LiTH-ISY-R-1974
Avoiding Mode Pairing when Updating Finite Element Models
Magnus Andersson
Department of Electrical Engineering Linkoping University, S-581 83 Linkoping, Sweden
www:http://www.control.isy.liu.se
email:magnusa@isy.liu.se
1997-09-01
REGLERTEKNIK
AUTOMATIC CONTROL
LINKÖPING
Technical reports from the Automatic Control group in Linkoping are available as UNIX-compressed Postscript les by anonymous ftp at the address 130.236.20.24 (ftp.control.isy.liu.se).
AVOIDING MODE PAIRING WHEN UPDATING FINITE ELEMENT MODELS
M. Andersson T. Abrahamsson
Dept. of Electrical Engineering, Linkoping University, Sweden, E-mail: magnusa@isy.liu.se
SAAB AB, Linkoping, Sweden, E-mail:Thomas.Abrahamsson@saab.se
Abstract:
Updating nite element models of complex mechanical structures requires some extra considerations. It is stressed that the two most important aspects on up- dating nite element models are parameter estimation properties and computational expenses. A novel mode-pairing free model updating formulation is found to have good parameter estimation properties. The computational expenses are reduced with a semi-xed modal basis, kept xed during several iterations.Keywords:
Parameter estimation mechanical systems nonlinear programmingmodel reduction.
1. INTRODUCTION
In aerospace applications it is for several reasons important that the FE model is accurate, i.e., that the model has a dynamical behavior close to the real structure. For instance, the security aspects on aero-planes require an experimental ground- test verication of the construction in a test set- up. The verication includes measuring of the low- frequency dynamical behavior of the structure for several congurations, e.g., for dierent amount of fuel in the wing-tanks. This measuring is very expensive and time-consuming. If an initial model of the structure can be updated/improved from the measurements of a single conguration, the other congurations can be simulated using the improved model. In order for these simulations to be valid, it is assumed that the improved model for the measured conguration is accurate, and that the dierences between the congurations are easily accounted for in the model. This is mostly the case since the dierences between the congurations are often known masses with known locations.
The specic task of improving an initial FE model using measurements is here denoted model updat- ing. An attractive approach is to parameterize the FE model physically and estimate the numerical values of these parameters. This corresponds to the minimization of a cost function
V
( ) 0 de- scribing the deviation between the model output and the measurements, = arg min2
V
( ) (1)where
denotes the open set of feasible values for the parameter vector . The minimization ofV
in (1) is in general a non-linear problem. An iterative procedure is therefore required in order to obtain an estimate^
of ,i+1= i+
i 0= init (2)where
^
= limi!1
i (3)
Hopefully, a consistent estimate of the parameters are achievable,
^
=.Updating FE models is a task of very high com- plexity. This is basically due to the size of the FE models and the high number of parameters
N
,but also the high number of actuators
m
and sen- sorsp
yielding a large amount of measurements.It is therefore important to reduce the computa- tional expenses for model updating. For instance, model reduction is inevitable when dealing with FE models of complex industrial structures.
However, the complexity of the models is not the main obstacle for obtaining a successful updating.
It is experimentally found that model updating formulations in general suers from local minima in the parameter space. In order to alleviate the risk of ending up in a local minima, the convexity properties of the cost function describing the dif- ference between the measurements and the model output must be enhanced.
2. BASIC THEORY
FE modeling of mechanical structures yield a mass matrixM and a stiness matrix Kof dimension
l
l
, respectively. The normal modes i and the resonance frequencies!
i of the undamped FE model represented by MK are determined via the generalized eigenvalue problem(K;
!
i2M)i= 0 TiMi= 1i
= 1::: l
where the normal modesiare mass normalized.(4) Notice that the modal matrix = 1:::
l] diagonalizesM andK: TM=ITK=
diag
(!
12::: !
l2) =2
where
diag
denotes a diagonal matrix. Two modes(5)i and j are considered to be near multiple if their corresponding resonance frequencies
!
i and!
j are close, i.e., if!
i!
j.A corresponding FE technique for modeling the damping is not available and ad hoc approaches to construct the damping matrix L is therefore inevitable. If
also diagonalizesL, TL=diag
(1
:::
l) =;
i= 2i!
i (6) the damping is so-called proportional. In (6), the i is the damping ratio of the normal mode i. The assumption of proportional damping is found to be valid for exible structures in practice as well as in theory, see (Geradin and Rixen, 1994).The input/output (I/O) behavior of a mechanical structure may be modeled
Mx
(t
) +Lx_
(t
) +Kx(t
) =Bu(t
)y(
t
) =Cx(t
) (7) where xis the vector of thel
degrees of freedom (dof.),uis the vector ofm
input forces, andythe vector ofp
displacement measurements. In (7),B (of dimensionl
m
) describes the positions of the actuators yielding the input forces and C (p
l
)the positions of the sensors measuring the output.
