Vous êtes sur la page 1sur 15

Mechanical Systems and Signal Processing (1999) 13(6), 839 } 853

Article No. mssp.1999.1248, available online at http://www.idealibrary.com on

ON-LINE MODAL MONITORING OF AIRCRAFT


STRUCTURES UNDER UNKNOWN EXCITATIONs
M. ABDELGHANI
Irisa-Inria, Campus de Beaulieu, 35042 Rennes Cedex, France
M. GOURSAT
INRIA, Domaine de Voluceau-Rocquencourt, BP. 105, 78153 Le Chesnay Cedex, France.
Email: maurice.goursat@inria.fr
AND

T. BIOLCHINI
SOPEMEA, AeH rodrome de Villacoublay, Zone AeH ronautique Louis Breguet, BP. 48,
78142 VeH lizy Cedex, France. Email: sopemea@worldnet.fr

(Received 16 December 1998, revised 16 July 1999, accepted 23 July 1999)

In this paper, we present the application of recently proposed subspace-based damage


detection and isolation algorithms in on-line structural monitoring of an airplane structure
under unknown excitation. The "rst step of the approach is a subspace-based modal analysis
of the safe structure, using output-only measurements. The monitoring step consists in the
computation of a s2-variable, which re#ects possible deviations of newly collected output-
only data w.r.t. the reference (safe) modal behaviour. No re-identi"cation of the possibly
damaged structure is required. An on-line version of the detection algorithm is proposed.
Monitoring certain critical modes can also be achieved through sensitivity tests. Experi-
mental results for the proposed monitoring algorithms are reported, in the case of the Paris
MS760 airplane, on which mass as well as sti!ness changes are introduced. Moreover, the
results of the output-only subspace-based modal analysis are compared with the modal
appropriation method.
( 1999 Academic Press

1. INTRODUCTION
In many applications, the structural operating conditions may signi"cantly di!er from
those applied during laboratory tests where the structure is installed and properly
excited. For structures under their natural working conditions, excitations cannot be
measured, they are usually non-stationary and of complex nature. Examples are o!shore
structures subject to the turbulent action of the swell, which depends on wind and weather
conditions; aircraft structures subject to strength, varying with #ight conditions; bridges
subject to wind, tra$c, etc. The analysis of structures under such conditions is of major
importance.
The problem of fault detection and isolation (FDI) is a crucial issue. Moreover, in-
operation damage monitoring and predictive maintenance of complex mechanical struc-
tures is of key importance in many industrial areas, such as aerospace applications. Using
the available data (e.g. accelerometers), it is of interest to detect small changes in the natural

s This work has been carried out within the framework of Eureka project no. 1562 SINOPSYS (Model-based
Structural monitoring using in-operation system identi"cation) coordinated by LMS, Leuven, Belgium.

0888}3270/99/110839#15 $30.00/0 ( 1999 Academic Press


840 M. ABDELGHANI E¹ A¸.

frequencies and modeshapes of the structure. A relevant approach to such a problem has
been shown to be based on the modelling of modes and modal shapes through state space
representations [1], the use of output-only, and covariance-driven identi"cation methods
[2], and the computation of speci"c s2-tests based on the so-called instrumental statistics
[3, 4].
Recently, new subspace-based FDI algorithms based on this class of identi"cation
methods have been proposed as a solution to the above-mentioned vibration monitoring
problem [5]. The successful application of these algorithms in damage detection and
diagnosis of mechanical structures has been reported in recent papers [6, 7].
In this paper, we propose an on-line version of such algorithms for continuous monitor-
ing of an airplane structure, under unknown excitation. As a "rst step, a detailed modal
analysis of the airplane structure under its nominal condition is performed using the
balanced realization (BR) algorithm, and the results are compared with those of the modal
appropriation technique. This step is done only once. No re-identi"cation on the damaged
structure is needed. Two types of damage are introduced: a mass change, by adding two
masses of 20 kg each, placed symmetrically around the wings, and a sti!ness change
introduced by unblocking the ailerons. In order to simulate an on-line change, the data sets
of the nominal structure and the damaged one are concatenated into a single data set. Using
the reference (nominal) identi"ed modal signature and the new data set, a global test is
computed as a function of time that can be monitored for detecting the damage.
The outline of the paper is as follows. Section 2 introduces the modelling issues and
the covariance-driven subspace-based identi"cation algorithm. Section 3 summarises the
subspace-based damage detection and isolation methods, and introduces an on-line version
of the detection algorithm. The experimental set-up and results are described in Section 4.

