Nonlinear ARX (NARX) based identification and fault detec-

tion in a 2 DOF system with cubic stiffness

J.S. Sakellariou and S.D. Fassois

Stochastic Mechanical Systems (SMS) Group
Department of Mechanical & Aeronautical Engineering


e-mail: sakj,fassois @mech.upatras.gr

web page: http://www.mech.upatras.gr/ sms

Abstract
This paper addresses the problem of system identification and fault detection in a two DOF nonlinear system
characterized by cubic stiffness. System identification is based upon Nonlinear ARX (NARX) models, while
a novel Functional Model Based Method is employed, for the first time within the context of a nonlinear
system, for tackling the combined problem of fault detection, identification (localization), and fault magnitude
estimation. The Functional Model Based Method utilizes Functional NARX (FNARX) models, which are
capable of accurately representing the system in a faulty state for the latter’s continuum of fault magnitudes, as
well as statistical decision theory tools. The results of the study indicate the effectiveness of both NARX based
identification and the Functional Model Based Method in detecting, identifying, and estimating the magnitude
of faults based upon only two measured signals.

    !
1 Introduction

 ( ) 2 3 / 4 0 5 6 78 9 * +- ,  . / 0 1 / 0
This paper is concerned with the problems of system
identification and fault detection, identification (lo- " #%$'&
calization), and magnitude estimation in a two DOF 
 

(Degree-of-Freedom) system characterized by local

polynomial nonlinearity (cubic stiffness ; figure 1).
The approach postulated is based upon discrete-
time NARX models [1, 2, 3], that is Nonlinear
AutoRegressive models with eXogenous excitation,
which constitute nonlinear extensions of the conven-
tional linear ARX models [4, 5]. Unlike alternative
nonlinear representations, such as those based upon
Volterra or Wiener series [1, 6, 7], describing func-
tions [1, 8], or neural networks [1], which are often
used in identification, NARX models offer a number
of advantages, including accuracy and compactness
of representation (the latter leading to improved sta-
tistical parsimony), physical significance, and direct
correspondence between the NARX and the physical
system parameters. NARX models also feature linear
regression based estimation and the availability of a
number of tools for model structure selection [9].
Once the feasibility and effectiveness of NARX
based system identification is demonstrated, the com-
bined problem of fault detection, identificaton (lo-
calization), and magnitude estimation is tackled via
a Functional Model Based Method. This method,
International Conference on Noise and Vibration Engineering
September 16-18, 2002 - Leuven, Belgium
 
Figure 1: Two DOF system with cubic stiffness.
which has been recently introduced by the authors
[10, 11], is used for the first time within the context of
a nonlinear system, cultivating upon the NARX based
system representation. The Functional Model Based
Method achieves fault detection, identification, and
magnitude estimation in a unified way, based upon
the novel class of stochastic functional models and
statistical decision theory tools. The stochastic func-
tional models, presently Functional NARX (FNARX)
models, play a very central role, as they are capable
of accurately representing the system in a faulty state
for the latter’s continuum of fault magnitudes.
The rest of this paper is organized as follows: The
system and the considered faults are described in sec-
tion 2, while NARX based identification is presented
in section 3. The Functional Model Based Method
 G =xT-O U>V V V O U W
:<;>=@? AB:DC-=FE E are considered; negative/positive
G ; =@? A G C =IH A G J I= HAA (kg) values indicate stiffening/loosening, respectively).
K ;>=ILME ANK J =IO A A (kN/m)
PRQ Q G X ;
The second mode corresponds to stiffness changes
G =@O A W
(Ns/m) in . Each individual fault is similarly represented as
. A single fault mangitude of (stiffness
Table 1: Physical system parameters.
reduction) is in this case considered.