Equivalently, (7) may be written
q(
t
) +; _
q(t
) +2q(
t
) =TBu(t
)y(
t
) =C(t
) (8) where q =x is a transformation to the modal coordinatesq.The frequency response function (FRF) of (7) is dened as the transfer functionG(
s
) evaluated on the imaginary axis,s
=j!
,G(
!
) =C;K+j!
L;!
2M;1B (9) where!
denotes frequency, and where G(!
) de- notes G(j!
) for simplicity. The FRF of (8) may be conveniently written asG(
!
) =Xli=1
CiTiB
!
i2;!
2+j
2i!
i!
(10) showing the interpretation of the triplefii!
igof modal parameters in the frequency domain.
An ecient reduction of the system matrices
M
LK to the low-frequency area can be per- formed asMr=
TrMr Lr=TrLr Kr=TrKr(11) using a reduced modal matrix
r= i1:::
ir].It remains to determine the
i
j's. Traditionally,ij are chosen through a pairing of the normal modesi of the model and the corresponding estimates of the structure
^
xj,ij $
^
xjj
= 1::: N
xq (12)where
N
xqdenotes the number of estimated modes of the structure. The actual pairing of modes often relies on the correlation between the^
i and thexj. The pairing of modes also implies a pairing of the corresponding triples of modal parameters,
fij
ij
!
ijg$f^
xj^xj
!
^xjgj
= 1::: N
xqgiving the possibility to construct a reduced(13) damping matrixLr, and incorporating deviation in the resonance frequencies in the cost function.
The mode-pairing in (12) is a combinatorial opti- mization problem of rather high complexity. For large exible structures with many modes in the model/structure, sub-optimal algorithms must be used in order to decrease the time for the mode- pairing.
The estimate G
^
x of Gx at the frequencies!
Rii
= 1::: N
!can be obtained asG
^kl(!
Ri) =y
k(!
Ri)u
l(!
Ri) (14) where ^G
kl is the entry of rowk
and columnl
of G^
x. In (14),y
k(!
Ri) denotes the Fourier 2transform of the output from sensor
k
, and whereu
l(!
Ri) denotes the Fourier transform of the sinusoidal input force applied with actuatorl
. In (Williams et al., 1985), the multi-variate mode indicator function (MMIF) is presented. It is used to detect (multiple) modes of the structure as well as compute an input force vectorUi. The input signal u(t
) = Uisin(^!
xit
) is then applied to the structure in order to obtain the estimate^
xi,^
xi= ImfGx(^
!
xi)Uig (15) fori
= 1::: N
xq, see (Andersson, 1997). TheUiis the eigenvector
imin imin= minj iji
= 1::: N
xq (16) where;
Ai;
ij(Ai+Bi)Uij= 0j
= 1::: m
(17) whereAi= RefG^
x(^!
xi)gTRefG^
x(^!
xi)gandBi= ImfG
^
x(^!
xi)gTImfG^
x(^!
xi)g.3. MODEL UPDATING FORMULATION In ground vibration testing, the estimate G
^
x ofGxis obtained in the low-frequency area. The es- timated triple of modal parametersf
^
xi^xi
!
^xigare also commonly available.
It is experimentally found that the previous de- scribed mode-pairing technique may introduce se- vere local minima in the parameter space of the cost function
V
. In particular, the assumption of a valid mode pairing is critical whenever the dynamical behavior of the model is not close to the structure's, i.e., when i 6 , or when near multiple modes of the model and/or the structure are present. An objective have therefore been to develop a mode-pairing free technique to update FE models of complex mechanical structures.It is also found that a combined modal/frequency domain cost function based on both deviations in resonance frequencies and deviations in properties dened by the the FRF has good parameter estimate properties (Andersson, 1997), i.e., an estimate
^
of where^
V
(^
)V
( 0) (18)where 0 is the initial value of the parameter vector .
3.1 Model reduction Dening
^
i as^
i=
q^
aii
= 1::: N
xq (19)where
q = 1:::
Nq] (20)^
ai= arg min
a2R N
q
k
qai;^
xik2 (21)where
N
qdenotes the number of computed normal modes. The matrix of reduction !ris then dened asr=
i
:::
Nx
q]
i= ; 1
(
^
i)TM^
i1=2^
i(22) where
i is the mass-normalized version of
^
i.Notice that the mass-normalization of
^
i in (22)is based on the assumption that the mass matrix
M of the model is approximating the mass distri- bution of the structure suciently accurate.
