Sie sind auf Seite 1von 16

Mechanical Systems and Signal Processing (1998) 12(1), 75–90

MODEL UPDATING IN STRUCTURAL DYNAMICS:


A GENERALISED REFERENCE BASIS APPROACH
R. K  Y. H
Faculty of Mechanical Engineering, Technion–Israel Institute of Technology, Haifa 32000 ,
Israel

(Received November 1996, accepted after revisions August 1997)

The paper considers the problem of updating an analytical model from experimental data
using the reference basis approach. In this general framework, certain parameters, e.g.
natural frequencies or modeshapes, are considered to be completely accurate and the others
are updated by solving a constrained optimization problem. The main results are
closed-form solutions to reported problems with general weighting matrices in the
optimisation criterion. The importance of this generalisation, in addition to its theoretical
value, is the ability to incorporate prior knowledge regarding the accuracy of the model
in specified areas into the method. This paper investigates also the geometrical
interpretation of the results, providing an insight to the mechanism of the updating process.
The advantages of the new updating schemes are demonstrated by means of examples.
7 1998 Academic Press Limited

1. INTRODUCTION
In modern analysis of structural dynamics, much effort is devoted to the derivation of
accurate models. These models are then used to predict the response of the system to
various excitations, boundary conditions and parameter changes. The first step is the
derivation of an analytical model, usually finite element, based on the assumed equations
of motion. When tests are performed to validate the analytical model, inevitably their
results, commonly natural frequencies and modeshapes, do not coincide with the expected
results from the theoretic model. These discrepancies stem from uncertainties regarding
the governing equations of the system, simplifying assumptions in the derivation and
inaccurate boundary conditions. Clearly one would like to have a better model, based on
both the theoretical and the experimental results. It should be noted that the experimental
results are always partial, so even if they are assumed to be absolutely accurate they cannot
be the sole source of the final model. The problem of how to modify the theoretic model
from the experimental results is known as model updating.
The model updating problem has been investigated thoroughly and many approaches
to it have been suggested. One extreme is the field of system identification which, to a
certain degree, completely disregards the analytical model. However, most approaches do
incorporate the analytical model since, as was mentioned earlier, the eigendata provided
by the experiment are incomplete. These approaches include sensitivity methods where the
dependence of the natural frequencies and the modeshapes on physical parameters is
calculated either numerically or analytically [1, 2] or by the mixed matrix approach [3]. In
[4], the rank of the error matrix, rather than the norm as in all other methods, is minimised.
Other methods use the complete frequency response function as their data [5]. An extensive
survey of model updating methods can be found in [6]. In reference basis methods [7–9],

0888–3270/98/010075 + 16 $25.00/0/pg970135 7 1998 Academic Press Limited


76 .   . 
certain quantities are assumed to be accurate and those that are free (generally stiffness
and mass) should satisfy the system’s relationships with minimal deviation from the
analytical model. In mathematical terms, the second requirement defines the optimization
criterion while the first one deals with the constraints.
The main advantage of reference basis methods is the mathematical and numerical
convenience. They provide closed-form solutions, requiring only elementary matrix
operations and inversion of low-dimension matrices. The main disadvantage is that they
do not account for any dependence of the matrices on physical properties, and cannot
incorporate any prior knowledge or engineering consideration or even preserve the
connectivity of the system. This trade-off is very clear when reference basis methods [7–9]
are compared with approaches that minimise a similar criterion yet impose certain
connectivity on the stiffness matrix [10–12]. The updated stiffness matrix in the latter
approach is normally closer to the true one, provided that the assumed connectivity is
correct. On the other hand, these methods involve the solution of a set of linear equation
whose dimension is equal to the number of non-zero elements in the stiffness matrix. This
may be orders of magnitude larger then the number of measured modes.
This paper presents improved and generalised versions of some of the main results in
the reference basis methods, presented in [7–9]. All these methods share the same structure:
minimisation of the distance from the analytical matrices multiplied by the mass matrix
which serves as a weighting function. The updating problem is stated then solved with
general weighting functions. In addition to its theoretical value, this generalisation is very
important from a practical point of view. One often has a priori knowledge regarding the
location of ‘suspicious’ degrees of freedom. This information can also be provided or
supported by error localisation methods such as the COMAC test [13]. General weighting
functions enable the engineer to use this information in the updating process so it links
the reference basis method to the error localisation approach [14] and also to intuitive
engineering practice. In addition, a geometric interpretation of these problems is presented,
which gives new insights to the updating process. It is shown that this interpretation relates
the method to the mixed matrix method [3]. The positive definiteness of the updated
matrices is investigated as well.
The material is organised as follows. The problem is stated in Section 2. Sections 3, 4
and 5 consider three problems: stiffness updating, sequential mass and stiffness updating,
and simultaneous mass and stiffness updating, respectively. In each of them a closed-form
solution is given and its physical interpretation is investigated. The advantages of the new
approach are demonstrated by means of an extended example in Section 6. The results
are summarised and discussed in Section 7.