Fault Description
PRQ Q S
mode
G C stiffness changes
3 NARX based identification

PRQ Q X G = T-O U>V V V O U W )

( @
The manipulation of the discretized system equations
G ; stiffness changes [y z {
of motion leads to the following relationship between
G
(reduction value =IO A W )
the force excitation
1 j'y z {
and the obtained displace-
ment response :
Table 2: The considered fault modes.
j'y z {a=} |`~ faY ‚]f y z {M\} |`ƒ faY ‚ f y z {N=a†
f €];' f €';'„ |`~ …
for fault detection, identification, and magnitude es-
timation is presented in section 4, and corresponding
=a†‡j'y z {=‰ˆ-Š>y z {nY ‹ (2)
results are summarized in section 5. The conclusions
where:
ˆpy z {]=
Œ y ‚`; y z { V V V ‚ y z { { ŠŽ 
of this study are summarized in section 6.
; ’ (3)
|`~ …a|`ƒ |`~ …g|`ƒ  ‘
2 The system and the faults
‹I=” Œ “ ; V V V ... ; V V V Š Ž  ; ’ (4)
  |>~ „ „ |`ƒ •
The two degree-of-freedom system considered is
GJ and z designates normalized discrete time (z‰=
|`~ …g|`ƒ  ‘
characterized by cubic stiffness in spring (figure
? h Unh Onh V V V ) with absolute time being ^ z]Te? _Y kal .
The regressors ‚ f y z {-^ i =‰? h V V V h : \@: _ are
1). The rest of the system elements are linear. The
system’s physical parameters are indicated in table 1.
The system dynamics are described by the differ- monomials of degrees – =—? h Unh O , as indicated in ta-
 „
and uncrosscorrelated with the excitation [y z { , noise
ential equations:
P
ble 3. With the addition of a zero mean, uncorrelated,
: ; Y ['Z ;]\I^ G ;]\ G C _ Y [';`T G CMY [aCJ \bK ; Y>['c ;-=
< term tMy z { in equation 2, the excitation-response rela-
:DCMY [aZ C>T G CMY [';]\ G CMY [aC]\ G J Y [ C \bK J Y ac[ Cd=eA (1) tionship assumes the NARX(˜ ,˜ ) (Nonlinear Au-
P  „
toRegressive with eXogenous excitation) form [1, 2]:
:<; , and [gf hRgc[ f h`[gZ f the i -th mass displacement, veloc-
with designating a force externally applied on mass
j'y z {=™ˆ Š y z {nY ‹š\btMy z { (5)

In this form ‹ represents the model parameter vec-

ity, and acceleration, respectively.
P
tor, with f ^ i =”? h V V V h : _ designating the i -th AR
System identification and fault detection are both
based upon measurement of the force excitation
[  
(AutoRegressive) parameter and f-^ i =›? h V V V h : _
[C
(subsequently designated as ) and the vibration dis-
the i -th X (eXogenous) parameter, and ˜
„ =„œ ,
j
placement response (subsequently designated as
˜ „ =xœ , the AR and X orders, respectively. The 
).

is ž =”œ , while the model includes linear, quadratic,

System simulation is based upon discretization of excitation-response delay (see the X term in table 3)

k'h E l A =I mnV m m mpo<? AMqnr>s t K

equations (1) via forward differencing [12, pp.13-22]

gree being max – =IO (see table 3).

and cubic terms, with the maximum nonlinearity de-
with time step
quency ul @
= ? ). A D
v w (sampling fre-

System identification. Identification of the

˜ ˜„
The faults. Two types of faults (fault modes) are

P QQ S G C
considered (table 2 and figure 1): The first mode healthy system is based upon the NARX( , )
corresponds to stiffness changes in . Each indi- model of the form of table 3. Estimation is accom-

GC
vidual fault is represented as , with the super- plished via minimization of the quadratic criterion

