cita scopus nnn

7/27/2019 Cita Scopus Nnn

1/15

Control Engineering Practice 12 (2004) 11511165

Fault diagnosis for a turbine engine $

Yixin Diao a, *, Kevin M. Passino ba IBM Research, IBM T.J. Watson Research Center, 19 Skyline Drive, Hawthorne, NY 10532, USA

b Department of Electrical Engineering, Ohio State University, 2015 Neil Avenue, Columbus, OH 43210, USA

Received 16 February 2000; accepted 28 November 2003

Abstract

Fault detection and diagnosis for jet engines is complicated by the presence of engine-to-engine manufacturing differences andengine deterioration during normal operation, the complexity of an accurate engine model, and our inability to directly measurecertain engine variables. Here, we work with a sophisticated component level model (CLM) simulation of a turbine engine(the General Electric XTE46) that can simulate the effects of manufacturing and deterioration differences, in addition to a variety of faults. To develop a fault diagnosis system we begin by using the CLM to generate data that is used by a LevenbergMarquardtmethod to train a TakagiSugeno fuzzy system to represent the engine. The resulting nonlinear model provides a reasonablyaccurate representation of manufacturing differences, engine deterioration, and fault effects. We use multiple copies of thisnonlinear model, each representing a different fault, to generate error residuals by comparing them to the engine output. In fact, wemanage the composition of the set of models with a supervisor that ensures that the appropriate models are on-line, and processesthe error residuals to detect and identify faults. Robustness and fault sensitivity of the proposed approach are studied in the paperand the component model level simulation of the XTE46 engine is used to illustrate the effectiveness of the fault diagnosis scheme.r 2004 Elsevier Ltd. All rights reserved.

Keywords: Fault diagnosis; Fuzzy systems; Multiple models; Robustness analysis; Jet engines

1. Introduction

The last two decades saw continuous improvement incontrol techniques resulting from the spectacularprogresses in control theory and computer technology.Meanwhile, stimulated by the growing demand forimproving the reliability and performance of systems,many fault diagnosis methods, in particular, model-based analytical redundancy methods, have been devel-oped, which have the capability of detecting theoccurrence of faults and determining the types (orlocations) of the faults. Survey papers by Gertler (1988) ,

Frank (1990) and Patton (1997) present excellent over-views of advances in model-based fault diagnosismethods. Due to the complexity and importance of jetengines, the study of fault diagnosis for jet engines hasbecome a very active research topic for both theoreticaland practical reasons. In Merrill (1985) , Merrill,DeLaat, and Bruton (1988) , and Merrill, DeLaat, andAbdelwahab (1991) , the authors studied sensor failuredetection for jet engines using a Kalman lter with ageneralized likelihood ratio testing-based scheme. InDuyar and Merrill (1992) and Duyar, Eldem, Merrill,and Guo (1994) the authors derived linearized models of jet engine systems via the a-canonical form parameter-ization identication method and applied a parameterestimation approach in fault detection and isolation(FDI) for the space shuttle main engine. In Patton andChen (1992, 1997) and Patton, Chen, and Zhang (1997) ,the authors studied fault detection of jet engine sensorsystems using an eigenstructure assignment technique todesign observer-based residual generators, and they alsostudied its robustness. The model based FDI techniqueshave also been applied to many other complex applica-tions. In Menke and Maybeck (1995) , the authors

ARTICLE IN PRESS

$ This work was supported by the NASA Glenn Research Centerunder grant NAG3-2084. The authors would like to thank Dr. S.Adibhatla at General Electric Aircraft Engines for providing the CLMfor the engine, technical advice on how to use it, and feedback on thedevelopment of the modeling and control strategies. (Also, this workwas done while the rst author was at the Dept. of ElectricalEngineering, The ohio state University, 2015 Niel Ave., Columbus,OH 43210, USA)

*Corresponding author. Tel.: +1-914-784-7226; fax: +1-914-784-6040.

E-mail addresses: [email protected] (Y. Diao), [email protected] (K. M. Passino).

0967-0661/$- see front matter r 2004 Elsevier Ltd. All rights reserved.doi:10.1016/j.conengprac.2003.11.012


2/15

detected sensor and actuator faults of ight controlsystems using multiple model adaptive estimation. Themultiple model fault diagnosis approach were alsostudied in Yen and Ho (2001) ; Fu, Shen, Zhang ,and Liu (2001) ; Patton, Toribio, and Simani (2001) ;Boskovic, Li, and Mehra (2001) ; and Tu, Pattipati, Deb,

and Malepati (2003) . In Gertler, Costin, Fang, Hira ,Kowalczuk, and Luo (1993) , the authors used orthogo-nal parity equations for automotive engines. In Tylee(1983) , the authors detected instrument failures innuclear power plant instrumentation using Kalmanlters, and in Wang, Kropholler, and Daley (1993) ,the authors applied robust observer-based methods to adistillation column. In the application of FDI, processdisturbances, sensor noise, and model inaccuracy areinevitable. Since residuals are not only affected by faultsbut also by these factors, robustness is a very importantproblem in model-based FDI ( Chen & Patton, 1999 ). Toreduce the sensitivity of the residuals to modeling errors,many methods have been developed by decoupling theeffects of faults from those of modeling error, which arebased on unknown input observers ( Seliger & Frank,1991 ; Chen & Zhang, 1991 ; Hou & Muller, 1994 ),eigenstructure assignment ( Patton & Chen, 1991 ,1997a ), or orthogonal parity relations ( Gertler & Luo,1989 ; Gertler & Singer, 1990 ). Another class of robustFDI method is achieved by making the thresholdadaptive to the input ( Emani-Naeini, Athter, & Rock,1988 ; Frank & Ding, 1994, 1997 ).

Recently, the research on FDI has focused more onnonlinear system fault diagnosis. Some nonlinear

techniques such as nonlinear observer methods ( Frank,1994 ; Garcia & Frank, 1997 ) and nonlinear parityrelations ( Krishnaswami, Luh, & Rizzoni, 1995 ) havebeen applied. Moreover, articial intelligence hasreceived increasing attention and been applied in theeld of nonlinear FDI (e.g., see the discussion inAntsaklis & Passino, 1993 ). The articial intelligentmethods such as fuzzy systems, neural networks, andexpert systems have the potential to learn the plantmodel from inputoutput data or learn fault knowl-edge from past experience, and they can be used asfunction approximators to construct the analyticalmodel for residual generation, or as supervisory schemesto make the fault analysis decisions. The nonlinearmodeling ability of neural networks has been utilized fornonlinear FDI problems ( Watanabe, Matsuura, Abe ,Kubota, & Himmelblau, 1989 ; Naidu, Zariou, &McAvoy, 1990 ; Sorsa, Koivo, & Koivisto, 1991 ; Sorsa& Koivo, 1993 ; Maki & Loparo, 1997 ; Kavuri &Venkatasubramanian, 1994 ). In addition, the learningability has also been studied and successfully used innonlinear robust FDI ( Polycarpou & Vemuri, 1995 ;Polycarpou & Helmicki, 1995 ; Vemuri & Polycarpou,1997a, b ; Vemuri, Polycarpou, & Diakourtis, 1998 ;Demetriou & Polycarpou, 1998 ), where robustness, fault

sensitivity and stability conditions of the learningscheme are also studied. Meanwhile, expert systems,fuzzy logic and genetic algorithms have been used inmodel based FDI ( Lopez, Benkhedda, & Patton, 1997 ;Fussel, Balle, & Isermann, 1997 ; Passino & Antsaklis,1988 ; Laukonen & Passino, 1995 ; Laukonen, Passino,

