Вы находитесь на странице: 1из 7

AN ALTERNATIVE APPROACH TO MODELING OF

STRUCTURES USING DYNAMIC STIFFNESS MATRICES

Hak-soo Kim

Kyung-Taek Yang

Structural Dynamics Lab.


Honam University
K wangjoo KOREA

Dept. of Mechatronic Engineering


Daelin College of Technology
Anyang KOREA

ABSTRACT. A nwdeling method is described to


provide a smaller structural dynamic nwdel which
can be used to compare analytical nwdel of a
structure with its experimental counter-part.
A
structural dynamic model ts assumed to be
represented by dynamic stiffness matrix. To validnte
a finite element mode~ it is often necessary to
condense a large degrees of freedom (dofs) to a
relatively small number of dofs. For these purpose,
static reduction techniques are widely used.
However, ermrs in these techniques are caused by
neglecting frequency dependent terms in the

NOMENCLATURE
[C]
[I]
[K]

[M]
[U]
[V]

[ 6,] : Singular value

c
T

-1
+1

functions relating slave dofs and master dofs. An


alternative method is proposed in this paper in
which the frequency dependent terms are considered
by expressing the reduced dynamic stiffness matrix
with ortlzogonal polynomials. The reduced nwdel ha_<;
finally a minimum set of dofs, such as sensor and
excitation points and it is under the same condition
as the physical system.
It is proposed that the
reduced model can be derived from finite element
nwdel and also can be identified from experimentally
measured frequency responses. By comparing these
nwdels,

it is possible

to check the differences

between experimental and theoretical models in


spatial coordinates.
The procedure is applied to
example structure and the results are discussed.

Damping matrix
Cnit matrix
Stiffness matrix
Mass matrix
Left singular vector
Right singular vector

s
m

: Condensed matrix
Transpose of a matrix
Inverse of a matrix
Generalized inverse of a matrix
Slave degrees of freedom
Master degrees of freedom

1. INTRODUCTION
Nowadays many finite element packages have been
developed, and they are widely used for structural
dynamic problems. However, quite often, the results
of finite element analysis differ from those of actual
tests. The discrepancy comes from the differences
between actual structure and a corresponding finite
element model. To validate a finite element model,
various techniques have been proposed in which
experimental
data
are
integrated
with
a
corresponding finite element model.
When a
structural system is complicated and includes many
different joints, the analytical model is difficult to
obtain by a theoretical method. To overcome the

674

computational burden. For the efficient structural


dynamic analysis, it is often necessary to reduce the
size of system matrices.
Equation (1) can be
expressed in partitioned form as

difficulties, great efforts have been made in the field


of experimentally obtaining mathematical models.
Although the experimental model can be compared
directly with its theoretical model, some difficulties
are caused by a large difference in the numbers of
dofs used in the two models.

(2)

For theoretical and experimental models, three


different representations are widely used as shown
in Fig.l. These are the Spatial Model, the Modal
Model and the Response Model. Theoretically, they
are equivalent forms, so it is possible to transform
one of them to another. However, when considering
the relation between experiments and analysis,
limitations imposed on each model make it
impossible to transform perfectly between these
models.

where Xm is the m.xl master set of dofs which


include exciting and measurement points, and Xs is
the (n-m)xl slave set of dofs. When a harmonic
exciting force is applied at master set of dofs,
equation(2) is expressed in frequency domain as

In this study, a modelling technique is described to


provide a means which can be used to compare
directly analytical model and its experimental
counter-part in physical coordinates. From a large
number of dofs finite element model of a structure,
it is reduced to a smaller number of dofs model for
subsequent comparison with
its
experimental
counter-part. A structural dynamic model is
assumed to be represented by dynamic stiffness
matrix. To consider a frequency dependent relations
between master and slave dofs which are ignored in
static reductions, the reduced dynamic stiffness
matrix is represented by orthogonal polynomials
instead of mass and stiffness matrices. In addition
to the reduction technique, an identification of
reduced dynamic stiffness matrix is also proposed.
From a measured FRFs at accessible points, the
reduced dynamic stiffness matrix is identified in
terms of orthogonal polynomials, which compensate
for the higher modes residing beyond the frequency
range of interest. By comparing these reduced
models, it is possible to check the 'differences
between theoretical and experimental models m
physical coordinates.