G
script indicating the fault mode and the subscript
the exact fault magnitude (changes in the range of
1
Lower case/capital bold face symbols designate vec-
tor/matrix quantities, respectively.
 Degree Ÿ' @¡ Ÿ' I¢ Ÿ' I£
Par. Monomial Par. Monomial Par. Monomial
AR term ¤n¥§¦`¥ ¨ © ªa e«'¨ ©]¬e¡ ª¤M­®¦­ ¨ © ªa e«'¨ ©]¬¯£ ªg° «'¨ ©¬<± ª²¤M³®¦³ ¨ © ªa e«'¨ ©¬š¢ ªg° «M´ ¨ ©]¬š± ª
” ” ¤ ´ ¦ ´ ¨ © ªa e«'¨ ©]¬¯¢ ª¤Mµ®¦µ ¨ © ªa e« ´ ¨ ©¬<± ª ¤ ¶·¦]¶M¨ © ªa e«'¸ ¨ ©¬š£ ªg° « ´ ¨ ©]¬š± ª
” ” ¤ ¸·¦]¸M¨ © ªa e«'¨ ©]¬¯£ ª ¤M¹®¦¹ ¨ © ªa e« ¨ ©¬<± ª
” ” ¤ º·¦]ºM¨ © ªa e«'¨ ©]¬š± ª
X term » ¥¼¦­ ¨ © ªa e½¨ ©]¬š± ª
AR order: ¾'¤¿ e± X order: ¾]»- e± Delay: Àp e±
AR terms: ÁD¤¿ I X terms: ÁD»- @¡

Table 3: The NARX(¾'¤gà ¾'» ) model structure.

Ä Å`Æ'ÇgÈ   ÅD¥ É ÅÊ Ë ¥gÌ ´ ¨ © ª , with Í designating the 0.3

length of the excitation-response signals used. Owing

to the linear dependence of the error term upon Ì ¨© ª 0.2

Î
the parameter vector , this leads to the linear regres- 0.1
sion estimator:
Displacemet (m)

Å ¥ Å 0

Ï Ñ Ñ
Îe ”Ð Í ¡ Ê Ë ¨ © ªg° ¨ © ª ÔDÕ °nÐ Í ¡ Ê Ë ¨ © ªg° «'¨ © ª Ô
¥'Ò -Ò Ó ¥'Ò
−0.1

Å (6) −0.2

ÏÖ × ´   ¡ Ñ Ê Ë
Í ÌÏ ´ ¨ © ª (7) −0.3

¥ −0.4

Identification is presently based upon ÍØ  0 2 4 6 8 10

Time (sec)
12 14 16 18 20

£ Ùnà ٠٠Ù>Ú ¤ ÁpÛgŸ Ì ¢ Ù>Ú Ì Ü

( ) long excitation and response
signals. A typical vibration response signal is de- Figure 2: System vibration displacement response.
picted in figure 2, whereas the corresponding iden- 0.2
tification result (model-based one-step-ahead predic-
Displacement (m)

0.1
tions and prediction errors) is, for a segment of that 0
signal, presented in figure 3. As it may be readily −0.1
observed, the model-based predictions practically co- −0.2
incide with the system response and the prediction er- −0.3
(a)
rors are very small.
−7
x 10

1
Residuals (m)

4 The fault detection and identi-

0
fication method
−1
The Functional Model Based Method consists of two (b)
2 2.5 3 3.5 4
phases (also see [10, 11]): The first (a-priori) phase Time (sec)

includes the baseline modeling (via identification) of

the healthy system’s dynamics, as well as the model-
ing of each fault mode, for its continuum of fault mag-
nitudes, via the novel class of stochastic functional
models.
The second (inspection) phase is performed pe-
riodically during the system’s service cycle, and in-
cludes the functions of fault detection, identification
(localization), and fault magnitude estimation.
Figure 3: (a) The actual (—) and predicted (- - -) dis-
placement responses; (b) the corresponding predic-
tion errors.
4.1 Baseline and fault mode modeling
(a-priori phase)
Baseline modeling. A single experiment is per-
formed, based upon which an interval estimate of
 ü
a NARX( Ý'Þgß Ý'à
) dynamical model, of the form of The FNARX model of equations (9)-(11), desig-
equation (5) representing the healthy system’s dy- nated as ø ýaú
, is thus parametrized in terms of the
namics, is obtained. parameter vector (to be estimated from the measured
Fault mode modeling. The notion of fault mode signals):
 .ì  !
 ì ø ânúé  ñ ý
refers to the union of faults of all possible magnitudes
(severities) originating from a single physical cause.
For the modeling of a fault mode, a series of á ý ñ ã Þ å  ... à å  ..
.  ..   ø ânú  ß" ßâ
experiments are performed (either physically or via
simulation). Each experiment is characterized by a The model of equation (9) may be then re-written
â as:
  
