Mechanical Systems and Signal Processing (2002) 16(1), 169}183

doi:10.1006/mssp.2000.1375, available online at http://www.idealibrary.com on

MODEL UPDATING USING ROBUST ESTIMATION

C. MARES
Department of Engineering (Mechanical Engineering Division), University of Liverpool,
Brownlow Hill, Liverpool L69 3GH, U.K.
M. I. FRISWELL
Department of Mechanical Engineering, University of Wales Swansea, Swansea SA2 8PP, U.K.
AND

J. E. MOTTERSHEAD
Department of Engineering (Mechanical Engineering Division), University of Liverpool,
Brownlow Hill, Liverpool L69 3GH, U.K.

(Received 6 July 2000, accepted 14 November 2000)

Model updating attempts to correct errors in a "nite element model using measured data.
However, the measurements are corrupted with noise and the "nite element model contains
errors. Most updating schemes try to "nd the best model in some mean sense. This paper
takes a di!erent view and uses robust estimation techniques to determine the most robust
parameter set. This parameter set is optimum in that the residuals are as small as can be, for
a range of bounded uncertainties on the model and measurements. The relationship between
this robust identi"cation and Tikhonov regularisation is explored. The use of the &L' curve
method and expected parameter uncertainty to determine the regularisation parameter are
shown. The approach is demonstrated using a simulated cantilever beam example, and
experimental data from the GARTEUR test structure.
 2002 Elsevier Science Ltd.

1. INTRODUCTION
Finite element model updating has become a viable approach to increase the correlation
between the dynamic response of a structure and the predictions from a model. In model
updating parameters of the model are adjusted to reduce a penalty function based on
residuals between a measurement set and the corresponding model predictions. Typical
measurements include the modal model (natural frequencies and mode shapes) and fre-
quency response functions. The choice of penalty function, and also the optimisation
approach, has been the subject of considerable research and are well covered by the authors'
survey paper, book and recent special issue of Mechanical Systems and Signal Processing
[1}3]. The treatment of the noise in the measurements and the errors in the "nite element
model is critical to obtain good estimates of the uncertain parameters. Often these errors are
estimated in terms of variances where the statistical distributions associated with the
parameters and measurements are assumed to be Gaussian. Once these probability distri-
butions are speci"ed the full power of probability theory may be brought to bear. However,
there is often insu$cient information to even specify the variance, let alone determine the
probability distribution.
Methods that specify uncertainties as bounds or convex sets have been popular in control
engineering for some time and are beginning to appear in structural dynamics [4}8]. This
paper is concerned with updating models of structures, where the errors are speci"ed as

0888}3270/02/010169#15 $35.00/0  2002 Elsevier Science Ltd.

170 C. MARES E¹ A¸.

bounds. One approach to the estimation problem is to determine the bounds on the
uncertain parameters, leading to the calculation of the feasible set of parameters which
invariably has complex boundaries [6, 7]. The approach adopted here is the estimation of
a single set of parameters, which is optimal in a min}max sense to be de"ned later [9}11].

2. BOUNDED DATA UNCERTAINTIES

The approach of Chandrasekaran et al. [9] will now be outlined. Much of their paper is
concerned with the proof of a theorem, which is stated below. Suppose the updating
problem has been linearised so that the solution is sought to the linear equation
Ax"b (1)
where b3RK contains the error in the measurements, A3RK;L is the sensitivity matrix and
x3RL the vector of parameter changes [2]. Equation (1) has m measurements, which
produce the observation vector, and n unknown parameters. It is assumed that m*n, that
is the problem is over-determined and there are more measurements than parameters.
Although the quantities in equation (1) are assumed real, complex equations may be
considered by separating the real and imaginary parts.
In equation (1) the sensitivity matrix, A, and observation vector, b, are assumed to be
known. In fact these quantities are subject to errors, and the true coe$cient matrix is
A#
A, and the true observation vector is b#
b.
A and
b arise from modelling errors
and parameter uncertainty and
b also arises from measurement noise. We suppose that the
upper bounds,  and  , of these errors are given, so that
@

A ) (2)

and

b ) . (3)
 @
We can now state the problem as follows: Given non-negative  and  determine the best
@
parameter estimate x; as the one that solves the following min}max problem:

min max (A#
A) x; !(b#
b) : 
A ), 
b ) . (4)

x;
   @
The signi"cance of equation (4) is not readily apparent and some more detailed interpreta-
tion will now be given. For any given x; , there are many residuals,
(A#
A) x; !(b#
b)  (5)

