Using Ontological Classification to Design Fault Detection and Isolation Architecture

Copyright to IFAC Fault Detection, Supervision and Safety for TechnicaI Processes, Kingston Upon Hull, UK, 1997

USING ONTOLOGICAL CLASSIFICATION TO DESIGN FAULT DETECTION AND ISOLATION ARCHITECTURE

L. G. Vela Valdes, D. Tbeilliol and D. Sauter.

CR4N - CNRS UR4 821 - Universite Henri Poincare - j'liancy 1. B.P. 239 54506 vandoeuvre Cedex France. Phone: (33) 3-83-91-20-69. Fax: (33) 3-83-91-20-30. e-mail: [email protected]

Abstract: In this paper the problem of knowledge modeling about Fault Detection Isolation is addressed. The proposed approach is based on the Ontology and construction of abstract models, applying the functional mode ling approach. The term abstract models is used to indicate the intended interpretations of behavior according to some criterion. A set of problem requirements in terms of goals and conditions are proposed in order to accomplish this criterion. The study was limited to a simulated system of a d.c. motor. The results and the diHerent abstracts models are compared. Copyright © 1998 IFAC

Keywords. Fault Detection Isolation, Classification, Fault Modelling, Function, d.c. motor.

1. INTRODUCTION

original generalized concept of the Fault Detection Isolation (FDI) abstraction problem is presented. Then. the main features of the proposed approach are sho\\TI in a simulated application, concerning a permanently exited d.c. motor. A sample ofFDI with parity space is illustrated and the advantages of the approach are discussed. Finally some conclusive remarks are reported and the future developments of the research are outlined.

Several authors (Abu-Hanna. et al. . 1991: Chitaro et al.. 1993) claim that Model Based Diagnosis systems depend on the functional model for their performance. Many device models use device structure and behavior to simulate behavior based on local interactions between primitive components. As the behavior becomes more complex. though, a more global view can provide a better insight into what a de\ice is doing. Concepts and abstractions of behavior reflect this global view.

2. BACKGROUND: USING ONTOLOGICAL CLASSIFICATION FOR MODELING TIIE KNOWLEDGE.

One important type of abstraction is the function, which reflects a purpose intended. The function is a subjective interpretation of structure and behavior. A pomp and a d.c. motor have a major objective. the power transformations. So their faults may be modeled in this way. The functional abstraction describes the goal of device at a level of abstraction that is of interest in the problem solving task. for example diagnosis. simulation. etc.

An abstraction identifies a principal characteristic or concept. The concepts are mental constructions which have a content and a scope. The basic relation between concepts is the entailed inheritance: a concept X has a concept Y as its characteristic or feature (Hautamaki, 1992). Fig. I shows two different abstractions for a RC network. In the domain of Electronic Engineering the designer looks a "filter". This abstraction identifies a desirable behavior in order to discriminate the high frequencies. In the other hand. in Automatic Control. the designer looks a compensation network or "controller" The RC

The aim of this paper is to specify a model in order to get a more natural fault representation. To this purpose. Fault Detection and Isolation in a d.c. motor is classed in electrical and mechanical faults . First. the conceptual background on knowledge ontological classification is outlined. Later. an 441

each other. Behavior knowledge is information which describes how components work and interact in terms of the physical quantities which characterize their state (variables and parameters) and the laws (equations) which rule their operation.

ABSTRACTIO:-; OBJI>CT

DESIGNER

: CO~C EPT

X ! CONC EPT Y I

I: t

!:

ELE CTRO}l; IC

R

i-", ~

FILTER E:"lGtNEERI~G

I; 1'OISCRIMI~A Fi~~UH~~y

]

! '

'

,

;:::=:====~,

!

I

I

AL'TOM ATIC

; CO MPESAT IOS

!

CONTROL

N ETWORK

TO

.

