School of Electrical and Computer Engineering. Georgia Institute of Technology. Atlanta, GA 30332 .... importance in manufacturing and automotive industry.
INTEGRATEDDIAGNOSISANDPROGNOSISARCHITECTURE FORFLEETVEHICLESUSINGDYNAMICCASE-BASED REASONING
AbhinavSaxena,BiqingWu,GeorgeVachtsevanos SchoolofElectricalandComputerEngineering GeorgiaInstituteofTechnology Atlanta,GA30332 404-894-4132 {asaxena,becky.wu,gjv}@ece.gatech.edu Abstract - This paper presents a hybrid reasoning architecture for integrated fault diagnosis and health maintenance of fleet vehicles. The aim of this architecture is to research, develop and test advanced diagnostic and decision support tools for maintenance of complex machinery. Artificial Intelligence based diagnostic approach has been proposed with particular reference to DynamicCase-BasedReasoning(DCBR).This system refines an asynchronous stream of symptom and repair actions into a compound case structure and efficiently organizes the relevant information into the case memory. Diagnosisiscarriedoutintotwostepsforfast and efficient solution generation. First the situation is analyzed based on observed symptoms (textual descriptions) to propose initial diagnosis and generate corresponding explanation hypothesis. Next, based on the generated hypothesis relevant sensor data is collected and corresponding data analysis modules are activated for data-driven diagnosis. This approach reduces the computational demands to enable fast experience transfer and more reliable and informed testing. This system also tracks the successratesofallpossiblehypothesesfora given diagnosis and ranks them based on statistical evaluation criteria to improve the efficiency of future situations. Since the system can interact with multiple vehicles it learns about several operating environments resulting in a rich accumulation of experiences in relatively very short time. A distributed and generic architecture of this system is outlined from technical implementation point of view which can be
0-7803-9101-2/05/$20.00 ©2005 IEEE
used for widespread applications where both qualitative and quantitative observations can be gathered. Further, a concept of expanding this architecture for carrying out prognostic tasksisintroduced.
INTRODUCTION A paradigm shift is emerging in system reliability and maintainability. The military and industrial sectors are moving away from the traditional breakdown and scheduled maintenance and adoptingconceptsreferredtoasConditionBased Maintenance. Significant efforts have been put in the past decade to automate the condition monitoring and health maintenance of industrial systems.Butatthesametimesystemcomplexity hasgrown out ofproportionleavingmodelbased and other extensive analytical techniques limited within the constraints of available time and computational power. In addition to signal processing, diagnostic and prognostic algorithms, these new technologies require storage of large volumes of both quantitative and qualitative information. Knowledge based systems on the other hand have shown a great promise in handlingsuchlargesystems.Thesesystemsoffer to use higher level information acquired from previous experience and reuse it for new but recurring situations. One of these approaches, called Case-Based Reasoning (CBR), retrieves old cases from case libraries and matches them withacurrentproblemtocomeupwithasolution similartotheoneappliedtooldcases.Manysuch knowledge based expert systems have been developed which depend on intelligent memory access in order to come up with fast solutions
96
Unlike traditional practices use of standardized language is being promoted in industrial environments for improved efficiency, accuracy and data interoperability [5]. A structured syntax andafixeddomainvocabularyreducethetaskof Natural Language Processing (NLP) significantly. The concept of standardized formal language offers several advantages over using nonstandardizedlanguageinindustrialenvironments. It not only helps reducing communication errors by avoiding ambiguities but also simplifies electronic textual data management and technologytransferbetweenmanufacturers,users and maintainers. Using a domain limited vocabulary and a well defined documentation format makes technical language globally interpretable and reduces multicultural linguistic barriers. For instance formal communication withintheaviationmaintenancedomainisdefined and regulated [6]. A hierarchy of written correspondenceisdefinedintheFederalAviation Regulations(FARs), whichincludesairworthiness directives (ADs), notices to airmen (NOTAMs), maintenance manuals, work cards, and other types of information, that are routinely passed among manufacturers, regulators, and maintenance organizations. The international aviation maintenance community has adopted a restricted and highly structured subset of the English language such as ATA-100 and AECMA SimplifiedEnglishtoimprovecommunication[5].
without having to explicitly solve complex model equations. In addition to analytical knowledge these complex systems also generate a lot of useful information in the form of descriptive knowledge. This knowledge so far has not been used to a large extent because of the lack of adequate text processing techniques which are often associated with high computational demands. To achieve better performance and a higher levelofautonomy thisknowledgemustbe utilized. This paper suggests one such approach based on Dynamic Case-Based Reasoning (DCBR). This approach offers to automate the troubleshooting process to a large extent by providing a decision support system using an extensive knowledge base and reduced computationaldemands.