specific fault magnitude , with the complete series
ã â ä`å æaß â ä`ç è é ÷ õ ã ö aé ñ$# õ ã ö &é %' ø ânú () ý* õ ã ö aé ñ
+ õ ã ö é  ý õ ã ö é
covering the range of possible fault magnitudes, say
, via a discretization ê ânë ß âMì ß í í í ß âMîðï
(in the sequel it is tacitly assumed, without loss of (12)
with:
ã  ë ø ânúaí í í  ø ânúMé  -
generality, that the healthy system corresponds to
â<ñ”ò ' , 
 ø n
â ú ñ  /. ë 0 
). This procedure yields a series of excitation-
response signal pairs (each of length ): ó
ôaõ ã ö é ßM÷ õ ã ö éxø ö ñ@ù ß í í í ß ó<úûø â¿ñIânë ß âMì ß í í í ß âMîðú  (13)

1
ý ñ  Þ ë 2 ë í í í Þ `
ä ç 2  . à ë 2 ë í í í à ä 2   - 3 ä`ç 4gä 5 6 /. ë 0
..
(8)
Based upon these, a proper mathematical descrip-
and % designating Kronecker product [13, pp. 27-28].
(14)
tion of the fault mode may be constructed in the form
of a stochastic Functional Model (presently Func- For model parameter estimation, the FNARX
tional NARX – FNARX – model). A FNARX model, equation (12) gives, following substitution of the data
being a generalization of a NARX model (5), is of the [equation (8)] corresponding to a single fault magni-
ü â
789 ÷ õ ã ù é : ; 897 + õ ã ù é : ;
form: tude :

ø ýaú>þÿ÷ õ ã ö aé ñ
õ ã ö é  dø ânú  õ ã ö ™
é â

789  õ ã ù é : ;
  (9) ã . < ñ
..

..
ã <  ý= ã
..
< (15)
õ÷ óDé õ+ óðé
.
õ .óðé
dø ânú = Þ ë ø ânúaí í í Þ ä`ç ø ânú ... à ë ø ânú'í í í à ä ø ânú 
(10) ñ?>A@ õ ñCB õ  ý=ED õ (16)

   ë Þ å   ø ânú—à å ø âgúñ   ë à å   ø ânú

Þ å ø ânúñ Stacking together these expressions for the
data corresponding to the discrete fault magni-
ê ânë ß â ì ß í í í ß âMîðï
(11)
â
õ ãö é ãö é
In these expressions designates the fault magnitude, tudes considered in the experiments

ôaõ ã ö é ÷ õ ã ö é
yields:
,
is defined analogously to [equation (3)],
designate the corresponding measured ex- @šñCB ý=ED
(17)

 õ ãö é
citation and resulting response signals, respectively,
788 @ ë : ;; 788 B¿ë : ;; 788 Dgë : ;;
with:
and the corresponding stochastic model resid-
ual (one-step-ahead prediction error). For an accu-
89 @ ì ; 89 Bpì ; 89 Dnì ;
rate model, the residual sequence is zero-mean, un-
correlated, with variance  ì ø ânú
, and uncrosscorre- F@ ñ ... < L B ñ . <
.. *D ñ ... <
G - K @ H îI J - K G 3B H I î J G - K D H îI J
lated with the corresponding excitation. Residual se-

î . ë0 î . ä>ç 4aä 5 6  0 î . ë0
quences corresponding to different fault magnitudes
are assumed uncrosscorrelated.
As equation (11) indicates, the AR and X parame-
Þ å ø ânú à å ø âgú Parameter estimation (determination of the param-
ý

ters , are modeled as explicit functions of
â
the fault magnitude , belonging to a -dimensional eter vector ) may be then based upon the Ordinary