2. SUBSPACE-BASED MODAL ANALYSIS


In this section, we brie#y describe the modelling and the output-only covariance-driven
subspace-based identi"cation algorithm we use.

2.1.1 MODELLING
Following [3, 1, 4], we assume that the behaviour of the mechanical structure can be
described by a stationary linear dynamical system, and that, in the frequency range of
interest, the input forces can be modelled as a stationary white noise. Consequently,
assuming a linear behaviour, the relevant model is the following matrix di!erential
equation:
MZ }Q (t)#KZ
}G (t)#CZ } (t)"l(t)
(1)
} (t)
>(t)"¸Z
where t denotes continuous time, M, C, K are the mass, damping and sti!ness matrices
respectively, (high-dimensional) vector Z } collects the displacements of the dofs of the
structure; the external (non-measured) forces l are modelled as a stationary white noise with
covariance matrix Q, measurements are collected in the (low-dimensional) vector >, and
matrix ¸ states where the sensors are located. The modal characteristics of the structure are
the vibration modes, or eigenfrequencies, denoted generically by k, and the modal shapes, or
eigenvectors, denoted generically by t . These quantities are solutions of the following
k
equation:
(Mk2#Ck#K)W "0, t "¸W , (2)
k k k
ON-LINE MODAL MONITORING OF AIRCRAFT STRUCTURES 841
Sampling model (1) at a rate 1/q yields the discrete-time model in state-space form
X "FX #e
k`1 k k`1 (3)
> "HX
k k
where

C D
Z
} (kq)
X" , > ">(kq).
k }Q (kt)
Z k

The state transition and observation matrices are

C D
0 I
F"eLq, L" , H"[¸ 0]
!M~1K !M~1C
and where state noise e is zero-mean, white, with covariance matrix Q. The modal
k`1
characteristics de"ned in equation (2) are equivalently found in the eigenstructure (j, u )
j
of F:
eqk"j, t "/ "Hu $%& .
k j j
Therefore, from now on we consider linear multivariable systems described by state-space
model (3), where state X and observed output > have dimensions m and r, respectively, with
r (much) smaller than m in practice. The state noise process (e ) is an unmeasured Gaussian
kk
noise sequence with zero mean and covariance matrix Q. We assume noise e to be
k
stationary, that is of constant covariance matrix; the issue of robustness with respect to
non-stationary excitation is addressed in [2, 8, 5].
The problem we consider is to monitor the observed system eigenstructure, which is
the collection of m pairs (j, / ), where j ranges over the set of eigenvalues of state transition
j
matrix F, / "Hu and u is the corresponding eigenvector. In what follows, we
j j j
assume that the system has no multiple eigenvalues, and thus that the j's and u 's
j
are pairwise complex conjugate. We stress that the collection of (j, / ) provides us
j
with a canonical parameterisation of the pole part of system (3). In particular, it does
not depend on the state-space basis. The set of the m modes (j, / ) is considered
j
as the system parameter h:

A B
$%& K
h" (4)
vec U.
where K is the vector whose elements are the eigenvalues j, U is the matrix whose columns
are the / 's, and vec is the column stacking operator. Parameter h has size (r#1)m. The
j
problem is to detect and isolate changes in parameter vector h. By isolation, we mean
deciding which mode(s) has (have) changed.

2.2. SUBSPACE-BASED IDENTIFICATION


For the identi"cation of a linear multi-variable system described by the state-space model
given in equation (3),s we consider a sequence of output covariances
$%&
R "E(> >T ) (5)
j k`j k

s Note that, in equation (3), a measurement noise with MA structure can be assumed as well as discussed in
[1, 5].
842 M. ABDELGHANI E¹ A¸.

For q*p#1, let H be the block-Hankel matrix:


p`1,q

A B
R R 2 R
0 1 q~1
R R 2 R
H " 1 2 q . (6)
p`1,q F F } F
R 2 2 R
p p`q~1
2.2.1. BR identixcation algorithm
It is well known [2, 9] that Hankel matrix H factorises as H"OC, where O is the
observability matrix:

AB
H
HF
O" .
F
HFp
From O, the pair (H, F) can easily be recovered, of which an eigenanalysis provides us
with the modes (j, u ). The actual implementation of what is called BR identi"cation
j
algorithm consists in substituting the empirical covariances