as each feasible
A and
b will give a di!erent residual. For a given x; there will be
a maximum value of these residuals. A second choice of x; will produce a di!erent maximum
residual. We want to choose the parameter estimate x; that produces the smallest maximum
residual. Thus, for all possible errors in the coe$cient matrix and residual vector we have
the smallest maximum possible error. The physical interpretation of this is that the residual
is the smallest one of all the cases of most adverse uncertainty. Chandrasekaran et al. [9]
gave a geometric interpretation of this optimisation. Of course if " "0 then the
@
min}max problem reduces to the standard least-squares problem. El Ghaoui and Lebret
[11] considered the same problem, but derived a di!erent method for its solution. As it
turns out, the solution to equation (4) is independent of  . This may be understood by using
@
the triangle inequality for vector norms to give
(A#
A) x; !(b#
b) )Ax; !b #x;  # . (6)
   @
 MODEL UPDATING 171
Chandrasekaran et al. [9] showed that the min}max problem, equation (4), always has
a solution if A is full rank, and gave a solution method. The solution is constructed as
follows:
(1) Compute the singular-value decomposition of A,




A"U V2 (7)
0
where U and V are orthogonal, and
"diag ( ,  ,2,  ) with the singular values
  L
arranged in descending order
 * *2* '0 (8)
  L
(2) Partition the vector U2b into


b
 "U2 b (9)
b

where b 3RL and b 3RK\L.
 
(3) It is useful to introduce the following function:

G()"b2(
!I)(
#I)\b ! b  . (10)
    
Chandrasekaran et al. [9] proved that G()"0 has a unique positive root. This root
will feature in the parameter solutions.
(4) De"ne

\b 
 "   (11)
 
\b 
 
and
A2b
 " . (12)
 b

The form of the solution will depend on the magnitude of the coe$cient matrix
uncertainty, , compared to  and  , and also on whether the residual vector b is in the
 
column space of A.

Case 1. If * then x; "0.


Case 2. If ( and b is not in the column space of A, then

x; "(A2A# I)\A2b (13)
where  is solution of G( )"0. This is the most common case, and the result is identical to
Tikhonov regularisation [12}14], where now the regularisation parameter is determined
from the uncertainty .
Case 3. If b is in the column space of A, then equation (13) is the solution only if
 (( . If ( then the solution is the standard least-squares solution, and if
  
" " there are in"nitely many solutions.
 

3. TIKHONOV REGULARISATION
The solution given by equation (13) is essentially a Tikhonov regularisation where the
regularisation parameter is obtained from the uncertainty in A given by . Tikhonov
172 C. MARES E¹ A¸.

regularisation minimises

J(x)"Ax!b #x (14)

 
where the parameter  is either given or must be determined. Notice that the essential
di!erence between Tikhonov regularisation and equation (4) is the use of squared norms as
opposed to norms, as highlighted by equation (6).
One useful approach to determining , when the bounds  and  are not available, is via
@
the &L' curve method of Hansen [15]. For a given value of  equation (14) can be minimised
to give the optimum estimate of the parameters as

x "[A2A#I]\A2b. (15)
?
Notice that the estimated parameters depend on . The &L' curve is a plot of the norm of
the side constraint, x , against the norm of the residual, Ax !b, for di!erent values of
? ?
. Hansen [15] showed that the norm of the side constraint is a monotonically decreasing
function of the norm of the residual. He pointed out that for a reasonable signal-to-noise
ratio and the satisfaction of the Picard condition, the curve is approximately vertical for
( , and soon becomes a horizontal line when ' , with a corner near the optimal
 
regularisation parameter  . The approach is called the &L' curve method because of the

shape of this curve. The optimum value of the regularisation parameter,  , corresponds to

the point with maximum curvature at the corner of the log}log plot of the &L' curve. This
point represents a balance between con"dence in the measurements and the analyst's
intuition.
If the bounds  and  are available then Prells [16, 17] suggested an approach to obtain
@
the optimum value of the regularisation parameter . He showed that  is a solution of

b2[AA2#I]\[I!AA2][AA2#I]\ b". (16)

@
Notice that in this approach, the upper bound on the measurement errors,  , also
@
in#uences the optimum value of .

4. THE APPLICATION TO MODEL UPDATING

The application of the robust estimation techniques to model updating will be demon-
strated using eigenvalue residuals. Other residuals could be used [18] and although the
parameters estimated would be di!erent the approach to the regularisation would be
equivalent. Suppose a "nite element model of the structure exists and the parameters to be
estimated using model updating have been selected. The purpose of model updating is to
minimise [2]