Industrialpractices In most situations health information is obtained from monitoring sensors and automatically generated activity logs. But most of the systems use only the quantitative information available from the sensors to automate the diagnosis task andalmostnoneorverylittleuseofthequalitative information is made. For instance [1] used CBR system for fault diagnosis in industrial robots using a case-base consisting only of acoustic signals that act as signatures of various faults. Another attempt on aircraft maintenance system focused on diagnosis of electronic ballast on the airplanes [2]. Several numerical features were includedandweightedusinggeneticAlgorithmsto calculate similarity. But any available qualitative information was again ignored. A diagnostic system was developed for aircraft fleet maintenanceusingfailureandwarningmessages generatedbyaircrafton-boarddiagnosticroutines [3]. The approach was based on formatted text messages for which trigram matching technique wasusedtoretrievesimilarcases.Thisapproach works well for a fixed text structure and when word order does not dictate the meaning of the phrase.Similarly[4]usedpartialmatching,based on key words and constraint-nets, to retrieve similar case from the case-base. This process also used only the textual information and no numerical analysis was required. Several other applicationscanbecitedwhichlieoneitherendof the spectrum. But in reality large industrial systems provide both quantitative and qualitative information which are important. This paper describesaninitiativeindevelopingasystemthat utilizesbothkindsofinformationtomakethetask ofdiagnosisefficient,fastandaccurate.
Similar to aviation industry the importance of standardized technical documentation is gaining importance in manufacturing and automotive industry. Efforts are being made to enhance the abilitytoupdatesupportdocumentsduringthelife cycleofasystemasitis maintained,modified or resold to form a valuable archive of knowledge concerning safe and reliable operation of the system. A lot of machine condition information is obtained in terms of operator observations expressedastextualdescriptionswhicharerarely used in automation of the health maintenance process. But these symptoms carry important information about the system which may not alwaysbeevidentfromsensorymeasurements.
DYNAMICCASE-BASEDREASONING Case-Based Reasoning (CBR) solves problems by adapting previously successful solutions that were used to solve old problems [7, 8]. DCBR is an extension over conventional CBR that include several dynamic components to enhance the
97
case base is sparse, populated with the most frequently encountered cases only. It may not cover the entire range of problems because they have either not been encountered so far or have been forgotten from the past. This knowledge is made available from the current experts of the systemormaintenancelogs.
capabilities of the system. In conventional CBR onlythecaselibraryisdynamicprimarilybyvirtue of its evolution as long as newer cases are encountered.DCBRinstantlygroupssimilarcases based on their spatial locations as they are encountered and compresses them into a dynamic case by creating a statistic-vector. Therefore, each case's contents change/evolve wheneverrelevantadditionalinformationbecomes available. The reasoning is dynamic based on whatpartoftheinformationincludedinthecases (case component) is given more importance for making inferences to produce an appropriate solution scheme. Furthermore several adaptation schemes and similarity metrics can be dynamically chosen based on what was more successfulinthepastforsimilarsituations.Asthe useofCBR isbeingexploredformoreandmore applications several variations in the structure have been suggested based on specific requirements. In [9] authors describe a system where time-tagged indexes and dynamic composite features make the CBR system dynamic.Inanotherapproachcasesareextracted and expanded dynamically based on context and the facts specified in advance [10]. Similarly in this approach the case contents are loaded dynamicallyandadaptationalgorithmsarechosen based on situation. A specific example has been described for diagnosis of fleet vehicles in the followingsections.
Enrichment Phase: This phase remains active until almost an entire range of cases have been documented and the new cases occur mostly as the repeated instances of already existing cases. Thisphasetendstofillthegaps,betweenvarious possibilities that were created during the initialization. Every time a problem case is presented(assumingafaulthasbeendetected),it can be either grouped with one of the old cases based on its similarity or considered as a new case.Ifanewcaseisdetected,thecaselibraryis updated by including this case. If this case matcheswithanalreadyexistingone,thestatisticvectorofthecorrespondingrepresentativecaseis updated.Thisstatistic-vectorcanbeusedfor: 1.FaultDetection:Decideifthepresenceoffault isindicatedbycurrentdataindicators. 2. Fault Identification and Isolation: Decide which fault(s)ispresentoutofseveralpossibilities.This may indicate the presence of several faults and hencemorethanonematchingcase(somewitha different degree of severity, while others with differentfaultcharacteristics).Aprobabilitycanbe assigned to them based on the confidence level calculated when categorizing the problem case intotherepresentativecases. The advantage of such a technique is that the similarity metric for retrieval is not calculated based on one representative case, which may or may not be the best representative for the corresponding problem, instead it is based on a statistic which has been preserved for all similar cases encountered in the past without explicitly storingallofthem.