$%& 1 n~j
RK " + > >T (7)
j n!j k`j k
k/1
for R in equation (6), then computing the SVD of H ) , truncating it for getting an estimate
j
O) of O, from which (HK , FK ), and then (j) , u( ), can be deduced as above.
j
2.2.2. System parameter characterisation
For designing a damage detection algorithm corresponding to the BR identi"cation
method, we need the following characterisation of the canonical system parameter h.
Choosing the eigenvectors of F as a basis for the state space of model (3) yields, the
following particular representation of the observability matrix [3]:

AB
U
UD
O (h)" (8)
p`1 F
UDp
where diagonal matrix D is de"ned as D"diag(K), and K and U are as in equation (4).
It is known [10] that the following property characterises whether a nominal parameter
h agrees with a given output covariance sequence (R ) :
0 jj
O (h ) and H have the same left kernel spaces with co-rank m. (9)
p`1 0 p`1,q
Property (9) can be checked as follows:
1. From h as in equation (4), form O (h ).
0 p`1 0
2. Pick an orthonormal basis of the left kernel space of matrix O (h ), in terms of the
p`1 0
columns of some matrix S of co-rank m such that
STS"I (10)
s
STO (h )"0. (11)
p`1 0

s The left kernel space of matrix M is the kernel space of matrix MT.
ON-LINE MODAL MONITORING OF AIRCRAFT STRUCTURES 843
Matrix S has dimensions (p#1)r]s, where s"(p#1)r!m, and it is not unique; it
can be obtained, for example, by the SVD-factorisation of O (h ). Because of
p`1 0
equation (11), matrix S depends implicitly on parameter h . We bring the reader's
0
attention to the fact that, even though S is not unique, in what follows we treat it as
a function of parameter h, denoted by S(h), which is fully justi"ed in [5].
3. The parameter h which actually corresponds to the output covariance sequence (R )
0 jj
of model (3) is characterised by
ST(h )H "0, (12)
0 p`1,q
Property (12) plays an important role in the design of the subspace-based damage
detection tests,s as we explain now.

3. SUBSPACE-BASED DAMAGE DETECTION


We now brie#y recall the detection algorithm we are going to use in the present study.
The reader is referred to Basseville et al. [5] for more details.

3.1. GLOBAL DETECTION


Assume we have at hand a nominal model h , and newly collected data > ,2, > . We
0 1 n
perform steps 1 and 2 described above, and we replace step 3 by what follows. We de"ne the
residual vector
$%&
f (h ) "Jn vec(ST(h )H K ) (13)
n 0 0 p`1,q
where H is the empirical block-Hankel matrix obtained by substituting RK for R in
p`1,q j j
)
equation (6), and S(h ) is computed as in equations (10)}(11).
0
From equation (12) we know that statistics f (h ) has zero mean in the absence of change
n 0
[5, 11]. Based on a data sample, testing whether this hypothesis is valid requires the
knowledge of the statistical properties of f (h ), which unfortunately is unknown, in general.
n 0
One way of circumventing this di$culty is to use a local approach, that is to assume close
hypotheses:
B
(Safe)H : h"h and (Faulty) H : h"h # (14)
0 0 1 0 Jn

where vector B is unknown. It turns out [5, 12, 11] that residual f in equation (13) is
n
asymptotically Gaussian distributed, under both H and H , making its evaluation easy.
0 1
More precisely, this residual is distributed as

G
N(0, R(h )) under H
f (h ) , 0 0 (15)
n 0 N(M(h )B, R(h )) under H
0 0 1
where M(h ) and R(h ) are the Jacobian and covariance matrices, whose computation and
0 0
estimation are described in [5]. Let M K , R) be consistent estimates of M, R. Under an
additional full rank assumption of M [5, 11], the detection problem, namely deciding that
residual f is signi,cantly di!erent from zero, can be achieved through
n
$%& R) ~1MK (M
s2 "fT K TR) ~1M
K )~1MK TR) ~1f (16)
n n n

s A generalised version of equation (12), involving pre- and post-weighting of Hankel matrix as in [9], can be
exploited in the same manner [5].
844 M. ABDELGHANI E¹ A¸.

which should be compared to a threshold. Test statistics s2 is asymptotically distributed as


n
a s2-variable with rank(M); hence the selection of a threshold.
K and R) are computed using the nominal model
We stress that, in equation (16), matrices M
h and the nominal data set, whereas the residual f is computed using the nominal model
0 n
and the newly collected data.