 ë ø ânú ß í í í ß   ø ânú
functional space spanned by the (mutually indepen- Least Squares (OLS) criterion:
K
M O î OP ãö
Þ å à å ã
dent) functions (functional basis).
The constants , designate the AR and X, respec- ñ Trace CN ov Dnéañ õ  ìõ é
tively, coefficients of projection. ë ë (18)
 Q RST
in which Cov designates sample covariance of the W m ul o n a u s c g e „ [ R i T&ˆ ‰ [ p q&r d e€d ™ S V l n j ]^ (23)
VXU W$Y Z [)Z,\]?^ S Z_[`W
indicated vector. This leads to the estimators:
f ^– — q›ššš ˜
d d 
] ^ d d „ — š Vl
with R i T defined by equation (3) and designating
W$a c d b e c g e R i T [ R i T j S a=c d b e c g e R i T k R i T j (19) the chosen fault mode’s vector of coefficients of pro-
^*f ^h h ^Xf ^h Since the healthy system corresponds to q
jection [of the form of equation (14)].
Wœ ,
d
m?l o/n p q&r Wtu s c g e v l n R i T for q W q ^ w x x x w q (20) fault ž Ÿ
detection may be based upon the hypothesis test-

f ^ b ing
ž : q/’ W
œ (No fault has occurred).
problem:
V l estimator is asymptotically (uzy|{ )
^ : q ’ W
  œ (A fault has occurred).
The

true parameter vector and covariance matrix }~ ,

Gaussian distributed with mean coinciding with the

based upon which interval estimates of the true pa-

following test at the ¡
WLœ x œ/¢ risk levelž Ÿ leads
which (based upon the previous results)
œ œ ¢ to the
( x prob-
rameter vector may be constructed [14].
ability of type I error, that is rejecting if it is cor-

ž=p Ÿ ¡ W
œ x œ/¢ r

rect):
4.2 Fault detection, localization, and es-
d
timation (inspection phase)
£ q l £/¤ s x ¥ ¦ m l W?§
Fault detection test
€R i T w kR i T i W s w &w x x x w u =
ž Ÿ is accepted
W?§ (no faultisisrejected
Let ( ) represent the excita-
detected).
tion and response signals, respectively, obtained from
Else
the system in its current (unknown) state.
(a fault is detected).
Fault detection. Fault detection may be based
upon the re-parametrized FNARX model of any fault Fault identification. Once fault occurrence

q m on
mode. Toward this end consider the re-parametrized has been detected, fault localization is based upon
(in terms of , , which are the parameters to be esti- the successive estimation and validation of the re-

qn
mated) FNARX model corresponding to the stiffness parametrized FNARX models [of the form of equa-
fault mode [notice that the basis functions and co- tion (21)] corresponding to the various fault modes.
efficients of projection are those of the chosen fault The procedure stops as soon as a particular model is
mode model; compare with equations (9) and (12)]:
‚ p q w mo n rƒ kR i T W
„ [ R i T&S … p q&r† v R i T W successfully validated; the corresponding fault mode
is then identified as current.
W$‡ „,[ R i T/ˆ‰ [ p qr Š S V † v R i T (21) Model validation may be based upon statistical
tests examining the hypothesis of excitation and resid-
ual sequence uncrosscorrelatedness, as well as resid-
The estimation of q , m o n based upon the current ex- ž=Ÿ
ual uncorrelatedness. The latter is presently examined
via the statistical hypothesis testing problem:
¨ ^ W ¨ n W S S S W ¨/© W
œ
citation and response signals is achieved via the non-
linear regression (Nonlinear Least Squares – NLS) es- ž :
(the fault mode is¤
timator (realized via golden search and parabolic in-
terpolation [15]): ^ /
¨ ª

W   œ 
« E
¤ ¬ r
identified as current).

d cg e v
: Some (1

q*l W‹ n R i T m l o n W u s c g e v l n R i T (22) « W s w &w x x x w ¬ r designates the residual

(the fault mode is not the current one).
arg ŒŽ 
in which ¨ ª p
f^ f^ «
series normalized autocorrelation at lag . It may be
uy‘{ ) Gaussian distributed, with mean
This estimator may be shown [14] to be asymptot- shown [16, p. 149] that the test statistic:
d l (underlying) d q value, say q/’ , and ­ W
u S p u †  r S c e © u¯¨ l ® ªn «
ically (

variance m n [ q
“$”•p q/’ w m n r ]. This may be in turn

equal to the true
ª ^
(24)

d m l n c g e v l nd e€d ™ ]?^
estimated as:
u designates the residual signal length (in
m?l n W u o a u s Ri T j W l
number of samples), ¨/ª the sample normalized resid-
in which
¬
f ^–/— ˜ L
q ual autocorrelation, and the maximum lag, follows
—
 5.847 −3.844
5 Fault detection and identifica-
5.846 −3.846
tion results