ImplementationandMaintenance ThelifecycleoftheDCBRsystemcanbedivided intothreemajorphases,asshowninFigure1. Initialization
Enrichment
Maintenance Time
Case-BaseMaintenance:Thisisarelativelyless frequentprocessthatcontinuesthroughoutthelife of the system to cope with any non-stationarity. Based on past successes and failures, a performance assessment can be carried out for various parameters that are being monitored beingcasecomponentsandareusedtodetecta fault.Weightscanbeassignedtocasesandtheir constituentcomponentstoexpresstheirrelevance in solving a particular fault. In a nut-shell this involves reorganization of data base to improve system performance based on data analysis and
Figure1.PhasesintheLife-CycleofDCBR
Initialization Phase: Like all other knowledgebasedsystemDCBRrequiresinitialbootstrapping with whatever knowledge is available initially. A
98
differentsensorsconstitutingthecasesoraneed for alternative sensors or sensor placement may beinferred.
following inferences. For example some sensors maybetaggedasredundantandothersasmore important after analyzing the success rates of Sensors
Feature Extraction
Diagnosis Module
Final Diagnosis
System
Repair Action + Explanation
Observations (Textual Data)
Information Extraction
Select Sensors
Select Features
Select Diagnostic Modules
Solution Revision
Initial Diagnosis
Knowledge Base
CBR Engine
Feedback Update Statistic
Solution Evaluation
Figure2.DCBRsystemforintegrateddiagnosisofindustrialsystems
DIAGNOSIS
SystemArchitecture
DCBR offers a good promise for diagnostic and decisionsupportsystemsbyemulatingthehuman reasoning process one step further. It narrows down the problem search space by dividing the diagnosistasksintosmallersteps.Inmostcases industry uses a two step procedure. First the problem is suspected by the operators due to unusual symptoms observed on the system or error logs. Maintenance experts try to come up with possible explanations for those symptoms. Then using known analytical techniques relevant diagnostic tests are run to confirm the problem. After a problem has been diagnosed experts use their experience to plan and execute the repair task. The DCBR system described here follows the same approach. The system refines an asynchronous stream of symptom and repair actions into a compound case structure and efficiently organizes the relevant information into the case memory. Thus not only the quantitative analytical knowledge needs to be considered but also the qualitative descriptive information should beprocessed.
Qualitativeinformationisusedastheinitialquery. Textual descriptions are converted into semantic networks which preserve the meaning of the text and at the same time convert the text into a defined structured for easier analysis [11]. The case-base is searched based on these semantic networks and the relevant hypotheses are generated. Based on the past experience these hypotheses are ranked and the most probable hypothesisistestedfirstbyactivatingtherelevant data acquisition and corresponding diagnostic modules.Ifthehypothesisisconfirmedtobetrue its solution is suggested for the current situation and its success rate is updated. Otherwise the next probable hypothesis is tested and corresponding success rates are updated. The procedure is repeated until a useful solution is obtainedoranewcaseisgeneratedandstoredin the case-base (Figure 2). For new cases approximate solutions are suggested based on closest matching cases. They are further revised based on feedbacks until a satisfactory solution hasbeenfound.
99
KNOWLEDGEREPRESENTATION The motivation for this kind of higher level CBR systemhasbeenderivedfromseveralfacts.First, a lot of useful knowledge is stored in the form of textual descriptions which must be used in additiontonumericaldata.Nextalotofanalytical knowledgehasbeendevelopedovertheyearsfor differentcomponentsofalargesystemthatneeds to be managed. Managing this knowledge will help in developing an integrated architecture for the whole system and not just the individual modules. This in turn will make this knowledge accessible for adaptation and transfer tasks betweensimilarbutdifferentcomponents.