2. PROBLEM STATEMENT
Let the ‘true’ equations of motion of the system be
MT ẍ(t) + KT x(t) = f (t) x $Rn (1)
where MT and KT are the exact mass and stiffness matrices, respectively, and n is the
number of degrees of freedom. The absolutely accurate model of the system is non-linear,
infinitely dimensional, etc., so ‘true’ means accurate enough for all practical purposes. The
natural frequencies and modeshapes of the system are vTi and FTi respectively. The
analytical model of the system, obtained by finite elements or any other method is given
by
MA ẍ(t) + KA x(t) = f (t) x $Rn (2)
     77
where in general MA $ MT and KA $ KT (MA and KA are the analytical mass and stiffness
matrices, respectively). The result of a modal test of the system is a partial set of m Q n
natural frequencies and modeshapes vEi and fEi respectively. These quantities are arranged
in matrix form as

VE = diag{vEi }

FE = [FE1 . . . FEm ] (3)

Two assumptions are made throughout this paper.


, The dimensions of the measured modeshapes in equation (3) and the analytical model
in equation (2) are the same. This is achieved either by order reduction of the analytical
model by a standard method, e.g. Guyan reduction [15] or by expanding the
measurements to the higher dimension [16]. Unlike the analysis in [10], this dimension
matching is independent of the updating process.
, The measured natural frequencies are correct, i.e.

VE = VT ,V (4)

The physical justification for this is that they are global variables, common to all
measurements in the system, and the errors in extracting them are relatively small.
The specific problems (or sub-problems) which are defined and solved in the next
sections are characterised by additional assumptions.
The problem discussed in this paper is how to combine the analytic information
(MA , KA ) and the experimental one (VE , FE ) and obtain a model which is more accurate,
i.e. closer to (MT , KT ).

3. SEQUENTIAL UPDATING OF MODESHAPES AND STIFFNESS MATRICES


The accurate quantity (the reference basis) in this case is the mass matrix, i.e. it is
assumed that MA = MT = M. The first step (which is discussed later) is correction of the
measured modeshapes FEX which results in a set F which is orthogonal and normalised
with respect to the mass matrix, hence

FTMF = I (5)

F is regarded in the sequel as the exact modeshapes. An important property, which is


typical to most structures, is that the stiffness matrix is symmetric and at least positive
semi-definite (psd). To enforce this requirement, KMLK LTK is defined. Clearly K e 0 and
if, in addition, LK has full row rank then K q 0. This is a deviation from previous works
where only the symmetry of K was enforced, and via a formal constraint. The minimisation
problem is thus given by

min J = >W−1/2
k (Lk LTk − KA )W−1/2
k >2F (6)

s.t. Lk LTk F = MFV2

where > >F is the frobenius norm defined by

>A>F = s s =aij =2 = trace(A*A)


i j
78 .   . 
The dimension of the constraint equation is nm but not all of these equations are
independent. To see this, multiply it by the square non-singular matrix T = [FT ( ]T where
( indicates arbitrary rows. Then one gets [FTLK LTK F ( ]T = [V2 ( ]T and since the upper
matrix is symmetric the total number of independent equations is only m (n − m/2 + 1/2).
Wk is a symmetric but otherwise general weighting matrix. This is the main distinction
between the current work and the results of [7–9] where WK = M was selected. While this
is a reasonable weighting matrix, there are other plausible choices such as WK = KA which
leads to a non-dimensional criterion or WK = I which is the absolute, unweighted case. The
most significant application of the flexibility in WK is emphasising substructures which are
known to suffer from poor modeling either as an a priori information or from tools like
COMAC [13]. From the cost function and the constraint in equation (6) the following
Lagrange function is constructed
C = 1/4 tr{W−1
k (Lk Lk − KA )Wk (Lk Lk − KA )} + tr{l (Lk Lk F − MFV )}
T −1 T T T 2
(7)
The first term is the Frobenius norm in ‘trace’ form (the 1/4 was added for convenience)
and l(n × m) is the Lagrange multiplier matrix.
The partial derivatives of C with respect to Lk and l are as follows
1C
k (Lk Lk − kA )Wk Lk + lF Lk + Fl Lk = 0
T
= W−1 −1 T T
(8)
1Lk
1C
= Lk LTk F − MFV2 = 0 (9)
1l
Assuming for the moment that Lk is non-singular, which means K q 0, we can
pre-multiply (8) by WK and post-multiply it by L−1
k Wk to obtain

Lk LTk = KA − Wk (lFT + FlT)Wk (10)