Krishnaswami, Luh, & Rizzoni, 1995 ; Gremling &Passino, 1997 ).This paper presents a multiple-model-based fault

diagnosis method that utilizes fuzzy modeling and anexpert supervisory scheme. Note that most existingmulti-model fault diagnosis methods are based on linearmodels, while nonlinear multiple models are rarely used.This paper presents a FDI method using multiplenonlinear models and the robustness and fault sensitiv-ity of the proposed method are also studied. Theanalytical system models in the form of TakagiSugenofuzzy systems are developed using nonlinear systemidentication techniques (Section 2). In Section 3 a faultdiagnosis system is described with a bank of multiplemodels and a supervisory scheme determines the propermodel bank to generate residuals and analyzes theresiduals to detect and identify faults. Robustness andfault sensitivity of the proposed approach are alsostudied in this section. In Section 4, the proposedmethod is applied to a turbine engine (the GeneralElectric XTE46) and its effectiveness is demonstrated viacomponent level model (LH) simulation examples.

2. Model development using TakagiSugeno fuzzy

systems

2.1. The XTE46 turbine engine

The General Electric XTE46 engine, as shown inFig. 1 , is a simplied, unclassied version of the originalIHPTET engine ( Adibhatla & Lewis, 1997 ). To developa model-based fault diagnosis scheme an accuraterepresentation of the engine dynamics is desired. Thismodel can be used to generate residuals by comparingits output to the engine output. However, modeling aturbine engine is undoubtedly a very difcult problem inthe fact that the jet engine system has an iterativestructure, which means that the model cannot be writtendown in a differentialalgebraic equation form. For-tunately, a thermodynamic simulation package, thecomponent level engine cycle model (CLM) of theXTE46 engine, was provided by General Electric Air-craft Engines (GEAE). This is a sophisticated highlynonlinear dynamic model where each engine componentis simulated. The CLM executes one pass within thedigital controls sampling time and thermodynamicstates are assumed to be in equilibrium after each passthrough the simulation. The CLM is a low-frequencytransient turbofan engine simulation and volume

ARTICLE IN PRESS

Y. Diao, K. M. Passino / Control Engineering Practice 12 (2004) 115111651152


3/15

dynamics and airow storage affects, which are high-frequency phenomenon, are not included. The operatingcondition of the engine is dened by the altitude (ALT),the Mach number (XM), the difference of temperature(DTAMB), and the throttle setting represented bypower code (PC). The health of the engine is describedby 10 quality parameters which include the ows andefciencies of the fan, the compressor and turbines. Themodel has three state variables, including the fan rotorspeed (XNL), the core rotor speed (XNH), and thetemperature at combustor inlet (TMPC). There are sixactuators, but the major control variables are thecombustor fuel ow (WF36), the exhaust nozzle area(A8), and the variable area bypass injector area (A16).(Refer to Table 1 for list of acronyms.)

In the course of developing fault diagnosis schemes,the use of analytical redundancy implies that amathematical model of the system is used to describethe inherent relationship (or redundancy) containedamong the system inputs and outputs, which can be usedto generate the residuals for fault diagnosis. Theresulting approaches are usually referred to as analy-tical redundancy-based fault diagnosis or model-based methods. From the point of view of theoreticalstudies in fault diagnosis, however, the CLM is toocomplicated. Although the CLM does provide a driverto trim the model to specied operating conditions andgenerate linearized models, studies show that theaccuracy of the linear models is not adequate for ourpurposes (where we consider faults with signicantnonlinear effects). Here, we developed a nonlinearmodel with TakagiSugeno fuzzy systems using a systemidentication methodology that utilized nonlinear tran-sient data generated by the CLM. Note that othernonlinear modeling methods such as neural networkscan also be used for modeling the engine dynamics. Inthis paper we choose to use the TakagiSugeno fuzzysystems because of its clear model interpretation,constructed from nonlinearly interpolated linear models.Besides modeling the engine behaviors at each operating

node, we also use TakagiSugeno fuzzy systems toconstruct the global engine model by interpolatingamong local models.

2.2. The TakagiSugeno fuzzy system

Developing mathematical models for nonlinear sys-tems can be quite challenging. TakagiSugeno fuzzysystems are capable of serving as the analytical modelfor nonlinear systems due to the universal approxima-tion property, that is, any desired approximationaccuracy can be achieved by increasing the size of theapproximation structure and properly dening theparameters of the approximator ( Passino & Yurkovich,1998 ). A TakagiSugeno fuzzy system can be dened by

y F ts x ; y PRi 1 gi xmi xP

Ri 1 mi x

; 1

ARTICLE IN PRESS

Fig. 1. Schematic of the XTE46 turbine engine.

Table 1Table of acronyms

CLM Component level modelXNL Fan rotor speedXNH Core rotor speedTMPC Temperature at combustor inletALT AltitudeXM Mach numberPC Power codeWF36 Combustor fuel owA8 Exhaust nozzle areaA16 Variable area bypass injector areaZSW2 Fan owSEDM2 Fan efciencyZSW7D Compressor tip owSEDM7D Compressor tip efciencyZSW27 Compressor hub owSEDM27 Compressor hub efciencyZSW41 High pressure turbine owZSE41 High pressure turbine efciencyZSW49 Low pressure turbine owZSE49 Low pressure turbine efciency

Y. Diao, K. M. Passino / Control Engineering Practice 12 (2004) 11511165 1153


4/15

gi x a i ;0 a i ;1x1 ? ai ;nxn ; 2

mi x Yn

j 1

exp 12

x j ci j s i j !

20@ 1A; 3where y is the output of the fuzzy system, x x1 ; x2 ; y ; xn ? holds the n inputs, and i 1; 2; y ; Rrepresent R different rules. The shapes of the member-ship functions are chosen to be Gaussian, and center-average defuzzication and product are used for thepremise and implication in the structure of the fuzzysystem. The gi x; i 1; 2; y ; R are called consequentfunctions of the fuzzy system, where the ai ; j areconstants. The premise membership functions mi x areassumed to be well dened so that PRi 1 mi xa 0 for allx : The parameters that enter in a nonlinear fashion areci j and s

i j ; which are the centers and relative widths of the

membership functions, respectively, for the j th inputs

and the i th rules. Actually, the TakagiSugeno fuzzysystem, where the consequent parts are chosen to beafne functions, is a special case of functional fuzzysystems ( Passino & Yurkovich, 1998 ). If other func-tions such as polynomial or neural networks are used asconsequent functions, different kinds of functional fuzzysystems will be generated.

The tunable parameter vector y in (1) can becomposed of both premise membership function para-meters ( ci j and s

i j ) and consequent function parameters

a i ; j : This is referred to as nonlinear in the parametercase. A nonlinear in the parameter TakagiSugeno fuzzysystem can be tuned by a variety of gradient methodssuch as the steepest descent method and Levenberg Marquardt method. Alternatively, the parameter vectory can be composed of only the consequent functionparameters so that y is a linear function of y: To tune thefuzzy approximator for this linear in the parameter case,a linear least-squares method will normally be suitable.

