationandintelligent retrieval fromparts catalogsetc. 7 Model Libraries Fig.4 showsthe concepthierarchyafter classification of the References aboveterminologies.In this section wewill showhowclassification can be used for the organizationand maintenance [1] F. Baader and E Hanschke.A Schemefor Integrating Concrete Domainsinto ConceptLanguages.Research of modellibraries. Thefollowingobservationsare important Report, DFKI,Kaiserslautern, Germany, 1991. w.r.t this aspect: [2] G.E. Collins. Quantifier Eliminationfor Real Closed Fields by Cylindrical Algebraic Decomposition.In Proc. of the SecondGI Conferenceon AutomataThepry and FormalLanguages,Springer, 1975. [3] P. Hanschke.A DeclarativeIntegration of Terminological, Constraint-Based,Data-driven,and Goal-directed Reasoning.Dissertation, Kaiserslautern,1993. [4] H. Hong.RISC-CLP(Real): Constraint Logic ProgramFigure4: Theclassified TBox mingover the real numbers.In Constraint LogicProgramming: Selected Research.MITPress, 1993. All conceptdefinitions are different frombottom.There[5] JoxanJaffar, Spire Michaylov,Peter J. Stuckey, and fore all definitions are satisfiable. This guaranteesthat no RolandH. C. Yap. The CLP(R)Languageand System. modelof behavioris mistakenlydefinedin a waythat there Technicalreport, 1987. exists no parametercombination that leads to this behavior. All conceptdefinitions are distinct fromeachother. This [6] Gerd Kampand Holger Wache.CTL- a description logic with expressiveconcrete domains.Technicalremeansthat there are no twomodelsof behaviorthat are equivport, LKI,1996. alent, somethingthat could easily happenwhentwo model [7] J. de Kleer and B. Williams. DiagnosingMultiple libraries are merged. Faults. Artificial Intelligence,32, 1987. Thestrong and weakslipping pairs in Fig.2e are modeled as specializationsof rot-pair andnot of slipping-rot-pair.Sit[8] Jean-Louis Lassez. Parametric queries, linear conuationslike this are verylikely if differentpeopleare simulstraints and variable elimination. In Prec. DISCO-90 taneouslydevelopingmodelsof behavior, or whenthe model 1990. librat’ies are complex. Theclassificationservicedetectedthe F. Reuleaux.TheKinematicsof machinery- outlines of [9] missingsubsumptionrelation betweenslipping-rot-pairand a theory of machines.Macmillan & Co,, 1876. strong-sUpping-rot-pair. In othersituationsit may bethecase P. Struss and O. Dressier. Physical Negation- Integrat[10] that a computed subsumption relation is not missingbut acing Fault Models into the General Diagnostic Engine.In cidentalin a sensethat it is causedbysomeerror or laxness Prec. IJCAI-89, 1989. in the descriptionof the modelsof behavior.SuchErrors are very likely in large and complexmodellibraries. Therefore [11] A. Tarski. A DecisionMethodfor ElementaryAlgebra andGeometry.In CollectedWorksof A. Tarski. the detectionof these errors is crucial for the development of 1° suchlibraries. Sinceall inferencesare soundandcomplete [12] VolkerWeispfenning.ApplyingQuantifier Elimination to Problemsin Simulationand optimization.Technical 1°atleast for linearsystems of inequalities in ourcurrentimpleReport,UniversitatPassau,1996. rnentation - 140-