Equations (9) and (10) now need to be solved. Notice that LK appears only in the product
LK LTk so K = LK LTk is used in the remainder of this section. Another observation from
equation (10) is that there is no unique solution for l. To see that define l* = l + FB,
where l is a solution. Substituting l* into (10) yields
K = KA − Wk (lFT + FlT)Wk − Wk (FBFT + FBTFT)Wk (11)
For any skew-symmetric B, the last term of equation (11) vanishes; hence for such B, l*
is a solution as well. This non-uniqueness is a direct consequence of the non-minimality
of the constraint equation. Nevertheless, this does not cause any practical problems since
K is unique.
To obtain a closed form solution for Lk substitute (10) into (9)
KA F − Wk (lFT + Fl T)Wk F = MFV2 (12)
and have
Wk l = (KA F − MFV2 − Wk FlTWk F)(FTWK F)−1 (13)
Substituting equation (13) (but not its transpose) into (10) leads to
K = KA − (KA F − MFV2)RT − WK FlTWK (I − FRT) (14)
where
R = Wk F(FTWk F)−1
     79
Substituting the transpose of equation (13) into (14), and recalling that R (I − FR ) = 0,
T T

leads to the solution for the optimisation problem (6) as


K = KA − (KA F − MFV2)RT − R(KA F − MFV2)T + R(FTKA F − V2)RT (15)
Note that the result in equation (15) can be obtained from Lemma 2.1 in [17] after some
redefinition of the matrices.
This solution was derived under the assumption that K = LK LTK . If K in equation (15)
is not psd than this assumption is violated and the expression is not a solution of the
problem. (It can be shown that this is still the solution for a similar problem where K is
required to be symmetric but not necessarily psd.) The solution still applies to the case
where K is psd but not pd since equation (10) holds even though a singular matrix has
been canceled. If K is indefinite, i.e. has at least one negative eigenvalue, then the updated
stiffness which minimises equation (6) while being psd, is the solution of the non-linear
equations (8) and (9). In such cases, Lk always turns out to be singular which means that
at least one of the natural frequencies of the system is zero. This property is characteristic
to rigid body dynamics.

3.1.     


Equation (15) can be used to get the numerical value of the optimal K yet the mechanism
of the optimisation is not obvious from that form of the result. To gain more insight into
the solution equation (15) can be rearranged as
K = MFV2FTM + (I − FRT)T(KA − MFV2FTM)(I − FRT) (16)
This form has a clear geometric interpretation. First notice that (I − FRT) is a projection
matrix since (I − FRT)2 = (I − FRT), and that (I − FRT)F = 0. Furthermore, let
FF = [F F ] and VF = diag[V V ] be the complete set of natural frequencies and
modeshapes. Then
KT = MFF V2F FTF M = MFV2FTM + MF V 2F TM (17)
Viewing K and KT as operators, the first and second terms in equations (16) and (17) are
their restrictions to the subspaces V = span(F) and its orthogonal complement V_ = span
(MF ), respectively. Thus, K has an accurate part (the first term), acting on the subspace
defined by the measured modeshapes F, and a correction part consisting of the difference
between the model stiffness matrix and the accurate part, restricted to V_ by the projection
matrix. In case KA is accurate, i.e. KA = KT , the difference is already restricted to that
subspace and therefore remains unchanged by the projection. WK affects only the ‘angle’
of the oblique projection into V_, but as demonstrated by examples later, this has a strong
effect on the results. Note also that the geometric structure of equation (16) bears some
resemblance to the mixed matrix method [3].
The relationship between K, KA and KT (which is of course unknown) is given by
K = KT + (I − FRT)T(KA − KT )(I − FRT) (18)
pre- and post-multiplying by W1/2
K and rearranging yields

W1/2 1/2 1/2 1/2


K (K − KT )WK = P1 WK (KA − KT )WK P1 (19)
Where P1 = I − W1/2K F(F WK F) F W1/2
T −1 T
K is a symmetric matrix, thus representing an
orthogonal projection. From the properties of such projections it follows that
>W1/2
K (K − KT )WK >2 E >WK (KA − KT )WK >2
1/2 1/2 1/2
80 .   . 
Hence the procedure improves the quality of the model by reducing (or at least not
increasing) its deviation from the accurate (and unknown) KT .

3.2.   W = M


Equation (15) gives the solution for any weighting function. Choosing WK = M one
obtains the case presented in [7] where the updated stiffness matrix for this special case
is given by
KS = KA − KA FFTM − MFFTKA + MFFTKA FFTM + MFV2FTM (20)
Lemma 1: If KA q 0 (KA e 0) and V q 0 (V e 0) then KS q 0 (KS e 0).
Proof: Equation (16) for this case is
KS = MFV2FTM + (I − MFFT)(KA − MFV2FTM)(I − FFTM) (21)
However, FTM(I − FFTM) = 0 and therefore
KS = MFV2FTM + (I − MFFT)KA (I − FFTM) (22)
Hence KS is a sum of non-negative definite matrices which makes it non-negative definite.
It is assumed now that KA q 0 and V q 0. The quadratic form xTKS x can be zero only
if x belongs to the nullspace of both terms in equation (22). The nullspace of the second
term is the column space of F but any vector which belongs to this subspace can be written
as x = Fa and
xTMFV2FTMx = aTV2a q 0 (23)
The strictly positive part of this Lemma was proved in the Theorem 2.1 in [17].

3.3.  