L
J (y)"  (y)!  (17)
G CG
G
where is the ith measured eigenvalue and is the ith eigenvalue predicted by the "nite
CG G
element model (and therefore depends on the parameter values). y is the vector of para-
meters to be updated.
The optimisation problem given by equation (17) is non-linear when eigenvalue residuals
are used. The standard approach adopted in model updating is to linearise the estimation
problem about the current parameter estimate, and to iterate until convergence [2]. There
are a number of issues relating to the conditioning of the estimation problem that must be
addressed.
 MODEL UPDATING 173
Weighting is applied to both the measured data and the parameters. Generally, the
weighting on the measured data is such that the entries in the eigenvalue residuals are
divided by the corresponding measured eigenvalue. Furthermore, the unknown parameters
are usually normalised so that the initial value of these parameters is unity. Using this
approach to the weighting of the measurements and the parameters leads to non-dimen-
sional quantities that help with the scaling of the &L' curve. However, increased weight may
be applied to eigenvalues that are accurately measured and to parameters that are not
expected to change signi"cantly from their initial values.
Often the number of uncertain parameters is large compared to the number of
measurements, resulting in an under-determined estimation problem. Even if there are
less parameters than measurements the resulting estimation problem can be ill-conditioned.
In this case extra information must be provided to produce a well-conditioned estima-
tion problem. The proper choice of these constraints is vital to ensure that the
updated parameters have physical meaning. This has be considered at length elsewhere
[12, 19] and is not considered further in this paper. Here we only consider the constraint
that the parameters deviate as little as possible from their initial values. This constraint
and the linearised form of equation (17) are easily transformed to the general form of
equation (1).
Of course since model updating is iterative, the robust estimation procedure must be
applied at each iteration. In robust updating the value of  is "xed at the outset (or perhaps
at the "rst iteration), and thus  changes at each iteration. In Tikhonov regularisation either
 is "xed, or else  is estimated at each iteration by the &L' curve method. However in
practice, with a suitable choice of parameters, convergence is fast and the change in these
regularisation parameters is small.
The choice of  is crucial to successful robust updating. One approach is to determine the
expected bounds directly by assuming that the unknown parameter vector lies in a given
convex set. For any particular parameter vector within this set, the matrix A is computed.
The norm of the di!erence between this matrix, and that obtained using the initial
parameter vector is obtained. This is repeated for all parameter vectors within the set and
the maximum found, which produces a numerical estimate of . If the relationship between
A and the parameters is well behaved, then it is often su$cient to look at the vertices in the
parameter space. This direct approach has some di$culty since the errors depend not only
on the parameter errors but also on model structure errors. For Tikhonov regularisation,
the &L' curve method is able to estimate a reasonable value of the regularisation parameter.
A similar approach for robust updating is desirable. One possibility is to "nd the corner of
the &L' curve at the "rst iteration, and use the roots of G()"0, now considered as
a function , to "nd the corresponding value of . Friswell et al. [20] estimated  using the
total least-squares approach suggested by El Ghaoui and Lebret [11]. Unfortunately, this
gave consistently low values of  and the results were very close to those obtained by
standard least squares.
No mention has been made about implementing these procedures for large-scale
models with many degrees of freedom. The great advantage of using eigenvalue residuals is
that to a large extent the size of the "nite element model is immaterial. The conditioning of
the estimation problem, and the quality of updated model, is totally dependent on the
information content of the measured data used, and the correct choice of parameters.
The parameters should be chosen so that their e!ect on the predictions of the measured
data vector are as linearly independent as possible. With large models of complex struc-
tures, where there are a huge number of candidate parameters, this choice is not easy.
However, the choice of parameters is beyond the scope of this paper, and will not be
consider further.
174 C. MARES E¹ A¸.

5. A SIMULATED CANTILEVER BEAM EXAMPLE

The robust updating approach was tested on a simulated cantilever beam example.
A simulated example has the advantage that the expected answer is known, and the
identi"cation can be performed for a number of nominal systems with bounded uncertainty.
Figure 1 shows the simulated beam with a discrete spring and mass. The beam is 1 m long,
2.5 cm thick and 5 cm wide and the material properties of the beam are those of steel. The
beam is simulated using 10 elements, and "ve eigenvalues are assumed to be measured. The
parameters are as follows

y "k/k , y "m/m , y "EI/EI (18)

     
where k "4 kN/m, m "0.2 kg and EI "13.7 kN m are the simulated spring sti!ness,
  
discrete mass and beam #exural rigidity, respectively. All identi"cation exercises start with
parameter values of 0.9, and clearly the &true' values of all parameters is 1. However, noise is
added to the &measurements', so it cannot be expected that the identi"ed parameters will
equal unity. The random noise is from a uniform distribution, which may be positive or
negative, with a maximum amplitude of 1% of the eigenvalue.
For this example  is estimated in two ways. First the three parameters are varied
simultaneously in the range 0.9}1.1, and for each set of parameter values the matrix A is
computed.  is then given by the maximum, over all parameter sets, of the di!erence
between this matrix and A evaluated when all the parameters are unity. Using this approach
"0.02303. This estimated  is possible because we have some idea of the variation in the
parameters. An alternative, as explained above, is to "nd corner of the &L' curve and then
compute the corresponding value of .
Figure 2 shows the &L' curve for this example for a typical noise sample, based on
Tikhonov regularisation, equation (14), for a range of values of . This &L' curve is for the
"rst iteration of the updating approach, and similar curves could be obtained at subsequent
iterations. The corner of the &L' curve, and also the points corresponding to the robust
updating approach with "0.02303 and the least-squares approach (equivalent to "0),
are marked. Clearly, the robust estimation places slightly less emphasis on the initial
parameter values than the regularised solution at the corner of the &L' curve. However, the
point corresponding to the robust estimation is quite close to the corner.
The alternative is to calculate the value of  directly, based on the upper bounds  and  ,
@
by solving equation (16). Of course, we now need to estimate  , although for the simulation
@
this may be approximated as 1% of the norm of b since a known level of noise has been
added. This approach gives  "0.002093. Table 1 shows the values of  obtained for
@
di!erent values of  and  . Clearly, with the estimated values of  and  ,  is very small and
@ @
the solution is close to the least-squares solution. This situation arises because the assumed
error bounds are small. As the error bounds increase then  changes signi"cantly. For