α3(k)
α2(k)

5.845 −3.848

5.844 −3.85
5.1 Baseline and fault mode modeling
5.843 −3.852

5.842 −3.854
(a-priori phase).
−40 −20 0 20 40 −40 −20 0 20 40

ÛÜÒ Û)Ý
k (%)
x 10
−11
k (%) Baseline modeling. The identification of the baseline
5
(healthy) system via a NARX( ) model of the
0.953
4
form of table 3 has been discussed in section 3.
0.952
α4(k)

b1(k)

± ² ² ³
Fault mode modeling. Fault mode modeling is
0.951
pursued only for the fault mode characterized by
°/· ±_² ² »
3
0.95

Þ Ê Âß
changes in the stiffness (notice that the fault
0.949 2
−40 −20 0 20 40 −40 −20 0 20 40 mode is not presently modeled). A total of
k (%) k (%)

° Ê ½¶ °·
experiments, one corresponding to the healthy sys-

± ² ² ³
Figure 4: Theoretical (- - -) and FNARX estimated tem ( variation in ) and the rest corre-

°
à$á Á,µ ¹/Ò µ ¹ ¶=â ã ° Êåä ¶

sponding to various fault magnitudes (faults with
° ±_² ² ³
(—) model parameters as functions of the fault mag-

¹&Ò ½ ½ ½
nitude (fault mode ). ; increment ), are carried

æ Ü ç=è?é ê æ
out. The signals obtained are, in all cases,
long.

ÛÜÒ ÛÝ ±_² ² ³

Test Incurred Fault The FNARX modeling procedure [14] leads to a
Case FNARX( ) fault mode model, with func-

è Ê ¹
tional basis consisting of the first two (0th and 1st
± ´ ² ³ (µ ¶
I No fault (healthy system)
II reduction in stiffness ) °/· degree, thus ) Chebyshev Type II polynomials
±_· ² ¸ ³ (¹ º ¶ °· [17]. The theoretical and FNARX-based estimates of
± ´ ² ¼ » (µ ½ ¶
III reduction in stiffness )
IV reduction in stiffness ) °/¾ °
certain of the model parameter trajectories (as func-
tions of the fault magnitude ) are compared in figure
Table 4: The four test cases. 4, from which excellent agreement is observed.

¿·
a chi-square ( ) distribution with degrees of ÀXÁCÂ 5.2 Fault detection and identification
(inspection phase).
Ã
freedom.

½
This leads to the test (at the risk level): Four test cases, as indicated in table 4, are presently
considered via Monte Carlo experiments ( runs per

Ä$Å ¿ ·¾ Æ&Ç&È É ÆFault

¾Ê?Ë Ì ¼ is accepted
identification test case).
Monte Carlo fault detection results are pictorially

Ì ¼ is rejected
(fault mode is current). presented in figure 5, fault identification (localization)
Else Ê?Ë results in figure 6, and a summary of the fault mag-
nitude estimation results (averages over runs per ½
(fault mode not current).

Ã Ê ½&Õ ½ Ö
case) is presented in table 5. In all statistical tests the
selected risk level is . Comments on each
Fault magnitude estimation. Once the current
test case follow.
fault mode has been determined, the interval esti-
° Ê ½
Test Case I (healthy system). In this case the fault
° Í Î Í ²· Î ·
mate of the fault magnitude is constructed based upon
½
magnitude interval estimate includes the value
Ï Ð °?Ò Ó Ô
Ñ Ä
Gaussianity and the , estimates [equations (22),
in each one of the runs [figure 5(a)], thus no fault
(23)] obtained from the model [equation
½
is (rightly) detected. In addition, the value of the
(21)] of the identified fault mode. Thus:
°
statistic is, for all runs, below the critical point [fig-