2.3. Fuzzy modeling for the XTE46 engine

The CLM for the XTE46 aircraft engine is acomplicated multiple-input multiple-output (MIMO)nonlinear system (involving schedules, look-up tables,

and partial differential equations). We assume that itsfundamental dynamic characteristics, however, can berepresented by a single-input single-output (SISO)system in the form

x f x ; u; c; p Zx ; u; c; p; 4

y hx ; u; c; p; 5

where x XNL ; XNH ; TMPC ? is the state vector, u WF36 is the input variable, y XN2 is the output of theengine, c ALT ; XM ; DTAMB ; PC ? represents theoperating condition of the engine, p [ZSW2, SEDM2,ZSW7D, SEDM7D, ZSW27, SEDM27, ZSW41, ZSE41,

ZSW49, ZSE49] ? represents the quality parametervector, f denotes the unknown function representingthe nonlinear characteristics of the engine, h XNLbecause the output variable XN2 is the measurementof the state variable XNL, and Z representsmodel uncertainty caused by describing the MIMO

engine by a SISO system.The analytical model of the engine is developed in twosteps. Fuzzy identication is applied rst to generate (agrid of) node models specied by operating condi-tions and quality parameters. Afterwards, the globalmodel can be constructed by fuzzy interpolation onthese node models. (We use this two-step method ratherthan identifying the model directly from the datacollected from the whole space of operating conditionsand quality parameters in that it is practically impossibleto train an approximator using such a large amount of data, due to the limitations of computing resources.)Given a specic operating condition ci and xed valuesof quality parameters pi ; the node model for theXTE46 engine can be obtained using nonlinear identi-cation technique as

dXNLdotXNLdot k F 1ts dXNLXNL k ; dXNHXNH k ; WF36 k ; y

1ci ; pi ;

6

dXNHdotXNHdot k F 2ts dXNLXNL k ; dXNHXNH k ; WF36 k ; y

2ci ; pi ;

7

dXNLXNL k 1

dXNLXNL k T s

dXNLdotXNLdot k ; 8

dXNHXNH k 1 dXNHXNH k T s

dXNHdotXNHdot k ; 9

dXN2XN2 k 1 dXNLXNL k 1; 10where two multiple-input single-output (MISO) Takagi Sugeno fuzzy systems, F 1ts and F

2ts ; are specied with

corresponding parameters y1 and y2 ; respectively. Thevariables dXNLXNL k and dXNHXNH k denote the estimatedvalues of XNL k and XNH k ( dXNLXNL 0 XNL 0 anddXNHXNH 0 XNH 0 to let the fuzzy model have thesame initial values as the engine), and dXN2XN2 k is theestimated value of XN2 k :

dXNLdotXNLdot k and

dXNHdotXNHdot k

are the outputs of the fuzzy systems. T s is the sample

time, which is 0 :02 s: The fuzzy systems are trained usingengine data generated by the transient driver of theCLM simulator of the XTE46 engine. One thousandengine inputoutput data pairs are collected whichreect the transient performance of the engine for 20 s(sampled every 0 :02 s) at a specic node (of operatingconditions and quality parameters). For the k th,experimental data pair, the input data are the statevariables XNL k and XNH k ; and the input variableWF36 k : The output data are XNLdot k andXNHdot k ; which denote the derivatives of XNL k and XNH k ; respectively.

ARTICLE IN PRESS



5/15

The structure of fuzzy approximators is selected byhaving the premise input be dXNLXNL k only because mostnonlinearity can be captured by dXNLXNL k (and thisselection will result in an afne model, which can bebenecial to fault tolerant control using stable adaptivefuzzy/neural approach). The TakagiSugeno fuzzy

system can be written in the form of

F ts dXNLXNL k ; dXNHXNH k ; WF36 k ; y PR j 1 a j ;0 a j ;1 dXNLXNL k a j ;2 dXNHXNH k a j ;3WF36 k m j dXNLXNL k PR j 1 m j dXNLXNL k :Using trial and error, fuzzy models with R 3 ruleseach were selected. The models are tuned by theLevenbergMarquardt method. Note that since theengine models are identied off-line, the stability andaccuracy of the models can be validated prior to their

applications to fault diagnosis.Notice that we did not use TMPC as one of the modelinputs because it is not measurable and the importanceof TMPC to the model accuracy is insignicant, which isveried by an input selection method referred to asregressor analysis (where the regression models areconstructed and the regression coefcients are analyzedto determine the importance of each input), so that it isnot necessary to estimate TMPC either. Also note thatthe fuzzy models are running in an open-loopmanner, i.e., the outputs of the models will be feededback into the models as the inputs, and only the basicbehavior of the engine can be obtained due to thedrifting effects caused by the accumulation of approx-imation errors. We use this approach because sometimesthis analytical model is desired to run as the truthengine in the simulation study, where the CLM andthus the engine states XNL and XNH are not available.Certainly, when we utilize the model in the faultdiagnosis of the CLM, we can use the (measurable) realengine states to be the inputs of fuzzy systems andexpect an improvement in model accuracy.

By nonlinearly interpolating between a grid of nodemodels obtained above from nonlinear system identi-cation, the global model can then be constructed which

is in the form

dXNLdotXNLdot k PN i 1 F 1ts dXNLXNL k ; dXNHXNH k ; WF36 k ; y1ci ; pi mi z

PN i 1 mi z;

11

dXNHdotXNHdot k PN i 1 F 2ts dXNLXNL k ; dXNHXNH k ; WF36 k ; y2ci ; pi mi z

PN i 1 mi z;

12

mi z Y14

j 1

exp 12

z j ci j s i j !

2

0@ 1A; 13where i 1; 2; y ; N represent N different models andz c? ; p? ? is the premise input vector including 14variables.

We choose to focus on fault diagnosis systemdevelopment for the climb region (which is denedas ALT A 12 500 ; 17 500 ; XM A 0:6; 0:8 ; DTAMB A 35 ; 35 ; and PC A 45 ; 50 ). We partition each operatingcondition variable into three regions to dene our grid.

In this way, we have 34

81 models to describe thenonlinearity presented in the climb region. The values of quality parameters are composed of four parts: thenominal value, the initial engine variation due tomanufacturing differences, the quality parameter ad- justment resulted from engine deterioration, and thequality parameter change due to the faults. Note thatthe effects of engine deterioration and faults are largerthan the initial engine variation, and we will try tocapture the characteristics of these two factors and leavethe effect of initial engine variation to be modeluncertainty. By assuming that the engine deteriorationaffects 10 quality parameters in the same way (so thatwe may have three grids to represent no deterioration,half deterioration, and full deterioration, respectively)and considering four sizes of faults (no fault, small,medium, or large fault), we have 3 4 4 48 modelsto describe the nonlinearity presented in the qualityparameters (for simplicity, we only consider twopossible faults, fan fault and compressor hub fault,here.) In total, we have 81 48 3888 (nonlinear) nodemodels to describe the nonlinearity in the climb region.We need this level of complexity to obtain a reasonablyaccurate design model for the development of ourfault diagnosis scheme.

The general form of the model can be described asx f x ; u; c; p Zx ; u; c; p; 14

where

f x ; c; p PN i 1 % f x ; ci ; pi mi c; pPN i 1 mi c; p; 15

ARTICLE IN PRESS

% f x ; ci ; pi

PR j 1 a j ;0ci ; pi a j ;1ci ; pi x1 a j ;2ci ; pi x2 a j ;3ci ; pi u % m j x1

PR j 1 % m j x1

; 16



6/15

where u WF36 and x x1 ; x2? dXNLXNL ; dXNHXNH

?:

The value of c is the known operating condition vector, p is unknown quality parameter vector, and ci ; pi are tospecify nodes where we establish node models. Inaddition, f and

% f represent 2 1 function vectors. The

model parameters a j ;0 ; a j ;1 ; a j ;2 ; a j ;3 are 2 1 vectors of linear parameters of TakagiSugeno fuzzy systems, and

% m j x1 are membership functions describing the non-linearity with respect to x1 : They are identied using theLevenbergMarquardt method, separately for eachnode. The function mi c; p are membership functionsof fuzzy interpolation between different operatingconditions and quality parameters, whose parametersare also identied using the LevenbergMarquardtmethod.