3.2. MODEL DIAGNOSTICS


Now, we address the problem of deciding which (subsets of ) components of the parameter
vector have changed. This allows us to formulate the diagnostics of changes in terms of
natural frequencies, damping coe$cients and modal shape vectors, and to decide which
modes (#exure, torsional, etc.) are a!ected by the damage.
Let us partition the fault vector as

A B
B
B" a
B
b
where B , B have known dimensions l , l , respectively (l #l "l), and partition the
a b a b a b
Jacobian and Fisher information matrices accordingly:
M"(M M )
a b
and

A B
I I
I"MTR~1M" aa ab .
I I
ba bb
The problem is now to detect changes in B , B being unknown and considered as a
a b
nuisance parameter. Our solution consists in projecting the deviations in B onto the
subspace generated by the components to be isolated. This is known as a sensitivity test,
and is easily solved using a GLR test between B"(0, 0) and B"(B , 0), where B O0. The
a a
resulting sensitivity test t is written as [13}15]
a
t "fT I~1 f (17)
a a aa a
where
f "MTR~1f.
a a
Under B O0, t is distributed as a s2-random variable with l dofs; hence the selection of
a a a
a proper threshold.

3.3. ON-LINE VERSION


We now consider the case where it is desired to monitor permanently the system under its
operating conditions. It is assumed that the system (structure) is stationary during its
operation.
We suppose that a nominal data set corresponding to the nominal state of the system is
available, which is used to compute the nominal model h , as well as the corresponding
0
Jacobian and covariance matrices. This step is done once for all at the design stage.
A straightforward solution to on-line detection is then to compute recursively the new
residuals based on a sliding window, whose size is chosen large enough to capture small
changes in the model parameters. The corresponding s2-test for a s-size data window I
writes
t "fT R) ~1MK (M
K TRK ~1MK )~1M
K TR) ~1f (18)
I I I
ON-LINE MODAL MONITORING OF AIRCRAFT STRUCTURES 845
where
f "Js vec(ST(h )H ) (19)
I 0 p`1,q
)
is the residual computed with the data in that window. Under no-change condition, t is
I
distributed as a s2 random variable with 2m(r#1) dofs. Computation of this variable as
a function of time (sliding window), allows the continuous monitoring of the system. Any
deviation of its mean value from zero is an indication of damage.
Modal sensitivity tests can also be computed using a sliding window. This allows the
continuous monitoring of a subset of components of the parameter vector, for instance
certain critical modes.

4. EXPERIMENTAL RESULTS
In this section, we describe the experimental set-up, and we report numerical results
regarding modal analysis and monitoring for this set-up.

4.1. EXPERIMENTAL SET-UP


The tested structure consists of a real aeroplane named Paris MS760 used at SOPEMEA
for research and validation purposes. A diagram of the aeroplane is given in Fig. 1. The
structure is suspended using specially calibrated sandows mounted in parallel, and with
a suspension frequency much below the "rst natural mode of the structure. As depicted in
Fig. 1, 104 accelerometers in the X, >, and Z directions are placed symmetrically around the
wings, fuselage and tail.
Two exciters are placed symmetrically around the wings (close to the tanks), and are used
to generate two independent white noise sequences. All the acceleration as well as the
excitation signals are simultaneously sampled at a rate of 200 Hz, and about 72 000 data
samples per channel are collected. The data acquisition is performed using the LMS-CadaX
system.
Three similar experiments are performed where, for each case, all the data are collected
and stored. First, the structure at its nominal state with blocked ailerons is considered as the
safe condition. Any change is to be compared to this con"guration. Then, in order to
introduce mass changes (2%), on each wing, a 20 kg mass is added near the 250 kg tank
located at its end (the total weight of the wing and the tank is 1000 kg). Finally, to introduce
a sti!ness change, the ailerons are released, after removal of the masses.

4.2. MODAL APPROPRIATION


In order to have a reference model that can be used for comparing the modal identi"ca-
tion results using outputs only, we consider the modal parameters of the same structure
under almost identical experimental set-up, computed using the modal appropriation
technique [16]. This technique is currently the most widely used and well recognised
method for accurate estimation of modal parameters for aeroplane structures. In contrast to
the traditional techniques, the modal appropriation method is based on a pure sinusoidal
excitation of the structure. Using appropriate excitation points as well as the intensity and
phase of the forces, the structure can be &appropriated' at one of its natural frequencies. This
is done by tuning the frequency manually and with the aid of the phase information between
the excitation forces and the responses. Once the structure is appropriated at one of its
modes, the modal parameters of that mode can be estimated: the natural frequency, mode
shape, damping ratio and generalised mass. Two techniques are usually used to compute
the damping and generalised mass: the force in quadrature method and the complex power
method.
846 M. ABDELGHANI E¹ A¸.