The reduced system matricesM
randK
rbecome
Mr=
TrM
r K
r=
TrK
r (23) 3.2 Damping model design
In order to obtain a reduced damping matrix
Lr, some assumptions about the damping dis- tribution in the structure must be made. Let
x = x1:::
xNqx] denote the "true" modes of the structure, and let Lx denote the "true"damping matrix of the structure. If the damp- ing distribution in the structure are proportional, then
TxLxxwill be diagonal. Assuming that^
xiis parallel withxi, and that
iis parallel with
^
xi:^
xi k xi
and^
rik^
xii
= 1::: N
xq (24)then also
TrLx
r will be diagonal. These as- sumptions indicate that L
rmay be modeled as
Lr=
diag
(:::
2^xi!
i:::
) (25) where!
i is the square root of the Rayleigh quo- tientR
inR
(i) =
TiK
i
TiM
i = (
!
i)2i
= 1::: N
xq (26)see (Craig, Jr, 1981).
However, the matricesM
randK
rare in general not diagonal, i.e., the damping is not proportional sinceL
ris diagonal. If a proportional damping is required during the model updating, a diagonal- ization ofM
r andK
r can rst be made as
Mdr=
TMr
K
dr=TK
r (27)
where
= 1:::
Nqx] (28) In (28), the basis vectorsi is dened by(K
r;
$
2iMr)i= 0
TiM
ri= 1 (29)
for
i
= 1::: N
xq, compare with (4)-(5). The proportional damping matrixLdr may be approx- imated by the diagonal ofTL
r:
Ldr=
diag
;TLr=
diag
(:::
i:::
) (30) Notice that the damping of the updated model becomes proportional if the parameter estimation is consistent (^
= ), and if the mode estimates are unbiased, i.e., if^
xi k xi. This is true even if non-proportional damping is used during the model updating.The reduced model (for proportional damping) becomes
Mdrq
+Ldrq
_
+Kdrq=(
dr)TBu
y
=Cdrq (31) where
dr=
r (32)
The FRF corresponding to (31) becomes
Gdr(
!
) =Cdr;Z
dr(
!
);1(dr)TB (33) where
Zdr(
!
) =diag
;::: $
2i ;!
2+j
i!:::
(34) 3.3 Cost functionUsing all the asssumed measurements avilable, a mode-pairing free cost function can be obtained as
V
( ) =21V
1( ) +22V
2( ) +32V
3( ) (35) whereiare cost function weights, and whereV
1( ) =XN!i=1
G
dr(
!
xRi);G^
x(!
xRi)2F (36)V
2( ) =Nx
q
X
i=1
!
i( );!
^xi!
^xi
2
2
(37)
V
3( ) =Nx
q
X
i=1
G
dr(
!
^xi)Uxi;Yxi22 (38) whereF
denotes the Frobenius norm, and!
xRide- notes the measured frequency points ofG^
x. Mini-mizing
V
1corresponds to the maximum-likelihood (ML) estimate of , see (McKelvey, 1995). TheV
2 is included in order to increase the convexity properties of the cost function, see (Abrahamsson et al., 1996) where the enhanced parameter esti- mation properties using the Rayleigh quotient in the cost function are studied. In (38), Yxi corre- sponds to the output from the structure for the inputu(t
) =Uxisin(^!
xit
),Yxi=Gx(^
!
xi)Uxi (39)3.4 Implementation
Using numerical dierence approximations of
@
M=@
and@
K=@
, approximating derivatives ofMr
K
r are obtained as
@
Mr
@
kTrM(
k+
kk);M(
k)
r (40)
@
Kr
@
kTrK(
k+
k);K(
k)
k
r (41)
If the modal basis of reduction
r and the lin- earization ofM
r andK
r,
Mlinr ( +
)Mr( ) +X
k
@
Mr( )
@
k #k
(42)
Klinr ( +
)Kr( ) +X
k
@
Kr( )
@
k #k are semi-xed, i.e., kept xed for several itera-(43) tions, the computational expenses can be reduced, see (Andersson, 1997).
In Fig. 1, a ow-chart of a general model updating strategy is shown.
Initial Model Measurements Parameterization
Computer
New Iteration Update Model New Model
Newr? Yes
LinearizeMr Kr No
Fig. 1. A ow chart of the general model updating procedure.
Let
V
( ) =v( )Tv( ) wherev( ) =vm( );vsis the deviation vector between the output of the modelvmand the measurements of the structure
vs. Linearizingv in ,
v( + )v( ) +J( ) (44) the parameter dierence vector
iin (2) can be determined as i= arg minn
J( i)
;d( i)22 (45)+
2 22o (46)where d( i) = vm( i) ; vs, and
denotes the regularization term, see (More, 1983). In 4(Andersson, 1997), a robust strategy for determin- ing the value of
is proposed. It is a modied version of Fletcher's strategy in (Wolfe, 1978), being more robust against ending up in undesired local minima.4. NUMERICAL SIMULATION EXAMPLE Fig. 2 shows a simple aircraft-like structure.