• TE '

,

,:

: . I 1

DEC!.~SE

I

Functional knowledge refers to a representation of the system in terms of interesting behaviors (i.e., tho~ behaviors which are significant to the goal for which the system has been designed). In the other hand teleological knowledge is devoted to described the legal goals for which the system can be used and how the system has to be operated in order to achieve them.

;

! i! STATE THE STEADY : ERR OR. i I

I

i

~------~. ~ . ------~

Fig. 1. Abstraction and Concept in a RC Network. network is used. for example, in order to decrease the steady state error.

2.2 The Ontological Models. An ontology allows to conceive an unified abstract scheme where knowledge classes may be defined, enumerated. and represented as generic terms (Ramoni. et al., 1992). For example. in Fig. 3 each knowledge model is classified in entities and relationships.

According to several authors (Chitaro, et al., 1993 : Jansweijer. et al., 1993) human abilitv to reason about physical systems depends cruci;'Uv on the types of knowledge used and how this kno'wledge is organized or classified. Given a physical system. it can be modeled in many ways. thus yielding a variety of different models. The model is a symbolic system designed to provide a representation of a physical system appropriate for a given purpose. So. a model is only a partial representation of reality and depends on subjective decisions of the model designer. For example. a physical model is used in order to represent a system. In the other hand. a diagnostic model provides a fault description or fault interpretation.

RJNDAME~rrAL

INTERPRETATIVE

1010WLEDGE

KNOWLEDGE

>=TTlES

, CadPON&'ITS

I,

, P AIU-"ETDtS

Ii I

v AJUA8I.ES

I

BEHAVIOR.\LM)OEL

I j ItE.UTIQofSH1!"5

The Knowledge Classes.

i

The knowledge model is separated in two classes as shown in Fig.2. The first class. the fundamental knowledge. include the structural and the behavioral model. The second class. the interpretative knowledge. is composed of functional and teleological models (Chitaro. et at. . 1993). FC"OA!\.iE~TAL

I~T ERPRETATI V E

KNOWl.EDGE

" "OWL ED G E

STRI:C T CRAL

~1 ()DEI.

BEHAVIORAL MODEL

:

£QliAnO~ !

Il'llT1lES :

: SIGND'lC"""T GOAL :

IREUTIa
,Ca
!

i •

ITELEOL
I ~~

' UGAL~

!

: CeJ'II"DIllONS

RE.I.AnooSHIPS

i

Fig. 3. The Knowledge Ontological Model. Thus in the fundamental knowledge. structural model treats components as entities and connections as relationships. Behavioral model defines parameters and variables as entities and equations as relationships. In the other hand. interpretative knowledge classes goals (Significant and legal) as entities and conditions to accomplish the goals as relationships.

F UNCTlO" AL !\.iOOH

' TELEOLOGICAL

I,

iREUTICMmPS 'CONNEcn=1

j l:'mTlES

2.1

R JNCllONAL M:>DEL .

SlRl..'Cl1JR.\L MODEL

1

I

In fact. fundamental and interpretative knowlegde may be used in order to represent physical model knowledge and / or diagnostic model knowledge.

~I ODEL

3. ONTOLOGICAL MODEL OF THE FAULT DETECTION ISOLATION.

Fig. 2. The Knowledge Classification.

The main strategy used in the Model Based Diagnosis (or Model Based FDI Diagnosis) systems is anal)1ical redundancy. the confrontation of measurement data \\'ith prior knov,,'ledge embodied

Structural knowledge is information about all possible pathways of interaction among components. more specifically. it describes which components constitute the system and how they arc connected to

442

in the mathematical model (Gertler. 1991). The anal~tical redundancy is used in order to provide residuals The residuals are quantities, ideally equal to zero. that reflect the presence of noise. modeling errors and occasional faults . The residuals are subjected to statistical tests in order to get a symptom. The statistical test is the comparison of a residual to a reference value. The symptom is a boolean vector obtained as a combination of residual values after the comparison with the reference value. The symptoms are analyzed. mostly by simple comparison to standard patterns, or fault signatures, in order to get the cause of the fault. or diagnosis. This general procedure is depicted in Fig. 4.