The function Z represents model uncertainty causedby using nite size approximators to approximate theengine dynamics. Note that we represent the relation-ship between the model and model uncertainty to beadditive but the actual relationship may not be so. Thisis because no matter what kind of systems we considerand what kind of model is obtained, the modeluncertainty Z can always be represented as thedifference between the system dynamics F and themodel dynamics f ; i.e. Z F f :

The resulting nonlinear model provides a reasonablyaccurate representation of engine dynamics (GE Air-craft Engines veried this for us). Here we give anexample at a point different from the nodes where wegenerated the model. The engine operating conditions

are ALT 16 000 ; XM 0:75 ; DTAMB 0; andPC 46 : The quality parameters are dened byconsidering some initial engine variation, nearly half engine deterioration, and a fan fault a little bit largerthan medium size. Fig. 2 compares system responses

between the nonlinear model and the CLM andindicates their similarity (where the solid lines representthe system response of the CLM and the dashed linesrepresent that of the analytical model). We alsoconducted many other such simulations to verify thequality of our design model; however, in the interest of brevity we do not include those plots here. Compared tomodeling engine dynamics using local linear models(either through a CLM driver to trim the model tospecied operating conditions and generate linearizedmodels, or by using fuzzy models with R 1), usinglocal nonlinear models generally provides more accuratemodeling results, especially when the input variableWF36 has a large variation magnitude and the statevariables XNL and XNH are far away from theoperating point.

3. Fault diagnosis by interfacing multiple models with asupervisory scheme

3.1. Residual generation using multiple models

The engine simulation can represent many situations.For example, it can simulate what happens for several

ARTICLE IN PRESS

0 5 10 15 201.05

1.06

1.07

1.08

1.09

1.1

1.11x 10

4

X N L

0 5 10 15 20-300

-200

-100

0

100

200

300

400

X N L d o

t

0 5 10 15 201.39

1.4

1.41

1.42

1.43

1.44x 10

4

Time (sec.)

X N H

0 5 10 15 20-300

-200

-100

0

100

200

300

Time (sec.)

X N H d o

t

Fig. 2. Illustrative example of the model performance.



7/15

types of faults, possibly with manufacturing differencesand engine deterioration. Usually, engine faults result inchanges of the inputoutput characteristics of thesystem so that the faults may be detected by evaluatingthe residuals generated by comparing the model outputand the engine output. Some faults, however, may affect

all such residuals, so that if we generate residuals withonly a single model, we can only detect but not isolatesuch faults. In such situations, a multiple modelmethod may be preferable. In this approach we utilizedifferent models to identify different fault situations,and isolate faults by comparing the residuals corre-sponding different models. The structure of the faultdiagnosis scheme using multiple models and a super-visory expert system is shown in Fig. 3 .

Assume that there are N possible fault situations andwe have N 1 models f M i gN i 0 ; where the nominalmodel M 0 corresponds to the situation where there is nofault, and the fault models M i ; i 1; 2; y ; N representthe i th fault situations. These models can be obtained byusing nonlinear system identication with Takagi Sugeno fuzzy systems (as explained in the previoussection). Note that the faults affect the engine dynamicsthrough engine quality parameters, and the enginequality parameters are also affected by initial enginevariation (due to manufacturing differences) and enginedeterioration, so that the quality parameter vector p in f x ; u; c; p of (14) is actually unknown (and unmeasur-able as well). However, we can represent the enginedynamics of different fault situations as f x ; u; c; p0i with p0i indicating the expected quality parameters with

respect to the corresponding fault situation, andconsider the effects of initial engine variation and enginedeterioration as model uncertainty. By running a bankof models on-line (which are selected by the expertsupervisory scheme as we will explain later), the

residuals ri are generated by ri y #

yi ; where y is theengine output and

#

yi is the output of the model M i which corresponds to the i th fault situation. (Morediscussion on the forms of the system and models will begiven later in the section of robustness analysis.)

3.2. Expert supervisory scheme

The determination of the on-line model bank compo-sition and the evaluation of residual signals areconducted using an expert supervisory scheme. Theadvantage of using an expert system is that the heuristicknowledge of faults and our experience in handlingfaults can be easily incorporated into the expert systemin the form of rules, and the operation of the expertsystem is transparent so that we can investigate theresidual evaluation process of the expert system.Basically, the rule-based expert supervisory system of our fault diagnosis scheme performs the followingfunctions:

1. Determine the on-line models to generate theresiduals : Our experience of working with the engineprovide us some a priori knowledge of the faults, forinstance, the possible types of the faults and the possiblecombinations of these faults. According to this, N 1models have been established that correspond todifferent no-fault or fault situations. Besides, we knowthere exist certain relationships among these situations.For instance, suppose fault i is happening at present,then the next possible fault situation can be the existenceof fault i together with fault j (a fault that could occur in

the presence of fault i ), but not the situations wherethere exist three faults under the assumption that no twofaults can occur simultaneously, or no fault exists if weknow that the fault cannot be self-recovered. Therefore,at each time instant only a subset of N 1 situations

ARTICLE IN PRESS

Plant

M

M

M

FaultDetector

IsolatorFault

Multiple Model Bank

Model Bank Determinator

Supervisory Scheme

I n d e x

G e n e r a t o r

M i n i m u m

R e s

i d u a l

F a u

l t I n d e x

G e n e r a t o r

.

.

.

.

.

..

..

.u y

F

I

ry

+-

+

-

+-

*

*

0

1

N

r

r

0

1

N

y

y1^

N^

0

Fig. 3. Structure of the fault diagnosis scheme.



8/15

can happen and thus only a subset of models arerequired to be on-line to generate the residuals, whichcan be summarized as the rule: IF ( fault situation i iscurrently in existence ) THEN (a subset of models S i will be used as the on-line model bank to generate residualsafterwards) . In particular, the rules can include the

following:* IF (there is no fault in existence) THEN (the model

bank consists of no fault model and all possiblesingle-fault models).

* IF (there is a single fault in existence) THEN (allsingle-fault models except the one that has beenisolated will be removed from the model bank, andthe multiple-fault models including the isolated faultand another possible fault will be added).

* IF (there are two faults in existence) THEN (theon-line model bank includes the model correspondingto the situation with these two faults, and the modelscorresponding to the situations with three faults, i.e.,these two faults and another possible fault).

The advantage of using the fault knowledge to select theon-line model bank is to reduce the computationalcomplexity caused by using this multiple model faultdiagnosis approach.

2. Fault detection : Once the residuals are generated,the fault decision logic is used to diagnose the faults. Asequence of minimum residual indices is rst gener-ated by

I k arg mini A S i

r i k ; 17

where I k denotes the index of the model whoseresidual is the minimum, which corresponds to the mostappropriate model to represent the engine at timeinstant k : Note that the change of fault situationsresults in the change of the inputoutput characteristicsof the system. Therefore, the change of index, i.e., thechange of most suitable model, may be used to indicatethe change of fault situations, i.e., the occurrence of newfaults.