13 26 15
8
12
9 10
11 24
1 3 7
6 5
4 14
25
16 17 18
19 23 27 28
20 21 22
2
z y
x 29
Fig. 2. The Garteur model.
In the examples below, 30 sensors are used to collect measurements. The cost function weights
i are determined so thatV
1= 1014V V
2= 114V V
3= 314V
(47) in (36)-(38) for the initial values of the parame- ters, = init.4.1 Example 1
If the actuators are located tof12z
14x 15z 20yg, the 9 low-frequency modes are estimated appro- priately, sincemini j(xi)T
^
xij2jxij2j
^
xij2 = 0:
99 (48)The parameter history then become as in Fig. 3. It is seen that the estimated parameter values
^
haveconverged to the global minima in 13 iterations.
The cost function history is shown in Fig. 4. It is seen that the linearization M
linr
K
linr is valid
since the circles () is close to the solid line. The
denotes the values of the cost function for a new computation of the modal basis of reduction
r. 4.2 Example 2
If the actuators are located to f12z
15z20yg, the 9 low-frequency modes are not estimated appropriately, since0 2 4 6 8 10 12 14
0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
iterationi
i k
=
ref k
Fig. 3. The history of the relative parameter values, Example 1.
0 2 4 6 8 10 12 14
-6 -5 -4 -3 -2 -1 0
iterationi
10log(V)
Fig. 4. The history of the cost function, Example 1.
mini j(xi)T
^
xij2jxij2j
^
xij2 = 0:
87 (49)The parameter history becomes as in Fig. 5. It can be stated that not appropriately estimated modes may turn up as bias in the parameter estimate.
0 2 4 6 8 10 12
0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5
iterationi
i k
=
ref k
Fig. 5. The history of the relative parameter values, Example 2.
However, even if the mode estimates
^
xi are notparallel with the true modes of the structure xi, the cost function decrease is substantial:
V
(^
) = 6:
710;5V
( init) (50)The relative large deviation in the estimated pa- rameter values
^
from is in view of (50) due to that the parameters are dependent, i.e., the deviation in a parameter value is compensated by the deviations in the other parameter values.5. CONCLUSIONS
It was pointed out that the two most important aspects on a model updating formulation is its parameter estimation properties and its compu- tational complexity. It was also noted that model updating in general suers from undesired local minima in the cost function. In order to reduce the risk of ending up in a local minima, the tra- ditional technique of pairing modes was circum- vented when constructing the reduced size model.
A novel combined frequency domain/modal do- main mode-pairing free cost function was in simu- lation examples shown to yield good parameter es- timates if the modes of structure are appropriately estimated. If the modes of the structure are not appropriately estimated, bias in the parameter estimates emerged. The value of the cost function are reduced considerably, though. A modal reduc- tion basis was used in several iterations in order to reduce the time-consuming computation of the normal modes.
6. REFERENCES
Abrahamsson, T., M. Andersson and T. McK- elvey (1996). A nite element model up- dating formulation using frequency responses and eigenfrequencies. In: Second Interna- tional Conference on Structural Dynamics Modelling - Test, Analysis and Correlation.
DTA/NAFEMS. pp. 293{305.
Andersson, M. (1997). Experimental design and updating of nite element models. Lic. thesis.
Automatic Control, Dept. of EE, Linkoping University, 581 83 Linkoping, Sweden.
Geradin, M. and D. Rixen (1994). Mechanical Vi- brations - Theory and Application to Struc- tural Dynamics. Wiley.
Craig, Jr, R. R. (1981). Structural Dynamics. An Introduction to Computer Methods. Wiley.
McKelvey, T. (1995). Identication of State-Space Models from Time and Frequency Data. PhD thesis. Linkoping University. Dept. of Electr.
Eng., Linkoping University, 581 83 Linkoping.
More, J. J. (1983). Recent developments inalgo- rithms and software for trust region methods.
In: Mathematical Programming. The State of the Art(A. Bachem, M. Grotchel and B. Ko- rte, Eds.). Springer-Verlag. pp. 258{287.
Williams, R., J. Crowley and H. Vold (1985).
The multivariate mode indicator function in modal analysis. In: Proceedings of the 3rd International Modal Analysis Conference.
pp. 66{70.
Wolfe, M. A. (1978). Numerical Methods for Unconstrained Optimization-an introduction.
Van Nostrand Reinhold.
6