In the previous subsections, the measured modeshapes were assumed to be accurate.
Unlike the natural frequencies, the modeshapes are not global variables thus they are more
susceptible to errors due to unsuitable equipment, wrong analysis of the measurements,
etc. Furthermore, errors in the modeshapes are detected easily since the true ones should
be orthogonal with respect to the mass matrix. Therefore, the first step should be solving
the following problem: find an orthogonal set which is closest to the measured set of
modeshapes. It is defined mathematically by
min J = >W1/2
1 (F − FEX )W2 >F
1/2 2

s.t. FTMF = I (24)


where FEX are the measured modeshapes, where each modeshape FEXi is normalised with
respect to the mass matrix. The constraint equation stems from the fact that the optimal
modeshapes F must be orthogonal and normalised with respect to the mass matrix. W1
(n × n) and W2 (m × m) are any symmetric matrices. W1 may be chosen as a diagonal
matrix where each element multiplies one row of the modeshapes matrix. This means
different weight to each degree of freedom. A diagonal W2 on the other hand multiplies
the columns of the modeshapes matrix, thus giving different weight to each modeshape.
The Lagrange function is
C = tr{W1 (F − FEX )W2 (F − FEX )T} + tr{a(FTMF − I)} (25)
where a(m × m) is the Lagrange multiplier matrix. This matrix is symmetric because
the constraint matrix equation is symmetric, and this fact plays an important role
     81
in the analysis that follows. The partial derivatives of C with respect to F and a are as
follows
1C
= W1 (F − FEX )W2 + MFa = 0 (26)
1F
1C
= FTMF − I = 0 (27)
1a
To eliminate a equation (26) is multiplied by FT and equation (27) is substituted to obtain
a = −FTW1 (F − FEX )W2 (28)
substituting equation (28) into (26) yields
(I − MFFT)W1 (F − FEX )W2 = 0 (29)
and for non-singular W2 ,
(I − MFFT)W1 (F − FEX ) = 0 (30)
The solution is characterized by the folowing matrix equations:
(1) equation (27), which includes m(m + 1)/2 independent equations;
(2) the expression for equation (28) should be symmetric, which adds m(m − 1)/2
independent equations;
(3) equation (29) contains (n − m)m independent equation. This is due to the fact that
when premultiplied by FT it becomes identically zero.
Thus the above contain nm independent equations, which is an equal number to the
number of unknowns in F, so this set of equations has a finite number of solutions.
In general the problem (24) does not seem to have a closed-form solution because both
the criterion and the constraint are quadratic. Therefore, the following cases are
considered.
Case 1 . W1 = M, and W2 is a general symmetric matrix.
In this case, the cost function can be written as
J = tr{M(F − FEX )W2 (F − FEX )T} = tr{(F − FEX )TM(F − FEX )W2 }
= 2tr{FTEX MFW2 } + tr{(I + FTEX MFEX )W2 } (31)
where the last equality is obtained by using FTMF = I. The second term is a constant and
the first one is linear in F. The problem is therefore equivalent to that of a quadratic cost
and a linear constraint which is known to have a closed-form solution. The partial
derivatives of C in this case are
1C
= M(F − FEX )W2 + MFa = 0 (32)
1F
1C
= FTMF − I = 0 (33)
1a
Premultiplying equation (32) by M−1 yields
F(W2 + a) = FEX W2 (34)
82 .   . 
Assuming that (W2 + a) is a non-singular, equation (34) becomes
F = FEX W2 (W2 + a)−1 (35)
Substituting equation (35) into (33) yields
(W2 + a) = (W2 FTEX MFEX W2 )1/2 (36)
Where for a symmetric A, B = A1/2 means a symmetric matrix such that B2 = A. The
solution for F in this case is given by
F = FEX W2 (W2 FTEX MFEX W2 )−1/2 (37)
Case 2 . General W1 , W2 = I, span(F) = span(FEX ).
The last restriction is a common assumption in modeshapes updating. In the previous case
it emerges naturally, but here it is imposed on the solution. Mathematically, it is expressed
as F = FEX D, with D square and non-singular. The optimisation problem (24) becomes
min J = >W1/2
1 FEX (D − I)>F
2
D

s.t. DTFEX MFEX D = I (38)