Figure3.Atypicallistofsensorsemployedinasystem. This information is usually available from the manufacturersortheoperators
Figure4.Atypicalsetoffeaturesusedforfaultdiagnosisinmechanicalsystems.Knowledge canbeeasilyorganizedinthisformatforeasyusebytheDCBRsystem. CaseExample. .ID .Component
B10 Bearing
.Location
Main_Transmission S_ID Symptom Wt Sementic_net Hypotheses 1 Noisy in neutral 0.57 SemNet1 h1, h2, h3 with engine running 2 Vibration 0.43 SemNet2 h1, h2, h4 H_ID Hypothesis Wt Diagnosis 1 Primary Gear worn 0.65 d1 2 Primary Bearing worn 0.15 d2, d4 3 Clutch Release Bearing worn 0.10 d3, d4 4 Lack of oil 0.10 d5 D_ID (Sensor,Feature) pairs Wt Solution 0.75 1 r1, r3 d1:{(S3,F1),(S3,F4)} 0.80 2 r2, r3 d2:{(S1,F15),(S1,F8)} 0.80 3 r3 d3:{(S2,F15),(S2,F8)} 0.6 4 r3 d3:{(S1,F13),(S4,F13)} R_ID Repair Wt 1 Change Primary Bearing 0.33 2 Replenish oil 0.33 3 Change Clutch Release Bearing 0.33 Last_Update Case_Quality Succ Fail Conditions 03:16:05 0.8 8 2 Full Load Windy
.Symptom
.Hypothesis
.Diagnosis
.Repair
.Version
Figure5.Anexamplecaseforautomotivevehiclediagnosisandtroubleshooting
100
AsshowninFigure3itispossibletoenlistallthe diagnostic sensors and their operating characteristicsdeployedinthesystem.Similarlya listofdiagnosticfeaturesfromthesensordatato diagnose known fault modes can be compiled. Relative importance of features and underlying philosophy can also be included in such a list (Figure 4). More knowledge of this kind can be added to the list as it is developed. Now all this informationcanbeorganizedintoacasestructure as shown in Figure 5. Based on the query semantic network relevant cases also matching these primary indexes are retrieved and a set of possible hypotheses is generated. Two approachescanbetakenatthisstage.Diagnosis can either be performed based on a set-covering algorithm to find a solution that explains most of the hypotheses [12] or different hypotheses can be tested one by one in order of decreasing successrates.
relationships between different words in the sentence. It was found that most sentences can be broken into smaller segments which independently define a relevant concept and can berepresentedastriads.AtriadIJisathree-tuple consistingoftwophrasesp1,p2andarelationship linkLbetweenthem(Figure7). L1 Triad
P1
P2
Transmission has no drive in reverse gear
Type-Itriad Type-IItriad
IN
TextProcessing The semi-structured nature of the industrial descriptions and a limited vocabulary within a domain resolves the meaning ambiguity problem to a large extent [13]. The proposed hypothesis maintainsthatthesimplerstructureofthetextcan compensate for the computational complexity, usually involved with NLP and can improve the degree of belief established by numerical data driven methods alone in addition to a more targeted data processing. In order to carry out NLP the words in the sentences are first tagged with corresponding parts-of-speech. A demo version of TreeTagger tool [14] developed at InstituteofNaturalLanguageProcessing(IMS)at Stuttgart University was used. The output of the program is three columns, first containing the original word as it appears in the sentence, second contains the tag abbreviation (e.g. NNP for proper noun, VB for verb etc) and the last columncontainsthestemoftheword(e.g.'run' for 'running'). Figure 6 shows a snapshot of theoutputfilefromTreeTagger.
AT
IS
GEAR
TRANSMISSION NO_DRIVE
REVERSE
Figure.7.Asemanticnetworksconsistsoftriads
Aphrasecanbeanoun,adjective,verboratriad itself. A triad not consisting of another triad is called Type-I triad and the ones that contain one or more triads are called Type-II triads. It was observedthatthreebasicrelations(IN,ATandIS) explain most of the relationships between words andtriadsandcan beusedasreducedgrammar for efficient computation purposes. For the purposesofmatchingsimilaritymetricshavebeen definedthatfirstcomputesimilaritybasedontypeI triads and then further breaks the tie between multiple matching cases by considering type-II triads.