Ã Ê ½&Õ ½/Ö ): ½
ure 6(a)]. The excellent accuracy of the estimates is
×
Fault magnitude interval estimate (
°=Í Á Â Õ º Ø Î Í ² Ò °_Í Ù
Â Õ º Ø Î Í ² Ú
confirmed by the average (over the runs) point and

± ´² ³ µ ¶ °/·
standard deviation estimates presented in table 5.
Test Case II (fault – reduction in the

½
stiffness). This is a small magnitude fault, yet fault de-
tection is accurate in all runs [the fault magnitude
 0.2
Fault magn.
k (%)

(a)
−0.2

3.5
Fault magn.
k (%)

(b)
2.5

29.5
Fault magn.
k (%)

29

(c)
28.5
Fault magn.

0.6
k (%)

0.4
0.2
(d)
0
1 2 3 4 5 6 7 8 9 10
Monte Carlo experiment

ë î ì í
ë ï ì ð í ë_î ì ò ñ
Figure 5: Fault detection results: (a) Test case I (healthy system); (b) test case II (fault ); (c) test case III

óô¯õ/ö õ/÷
(fault ); (d) test case IV (fault ) [10 Monte Carlo runs per case; the solid horizontal lines designate true
fault magnitude, the circles corresponding point estimates, and the boxes interval estimates at the
level].

þý ì
Test Case Fault True Fault Average Point Average Standard
ø/ù,ú ûŽü øý î ï
ë òì í
Magnitude Estimate Deviation Estimate

ÿ /ö ÷  õ  7.20   õ ï
ë î ì í
I 0
7.12   õ ï
ë ï ì ð í
II 3 3.02
III 29 29.02 7.18   õ
Table 5: Fault magnitude estimation results (averages over 10 Monte Carlo runs).

øŽô
õ ë_ï ì ð í
–  û øï
ë ì ì í
interval estimates do not include the value; fig- Test Case III (fault reduction in the

õ õ
ure 5(b)]. The fault mode is also correctly iden- stiffness). This is a larger magnitude fault, the detec-
tified, as the value of the
statistic is, for all  runs, tion of which is also without problems in all  runs

ø øô$õ
below the critical point [figure 6(b)]. The excellent [the fault magnitude interval estimates do not include

õ
accuracy of the estimates is, once again, confirmed the value; figure 5(c)]. Fault mode identifica-
by the average (over the  runs) point and standard tion is also accurate, as the value of the
statistic is
deviation estimates presented in table 5. õ
below the critical point for all  runs [figure 6(c)].
 60
Q−statistic

40
20
(a)
0

60
Q−statistic

40
20
(b)
0

60
Q−statistic

40
20
(c)
0

2000
Q−statistic

1000

(d)
0
1 2 3 4 5 6 7 8 9 10
Monte Carlo experiment

Figure 6: Fault identification results:  statistic (bars) and the critical point (- - -) at the     level (the
fault mode   is identified as current if  is lower than the critical point). (a) Test case I (healthy system); (b)

Test case II (fault    ); (c) test case III (fault     ); (d) test case IV (fault     ) [10 Monte Carlo runs per case;
   ].