3. Fault isolation : The change of index can serve as afault detector, but not a fault isolator. Actually, the newindex may not indicate that the new fault is the onecorresponding to this index. This is because during thetransient phase the residuals corresponding to the on-line models may change drastically and some of themmay happen to be very small for a short time (andbecome large afterwards since they are not the onecorresponding to the new fault situation). In order toisolate the fault situations correctly, some logic rules areused to guarantee that a fault will be isolated only if itscorresponding index is the minimum index at least forT 0 seconds (to indicate the suitability of the model). Thetime delay term T 0 will be used to ensure robustness of fault diagnosis (which we will discuss later), and its

value can be obtained by trial and error to balance therobustness and sensitivity of fault isolation. The resultsof isolation is represented by a fault index

F k arg mini A S i ;k T 0p j p k

r i j ; 18

where F k denotes the index of the model whose

residual is minimum for a period of T 0 : Actually, thefault diagnosis strategy of using a time delay term mayaffect fault sensitivity to some extent but it is oftenuseful in isolating faults accurately (which is moreimportant), and it generally does not affect thesensitivity of fault detection. As a result, we will havea fault diagnosis scheme that can detect a fault relativelyfast, but isolate the fault type a little more slowly butaccurately, which is quite reasonable.

3.3. Robustness analysis

The robustness of fault diagnosis refers to the abilityto prevent false alarms in the presence of modelinguncertainties, i.e., if the system is in the i th faultsituation, the fault diagnosis system should indicate thei th fault situation rather than the j th fault situationwhere j a i : The robustness of the proposed faultdiagnosis system is achieved by using the time delayterm T 0 in the fault isolation scheme.

We study the nonlinear dynamic system in the i thfault situation that can be described by

x f x ; u; c; p0i Zi x ; u; c; pi ; 19

y Cx ; 20

where C c1 ; c2 ; y ; cn ; cl > 0; l 1; 2; y ; n and f f 1 ; f 2 ; y ; f n

? denotes the known model dynamics,which are characterized by the developed Takagi Sugeno model with nominal quality parameter vector p0i (corresponding to the i th fault situation but withoutconsidering initial engine variation and engine deteriora-tion). The functions Zi Zi ;1 ; Zi ;2 ; y ; Zi ;n

? representmodeling uncertainties in the system (e.g., caused byignoring initial engine variation and engine deteriora-tion, and by representing the system dynamics withnite size approximators).

The model corresponding to the j th fault situation canbe represented by

#

x j f #

x j ; u; c; p0 j ; 21

#

y j C #

x j ; 22

where #

x j is the estimated state vector of the j th modeland

#

y j is the estimated output.To study the robustness of the fault diagnosis scheme

for the above nonlinear system, we have the followingassumptions.

(A1) The modeling errors are additive, unstructured,and bounded, and the accumulation effect of

ARTICLE IN PRESS



9/15

modeling uncertainties can be described by

Z t0 Zi ;l x ; u; c; pi dt p Z0i ;l ; 23where l 1; 2; y ; n; and Z0i Z

0i ;1 ; Z

0i ;2 ; y ; Z

0i ;n

?

are constant bounds on modeling uncertainties.

(A2) The i th model can represent the system dynamicsin the i th fault situation accurately enoughcompared to the modeling uncertainties, i.e.,

Z t0 f l x ; u; c; p0i f l #x i ; u; c; p0i dt p % aZ0i ;l : 24(A3) The difference between the j th and i th model j a i

is large enough compared to the modelinguncertainties, i.e.,

maxt T 0p t0p t Z t00 f l x ; u; c; p0i f l #x j ; u; c; p0 j d t

X % aZ0i ;l : 25

The constants % a and % a are parameters that satisfy theinequality % aX

% a 2 (which will be used later forrobustness analysis). We use max to represent thatjR t00 f l x ; u; c; p0i f l #x j ; u; c; p0 j dt j may only be smallerthan % aZ0i ;l for no longer than T 0 seconds, which couldhappen when the estimated output of a wrong modelapproaches the engine output for a short period of timeand then leaves again. Note that the effects of the faultson the quality parameters are usually larger than thoseof initial engine variation and engine deterioration, andthe inaccuracy that arises from the accumulation of drifting errors is usually smaller than that frommodeling uncertainties; hence, the above assumptionscan be satised in real applications.

With the above assumptions the robustness problemof the fault diagnosis system is that in the i th faultsituation, the residual of the j th model r j y

#

y j ; j a i cannot be smaller than the residual of the i th modelr i y

#

yi for a period of T 0 time, i.e.,

maxt T 0p t0p t

jr i t0jp maxt T 0 p t0p t

jr j t0j; 26

so that the fault diagnosis system will still indicate i asthe current fault situation. The following analysisensures robustness of the proposed fault diagnosisscheme. Related analysis, but for a different problem,can be found in Vemuri and Polycarpou (1997a) .

For the i th model, dening the state estimation errorto be ei x

#

x i ; the error dynamics are described by

ei f x ; u; c; p0i f #

x i ; u; c; p0i Zi x ; u; c; pi ; 27

r i Ce i : 28

By processing the differential equation we obtain

jr i tj C Z t0 f x ; u; c; p0i f #x i ; u; c; p0i Zi x ; u; c; pi dt

p C Z t

0 f x ; u; c; p0i f

#

x i ; u; c; p0i dt

C Z t0 Zi x ; u; c; pi dtp C % aZ0i C Z

0i

% a 1C Z0i

and thus

maxt T 0 p t0p t

jr i t0jp % a 1C Z0i : 29

The error dynamics of the j th model j a i can be

described bye j f x ; u; c; p0i f

#

x j ; u; c; p0 j Zi x ; u; c; pi ; 30

r j Ce j : 31

From the differential equation we obtain

maxt T 0 p t0p t

jr j t0j maxt T 0p t0p t

C Z t00 f x ; u; c; p0i f

#

x j ; u; c; p0 j Zi x ; u; c; pi dt

X C maxt T 0p t0p t Z t0

0 f x ; u; c; p0i

f #

x j ; u; c; p0 j d t

C maxt T 0p t0p t Z t00 Zi x ; u; c; pi dt

X % aC Z0i C Z

0i

% a 1C Z0i :

Therefore, max t T 0 p t0p t jr i t0jp max t T 0 p t0p t jr j t

0j when

% aX % a 2; which means the robustness of the faultdiagnosis system is guaranteed and no false alarms willbe generated.