Figure 1. The cantilever beam example.

 MODEL UPDATING 175

Figure 2. The &L' curve for the cantilever beam example. 0, L curve; 䊊, Corner of the L curve; #,  from
parameter changes; *, Least squares solution.

TABLE 1
¹he calculated values of  using equation (16) for di+erent values of  and
 ( "0.02303 and  "0.002093)
@  @

@
  2 5 10
@ @ @ @
 0.000185 0.000200 0.000409 0.29812

2 0.000514 0.000547 0.000984 0.31183

5 0.002931 0.003145 0.006362 0.39424

10 0.27158 0.28490 0.36870 0.60073


comparison "0.002505 for the robust estimation. The sensitivity of , and therefore the
parameter estimates, to the bounds makes it di$cult to justify choosing a single .
The estimation process was run for 200 noise samples using least-squares estimation,
robust least-squares estimation with "0.02303 and using  obtained from the corner of
the &L' curve at the "rst iteration. Table 2 shows the mean and standard deviation of the
parameter estimates obtained at convergence. Also shown are the errors in the eigenvalues
on convergence. As expected, the higher the value of , or , the more weight is placed on the
initial parameters. This in turn means that the estimated parameters are closer to the initial
parameters, and the errors in the predicted eigenvalues are greater. Thus, in strictly
statistical terms, robust updating produces estimated parameters with a lower standard
deviation, but increased bias. This bias arises because of the con"dence in the initial
parameters, and not because the method has performed poorly. Interestingly, parameter 3 is
always updated accurately, mainly because the eigenvalues are very sensitive to this
parameter. What is clear from the results is that the standard deviation of the parameters
estimated from the least-squares solution are much larger than those estimated from robust
least squares. In a practical example there will be only one set of measurements, and thus
176 C. MARES E¹ A¸.

TABLE 2
¹he statistics of the identi,ed parameters and eigenvalue errors (in %) for the beam example

Robust least squares

Least squares Parameter variation &L' curve

Mean SD Mean SD Mean SD

k 0.9978 0.3519 0.8942 0.0423 0.8939 0.0073

m 1.0303 0.2096 0.9361 0.0683 0.9050 0.0115
EI 0.9984 0.0201 0.9937 0.0038 0.9937 0.0025
Mode 1 0.0000 0.0006 !0.1059 0.2043 !0.2516 0.4553
Mode 2 0.0122 0.2934 0.1741 0.4358 0.1759 0.4532
Mode 3 !0.0555 0.4707 0.0000 0.4766 0.0008 0.4828
Mode 4 0.0698 0.5021 0.0370 0.4913 0.0373 0.4949
Mode 5 !0.0335 0.3626 !0.1381 0.4878 !0.1379 0.5190

the identi"ed parameters are very likely to be a long way from the mean, and therefore the
identi"ed model will be relatively inaccurate.

6. THE GARTEUR SM-AG19 TEST STRUCTURE

The Structures and Materials Action Group (SM-AG19) of the Group for Aeronautical
Research and Technology in EURope (GARTEUR) was initiated in 1995 with the purpose
of comparing a number of current measurement and identi"cation techniques applied to
a common structure [21}23]. The testbed was designed and manufactured by ONERA and
has now been accepted as a benchmark structure by the working group on Finite Element
Model Updating of the COST Action F3 on Strucural Dynamics [24, 25]. The testbed
represents a typical aircraft design with a fuselage, wings and a tail. Realistic damping levels
are achieved by the application of a viscoelastic tape bonded to the upper surface of the
wings and covered by a thin aluminium constraining layer. The testbed has a wingspan of
2 m and a length of 1.5 m.
The experimental data-base (supplied by the University of Manchester and DLR)
consists of 24 frequency response functions for an excitation at the right wing tip in the
range 0}80 Hz, 24 frequency response functions for excitations at each of the right and left
wing tips in the range 4}65 Hz, and 14 normal modes in the range 6.38}151.32 Hz.