The Lagrange function is
C = tr{W1 FEX (D − I)(DT − I)FTEX } + tr{aT(DTFTEX MFEX D − I)} (39)
and the partial derivatives of C with respect to D and a are as follows
1C
= w1 (D − I) + mDa = 0 (40)
1D
1C
= DTmD − I = 0 (41)
1a
where w1 = FTEX W1 FEX and m = FTEX MFEX . Premultiplying equation (40) by DT gives
a = −DTw1 (D − I) (42)
Notice that a in equation (42) must be symmetric and that implies that the term DTw1
should be symmetric as well. The problem can now be formulated like this: find D which
satisfies the following requirements
DTmD = I (43)
DTw1 = (DTw1 )T (44)
A solution for this problem is
D = w−1 −1 −1 −1/2
1 (w1 mw1 ) (45)
and the updated modeshapes are given as
F = FEX (FTEX W1 FEX )−1[(FTEX W1 FEX )−1(FTEX MFEX )(FTEX W1 FEX )−1]−1/2 (46)
In the special case W1 = M and W2 = I which was considered in [7], the solution for the
updated modeshapes is given as
F = FEX (FTEX MFEX )−1/2 (47)
This result can be obtained either from the first case [equation (37)] by setting W2 = I or
from the second case [equation (46)] by setting W1 = M.
     83
4. SEQUENTIAL UPDATING OF THE MASS AND STIFFNESS MATRICES
In this method, the measured modeshapes are assumed to be accurate, i.e,
FEX = FT = F. The first step is mass updating. As before, the mass matrix has to be
symmetric and psd (actually pd) and therefore MMLm LTm is defined. The minimisation
problem is given by
min J = >W−1/2
m (Lm LTm − MA )W−1/2
m >2F
s.t. FTLm LTm F = I (48)
where Wm is a symmetric weighting matrix which is selected in a similar manner to WK
in Section 3. The Lagrange function is
C = 1/2 tr{W−12
m (Lm Lm − MA )Wm (Lm Lm − MA )} + tr{a(F Lm Lm F − I)}
T −1 T T T
(49)
The matrix of the Lagrange multipliers a(m × m) is symmetric because the constraint is
symmetric. Partial derivatives of C with respect to Lm and a yield
1C
m (Lm Lm − MA )Wm Lm − FaF Lm = 0
T
= W−1 −1 T
(50)
1Lm
1C
= FTLm LTm F − I = 0 (51)
1a
If Lm is non-singular, equation (50) can be rearranged as
M = Lm LTm = MA − Wm FaFTWm (52)
Substitution of equation (52) into (51) gives
a = q−1 −1
m (I − ma )qm (53)
where qm = FTWm F and ma = FTMA F. Substituting this into equation (50) gives the
updated mass matrix
M = Lm LTm = MA + Wm Fq−1
m (I − ma )qm F Wm
−1 T
(54)
This formula holds as long as M e 0. Otherwise, analogous to stiffness updating, the
non-linear equations (50) and (51) have to be solved for Lm and a and the resulting Lm
is singular. The next step is updating the stiffness matrix. This is done similar to equation
(15) but using the optimal M obtained from equation (54) instead of the ‘true’ mass matrix,
which is not available in this case.

4.1.     


The structure of equation (54) has similarities with that of sequential least squares. The
new estimate (M) is equal to the old estimate (MA ) plus the new information in the
measurement, i.e. its deviation from what was expected (I − ma ) multiplied by some factor.
To gain better understanding of the geometric meaning of the updated mass matrix,
equation (54) can be rearranged as
M = MA + Wm Fq−1
m F (M
T
 − MA )Fq−1
m F Wm
T
(55)
Where M  is any matrix such that FTM
 F = I. Natural, but certainly non-unique, selections
 = F(FTF)−2FT or M
are M  = M1/2
A F(F MA F)
T
F M1/2
−2 T
A . As can be seen, Wm Fqm F is a
−1 T

projection matrix. Thus the correction term is the difference between M  and the model
matrix pre- and post-multiplied by a projection into span(F). Notice that this
84 .   . 
interpretation is somewhat different from the geometric interpretation of the updated
stiffness matrix. In equation (16), K consists of the measured part corrected by the
difference between the measured and the model sitffness matrices. Here, the updated mass
matrix consists of the model matrix and corrected by the difference between the measured
part and the model matrix MA . In terms of sequential least squares, this is equivalent to
deciding which one of the two sources of information, i.e. the analytical model and the
measured modeshapes, is regarded as the ‘old estimate’ and which one is the ‘new
measurement’.

4.2.   Wm = MA


By choosing Wm = MA , the optimal mass matrix, which is presented in [8], is obtained.
In that case
Ms = MA + MA Fm−1
a (I − ma )ma F MA
−1 T
(56)
By using the matrix inversion lemma and some algebraic manipulation, one can obtain
from equation (56)
s = FF + (I − Fma F MA )(MA − FF )(I − Fma F MA )
M−1 T −1 T −1 T −1 T T
(57)
This form has a geometric interpretation similar to that of the updated stiffness matrix
s = FF FF . Thus one can recognize the accurate part and the
T
in equation (16) because M−1
A − FF ) in equation
correction part restricted to an orthogonal subspace. Notice that (M−1 T

(57) can be replace by (M−1 A ) with no change in the result


Lemma 2: If MA q 0 then Ms q 0.
Proof: equation (56) can be rearranged as
Ms = (I − MA Fm−1
a F )MA (I − MA Fma F ) + MA Fma F MA
T −1 T T −2 T
(58)
The proof then continues as for Lemma 1.