3.4. Fault sensitivity analysis

Fault sensitivity of fault diagnosis refers to the abilityto correctly determine the existence of a fault and thenisolate its type. In particular, if a fault happens at timeT f and thus the system changes from the i th faultsituation to the k th fault situation k a i at that time,the fault diagnosis system can isolate the k th situationafter an isolation time T i if for the k th model and

ARTICLE IN PRESS



10/15

tX T f T i T 0

Z t0 f l x ; u; c; p0k f l #xk ; u; c; p0k d t p % bZ0k ;l 32and for j th model ( j a k but j may be equal to i ) and

T f T i T 0p tp T f T i

Z t0 f l x ; u; c; p0k f l #x j ; u; c; p0 j dt X % bZ0k ;l ; 33where % b and % b are parameters that satisfy the inequality

% bX% b 2 (which will be used later for fault sensitivity

analysis). Under above conditions, using the triangleinequality the residual corresponding to the k th modelhas the following relation (when tX T f T i T 0):

jrk tj C

Z t

0

f x ; u; c; p0k f #

xk ; u; c; p0k

Zk x ; u; c; pk dt

p C Z t0 f x ; u; c; p0k f #xk ; u; c; p0k dt C Z t0 Zk x ; u; c; pk d t

p C % bZ0k C Z0k

% b 1C Z0k

and for T f T i T 0p tp T f T i we have

jr j tj C Z t0 f x ; u; c; p0k f #x j ; u; c; p0 j Zk x ; u; c; pk d t

X C Z t0 f x ; u; c; p0k f #x j ; u; c; p0 j dt C Z t0 Zk x ; u; c; pk dt

X % bC Z0k C Z

0k

% b 1C Z0

k :

Therefore, in the time interval [ T f T i T 0 ; T f T i ],we have jrk tjp jr j tj since % bX

% b 2; which implies thatthe fault isolation scheme can indicate the k th faultsituation at the time T f T i : (Related analysis, but for adifferent problem, is given in Vemuri and Polycarpou(1997a) .) Note that the above analysis on robustnessand fault sensitivity are obtained under worst-caseconditions, which are, in other words, sufcient butnot necessary conditions to guarantee the robustnessand fault sensitivity of the fault diagnosis system, so thatthe analysis are quite conservative.

4. Component level model simulation

To study the effectiveness of the proposed faultdiagnosis method we pick an engine in the climb region.Specically, the operating condition variables arechosen to be ALT 15 000 ; XM 0:7; DTAMB 0;

and PC 46: For the quality parameters of theengine, we set the initial engine variation to be piev 0:1% ; 0:1% ; 0:2% ; 0:1% ; 0:1% ; 0; 0:3% ; 0:3% ; 0:1% ;0:1% ; and the engine deterioration index to be 0.1. Forsimplicity, we only consider two fault types: fan faultand compressor hub fault, and each fault may havethree different levels: small, medium, and large. UsingTakagiSugeno fuzzy systems we generated 16 modelsincluding one nominal model (no fault), six single faultmodels (three models corresponding to small, medium,and large fan fault, respectively, and three models of small, medium, and large compressor hub fault), andnine double fault models (small fan fault with smallcompressor hub fault, small fan fault with mediumcompressor hub fault, etc.). The fault diagnosis resultsfor four different fault situations are given as below. Wealso conducted many other such simulations to verifythe effectiveness of our fault diagnosis scheme; however,in the interest of brevity we do not include them here.

4.1. No fault

First, we study the case where there is no fault. Theexpert supervisory scheme determines the initial bank of on-line multiple models to consist of six single-fault

models and one no-fault model. The residuals will begenerated by comparing the outputs of these sevenmodels and the engine output. Afterwards, the residualsare analyzed by the expert supervisory scheme todiagnose whether there are any faults in the engine.The output of the engine (XN2) and the estimatedoutputs (and residuals) of multiple models are shown inFig. 4 , where the wide solid lines represent the engineoutput, the thin solid lines correspond to the nominalmodel (model 0, or index 0), the thin dotted linesrepresent the small fan fault model (model 1, or index 1),the thin dashdotted lines represent the medium fanfault model (model 2, or index 2), the thin dashed linesrepresent the large fan fault model (model 3, or index 3),the wide dotted lines represent the small compressor hubfault model (model 4, or index 4), the wide dashdottedlines represent the medium compressor hub fault model(model 5, or index 5), and the wide dashed linesrepresent the large compressor hub fault model (model6, or index 6). Note that there is some oscillation inthe beginning of the simulation, which can be seen in theplot of minimum residual index. However, theoscillation has been attenuated by the isolation scheme,where the time delay term is designed to be T 0 1:2 s bytrial and error in order to let the fault diagnosis scheme

ARTICLE IN PRESS



11/15

be robust to the modeling uncertainties and sensitive tofaults. Thus, the fault index, as seen in Fig. 4 , isalways zero, which indicates that the nominal (no-fault)model always serves as the most suitable model and theengine is running in the no-fault situation.

From this case (no-fault, i 0) we can also study how

the assumptions for robustness analysis (A1, A2, andA3) are hold. From the Residual plot in Fig. 4 , wenotice that jr0tjp 9; i.e., we should be able to nd % aand Z00 such that % a 1Z

00X 9: For the residuals

generated from other models j 1; 2; y ; 6 we havemax t T 0 p t0p t jr j t

0jX 10; which implies that % a 1Z00p 10: We also know that the residual between thenominal (no-fault) model and the nominal engine (notshown in this gure) is smaller than 2.8, which means weshould have % aZ00X 2:8: Therefore, if we pick Z

00 6; % a

0:5; and % a 2:6; all the above inequalities will hold and

% aX% a 2 also holds. Studies of the validity of assump-

tions for other cases and for the assumptions of faultsensitivity analysis (32) and(33) are similar to the aboveanalysis.

4.2. Small fan fault

Next, we study the case where only the small fan faultoccurs in the take-off region and there is nocompressor hub fault in the climb region. The initialbank of on-line multiple models is chosen to becomposed of six single-fault models and one no-faultmodel. As shown in Fig. 5 , the fault is isolated at 2 :44 s:Note that the residuals corresponding to the small fan

fault model and the small compressor hub fault modelare interlaced, which imply that these two small faultshave some similar signatures. However, since the smallfan fault is isolated at 2 :44 s; the small compressor hubfault model will not be used after that to generate theresidual so that it will not cause any confusion. (Here,

we provide them just for the purpose of illustration.)Also note that the multiple-fault models are not usedafter the small fan fault was isolated. This is because wewant to keep this example simple and thus assume thatonly single fault can occur here. Studies of the multiple-fault cases will be given below.

4.3. Multiple faults: case 1

The effectiveness of the fault diagnosis scheme formultiple-fault situations is illustrated in the example,where there is no fault for the rst 5 s ; a smallcompressor hub fault occurs and lasts for 10 s ; and at15 s the size of the compressor hub fault changes fromsmall to large. Here, six single-fault models and one no-fault model are used as the initial bank of on-linemodels. As shown in the minimum residual index plotin Fig. 6 , a fault (index 1, small fan fault) is detected at5:3 s: However, this is just the transient phenomenonthat occurred after the change of the fault situation.After it lasts for 0 :86 s; the minimum residual indexchanges from 1 to 4, which indicates a small compressorhub fault. This index is the minimum for more than1:2 s: At 7 :38 s the small compressor hub fault is claimedto be isolated by the fault isolation scheme (using the

ARTICLE IN PRESS

0 5 10 15 201.055

1.06

1.065

1.07

1.075

1.08

1.085

1.09x 10

4

X N 2

0 5 10 15 200

20

40

60

80

100

R e s

i d u a

l

0 5 10 15 20

0

1

2

3

4

5

6

M

i n u m u m

R e s

i d u a

l I n d e x

Time (sec.)0 5 10 15 20

0

1

2

3

4

5

6

F a u

l t I n d e x

Time (sec.)

Fig. 4. No-fault situation.



12/15

time delay term T 0), as shown in the fault index plot.Afterwards, only three single-fault models (small,medium, and large compressor hub fault models) willbe used as on-line models to generate the residuals andat 15 :7 s a fault is declared to be detected, as shown in

the minimum residual index plot and at 17 :04 s thelarge compressor fault is isolated, as shown in the faultindex plot. Clearly, a fault can be detected shortly afterit occurs, and be isolated correctly a bit later to helpmaintain robustness.