The accuracy of the ! estimates is similarly excellent 6 Conclusions

(table 5).
Test Case IV (fault     – "  # reduction in the
! $ stiffness). This is a somewhat different case, as the This paper was concerned with system identification
fault considered does not belong to the modeled   and fault detection in a two DOF nonlinear system
fault mode (for this reason fault magnitude estimation  characterized by cubic stiffness. System identifica-
is not addressed). tion was based upon Nonlinear ARX (NARX) mod-
Yet, the obtained fault detection results are very els, while a novel Functional Model Based Method,
good, as the fault magnitude interval estimates do not, using Functional NARX (FNARX) models, was, for
in all %  runs, include the !&
' value [figure 5(d)]. the first time, employed for fault detection, identifica-
Moreover, the fault mode identification results defi- tion, and fault magnitude estimation within the con-
nitely suggest that the present fault does not belong text of a nonlinear system.
to the    mode, as the value of the  statistic is far The results of the study confirmed: (a) The effec-
above the  critical point for all %  runs [figure 6(d)]. tiveness and accuracy of NARX based identification
for the system at hand; (b) the effectiveness and accu-
racy of the Functional Model Based Method for tack-
ling the combined problem of fault detection, identi-
fication, and fault magnitude estimation.
 The Functional Model Based Method was specif-
ically demonstrated to accurately detect, identify (lo-
calize), and estimate even small magnitude faults (a
fault as small as ( ) stiffness reduction was consid-
ered) in the presence of stochastic uncertainty and
only two measured signals.
Acknowledgements
The authors acknowledge the financial support of this
study in part by the General Secretariat for Research
and Technology – Greece and the European Social
Fund (PENED99 Project #580), and in part by the
European Commission (Growth Project GRD1-2000-
25261 – ADFCSII).
References
[1] K. Worden, G.R. Tomlinson, Nonlinearity in
Structural Dynamics: Detection, Identification
and Modeling, Institute of Physics Publishing,
2001.
[2] I.J. Leontaritis, S.A. Billings, Input-output para-
metric models for non-linear systems: parts I and
II, Intl. Journal of Control, Vol. 41(2), Taylor and
Francis Ltd (1985), pp. 303-344.
[3] P. Palumbo, L. Piroddi, Harmonic analysis of
non-linear structures by means of generalized
frequency response functions coupled with NARX
models, Mechanical Systems and Signal Process-
ing, Vol. 14(2), Academic Press (2000), pp. 243-
265.
[4] L. Ljung, System Identification: Theory for the
User, Second Edition, Prentice Hall PTR (1999).
[5] S.D. Fassois, Parametric Identification of Vibrat-
ing Structures, in Encyclopedia of Vibration, S.G.
Braun, D.J. Ewins, S.S. Rao, editors, Academic
Press (2001), pp. 673-685.
[6] A.A. Khan, N.S. Vyas, Non-linear parameter
estimation using Volterra and Wiener theories,
Journal of Sound and Vibration, Vol. 221(5), Aca-
demic Press (1999), pp. 805-821.
[7] A.A. Khan, N.S. Vyas, Nonlinear bearing stiff-
ness parameter estimation in flexible rotor-
bearing systems using Volterra and Wiener ap-
proach, Probabilistic Engineering Mechanics,
Vol. 16, Elsevier Science Ltd (1999), pp. 137-
157.
[8] Y. Benhafsi, J.E.T. Penny, M.I. Friswell, A pa-
rameter identification method for discrete non-
linear systems incorporating cubic stiffness ele-
ments, Intl. Journal of Analytical and Experimen-
tal Modal Analysis, Vol. 7(3) (1992), pp. 179-
195.
[9] S. Chen, S.A. Billings, W. Luo, Orthogonal least
squares methods and their application to non-
linear system identification, Intl. Journal of Con-
trol, Vol. 50, Taylor and Francis Ltd (1989), pp.
1873-1896.
[10] J.S. Sakellariou, D.D. Rizos, S.D. Fassois, Fault
detection and magnitude estimation for an air-
craft skeleton structure via a Functional Model
Based Method, First European Workshop on
Structural Health Monitoring, Paris, France,
(2002), in press.
[11] J.S. Sakellariou, K.A. Petsounis, S.D. Fassois,
On board fault detection and identification in
railway vehicle suspensions via a Functional
Model Based Method, Proceedings of ISMA
2002, Leuven, Belgium.
[12] A.J. Jerri, Linear Difference Equations With
Discrete Transform Methods, Kluwer Academic
Publishers (1996).
[13] J.R. Magnus, H. Neudecker, Matrix Differential
Calculus, John Wiley and Sons (1988).
[14] J.S. Sakellariou, S.D. Fassois, Stochastic dy-
namical functional models: properties and esti-
mation, under preparation for publication (2002).
[15] G.E. Forsythe, M.A. Malcolm, C.B. Moler,
Computer Methods for Mathematical Computa-
tions, Prentice-Hall (1976).
[16] W.W.S. Wei, Time Series Analysis: Univariate
and Multivariate Methods, Addison-Wesley Pub-
lishing Company (1990).
[17] M. Abramowitz, I.A. Stegun, Handbook of
Mathematical Functions, Dover (1970).