DECISION SYMPTOMS

I RESIDUAL 1 I EVALt:ATION ,

: -

-

'

R.[ stDCALS

RESIDUAL ; GE!'ERATION ~ ot,;TP~ TS

I~Pt."T~

_ _ _ _ _~ , ;.S~Y~ST~E~~~'~ , - - - - -:o.;otSE

,

Ontological Model of the FDI allows to describe diagnostic knowledge. The objectif is to choose an appropiate concept in order to get a more natural fault representation 3.2 Ontological Models of Three Fault Detection and Isolation Schemes.

OlAGSOSIS

1

Ontological framework. So only entities and relationships are used in order to describe the Ontological Model. The Structure defines measurement data as entities and residual equations as relationships. The Behavior treats residuals as entities and signatures as relationships. In the other hand, the Function classifies significant goal as entities and conditions as relationships. Finally the Teleology looks legal goal as entities and conditions as relationships.

DISlUR.BA;'I:C!' fAt.:LTS

As it was mentioned before. structure and behavior into Ontological Model of FDI are defined by measurement data, residuals, equations and fault signature. In this section a general procedure is outlined in order to define function and teleology. Teleology or legal goal may be looked as general objective and Function or significant goal may be defined as particular objective. So in a Ontological Model of FDL Teleology and Function may be represented as follows . TELEOLOGY Legal Goal

Fig. 4. Fault Diagnosis 3. 1 Falllt Detection Isolation and Ontology.

The basic definitions for knowledge classification and diagnosis illustrated above, namely Ontology and FDL provide the background for the abstraction and classification of faults which characterize the proposed approach. The notions of concept. abstraction and ontology may concern any application. The Ontological Model allows to conceive an abstract scheme where Fault Detection Isolation may be represented as generic tenns. For example, in Fig. 5 the Fault Detection Isolation is classified in an

FUNCTION

Significant Goal

- model structure. - model parameter. - type of fault. - type of input. As it was mentioned before, Functional and Teleological knowledge are subjective interpretations in terms of goals and conditions. In fact. following definitions are based in these features. So "model structure" and "model parameter" are taked as conditions in the Teleological Model. In the other hand. Functional Model defines "type of fault" as condition I and "type of input" as condition 2.

:.IEASI:J\E\!E:\T DATA

; Ra,\l1Cl'\sHif'S

REsIDt:Al

EQCAnoNS

REl.AnONSHlPS . COl'.nm o:-
TELE<.>L
, R.f:l..\nt f.'.:\ i!l?S

FAl:LT

TUl..An()!'>.;sHIPS

: Detection and isolation

In Isermann (1994). the main features of three FDI schemes (parity Space, Observers and Parapeter Estimation) are presented. In order to define the teleological and functional conditions, common characteristics between these three schemes will be finded. This characteristics or features can be classed in four groups.

F1.;:\CTlO).' E:\~~

: Point out an unacceptable change

This choice is only a subjective interpretation of the three FDI schemes presented. Table I shows the Ontological Models of these schemes and summarizes the results presented in this section

("'()r-.'Omcr.-.;s

Sl~n.'ltE

Fig 5 Ontological Model of the FDI .

443

Table I Teleological and Functional Models of Three FDI Schemes. F

decoupled from one input or output signal.

, lHEClCXIY : KlN[aJT l.N\cx:H'fAElE ~ PARA"fTIR fSI1M\TIOI

PARm'

!

S?.>£:E

'

a : MXU

i

CllSERVCR

: a: i'>UE.

SIRlcnm:

smnm:

rl(t) = La la(t) + Ra la(t) + Fl w(t) - Va(t )

(I)

r2(t) = J w(t) + Fl la(t) + Mt1 w(t) + MI(t)

(2)

.