ARTICLE IN PRESS

0 5 10 15 201.06

1.065

1.07

1.075

1.08

1.085

1.09x 10

4

X N 2

0 5 10 15 200

20

40

60

80

R e s

i d u a

l

0 5 10 15 20

0

1

2

3

4

5

6

M

i n u m u m

R e s

i d u a

l I n d e x

Time (sec.)0 5 10 15 20

0

1

2

3

4

5

6

F a u

l t I n d e x

Time (sec.)

Fig. 5. Small fan fault situation.

0 5 10 15 201.055

1.06

1.065

1.07

1.075

1.08

1.085

1.09x 10

4

X N 2

0 5 10 15 200

20

40

60

80

100

R

e s

i d u a

l

0 5 10 15 20

01

2

3

4

5

6

M i n i m u m

R e s

i d u a

l I n d e x

Time (sec.)0 5 10 15 20

01

2

3

4

5

6

F a u

l t I n d e x

Time (sec.)

Fig. 6. Small compressor hub fault followed by large compressor hub fault.



13/15

4.4. Multiple faults: case 2

Another multiple-fault case we have studied is shownin Fig. 7 , where the medium fan fault exists in thebeginning and the medium compressor hub fault occurs

at 10 s : The medium fan fault is isolated by the faultdiagnosis system at 2 :34 s: Note that there are signicantoscillations for the rst 2 s : After that, a bank of fourmodels (which correspond to medium fan fault only,medium fan fault with small compressor hub fault,medium fan fault with medium compressor hubfault, and medium fan fault with large compressor hubfault) are used to generate residuals. A fault is thendetected at 10 :3 s and the medium compressor hub faultis isolated at 12 :12 s:

4.5. Robustness and sensitivity issues

To obtain a sensitive fault detection scheme, anychanges in the minimum residual index (i.e., the changeof the most suitable model) will be considered asindicating the occurrence of new faults, and theminimum residual index is used to detect the faultpromptly. In our simulation examples, the detectiondelay time is less than 1 s : Furthermore, considering thepresence of modeling uncertainties, a time delay term T 0is included to achieve the robustness of the faultisolation scheme. In general, there is a trade-off betweenrobustness and sensitivity of fault isolation. Improvingthe robustness to modeling uncertainties may cause the

fault isolation scheme to be insensitive. Here, T 0 ischosen to be 1 :2 s via a priori knowledge on modelingerrors, and the isolation delay in our simulation is lessthan 2 :5 s:

5. Conclusion

A fault diagnosis method that interfaces fuzzy modelswith an expert supervisory scheme has been presented.This method can recognize the current fault situation of the plant by comparing a bank of nonlinear analyticalmodels formed by TakagiSugeno fuzzy systems. Underthe supervision of an expert system, the proper bank of models are applied to generate residuals which areanalyzed on-line to indicate the occurrence of faults andidentify them. Both robustness and fault sensitivity areanalyzed and the XTE46 engine simulation example isused to show the reliability of the method.

The effectiveness of the proposed fault diagnosismethod has been demonstrated via the CLM simulationof the XTE46 engine. Unlike the typical engine modelsthat are used in some of the literature, this XTE46simulator has been developed by GEAE to be verycomplicated and accurate (through thermodynamicsimulation for each engine component) so that thesimulation conducted on this simulator is very close tothat on the real engine for actual ights.

Nonlinear modeling is important to model-based faultdiagnosis for nonlinear systems. Having a hierarchical

ARTICLE IN PRESS

0 5 10 15 201.06

1.065

1.07

1.075

1.08

1.085

1.09

1.095x 10

4

X N 2

0 5 10 15 200

20

40

60

80

100

120

R e s

i d u a

l

0 5 10 15 20

0

2

4

6

8

M i n i m u m

R e s

i d u a

l I n d e x

Time (sec.)0 5 10 15 20

0

2

4

6

8

F a u

l t I n d e x

Time (sec.)

Fig. 7. Medium fan fault followed by medium compressor hub fault.



14/15

modeling structure and interpolating among nonlinearlocal models helps to model a complex system in amanageable way. A similar hierarchical structure canalso be used to organize the multiple models for faultdiagnosis so as to scale up for the cases where a largenumber of faults are involved.

Acknowledgements

We would like to thank S. Adibhatla at GeneralElectric Aircraft Engines for his assistance and adviceon many aspects of this project. Moreover, we wouldlike to thank T.H. Guo at NASA Glenn ResearchCenter for his support.

References

Adibhatla, S., & Lewis, T. (1997). Model-based intelligent digitalengine control. In AIAA-97-3192 , 33rd joint propulsion conference .

Antsaklis, P. J., & Passino, K. M. (Eds.), (1993). An introductionto intelligent and autonomous control . Norwell, MA: KluwerAcademic Publishers.

Boskovic, J., Li, S.-M., & Mehra, R. (2001). On-line failure detectionand identication (fdi) and adaptive recongurable control (arc) inaerospace applications. In Proceedings of the American control conference , Arlington, VA.

Chen, J., & Patton, R. J. (Eds.), (1999). Robust model-based faultdiagnosis for dynamic systems . Norwell, Massachusetts: KluwerAcademic Publishers.

Chen, J., & Zhang, H. (1991). Robust detection of faulty actuators via

unknown input observers. International Journal of Systems Science , 22(10), 18291839.

Demetriou, M., & Polycarpou, M. (1998). Incipient fault diagnosis of dynamical systems using on-line approximators. IEEE Transac-tions on Automatic Control , 43(11), 16121617.

Duyar, A., Eldem, V., Merrill, W. C., & Guo, T. (1994). Faultdetection and diagnosis in propulsion systemsa fault parameterestimation approach. Journal of Guidance, Control and Dynamics ,17 (1), 104108.

Duyar, A., & Merrill, W. C. (1992). Fault diagnosis for the spaceshuttle main engine. Journal of Guidance, Control and Dynamics ,15(2), 384389.

Emani-Naeini, A., Athter, M., & Rock, S. (1988). Effect of modeluncertainty on failure detection: The threshold selector. IEEE Transactions on Automatic Control , 33(12), 11061115.

Frank, P. (1994). Online fault detection in uncertain nonlinear systemsusing diagnostic observersa survey. International Journal of Systems Science , 25(12), 21292154.

Frank, P., & Ding, X. (1994). Frequency domain approach tooptimally robust residual generation and evaluation for modelbased fault diagnosis. Automatica , 30(4), 789804.

Frank, P., & Ding, X. (1997). Survey of robust residual generation andevaluation methods in observer-based fault detection systems.Journal of Process Control , 7 (6), 403424.

Frank, P. M. (1990). Fault diagnosis in dynamic systems usinganalytical and knowledge-based redundancya survey and somenew results. Automatica , 26 , 459474.

Fu, Q., Shen, Y., Zhang, J. Q., & Liu, S. (2001). An approach to faultdiagnosis based on a hierarchical information fusion scheme[and turbine application]. In Proceedings of the 18th IEEE

instrumentation and measurement technology conference , Budapest,Hungary (pp. 875878).

Fussel, D., Balle, P., & Isermann, R. (1997). Closed loop faultdiagnosis based on a nonlinear process model and automatic fuzzyrule generation. In IFAC fault detection , supervision and safety fortechnical processes , Kingston Upon Hull, UK (pp. 349354).

Garcia, E., & Frank, P. (1997). Deterministic nonlinear observer-based

approaches to fault diagnosis: A survey. Control EngineeringPractice , 5(5), 663670.Gertler, J. (1988). Survey of model-based failure detection and

isolation in complex plants. IEEE Control Systems Magazine ,8(6), 311.