Figure 1. PARIS MS760 aeroplane and sensor locations (SOPEMEA).

The exciting forces level used in this test is slightly smaller than that used in the modal
appropriation test. Therefore, we expect small frequency shifts in the identi"cation results.

4.3. SUBSPACE-BASED MODAL IDENTIFICATION


Based on a previous study [17], it was decided to consider only pools with four sensors.
The choice of sensor pools is completely random, in order to get an idea of the e!ect of
sensor selection on the numerical results. No prior optimal sensor selection for damage
detection, as done in [18, 14], has been performed here. The sensor pools used are given in
Table 1.
The balanced realisation (BR) algorithm described above is used for processing the data
and for modal identi"cation. The analysis is done according to the practical use of the
method as outlined in [17, 7].
Data samples containing 48 000 points are used for computing 256 covariance lags. In the
frequency range of interest (0}100 Hz), the total number of modes searched for is about 30.
In order to use as many available covariance lags as possible, a square block Hankel matrix
of size 400 is built, and an SVD decomposition with truncation order up to 200 is
ON-LINE MODAL MONITORING OF AIRCRAFT STRUCTURES 847
TABLE 1
PARIS MS760 aeroplane: sensor pools used in the numerical experiments

Pool 1 Pool 2 Pool 3


Sensor Sensor name Sensor Sensor name Sensor Sensor name

05 Left wing 09 Left wing 22 Left wing


59 Left horizontal "n 22 Left wing 24 Left wing
77 Vertical "n 63 Left horizontal "n 91 Left fuselage
91 Left fuselage 89 Right tank 42 Right wing

TABLE 2
Modal identi,cation of the safe aeroplane: modal appropriation technique ( ,rst column),
BR method for di+erent sensor pools (column 2}4)

Appropriation BR: Sensor pool 1 BR: Sensor pool 2 BR: Sensor pool 3
Freq. Damping Freq. Damping Freq. Damping Freq. Damping
(Hz) ratio (%) (Hz) ratio (%) (Hz) ratio (%) (Hz) ratio (%)

4.87 1.05 4.82 0.64 4.82 0.72 4.82 0.67


8.16 0.75 8.23 0.56 * * * *
10.87 1.20 10.92 1.30 10.92 1.24 10.92 1.29
11.29 1.30 11.28 1.38 11.27 1.39 11.27 1.36
11.40 1.28 11.79 0.99 11.78 0.95 * *
13.86 3.70 * * * * * *
14.94 1.80 14.14 1.79 * * 14.12 1.67
15.90 1.25 15.43 1.42 * * 15.44 1.60
15.92 1.37 * * * * * *
15.92 1.35 * * * * * *
16.54 1.50 * * * * * *
17.78 1.79 * * * * * *
17.93 1.80 18.35 1.13 18.34 1.09 18.34 1.11
19.49 3.10 19.70 1.77 19.70 2.00 19.68 1.71
25.90 1.10 27.17 1.09 * * * *
28.77 1.20 28.24 2.27 28.29 2.45 28.26 2.40
29.03 2.80 * * * * * *
29.82 1.98 29.52 1.00 29.56 0.74 29.53 0.97
34.68 1.38 34.71 1.50 34.66 1.51 34.74 1.35
35.80 8.70 34.95 1.32 * * * *
42.30 2.70 43.18 2.27 42.51 0.67 42.73 1.31
48.59 2.20 47.07 0.91 * * * *
49.44 3.40 52.07 1.26 52.03 1.88 50.91 1.51
63.21 4.00 62.36 1.24 61.87 1.64 * *
65.40 4.80 64.87 0.54 66.60 1.08 68.05 0.59
72.20 1.23 * * 68.92 1.08 68.87 0.97

performed. This over-determination of the Hankel matrix size is necessary in order to


identify closed modes. More details on the practical use of these algorithms are given in
[1, 14, 17, 7].

4.3.1. Safe structure


The modal identi"cation results for the di!erent sensor pools are given in Table 2,
together with the results obtained by the modal appropriation method mentioned above.
848 M. ABDELGHANI E¹ A¸.