D

r3(t) = J La Ia(t) + (La Mfl + J Ra) la(t) + (F1 2 + Ra Mfl) Ia(t) - J Ua(t) - Mfl Ua(t)

·1

U(AClLY K , C! :

~I:UL

C! : MDL

Q : :\f:QL

p'.\R~'£IER

P'<\R~'£IER

INl."O\!\'

"'~

:

- Fl MI(t)

;

(3)

.

r4(t) = J La w(t) + (La Mt1 + J Ra) w(t) + (Fl 2 + Ra Mt1) w(t) - Fl Ua(t) + : a : :l-I.LlIHlOOlVE a

.

La MI(t) + Ra MI(t)

:.~F.-\lU

(4)

F.-\lU : C! : EXlL-UXN BV mE l\P{,T lS l'IIEI&\RY

~.

Q : EXlT.6J1(]'iBVlHE

N'LTISXJT

C! : DcrTAllCNBV lHEL"l'LTlS

i : :

The above d.c. motor has been simulated and Parity Space approach has been used in order to get residuals of additive and multiplicative faults. The input Ua is kept constant. Each fault is simulated by an abrupt variation of + 10 % in your nominal value. The mean of residuals are subjected to simple threshold test in order to get the symptoms of each fault. Table 2 shows the fault signatures and summarizes the results of simulation and residuals evaluation.

~.

~~'

A SIMULATED EXAMPLE: D.e. MOTOR

Sensor and parameter faults with parity space has been already regarded in other works (Isermann. 199~: Hofling and Pfeufer. 1994). In this section an original approach to FDI is presenred. which is based in functional abstractions of Ontological ModeL A functional abstraction is only a partial representation of the reality and depends on subjective interpretations of the FDI goaL In next section the kernel of this approach ,,,ill be developcd.

Table 2 Fault Signatures with Paritv Space.

ADDITIVE

S

A typical d.c. motor block diagram is shown in Fig. 6. where Ua. la. ML ware armature voltage. armature current. load torque and velocity. The mathematical model and the nominal parameter vallles are described in Hofling and Pfeufer (1994).

~o

MI 0

:
0

, ! :,2

0

s3

0

,4

0

MULTlPLICATIVE

FAULTS

u.

w

la

FI

Ra

La 0

0

0

0

FAULTS

J 0

Mn : 0

0

0

0

MI -1.1

Parity Space and Functional Abstractions in a D.e Motor.

In order to diagnose a speed sensor fault in a d.c. motor. it may be useful to consider other abstractions besides the additive and multiplicative ones. For example. the type of phenomenons in the motor can be classed in mechanical and electrical power transformations.

-.--------EJf..-----Fig 6. Block Diagram of The D.e. Motor.

The concept of functional abstractions. in a FDI Ontological modeL is the kernel of this research. This concept provides and original and generalized interpretation of fault types. In this context. a fault depends also on subjective interpretations of the designer.

For the detection and isolation of sensor (output) and actuator (input) faults. a set of structured residuals is appropriate (Gertler. 199 I) In Hofling and Pfeufcr (191)'+) a continuous time parity space was proposed for a de. motor. whcre each parity cquation is

444

Fig. 7 shows two different abstractions of the Fault Detection Isolation Function. In the first row. the concept X looks this function as "Fault Detection and Isolation in components" . The concept Y classifies the faults in fault sensor and fault parameter.

Fig. 8 provides this functional abstraction. or significant goal. which classifies the faults of system . in components (see Fig. 7). 1\11. Mn l :. , R. la

'- ~

I ,

. I

DESIGNER

i

: COKCEPT X : OF F D I

D.e

FAULT

COKCEPT Y , OF F D I '

,'

•

i FAt."LTOE'TEC110S :

CO~IPONENTS

~OTOR

ISOLATIO N

OF

(SOLATIO}:

OF

SPACE

i

FAULT

FA t: L.TS

6

F.\t.:LTS

L : D C

!.--'--' ;

: PHENOMENONS ;

~';===~': i I!': THE SYSTEM

:

,

'

i FAUtT Drn:cnOS

ii

ELECTRI CAL

lSOLA710 S IS

ii

FAULT

'OWE».