Gertler, J., Costin, M., Fang, X., Hira, R., Kowalczuk, Z., & Luo, Q.(1993). Model-based on-board fault detection and diagnosis forautomotive engines. Control Engineering Practice , 1(1), 317.

Gertler, J., & Luo, Q. (1989). Robust isolable models for failurediagnosis. American Institute of Chemical Engineers Journal , 35(11),18561868.

Gertler, J., & Singer, D. (1990). A new structural framework for parityequation based failure detection and isolation. Automatica , 26 (2),381388.

Gremling, J., & Passino, K. M. (1997). Genetic adaptive failure

estimation. In Proceedings of the American control conference ,Albuquerque, NM (pp. 908912), June.

Hou, M., & Muller, P. (1994). Fault detection and isolation observers.International Journal of Control , 60(5), 827846.

Kavuri, S., & Venkatasubramanian, V. (1994). Neural networkdecomposition strategies for large scale fault diagnosis. Interna-tional Journal of Control , 59(3), 767792.

Krishnaswami, V., Luh, G., & Rizzoni, G. (1995). Nonlinear parityequation based residual generation for diagnosis of automotiveengine faults. Control Engineering Practice , 3(10), 13851392.

Laukonen, E. G., & Passino, K. M. (1995). Training fuzzy systems toperform estimation and identication. Engineering Applications of Articial Intelligence , 8(5), 499514.

Laukonen, E. G., Passino, K. M., Krishnaswami, V., Luh, G.-C., &Rizzoni, G. (1995). Fault detection and isolation for anexperimental internal combustion engine via fuzzy identication.IEEE Transactions on Control Systems Technology , 3, 347355September.

Lopez, C. J., Benkhedda, H., & Patton, R. J. (1997). Fuzzy observersfor FDI: Application to bilinear systems. In IFAC fault detection ,supervision and safety for technical processes , Kingston Upon Hull,UK (pp. 11471151).

Maki, Y., & Loparo, K. (1997). A neural network approach to faultdetection and diagnosis in industrial processes. IEEE Transactionson Control Systems Technology , 5(6), 529541.

Menke, T., & Maybeck, P. (1995). Sensor/actuator failure-detection inthe vista F-16 by multiple model adaptive estimation. IEEE Transactions on Aerospace an Electronic Systems , 31(4), 12181229.

Merrill, W. C. (1985). Sensor failure detection for jet engines using

analytical redundancy. Journal of Guidance ; Control and Dynamics ,8(6), 673682.Merrill, W. C., DeLaat, J., & Bruton, W. (1988). Advanced detection,

isolation, and accommodation of sensor failuresa real-timeevaluation. Journal of Guidance ; Control and Dynamics , 11(6),517526.

Merrill, W. C., DeLaat, J. C., & Abdelwahab, M. (1991). Turbofanengine demonstration of sensor failure-detection. Journal of Guidance, Control and Dynamics , 14(2), 337349.

Naidu, S., Zariou, E., & McAvoy, T. (1990). Use of neural networksfor failure detection in a control system. IEEE Control SystemsMagazine , 10, 4955.

Passino, K. M., & Antsaklis, P. J. (1988). Fault detection andidentication in an intelligent restructurable controller. Journal of Intelligent and Robotic Systems , 1, 145161.

ARTICLE IN PRESS



15/15

Passino, K. M., & Yurkovich, S. (Eds.), (1998). Fuzzy control . MenloPark, CA: Addison-Wesley Longman.

Patton, R., & Chen, J. (1991). Robust fault detection usingeigenstructure assignment: A tutorial consideration and somenew results. In Proceeding of the 30th IEEE conference on decisionand control , Brighton, UK (pp. 22422247).

Patton, R., & Chen, J. (1992). Robust fault detection of jet engine

sensor systems using eigenstructure assignment. Journal of Gui-dance, Control and Dynamics , 15(6), 14911497.Patton, R., & Chen, J. (1997). Modelling methods for improving

robustness in fault diagnosis of jet engine system. In Proceeding of the 31st IEEE conference on control and decision , Tucson, Arizona(pp. 23302335).

Patton, R., Chen, J., & Zhang, H. Y. (1997). Observer-based faultdetection and isolation: Robustness and applications. Control Engineering Practice , 5(5), 671682.

Patton, R., Toribio, C. L., & Simani, S. (2001). Robust fault diagnosisin a chemical process using multiple-model approach. In Proceed-ings of the 40th IEEE conference on decision and control , Orlando,FL, (pp. 149154).

Patton, R. J. (1997). Fault-tolerant control: The 1997 situation. InIFAC fault detection , supervision and safety for technical processes ,

Kingston Upon Hull, UK (pp. 10291051).Patton, R. J., & Chen, J. (1997a). Observer-based fault detection andisolation: Robustness and applications. Control EngineeringPractice , 5(5), 671682.

Polycarpou, M., & Helmicki, A. (1995). Automated fault detection andaccommodationa learning systems approach. IEEE Transactionson Systems, Man, and Cybernetics , 25(11), 14471458.

Polycarpou, M., & Vemuri, A. (1995). Learning methodology forfailure detection and accommodation. IEEE Control SystemsMagazine , 15(3), 1624.

Seliger, R., & Frank, P. M. (1991). Fault diagnosis by disturbancedecoupled nonlinear observers. In Proceedings of the 30th IEEE conference on decision and control , Brighton, UK (pp. 22482253).

Sorsa, T., & Koivo, H. (1993). Application of articial neural networksin process fault diagnosis. Automatica , 29(4), 843849.

Sorsa, T., Koivo, H., & Koivisto, H. (1991). Neural networks inprocess fault diagnosis. IEEE Transactions on Systems ; Man and Cybernetics , 21(4), 815825.

Tu, F., Pattipati, K., Deb, S., & Malepati, V. (2003). Computationallyefcient algorithms for multiple fault diagnosis in large graph-

based systems. IEEE Transactions on Systems ; Man and Cyber-netics , 33(1), 7385.Tylee, J. (1983). On-line failure detection in nuclear power plant

instrumentation. IEEE Transactions on Automatic Control , 28(3),406415.

Vemuri, A., & Polycarpou, M. (1997a). Robust nonlinear faultdiagnosis in inputoutput systems. International Journal of Control ,68(2), 343360.

Vemuri, A., & Polycarpou, M. (1997b). Neural network based robustfault diagnosis in robotic systems. IEEE Transactions on Neural Networks , 8(6), 14101420.

Vemuri, A., Polycarpou, M., & Diakourtis, S. (1998). Neuralnetwork based fault detection and accommodation in roboticmanipulators. IEEE Transactions on Robotics and Automation ,14(2), 342348.

Wang, H., Kropholler, H., & Daley, S. (1993). Robust observer basedFDI and its application to the monitoring of a distillation column.Transactions of the Institute of Measurement and Control , 15(5),221227.

Watanabe, K., Matsuura, I., Abe, M., Kubota, M., & Himmelblau, D.(1989). Incipient fault diagnosis of chemical processes via articialneural networks. American Institute of Chemical Engineers Journal ,35(11), 18031812.

Yen, G., & Ho, L.-W. (2001). On-line multiple-model based faultdiagnosis and accommodation. In Proceedings of the IEEE international symposium on intelligent control , Mexico City, Mexico(pp. 7378).

ARTICLE IN PRESS


cita scopus nnn

Documents