TABLE 3
Modal BR identi,cation of the sate (column 1) and damaged (columns 2 and 3) aeroplane,
for sensor pool 1. ¹he MAC coe.cient is computed with respect to the nominal parameters
given in column 1

BR: Nominal BR: Mass change BR: Sti!ness change


Freq. Damping Freq. Damping MAC Freq. Damping MAC
(Hz) ratio (%) (Hz) ratio (%) (%) (Hz) ratio (%) (%)

4.82 0.64 4.80 0.49 99.99 4.82 0.49 100.00


8.23 0.56 8.38 0.83 99.87 8.27 0.70 99.24
10.92 1.30 10.83 1.36 99.99 10.89 1.23 99.97
11.28 1.38 11.19 1.26 92.09 11.26 1.16 90.93
11.79 0.99 11.80 1.02 99.97 11.75 1.12 99.99
14.14 1.79 14.04 1.84 99.93 14.11 1.81 99.95
15.43 1.42 15.39 1.62 99.94 15.45 1.55 99.97
18.35 1.13 18.31 1.10 99.99 18.36 1.23 99.99
19.70 1.77 19.60 1.71 100.00 19.70 1.57 99.99
27.17 1.09 27.10 0.86 99.50 27.12 0.90 99.79
28.24 2.27 * * * 28.10 1.96 99.92
29.52 1.00 29.35 1.28 99.38 29.46 0.86 99.83
34.71 1.50 33.44 0.85 94.91 34.74 1.20 99.85
34.95 1.32 * * * * * *
43.18 2.27 42.65 1.18 55.58 43.19 2.90 98.57
52.07 1.26 * * * 51.36 1.83 99.31
62.36 1.24 62.31 1.77 96.84 62.68 1.58 99.03
64.87 0.54 64.74 0.81 99.92 68.63 0.64 96.05

Inspection of this table reveals that most of the modes are well identi"ed for all the sensor
pools used. However, pool 1 seems to give the best results. Nevertheless, some modes
remain non-identi"able. This is likely to be due to the sensors used.

4.3.2. Damaged structure


In order to assess the in#uence of the modi"cations on the structure (added masses and
sti!ness change), modal analysis is performed on the damaged data sets. For the BR
algorithm, the same parameters as for the nominal data set are used. The modal identi"ca-
tion results are summarised in Table 3. Inspection of this table reveals several facts: (i) The
induced damages on the structure (mass as well as sti!ness) have very little e!ect on the
modal characteristics; this is especially true of the sti!ness change. (ii) It is extremely di$cult
to assess whether the modal changes are due to a damage or rather to the uncertainty
associated with estimation and operator use.

4.4. DAMAGE DETECTION


In this section, we study the performance of the proposed subspace-based tests for on-line
monitoring of the airplane structure. As mentioned above, we limit the study to the
BR-based algorithm. We also study the e!ect of a reduced modal signature.
We recall that modal analysis is performed only once on the nominal structure. No
re-analysis is needed for the damage case. Moreover, the nominal data set (48 000 data
samples) is also used to compute the Jacobian and covariance matrices of the s2-test.
In order to simulate a sudden change to be monitored on-line, the signals corresponding
to the nominal state (48 000 points) and a damage state (40 000) were concatenated to
ON-LINE MODAL MONITORING OF AIRCRAFT STRUCTURES 849

Figure 2. Normalised global test as a function of time for di!erent sensor pools: change in mass. For each sensor
pool, the full modal signature is given in Table 2, columns 2}4, respectively: ** sensor pool 1; } } } sensor pool 2;
} ) } ) sensor pool 3.

Figure 3. Normalized global test as a function of time for di!erent sensor pools: change in sti!ness. For each
sensor pool, the full modal signature is given in Table 2, columns 2}4, respectively: ** sensor pool 1; } } } sensor
pool 2; } ) } ) } sensor pool 3.

produce one single signal that is used for the analysis. A window size of 8000 samples has
been selected, and an overlap of 1000 samples was used. Consequently, in Figs. 2}7, the
time index is the block window number.

4.4.1. Global monitoring


Figures 2 and 3 show the values of the normalised global test as a function of time, for
mass and sti!ness changes and for di!erent sensor pools. The normalisation is done with
the number of dofs (2m(r#1)) of the test, so that the theoretical value of the test under
no-change condition is one.
850 M. ABDELGHANI E¹ A¸.

Inspection of these "gures reveals a signi"cant jump in the value of the global test with
respect to its value under nominal conditions (a contrast of up to 1000). This jump is an
indication of the damage, mass as well as sti!ness, and that for all the sensor pools
considered. However, it is interesting to note that, in the case of sti!ness change in
Fig. 3, the global test is less sensitive to the changes for the sensor pools 2 and 3. Sensor
pool 1 seems to give the best results. This illustrates the importance of the sensors
location in detecting damage. Moreover, it is obvious that it is much easier to detect the
mass change than the sti!ness change. Recall that the sti!ness change is introduced only
by unblocking the ailerons, while the mass change is obtained by adding 20 kg masses on
each wing.
The global test plotted in Figs. 2 and 3 can be considered as a novelty index [19], that
captures the condition of the structure as a function of time. Under no change, the value of
the test #uctuates around its theoretical value (one). It is important to stress that this test
implicitly accounts for the uncertainty in the estimated modal parameters [12, 11]. Note
that the statistical distribution of this global test (s2) is known, and hence a s2-table can be
used to set up alarm levels. When an alarm is "red, a diagnostics procedure can be activated.
Morevoer, this global test does not require any human processing.