5. SIMULTANEOUS UPDATING OF MASS AND STIFFNESS MATRICES


In this method, the reference basis quantity is the measured modeshapes, but unlike in
Sections 3 and 4, one combined criterion is used to update both K and M. The motivation
for this formulation is the fact that the mass and stiffness matrices are related by the
eigenvalue equation and therefore it is conceivable that a simultaneous correction will yield
better results then sequential correction. Again, to enforce the requirement that the
updated mass and stiffness matrices are symmetric and psd, it is stated that MMLm LTm and
KMLK LTK , respectively. The optimisation problem is
min J = >W−1/2
m (Lm LTm − MA )W−1/2
m >2F + >W−1/2
k (Lk LTk − KA )W−1/2
k >2F
s.t. FTLm LTm F = I
Lk LTk F = Lm LTm FV2 (59)
In this method, using general weighting is even more advantageous since Wm $ Wk . Thus
one can give different weighting to each part of the criterion and normalise between the
two parts. The Lagrange function is
C = 1/2 tr{(Lm LTm − MA )W−1
m (Lm Lm − MA )Wm }
T −1

k (Lk Lk − KA )Wk }
+ 1/2 tr{(Lk LTk − KA )W−1 T −1

+ tr{aT(FTLm LTm F − I)} + tr{lT(Lk LTk F − Lm LTm FV2)} (60)


     85
where a and l are the Lagrange multipliers and a is symmetric. The partial derivatives
of C with respect to Lm , LK , a and l are as follows:
1C
m (Lm Lm − MA )Wm Lm − lV F Lm − FV l Lm + FaF Lm = 0
T
= W−1 −1 2 T 2 T T
(61)
1Lm
1C
k (Lk Lk − KA )Wk Lk + lF Lk + Fl Lk = 0
T
= W−1 −1 T T
(62)
1Lk
1C
= FTLm LTm F − I = 0 (63)
1a
1C
= Lk LTk F − Lm Lm FV2 = 0 (64)
1l
Assuming that Lm and LK are non-singular matrices, i.e. M and K are pd, it is possible
to obtain the following expressions from equations (61) and (62):
M = Lm LTm = MA + Wm (lV2FT + FV2lT − FaFT)Wm (65)
K = Lk L = KA − Wk (lF + Fl )Wk
T
k
T T
(66)
For the derivation in this subsection, the following matrices are defined
P = Wm F(FTWm F)−1, ma = FTMA F, Qm = FTWm F, Qk = FTWk F
To eliminate the Lagrange multiplier a, substitute equation (65) into (63) and then obtain
M = MA + (I − PFT)Wm lV2FTWm + Wm FV2lTWm (I − PFT)T − P(ma − I)PT (67)
Substitution of equations (66) and (67) into (64) yields
Wk lQk + Wk FlTWk F = KA F − MA FV2 − (I − PFT)Wm lV2Qm V2 + P(ma − I)PTFV2
(68)
As can be seen, equation (68) is linear in l. It can be transformed into the standard form
of linear equations by stacking the columns of l and the right-hand side into (mn × 1)
vectors and using Kronecker products to construct the (mn × mn) coefficients matrix. It
should be noted that as in the previous cases the coefficient matrix is singular but the
equations are consistent. Therefore, one can obtain an exact, yet non-unique, solution for
l. However, substituting any such solution into equations (66) and (67) gives the unique
optimal matrices M and K.

5.1.   Wk = aWm


A closed-form solution for the simultaneous updating problem can be found for the case
where Wk = aWm where a is a scalar. From equation (68),
Wm l = [KA F − MA FV2 − a 2Wm FlTWm F + PFTWm lV2QV2 + P(ma − I)ATFV2]E−1
(69)
where Q = Qm and
E = a2Q + V2QV2
By substitution of equation (69) into (67), the terms that depend on l on the right of
equation (67) disappear due to the fact that the projection (I − PFT) is orthogonal to
Wm F. The solution for the updated mass matrix is obtained as
M = MA + (I − PFT)(KA F − MA FV2)E−1V2FTWm
+ (E−1V2FTWm )T(KA F − MA FV2)T(I − PFT)T − P(ma − I)PT (70)
86 .   . 
For a given M, equations (64) and (66) are identical with (9) and (10) which characterise
the solution in Section 3. Hence using the optimal M from equation (70) in (15) yields the
optimal K. To obtain an explicit expression for the updated stiffness matrix K, one should
substitute equation (70) into (15). After some algebraic manipulation, the solution is
given as
K = KA − a 2(I − PFT)(KA F − MA FV2)E−1FTWm
− Wm FE−1(KA F − MA FV2)T(I − PFT)T − P(FTKA F − V2)PT (75)
If either the updated stiffness matrix K or the updated mass matrix M are not psd, then
the expressions in equations (74) and (75) are not the solution for the problem (they are
still the solution for a similar problem where K and M are required to be symmetric). In
this case, the psd matrices which minimise equation (59) are the solution of the non-linear
equation in equation (61)–(64). Lm or Lk in such cases are always singular. Note that even
if only one of those matrices is singular, the solution of the other pd matrix changes.