6.1. MODELING ASPECTS

The test structure [21] consists of aluminium beams of rectangular cross-section. The
modelling uncertainty, which we aim to reduce by model updating, is mainly concentrated
at the joints and, to a lesser extent, at the constrained viscoelastic layer that runs over the
length of the wings. In the physical structure the joint between the fuselage and the wing is
achieved by screwed connections through a small plate which is sandwiched between the
two. In several modes this joint is at a vibration node in which case its representation in the
"nite element model is not critical. In other modes, however, the joint is strained consider-
ably and the form of the model in the region of the joint may then be important. This joint is
the main one that in#uences the choice of beams or plates for the "nite element model. The
wings and fuselage are long slender structures which might suggest the use of beam
 MODEL UPDATING 177

Figure 3. The "nite element model of the GARTEUR test structure.

elements. However, they are not connected at the neutral axes. Therefore, locally at the joint
there will almost certainly be deformations that cannot be properly represented by the
assumption of plane sections remaining plane. The counter argument, for beam-like
structures, is that a "ne mesh of plate elements would be needed to achieve the same
accuracy that can be readily obtained from a small number of Hermitian beams. We choose
to construct a "nite element model consisting of 126 beams, 26 rigid elements and 5 lumped
masses. The complete model, shown in Fig. 3, has 534 degrees of freedom. An equivalent
beam model is used to represent the joint between the fuselage and wings. The fuselage
imparts an added sti!ness to the wings at the joint and similarly the wings have a sti!ening
e!ect on the fuselage as they pass over it. Beam o!sets are used to model these sti!ening
e!ects. The damping and constraining layers on the wings are each represented by beams
with nodes o!set to the neutral axis of the main wing-beam. Rigid elements are used to
connect the winglets at the wing-tips. O!sets are used at the connection between the "n and
the tail-plane and between the "n and the fuselage. A point mass (and a point inertia) at the
"n/tail-plane joint represents the mass of additional parts used in the construction of the
joint.

6.2. THE SELECTION OF UPDATING PARAMETERS

The choice of updating parameters is an important aspect of the "nite element model
updating process [1, 2]. The sensitive parameters chosen must also be justi"ed by engineer-
ing understanding of the test structure [26}30]. The #exural and torsional rigidities
(EI and GI ) of the wings are of course very sensitive because the wings are the most active
W R
components in all of the lower modes. Their use as updating parameters seems reasonable
mainly because of uncertainty in the thickness of the viscoelastic and constraining layers.
The beam-o!set, (a ) , used to sti!en the wings at the joint with the fuselage represents an
@U
unrealistic rigid joint and must be adjusted. They are sensitive for the same reason that the
wing rigidities are sensitive. At the seventh mode in-plane wing bending occurs with the two
wing-tips in anti-phase. This puts the equivalent beam separating the wings and fuselage
into torsion and it is sensitive for that mode. The torsional eigenvalue, (q ) , of the sti!ness
 
matrix (equivalently the torsional rigidity) of the equivalent beam is updated. Two updating
parameters at the tail are used. These are the o!set, (a ) , at the tail-"n/fuselage joint and the
@R
"rst in-plane bending eigenvalue, (q ) , of the substructure sti!ness matrix for the three
R
elements at the tail-plane/"n joint. Mares et al. [31] described the choice of parameters in
more detail.
In summary the six parameters used for updating were:

E The "rst in-plane bending eigenvalue of the substructure sti!ness matrix for the three
elements at the tail-plane/"n joint, (q ) .
R
178 C. MARES E¹ A¸.

E The beam o!set at the wing/fuselage joint, (a ) .

@U
E The #exural rigidity of the wing, EI .
W
E The torsional rigidity of the wing, GI .
R
E The beam o!set at the tail-"n/fuselage joint, (a ) .
@R
E The torsional eigenvalue of the sti!ness matrix of the beam joining the wing and fuselage,
(q ) .
 

7. ROBUST UPDATING OF THE GARTEUR STRUCTURE

The six uncertain parameters were updated using the lower nine measured natural
frequencies ( f !f ). The remaining "ve natural frequencies ( f !f ) were used only for
   
checking the capacity of the updated model to predict natural frequencies outside the
frequency range, and thus gauge the quality of the updated model. The measured residuals
were weighted with the inverse of the measured natural frequencies, and the parameters
normalised so that the initial parameter values were unity. The residuals were weighted
further based on the estimated accuracy of the measurements by engineering judgement.
Figure 4 shows how the value of  obtained from equation (10) varies with  for the "rst
iteration in this example. As expected a single value of  is obtained for each value of  and
 is monotonically increasing. At each value of , we may obtain the corresponding value of
, and then "nd the updated parameter values, given by equation (13). These identi"ed
parameter values are shown in Fig. 5, which shows the changes from the initial parameter
values of unity for the "rst iteration. As  increases, so does  and thus the initial parameter
values are more heavily weighted. This causes the changes from the initial parameter values
to be smaller as  increases.
The question of how to choose  then arises. The &L' curve method is an established
approach for choosing or  in Tikhonov regularisation [15]. The resulting  produces the
best compromise between the weighting of the residuals and the parameter changes.
Figure 6 shows the &L' curve for the "rst iteration and shows a corner at about "0.4863.
This would be a good choice of , which was then converted, by solving G()"0, to give
"0.7137. Table 3 shows the initial, measured and updated natural frequencies for three
di!erent values of  close to this corner. Unfortunately, the equations were too ill-
conditioned for the standard least-squares solution to be convergent. However, the results
for "0.1 are included, which is close to the smallest  for which the parameters converge.