';

, !

[",",SSFo.. ""noss:

I

MOTOR

•

I

Fig. 8. Fault Detection Isolation in Components.

:vtECHA:-IICAL . fAt: LT

It is important to point out that CONDITION I defines exactly the type of component. In this case. CONDITION I treats only sensor (additive) faults. Which is the classic goal in a Parity Space scheme (see Table 1). In other words. only additive fault signatures will be considered

Fig. 7. Functional Abstractions in a D.e. Motor. In the other hand in the second row. the concept X looks the FOI function as "Fault Detection and Isolation in power transformations". The concept Y classes the faults in electrical and mechanical ones. The functional abstractions in FOI Ontological ModeL are not new models. but ways looking at existing ones from a given perspective using an appropriate concept. -I.]

. DETECTION

c:

j

F D I

SYSTEM

THE

I

PARITY

SE:-ISOR fAL: LT

COMPONE:-."TS

; ~THE S YSTE:vt ;

I

';

IS

~ . \.

PARA~IETER

-',-

! ,

Of

COMPONENTS

I LOOKS THE

CLASSIFICATION

SESSOR.

ABSTRACTION OBJECT

w

/\,

/\

In this context the functional abstraction illustrated above, provides "detection" in six faults (Ml, Mfl Ua. Ra. la and w) and isolation in only two faults (la and Ra). In the other hand. the second functional abstraction in the ontological model describes the FOI scheme as follows:

A Comparative Study of Two Functional Ahstractions Components and Power Tral1!>1ormations.

FOI SCHErvtE : PARITY SPACE SIGNIFICANT GOAL : Fault Detection Isolation (FUNCTION) in Power Transformations. CONDITION I :Electrical and mechanical faults. Excitation by the input is CONDITION 2 not necessary.

The sensor Ua and the parameter Ra have the same fault signature (see Table 2) An similar situation is produced by the sensor MI and the parameter Mfl. Functional abstractions may be extremely useful in differentiating the effects between senso; (additive) and parameter (multiplicative) faults .

Fig. 9 shows this functional abstraction. or significant goal. which classifies the faults of system in power transformations (see Fig. 7). In this case CONDITION 1 treats electrical and mechanical faults. So this condition allow us to use additive and multiplicative fault signatures.

The objective in this section is to compare the performance of two functional abstractions (classifications) in a FOI scheme. Parit}' Space is used in order to detect and to isolate faults in a d.c. motor. These functional abstractions are based on the Ontological Model. The criteria for comparison is the quantity of faults detected and isolated.

• • OF

C :"A SSIFICA 7 IO:-':

!

The d.c. motor system consists of four sensors Ml. Ua. wand la and five parameters flux. Ra. La. J and Mfl A description of the first functional abstraction in the ontological model is the following

ELECTRICAL

)lE C HANI C AL

fAt;L T

fACLI

--,

, POWER I~

TRASSFOR~ATiONS i

THE

SYSTEM

DE TE CTI O";

OF

~

FA t: L T S

IS O LA T : O:O-:

OF

~

F Al:L TS

P .\.JtITY SPA C E

: PARITY SPACE FDI SCHErvtE SIGNIFICANT GOAL : Fault Detection Isolation (FUNCTION) in components. CONDITION I Sensor fault. CONDITION 2 Excitation by the input is not necessary.

D ('

\1 0TO R

Fig. 9. Fault Detection Transformations

445

Isolation

in

Power

Now, the functional abstraction provides "detection" in seven faults (Ft la, Ua. Ra. Mfl, Ml and w) and isolation in five faults (flux. la, Ua- Ra. Mfl-Ml and

motor. electrical and mechanical faults represent faults in power transformations, which is the major and natural goal of a motor.

w) .

Future research activity will focus on refinement and e;...1ension of the conceptual approach. For example, where several FDI schemes ( Parity Space, Observers and Parameter Estimation), can be combined.