4.4.2. Modal diagnostics


While the global test considers all the individual changes in all the modal parameters, it is
sometimes desired to focus the monitoring on some modes that are judged critical, while
considering the changes in the other modes as nuisance parameters. This is easily done
using the sensitivity tests described above. The modes which have been judged critical, and
are considered in the sensitivity tests are: 4.82, 11.28, 28.24 and 43.18 Hz.
Figures 4 and 5 illustrate the sensitivity tests as a function of time for the di!erent sensor
pools, and for mass and sti!ness changes, respectively. Note that these tests are not
normalised. Inspection of these "gures reveals that even though the sensitivity tests tend to
#uctuate as a function of time under no change condition compared to the global test,
a signi"cant jump in the test value occurs, which suggests the presence of a damage.

Figure 4. Modal sensitivity test for di!erent sensor pools: change in mass. The monitoring is focussed on the
modes with eigenfrequencies 4.82, 11.28, 28.24 and 43.18 Hz: ** sensor pool 1; } } } sensor pool 2; } ) } ) } sensor
pool 3.
ON-LINE MODAL MONITORING OF AIRCRAFT STRUCTURES 851

Figure 5. Modal sensitivity test for di!erent sensor pools: change in sti!ness. The monitoring is focussed on the
modes with eigenfrequencies 4.82, 11.28, 28.24 and 43.18 Hz: ** sensor pool 1; } } } sensor pool 2; } ) } ) } sensor
pool 3.

Figure 6. Global test for full and reduced modal signatures (sensor pool 1): change in mass. The full modal
signature is given in Table 3, column 1, where the reduced one appears in bold face: ** full signature; } } }
reduced signature.

A contrast of up to 1000 is achieved for the mass change. These tests indicate that a
signi"cant change in the selected modes has occured. Although these tests do not give any
quantitative answer regarding the damage (location and nature), they should be used along
with the global test as extra tools for performing modal diagnostics. Actually, it is
sometimes very important to know which vibration modes have changed: this is the case in
#utter analysis.

4.4.3. Reduced monitoring


It is sometimes very di$cult or impossible to identify the full model signature of the
structure, especially for structures under operating conditions where some modes might be
852 M. ABDELGHANI E¹ A¸.

Figure 7. Global test for full and reduced modal signatures (sensor pool 1): change in sti!ness. The full modal
signature is given in Table 3, column 1, where the reduced one appears in bold face: ** full signature; } } }
reduced signature.

poorly excited. Therefore, only a reduced modal signature is available. We now investigate
the in#uence of such a model reduction on the global test for damage detection.
Figures 6 and 7 show the normalised global test as a function of time for sensor pool 1 for
the full signatures as well as a reduced one, for mass and sti!ness changes, respectively. The
reduced signature is made of the four modes used in the sensitivity tests above. The "gures
clearly show that a damage has been detected in both cases, even though the full signature
seems to give higher contrast.

5. CONCLUSIONS
We have presented the application of recently proposed subspace-based damage detec-
tion and diagnostics algorithms in on-line modal monitoring of an aircraft structure under
unknown excitation. The study is based on the analysis of the Paris MS760 airplane to
which mass and sti!ness changes are introduced. The problem of damage detection and
monitoring is transformed into the universal problem of monitoring the mean value of
a Gaussian variable with known covariance matrix; a problem that is easily solved using
a s2-test. A global test monitoring the state of the structure is produced, in which jumps in
mean indicate a damage. Sensitivity tests can also be used in order to focus the monitoring
on speci"c critical modes.
In contrast to most other techniques, we perform modal analysis only once on the
nominal structure. Furthermore, we show that using our technique, we can detect small
changes in the structure's modal characteristics, which would not be visible on spectra.
Several sensor pools have been used to illustrate the importance of the choice of sensor
locations in damage detection.
Further investigation of this technique for the modal analysis and monitoring of air-
planes under their operating (#ying) conditions is currently under progress.