Figure 4. The relationship between  and  for the "rst iteration.

 MODEL UPDATING 179

Figure 5. The updated parameters after the "rst iteration for di!erent values of . 0, (q ) ; + +, (a ) ; ...... , EI ;
R @U W
. - . -, GI ; *, (a ) ; } }, (q ) .
R @R  

Figure 6. The &L' curve for the "rst iteration.

Table 4 gives the corresponding updated parameters. Figure 7 shows the evolution of the
estimated parameters for four iterations using a constant value of "1. Convergence was
very fast and the updated model not only improves the estimates of the "rst nine natural
frequencies, it also signi"cantly improves the estimates of the "ve higher natural frequencies
not used in the updating. As expected as  increases more weight is given to the initial
parameter values, and the changes from their initial values is smaller.
Tables 5 and 6 show the error in the natural frequencies and the parameter estimates after
updating for di!erent values of  and  . The value of  was computed using equation (16).
@
The magnitude of the error bounds may be compared to A "20.68 and b "
 
1.27;10. Thus, the values of  of 10 and 2.5;10 represent 8 and 19% of the norm of b.
@
The results are better for the smaller value of  for the same value of , both in terms of the
@
"t to the "rst nine natural frequencies and also in the prediction of the "ve higher natural
frequencies. This may be expected as the errors in the measured data are likely to be nearer
8 than 19%.
180 C. MARES E¹ A¸.

TABLE 3
¹he natural frequencies for the GAR¹E;R structure

Natural frequencies (Hz) Errors (%)

Mode Test Initial "0.1 "0.5 "1 "4 Initial "0.1 "0.5 "1 "4

f 6.38 6.61 6.47 6.46 6.42 6.24 3.53 1.47 1.23 0.70 !2.26

f 16.10 16.95 16.30 16.28 16.31 16.22 5.25 1.24 1.14 1.33 0.72

f 33.10 35.98 33.62 33.69 33.51 33.77 8.70 1.56 1.80 1.24 2.02

f 33.53 36.45 33.97 34.08 33.77 33.92 8.72 1.31 1.63 0.71 1.17

f 35.65 37.22 35.05 35.07 35.13 35.81 4.40 !1.69 !1.64 !1.45 0.44

f 48.38 50.88 48.52 48.49 48.49 47.66 5.17 0.28 0.24 0.23 !1.49

f 49.43 53.94 49.15 49.03 48.77 52.18 9.12 !0.57 !0.81 !1.33 5.56

f 55.08 59.30 55.20 55.20 55.26 57.01 7.66 0.21 0.22 0.32 3.51

f 63.04 64.48 63.77 63.67 63.63 63.72 2.29 1.15 1.00 0.93 1.08

f 66.50 66.94 64.35 64.35 64.35 64.61 0.66 !3.23 !3.24 !3.24 !2.85

f 102.90 105.71 98.94 98.94 100.44 104.40 2.73 !3.85 !3.85 !2.39 1.46

f 130.54 133.09 127.53 127.51 128.35 130.49 1.95 !2.30 !2.32 !1.68 !0.04

f 141.38 146.33 140.97 140.63 139.93 137.17 3.50 !0.29 !0.53 !1.03 !2.98

f 151.32 158.68 149.17 148.91 148.76 150.41 4.86 !1.42 !1.59 !1.69 !0.60


TABLE 4
¹he updated parameters for the GAR¹E;R structure

Parameter "0.1 "0.5 "1 "4

(q ) 0.911 0.908 0.949 0.975

R
(a ) 0.050 0.050 0.064 0.481
@U
(EI ) 1.098 1.091 1.073 0.934
WU
(GI ) 0.849 0.856 0.836 0.843
RU
(a ) 0.544 0.553 0.614 0.907
@R
(q ) 0.384 0.383 0.381 0.802
 

Figure 7. The evolution of the updated parameters with a constant . 0, (q ) ; + +, (a ) ; ...... , EI ; . - . -, GI ; *,

R @U W R
(a ) ; } }, (q ) .
@R  
 MODEL UPDATING 181
TABLE 5
¹he natural frequency errors for the GAR¹E;R structure using equation (16) to calculate .
¹he values of  should be multiplied by 1000
@
Natural frequencies
(Hz) Errors (%)