The above example shows the usefulness of the proposed approach to diagnosis when Fault Detection Isolation is classified in an Ontological Framework. The use of different functional abstractions can produce a solution to problem of interaction between sensor (additive) and parameter (multiplicative) faults. This has been shown in the example.

In addition, some complementary research directions will be e;"'l'lored. which include among others : - Feedback Systems. - Multi-Input Multi-Output (MIMO) Systems.

The second functional abstraction has certain advantages concerning the fault detection and isolation in a d. c. motor. But no claim is made for completeness. So the proposed approach needs make total fault isolation Ua-Ra. MI-Mfl and fault detection isolation in La-1. Fig. 10 shows a possible solution, which use a Parameter Estimation scheme in parallel with the Parity Space one. FI

I;>

{j";>

f f t

;>;OFAt:LT

Ra La

t t

Mn

J

COl'iDITlON I I

EUcrRJC.u. FAL"LTS

t

t

/\ :!

!I

t t

FAULTS

PARAMETER ESTIMATION !

Abu-Hanna A. , R. Benjamins and W. Jansweijer (1991) . Device Understanding and Modeling for Diagnosis. IEEE E,(PERT. Vol. 6. No. 2. pp. 26-32. Chitaro L.. G. Guida., C. Tasso and E. Toppano (1993) . Functional and Teleological Knowledge in the Multimodeling Approach for Reasoning About Physical Systems : A Case Study in Diagnosis. IEEE Transactions on Systems, Man, and (vbernetics. Vol. 23, No. 6, pp. 1718-1751. Gertler 1.1. (1991). Analytical Redundancy Methods in Fault Detection and Isolation : Survey and Synthesis. IFAC SAFE PROCESS '91, BadenBaden Germany. September 10-13. Vol. 1 pp. 9-21. Hautamaki A (1992). A Conceptual Space Approach to Semantic Networks. Computer .".lath. Aplic. Vol. 23. No. 6-9. pp. 517-525. Hofling T. and T. Pfeufer (1994). Detection of Additive and Multiplicative Faults - Parity Space vs. Parameter Estimation. IFAC SAFE PROCESS '94. Espoo Finland. June 13-16. Vot. 2 pp. 539-544. Isermann R. (1994). Integration of Fault Detection and Diagnosis Methods. IF.4C SAFE PROCESS '94. Espoo Finland. June 13-16. Vot. 2 pp. 597-612. Jansweijer W . A. Abu-Hanna and R. Benjamins (1993). Configuration of Diagnostics Applications. TOOLDJAG'93 International Conference on Fault Diagnosis. Toulouse France. April 5-7. Vol. 2 pp. 531-540. Ramoni M. , M Stefanelli. L. Magnani and G. Barosi (1992). An Epistemological Framework for Medical Knowledge-Based Systems. IEEE Transactions on .~vstems, Afan, and (vbernetics. Vol. 22. No. 6. pp. 1361-1375.

w

MI:.CHk"lfiCAL

I

!\.

MI

6. REFERENCES.

zs

!

I!

I

D.e. !i

MOTOR:

I

I

, '

:

;

:i

, I

'-- -'

iI ,i ! i

PARITY

SPACE

Fig. 10. Electrical and Mechanical Faults with Parity Space and Parameter Estimation.

5. CONCLUSION.

In this paper. a Fault Detection and Isolation Ontological Model has been developed. The approach used to realize this model is based on Ontology and functional abstractions. With this approach. faults can be described in a more natural way. In fact. functional abstractions in a Ontological Model of the FDI allows to look the model knowledge and the diagnostic knowledge from another perspective. in order to get an appropiate fault abstraction. The advantage of this approach is emphasized throught the comparison of two functional abstractions with Parity Space. The approach proposed shows t\....o major results. First. functional abstractions may be useful in differentiating the effects between additive and multiplicative faults. Second. a functional abstraction let to find a more natural fault representation. For example. in a d.c.

446