6. EXAMPLES
To demonstrate the advantages of the new methods, consider the six degrees of freedom
system (n = 6), presented in [12] (Fig. 1).
The true values of the masses and springs are: m1 = 1, m2 = 0.2, m3 = 0.1, m4 = 1, m5 =
0.1, m6 = 0.1; k1 = 10 000, k2 = 1500, k3 = 1250, k4 = 10 000, k5 = 10 000, k6 = 10 000,
k7 = 5000, k8 = 7000.

6.1.  1
The errors in the model are at the springs k1 (k1A = 12 000, 20% error). This example
can represent a case where the ground connections cannot be modeled exactly. One, two

k1

m1

k2 k3

m3 k6 m2

k4 k5

m4

k7 k8

m5 m6

k1 k1

Figure 1. System used in the examples.


     87
and three modeshapes and natural frequencies are ‘measured’ and assumed to be true, i.e.
FEX = F and VEX = V. This example considers five cases:
Case 1 . WK = MA . This yields the cases presented in [7, 8].
Case 2 . In this case, a priori knowledge is assumed and WK = WAP which is a diagonal
matrix with large values in the first, fifth and sixth diagonal elements, which correspond
to the location of the error. This means small weighting to changes in the substructure
which suffers from poor modeling. In this case
WAP = diag{10 1 1 1 10 10}
Case 3 . Unlike case 2, where prior knowledge regarding the modeling error was assumed,
here an error localisation method is used. The COMAC method [13] was used but similar
results can be obtained from any other method. The COMAC compares the modeshapes
from the model and from the test and produces a set of values between zero and one which
indicate the quality of modeling (zero for the worst case). The weighting matrix in this
case is
WCOM = diag{1 − COMAC}
Case 4 . WK = KA . This weighting function normalises the stiffness matrix.
Case 5 . WK = I, i.e. Unweighted optimisation.
Uniform cost was defined for all cases which describes the distance of the updated matrix
from the true one as
cost k = >K − KT >2F
The values of the COMAC were calculated for two and three modeshapes and their
deviations from unity are given by:
1 − COMAC (2) = {42 43 43 41 44 44} × 10−4
1 − COMAC (3) = {42 7 26 36 40 40} × 10−4
As can be seen with two modes, the COMAC does not provide any useful information
and only with three modes it points out clearly at the location of the error.
In general, the COMAC is more effective in cases where the number of degrees of
freedom involved with the model inaccuracies is relatively small as compared to the overall
dimension of the system. In that respect, the example here is not best suited to demonstrate
its potential for automatic selection of the weighting matrices. For a more convincing
example see [18].
The updated stiffness matrix was calculated from equation (16) and the results for the
five cases are given in Table 1.

T 1
Cost k for example 1
Wk One mode (×107) Two modes (×106) Three modes (×106)
MA 1.1410 9.1890 9.8981
WAP 0.8076 4.2735 4.1026
WCOM – 7.2252 6.5785
KA 1.0290 8.1234 8.6360
I 1.0803 7.2725 6.9318
88 .   . 
T 2
Cost m for sequential updating, example 2
Wm One mode Two modes Three modes
MA 0.0842 0.0019 0.0019
WAP 0.0735 0.0015 0.0013
WCOM – 0.0041 0.0030
I 0.0911 0.0035 0.0034

The minimal cost is always achieved in case 2, where Wk was constructed with a priori
knowledge about the modeling error at the spring K1 . If this information is unavailable,
the COMAC can be used. The costs using WCOM , KA , and I are much smaller than the
first case where WK = M, for all numbers of measured modes. Another interesting
observation is that the rate of improvement with the number of measured modes is greatest
in the COMAC based method. This is because in addition to the improvement due to the
new measurement, the weighting matrix also becomes more appropriate for the physical
situation.

6.2.  2
In this example, based on the same system, it is assumed that both MA and KA are
inaccurate. The errors in the model are: K1A = 12 000 (20% error), M1A = 1.3 and
M5A = M6A = 0.13 (30% error). The values of the COMAC for two and three modes are
1 − COMAC (2) = {64 80 93 58 58 59} × 10−4
1 − COMAC (3) = {64 13 55 50 66 63} × 10−4
The solutions are shown using the sequential and the simultaneous approaches (Sections
4 and 5, respectively). Starting with the sequential approach the first step is to find the
best weighting function for mass updating. Four weighting matrices are examined and the
results are given in Table 2. Similar to example 1,
cost m = >M − MT >2F
The results show that the best weight function for the mass updating is the analytical
mass matrix (case 1), except of course the physical weight function (case 2). The updated
mass matrix obtained in case 1 is then used for the updating of the stiffness matrix. The
same five cases of example 1 were investigated and the results are given in Table 3. The
conclusions are very much the same as those from Table 1, namely that the weighting
function based on a priori knowledge is the best choice for the stiffness updating and in
case it is unavailable the COMAC-based weighting can be used instead.