REFERENCES
1. M. PREVOSTO, M. OLAGNON, A. BENVENISTE, M. BASSEVILLE and G. LEVEY 1991 Journal of Sound
and <ibration 148, 329}342. State-space formulation, a solution to modal parameter estimation.
ON-LINE MODAL MONITORING OF AIRCRAFT STRUCTURES 853
2. A. BENVENISTE and J.-J. FUCHS 1985 IEEE ¹ransactions on Automatic Control AC-30, 66}74.
Single sample model identi"cation of a non-stationary stochastic process.
3. M. BASSEVILLE, A. BENVENISTE, G. MOUSTAKIDES and A. ROUGED E 1987a Automatica 23,
479}489. Detection and diagnosis of changes in the eigenstructure of non-stationary multi-
variable systems.
4. M. BASSEVILLE, A. BENVENISTE, B. GACH-DEVAUCHELLE, M. GOURSAT, D. BONNECASE, P.
DOREY, M. PREVOSTO and M. OLAGNON 1993 Mechanical Systems and Signal Processing 7,
401}423. Damage monitoring in vibration mechanics: issues in diagnostics and predictive
maintenance.
5. M. BASSEVILLE, M. ABDELGHANI and A. BENVENISTE 1999 Automatica 35. Subspace-based fault
detection algorithms for vibration monitoring, to appear.
6. M. ABDELGHANI, M. BASSEVILLE, A. BEND VENSITE, L. MEVEL, E. BALMES, L. HERMANS and
H. VAN DER AUWERAER 1999 Proceedings of the IMAC-X<II, International Modal Analysis
Conference Kissimmee, F¸, ;SA. Assessment of subspace fault detection algorithms on a realistic
simulator-based example.
7. L. MEVEL, M. BASSEVILLE, A. BENVENISTE, M. GOURSAT, M. ABDELGHANI and L. HERMANS
1999 Proceedings of IMAC-X<II, International Modal Analysis Conference, Kissimmee, F¸.
Application of a subspace-based fault detection method to industrial structures.
8. G. MOUSTAKIDES and A. BENVENISTE 1986 Stochastics 16, 137}155. Detection changes in the AR
part of a non-stationary ARMA process.
9. P. VAN OVERSCHEE and B. DE MOOR 1996 Subspace Identi,cation for ¸inear Systems:
¹heory2Implementation2Methods. Dordrecht: Kluwer.
10. M. VIBERG, B. WAHLBERG and B. OTTERSTEN 1997 Automatica 33, 1603}1616. Analysis of state
space system identi"cation methods based on instrumental variables and subspace "tting.
11. M. BASSEVILLE 1998 Automatica 34, 1391}1416. On-board component fault detection and
isolation using the statistical local approach.
12. A. BENVENISTE, M. BASSEVILLE and G. MOUSTAKIDES 1987 IEEE ¹ransactions on Automatic
Control AC-32, 583}592. The asymptotic local approach to change detection and model
validation.
13. M. BASSEVILLE, A. BENVENISTE and G. MOUSTAKIDES 1986 IEEE ¹ransactions of Information
¹heory IT 32, 412}417. Detection and diagnosis of abrupt changes in modal characteristics of
nonstationary digital signals.
14. B. DEVAUCHELLE-GACH 1991 ¹hesis, Paris IX Dauphine ;niversity (in French). Diagnostic
MeH canique des Fatigues sur les Structures Soumises à des Vibrations en Ambiance de Travail.
15. M. BASSEVILLE 1997 Automatica 33, 783}803. Information criteria for residual generation and
fault detection and isolation.
16. S. NAYLOR, J. E. COOPER and J. R. WRIGHT 1997 Proceedings of the IMAC-X<, International
Modal Analysis Conference, Orlando, F¸. Modal parameter estimation of non-proportionally
damped systems using force appropriation.
17. M. ABDELGHANI, M. GOURSAT, T. BIOCHINI, L. HERMANS and H. VAN DER AUWERAER 1999
Proceedings of the IMAC-X<II, International Modal Analysis Conference, Kissimmee, F¸, ;SA.
Performance of output-only identi"cation algorithms for modal analysis of aircraft structures.
18. M. BASSEVILLE, A. BENVENISTE, G. MOUSTAKIDES and A. ROUGED E 1987b IEEE ¹ransactions on
Automatic Control, AC-32, 1067}1075. Optimal sensor location for detecting changes in dynam-
ical behavior.
19. K. WORDEN 1997 Journal of Sound and <ibration 201, 85}101. Structural fault detection using
a novelty measure.

Vous aimerez peut-être aussi