"0.5 "1.0 "4.0 "0.5 "1.0 "4.0

Mode Test Initial Initial  "10  "10  "10  "25  "25  "25
@ @ @ @ @ @
f 6.38 6.61 3.53 1.43 !1.54 !2.25 !1.67 !1.80 !2.28

f 16.10 16.95 5.25 1.02 1.01 0.73 0.98 0.94 0.71

f 33.10 35.98 8.70 2.02 0.98 2.02 1.07 1.18 2.16

f 33.53 36.45 8.72 2.14 0.14 1.17 0.22 0.33 1.31

f 35.65 37.22 4.40 !1.65 !0.33 0.43 !0.21 !0.08 0.48

f 48.38 50.88 5.17 0.25 !1.19 !1.48 !1.26 !1.33 !1.47

f 49.43 53.94 9.12 !0.60 1.61 5.54 2.10 2.73 5.78

f 55.08 59.30 7.66 0.22 1.56 3.51 1.83 2.12 3.69

f 63.04 64.48 2.29 0.98 0.58 1.08 0.64 0.71 1.12

f 66.50 66.94 0.66 !3.24 !3.12 !2.85 !3.10 !3.06 !2.82

f 102.90 105.71 2.73 !5.02 0.01 1.45 0.20 0.42 1.56

f 130.54 133.09 1.95 !2.76 !0.65 !0.04 !0.57 !0.48 0.00

f 141.38 146.33 3.50 !0.33 !2.86 !2.98 !2.90 !2.95 !2.95

f 151.32 158.68 4.86 !1.55 !1.70 !0.60 !1.56 !1.40 !0.49


TABLE 6
¹he updated parameters for the GAR¹E;R structure using equation (16) to calculate .
¹he values of  should be multiplied by 1000
@
"0.5 "1.0 "4.0 "0.5 "1.0 "4.0
Parameter  "10  "10  "10  "25  "25  "25
@ @ @ @ @ @
(q ) 0.868 0.966 0.975 0.966 0.966 0.977
R
(a ) 0.050 0.230 0.481 0.265 0.302 0.503
@U
(EI ) 1.097 0.986 0.935 0.978 0.969 0.931
WU
(GI ) 0.867 0.823 0.843 0.825 0.827 0.845
RU
(a ) 0.527 0.812 0.907 0.829 0.846 0.913
@R
(q ) 0.383 0.528 0.799 0.551 0.587 0.818
 

8. CONCLUSIONS
This paper has introduced the robust model updating approach, where the uncertainty in
the model and measurements are speci"ed as bounds on the norms of the coe$cient matrix
and the residual vector. This is shown to be equivalent to Tikhonov regularisation and the
relationship between the uncertainty bound and the Tikhonov regularisation parameter has
been explored. Two approaches to obtaining the uncertainty bound were given, namely
using the corner of the &L' curve, and using the expected parameter uncertainty. The robust
updating approach was demonstrated on a simulated cantilever beam example in order to
highlight the statistical properties of the approach, and also on the experimental data from
the GARTEUR test structure.
182 C. MARES E¹ A¸.

ACKNOWLEDGEMENTS
This research is supported by EPSRC grant GR/M08622. Prof. Friswell gratefully
acknowledges the support of the EPSRC through the award of an Advanced Fellowship.