T 3
Cost k for sequential updating, example 2
Wk One mode (×107) Two modes (×106) Three modes (×106)
MA 1.2670 8.7283 8.7296
WAP 0.9381 4.3715 4.3560
WCOM – 7.5999 6.9747
KA 1.1959 7.9431 7.9231
I 1.1284 7.3110 7.1730
     89
T 4
Cost k and Cost m for simultaneous updating, example 2
Wm = MA Wm = WAP
a Cost m Cost k (×10 ) 6
Cost m Cost k (×106)
a0 0.0021 8.6024 0.0015 4.3831
5a0 0.0019 8.7225 0.0013 4.3529
10a0 0.0019 8.7278 0.0013 4.3549
1 0.0319 8.2906 0.0127 5.5498

Consider now the simultaneous method. Two cases were investigated:


, Wm = MA , Wk = aWm
, Wm = WAP , Wk = aWm
One of the roles of a is to nomralise the two parts of the optimisation criterion in equation
(59). Therefore, it is reasonable to choose it in terms of
>KA >F
a 0M
>MA >F
As can be seen from Table 4, the results are only mildly affected by the change of a and
are very close to those obtained by sequential updating. When a = 1, the two parts of the
cost function are magnitudes apart and leading to worse updated mass matrices because
their contribution to the cost is marginal. The relationships between the two types of
updating and the correct way to choose the weighting in the simultaneous approach are
a matter of current investigation.

7. SUMMARY
The generalised reference basis model updating problems were defined and solved in this
paper. The generalisation is in using general weighting matrices in the optimisation
criterion while previous solutions were restricted to the mass matrix as the weight. The
free choice of the weights enables incorporation of prior knowledge and engineering
practice, or in broader terms user’s intervention, into an otherwise pure mathematical
operation. In particular, the outcome of error localisation methods can be used to
determine the weighting matrices. This brings the reference basis closer to sound
engineering considerations while keeping its simplicity and closed form solutions. The
examples in Section 6 demonstrate this point.
Another issue is the reconfiguration of the results in a meaningful form and their
interpretations. Two points of view were considered. One is geometric, involving subspaces
defined by the measurements and projection matrices into them such as FRT in Section
3 and Fq−1 m F WM in Section 4. These projections are oblique and their null spaces, or
T

‘angles’ in simpler terms, are defined by the weighting matrices. Secondly, one can think
of the two sources of data as arriving sequentially and indeed equations (16) and (55) have
the structure of recursive least squares.
The issue of positive definiteness of the updated matrices has received special attention
and the solutions guarantee this property. This was achieved by using a multiplicative
parametrisation of these matrices, rather than the additive parametrisation in previous
works which imposes only symmetry.
90 .   . 
REFERENCES
1. G. L, Q. Z, R. F and J. P 1987 Proceedings of the 5th IMAC,
1183–1190. A complete procedure for the adjustment of a mathematical model from the
identified complex modes.
2. G. L, J. P and S. C 1991 Proceedings of the 9th IMAC, 363–368.
Parametric correction of finite element models by minimization of an output residual:
improvement of the sensitivity method.
3. A. R. T 1972 AIAA Company Paper, 72–346. Derivation of mass and stiffness matrices
using vibration test data.
4. D. C. Z and M. K. K 1994 ASME Journal of Vibration and Acoustics 116,
222–231. Structural damage detection using a minimum rank theory update theory.
5. C.-P. F and S. Z 1991 Computers and Structures 20, 475–486. Updating of finite
element models by means of measured information.
6. J. E. M and M. I. F 1993 Journal of Sound and Vibration 16, 347–375. Model
updating in structural dynamics: a survey.
7. M. B and Y. B I 1978 AIAA Journal 16, 346–351. Optimal weighted
orthogonalization of measured modes.
8. A. B and E. J. N 1983 AIAA Journal 21, 1168–1173. Improvement of a large analytical
model using test data.
9. F.-S. W 1989 Proceedings of the 7th IMAC, 562–567. Structural dynamic model modification
using vibration test data.
10. F. H and C. F 1993 AIAA Journal 31, 1702–1711. Updating finite element dynamic
model using an element-by-element sensitivity methodology.
11. A. M. K 1985 AIAA Journal 23, 1431–1436. Stiffness matrix adjustment using mode data.
12. D. C. K 1988 AIAA Journal 26, 104–112. Optimum approximation of residual stiffness
in linear system identification.
13. R. J. A and D. L. B 1983 Proceedings of the 1st IMAC, 110–116. A correlation
coefficient for modal vector analysis.
14. N. A. L and D. J. E 1990 Proceedings of the 8th IMAC, 768–773. Error localization
and updating finite element models using singular value decomposition.
15. R. J. G 1965 AIAA Journal 3, 380. Reduction of stiffness and mass matrices.
16. H. P. G 1990 Proceedings of the 8th IMAC, 363–368. Comparison of expansion methods
of FE modeling localization.
17. C. A. B and S. W. S 1992 Journal of Optimization Theory and Application 74, 23–56.
Optimal matrix approximants in structural identification.
18. R. K and Y. H 1997 ASME Design Engineering Technical Conference, no.
VIB-4151. Generalised reference basis model updating in structural dynamics.

Das könnte Ihnen auch gefallen