EXPANSIONFORPROGNOSIS
Figure.6.TreeTaggeroutput
After all the words have been tagged a set of syntactic rules are employed to extract
Thesystemdescribedabovecanbeexpandedfor prognosistaskwithlittleefforts.Themoststraight forward expansion is based on the fact that already existing prognostic algorithms can be integrated in a similar fashion as diagnostic algorithms to get activated once the fault has been localized. Fault is first identified by
101
diagnostic routines followed by activating the correspondingprognosticroutines. The case structure itself can also be used as a higher levelprognosticplatform.Atimehistory of the situations can be included as a part of the case. Such a history can be implemented as Traces.Tracesnotonlykeeptrackofcurrentstate ofthesystembutalsotheevolutionofthestatein therecentpast.Similarlythetime-taggedindexes as described in [9] can be used for generating trends. These trends can be used to make subsequentprognosis.
CONCLUSIONS This paper has described a novel approach for integrated diagnosis/prognosis of systems. The suggested architecture enables encoding of analyticaltechniquesfromasystemspointofview and its expansion for prognosis tasks under the same structure. The performance of such knowledgebasedsystemdependsonthedegree of completeness of its knowledgebase. Since the systemcaninteractwithmultiplevehiclesitlearns aboutseveral operatingenvironmentsresultingin a rich accumulation of experiences in relatively very short time. At the same time it also serves multiple systems. A natural language processing technique has been developed to extract information from the textual descriptions that is less computationally expensive than the usual NLPtechniquesandstillpreservesthemeaningof the text. The test data is being gathered for the experiments from the domain of automobiles to showthecapabilityofthesystem.
REFERENCES [1] Olsson,E.,Funk,P.,Bengtsson,M.:Fault DiagnosisofIndustrialRobotsUsingAcoustic SignalsandCase-BasedReasoning,In:Funk,P. andGonzalez,P.A.(eds.):ECCBR2004,LNAI 3155,Springer-VerlagBerlinHeidelberg,pp.686701,2004 [2] Chiu,C.,Chiu,N.-H.,Hsu,C.-I.:Intelligent AircraftMaintenanceSupportSystemUsing GeneticAlgorithmsandCase-BasedReasoning, InternationalJournalofAdvancedManufacturing Technology,24:pp.440-446,Springer-Verlag LondonLimited2004
ProceedingsoftheAAAI'98Workshop:CasebasedReasoningIntegrations.TechnicalReport WS-98-15.Madison,Wisconsin,USA.July27, 1998.pp.91-95.NRC43577 [4] Lee,S.G.,Ng,Y.C.:HybridCase-Based ReasoningforOn-LineProductFaultDiagnosis, InternationalJournalofAdvancedManufacturing Technology,Springer-VerlagLondonLimited 2005 [5] Greenough,R.:e-Smart-ElectronicSupport ofManufacturingTechnologyWorkPackage Report,CranfieldUniversity,UK [6] Drury,C.G.,Ma,J.:LanguageErrorsin AviationMaintenance,Year1InterimReportfor FederalAviationAdministration,StateUniversity ofNewYorkBuffalo2003 [7] Aamodt,A.,Plaza,E.:Case-Based Reasoning:FoundationalIssues,Methodological Variations,andSystemApproaches,AI- Communications,7(i),pp.39-59.1993 [8] Kolodner,J.L.:Case-BasedReasoning,San Mateo,CA,MorganKaufmannPublishers,Inc 1993 [9] Xia,Q.,Rao,M.;DynamicCase-Based ReasoningforProcessOperationSupport Systems,EngineeringApplicationsofArtificial Intelligence,vol.12,pp.343-361,1999 [10] Mostek,T.A.,Forbus,K.D.,Meverden,C.: DynamicCaseCreationandExpansionfor AnalogicalReasoning,ProceedingsofAAAI-2000 [11] Winston,P.H.:ArtificialIntelligence,3rded., Addison-WesleyPublishingCo.,1993 [12] Stefik,M.:IntroductiontoKnowledge Systems,MorganKaufmannPublishersSan Francisco,CA,1995 [13] Lenz,M,hubner,A.,Kinze,M.:TextualCBR, In:Lenz,M.,Bartsch-Sporl,B.,Burkhard,H., Wess,S.,(Eds.):Case-BasedReasoning Technology:FromFoundationstoApplications, LectureNotesinArtificialIntelligence1400, Springerverlag,1998 [14] Schmidt,H.:ProbabilisticPart-of-Speech TaggingUsingDecisionTrees,inInternational ConferenceonNewMethodsinLanguage Processing,Manchester,UK,1994
[3] Lehane,M.,DubÈ,F.,Halasz,M.,Orchard, R.,Wylie,R.,Zaluski,M.:IntegratedDiagnostic System(IDS)forAircraftFleetMaintenance,
102