REFERENCES
1. J. E. MOTTERSHEAD and M. I. FRISWELL 1993 Journal of Sound and <ibration 162, 347}375.
Model updating in structural dynamics: a survey.
2. M. I. FRISWELL and J. E. MOTTERSHEAD 1995 Finite Element Model ;pdating in Structural
Dynamics. Dordrecht: Kluwer Academic Publishers.
3. J. E. MOTTERSHEAD and M. I. FRISWELL (Guest eds), 1998 Mechanical Systems and Signal
Processing 12, 1}224. Special Issue on Model Updating.
4. Y. BEN-HAIM, S. COGAN and L. SANSEIGNE 1998 Mechanical Systems and Signal Processing 12,
121}134. Usability of mathematical models in mechanical decision processes.
5. Y. BEN-HAIM and I. ELISHAKOFF 1990 Convex Models of ;ncertainty in Applied Mechanics.
Amsterdam: Elsevier Science Publishers.
6. J. P. NORTON 1987 Automatica 23, 497}507. Identi"cation and application of bounded-parameter
models.
7. J. P. NORTON 1999 2nd International Conference on Identi,cation in Engineering Systems, Swan-
sea, 33}47. Worst-case identi"cation for robust control.
8. R. S. SMITH and J. C. DOYLE 1992 IEEE ¹ransactions on Automatic Control 37, 942}951. Model
validation: a connection between robust control and identi"cation.
9. S. CHANDRASEKARAN, G. H. GOLUB, M. GU and A. H. SAYED 1998 SIAM Journal on Matrix
Analysis and Applications 19, 235}252. Parameter estimation in the presence of bounded data
uncertainties.
10. S. CHANDRASEKARAN, G. H. GOLUB, M. GU and A. H. SAYED 1998 SIAM Journal on Matrix
Analysis and Applications 20, 839}859. An e$cient algorithm for a bounded errors-in-variables
model.
11. L. EL GHAOUI and H. LEBRET 1998 SIAM Journal on Matrix Analysis and Applications 18,
1035}1064. Robust solutions to least squares problems with uncertain data.
12. H. AHMADIAN, J. E. MOTTERSHEAD and M. I. FRISWELL 1998 Mechanical Systems and Signal
Processing 12, 47}64. Regularisation methods for "nite element model updating.
13. P. C. HANSEN 1994 Numerical Algorithms 6, 1}35. Regularisation tools: a MATLAB package for
analysis and solution of discrete ill-posed problems.
14. A. TIKHONOV and V. ARSENIN 1977 Solutions of Ill-Posed Problems. New York: Wiley.
15. P. C. HANSEN 1992 SIAM Review 34, 561}580. Analysis of discrete ill-posed problems by means
of the L-curve.
16. U. PRELLS 1995 Ph.D. thesis, ;niversita( t Hannover, Germany. Eine Regularisierungsmethode fuK r
die Lineare Fehlerlokalisierung von Modellen Elastomechanischer Systeme.
17. U. PRELLS 1996 Inverse Problems in Engineering 3, 197}217. A regularization method for the
linear error localization of models of elastomechanical systems.
18. H. G. NATKE, G. LALLEMENT, N. COTTIN and U. PRELLS 1995 Inverse Problems in Engineering 1,
329}348. Properties for various residuals within updating of mathematical models.
19. M. I. FRISWELL, J. E. MOTTERSHEAD and H. AHMADIAN 2001 Proceedings of the Royal Society
of ¸ondon, Series A: Mathematical, Physical and Engineering Sciences, 359, No. 1778,
169}186. Finite element model updating using experimental test data: parameterisation and
regularisation.
20. M. I. FRISWELL, J. E. MOTTERSHEAD and T. SHENTON 2000 Proceedings of the 18th International
Modal Analysis Conference, San Antonio, TX, U.S.A., 987}991. Robust model updating.
21. M. DEGEUER and M. HERMES 1996 D¸R, Deutsches Zentrum fur ¸uft-und Raumfahrt, Goettingen,
Inst. fur Aerolastik, Bericht IB 232-96 JO8. Ground vibration test and "nite element analysis of the
GARTEUR SM-AG19 testbed
22. E. BALMES 1997 Proceedings of the 15th International Modal Analysis Conference, Orlando,
U.S.A., 1346}1352. GARTEUR group on ground vibration testing results from the tests of a single
structure by 12 laboratories in Europe.
23. E. BALMES 1998 Proceedings of the 16th International Modal Analysis Conference, Santa
Barbara, U.S.A, 558}564. Predicted variability and di!erences between tests of a single
structure.
 MODEL UPDATING 183
24. M. LINK and T. GRAETSCH 1999 Proceedings of the Second International Conference on Identi,ca-
tion in Engineering, Swansea, UK, 48}62. Assesment of model updating results in the presence of
model structure and parametrisation errors.
25. S. KEYE 1999 Proceedings of the Second International Conference on Identi,cation in Engineering,
Swansea, U.K., 122}134. Prediction of modal and frequency response data from a validated "nite
element model.
26. J. E. MOTTERSHEAD, M. I. FRISWELL, G. H. T. NG and J. A. BRANDON 1996 Mechanical Systems
and Signal Processing 10, 171}182. Geometric parameters for "nite element updating of joints and
constraints.
27. H. AHMADIAN, J. E. MOTTERSHEAD and M. I. FRISWELL 1997 Proceedings of the 15th Interna-
tional Modal Analysis Conference, Orlando, U.S.A., 142}146. Parametrisation and identi"cation of
a rubber-seal.
28. J. E. MOTTERSHEAD, C. MARES, M. I. FRISWELL and S. JAMES 2001 Mechanical Systems and
Signal Processing 15, 87}100. Selection and updating of parameters for an aluminium space-frame
model.
29. G. M. L. GLADWELL and H. AHMADIAN 1996 Mechanical Systems and Signal Processing 9,
601}614. Generic element matrices for "nite element model updating.
30. H. AHMADIAN, G. M. L. GLADWELL and F. ISMAIL 1997 Journal of <ibration and Accoustics 119,
37}45. Parameter selection strategies in "nite element updating.
31. C. MARES, J. E. MOTTERSHEAD and M. I. FRISWELL 2000 Proceedings of ISMA 25, Leuven,
Belgium, 635}640. Selection and updating of parameters for the GARTEUR SM-AG19 testbed.