2. THEORETICAL BASIS
When a mechanical structure is discretized into n
dof system, the equations of motion can be written
as

where Xm( (LJ), X.( (LJ) and Fm( (LJ) are the Fourier
transforms of Xm(t), x.(t) and fm(t), respectively.
Dii( (LJ) is the dynamic stiffness components : i. e.,
Dii( w)

= Kti(

w) -

w 2 M;i( (LJ)

=m

for i, j

By solving the lower part of equation (3) for X.( w),


we find that :
{X.( (LJ)}

= -[D ( (LJ)r'

= -[[Ks.J- W

[D.m( w)] {Xm( w)}

(4)

[M Jr' HKsmJ- (LJ [MsmJJ {Xm( (LJ)}

After substituting equation (4) into the upper part of


equation (3), the equation of motion in reduced
coordinates can be expressed as

(5)
where
[DR( W)]

[Dmrn(

(LJ)] -

[Dm,( W)] [Dss(

(LJ)r 1

[D,m(

(LJ)] (6)

Here the mxm reduced dynamic stiffness matrix,


[DR( (LJ)], has n eigen frequencies and all the
dynamic characteristics of original systems. Static
reduction, widely used in finite element packages,
uses a transformation derived from static solution of
equation (3). The relationship between Xm and X. IS
much simplified and expressed as follows [1]
{X.( w)} = - [Ks.r' [Ksm] {Xm( w)}

[MJ {X} +[C]{ X} +[K] {X}= {f}

or s

(7)

(1)

The analytical model such as finite element model


has a large number of dofs and it takes a huge

675

As the relationship defined in equation (7) ignores


the frequency dependent inertia terms, it is valid in
low frequency regions and is not in high frequency

regions.
Also the frequency range where the
equation (7) is valid, is limited to about a half of

wG which is the smallest eigenvalue of the

From equation (11), assuming symmetric properties


[Ai]=[Ai]T, the coefficient matrices are obtained by
solving the following equation.

following eigenvalue problem [2].


To( Wt)(I]----

0
Tp( w1)[ 1] [ [A ]

To(w2)[1]---- Tp(w2)[1]

[ [KssJ- (J/[MssJJ { } = { 0}

(8)

(11')

[AP]

[DR(wQ)]

To(wQ)[l]---- Tp(WQ)[l]

[ [ DR(wl)]
=

It is thus important to select the master set of dofs

so that WG will be maximum, and some researches


[3) have been made to select optimum master sets
of dofs. However this criteria is difficult to apply
to real structures and the selected positions have no
physical meaning.
To consider a frequency
dependent term, many researches [4, 5] have been
made and Kidder [6] expanded the inverse term in
equation (4) as even order polynomials.
In this research, an alternative method has been
developed so as to consider the frequency
dependency, which is the results of dynamic
reduction. By use of Chebyshev polynomials, the
reduced dynamic stiffness matrix in equation (5) can
be expressed in the interesting frequency range (

<

(J)

< WR) as equation

If

(J) Q

is larger than the "cut off"

WG

defined in

equation (8), it is difficult to construct the equation


(11'). That is : when (J) approach to WG, [DR( w)]
can not be obtained by use of equation (6) due to

the inversion of ill-conditioned matrix, [Dss< (J) )). To


overcome such a problem, S.V.D (singular value
decomposition) is introduced.

For

(J)

close to the

"cut off" WG, the S.V.D of a matrix [D ( w)] results


in three component matrices as follows [7]

[Dss(w)] = [KsJ- w2[Mss]

Wt

= ( U](l')( ~1

(12)

(9),

where [ l] is a diagonal matrix of scalars called the


(9)

where [Ai) is the coefficient matrix and T;( (JJ)


represents an orthogonal polynomial of order i
having the orthogonal property that :

singular values ( O;), which are assumed to be


arranged in descending order. [U] and [V] are called
the left and right singular vectors, respectively and
they have the following orthogonal properties.
(13)

(10)

ll,f.J. being non-negative integers and

Q being the

From
the
orthogonal
properties,
it
follows
immediately that the generalized inverse of a matrix
[D ( w)] is

number of frequency components.

[Dss(w)] + = [ V][l'] + [
With the given analytical model such as finite
element model, exact reduced dynamic stiffness are
obtained by use of equation (6). For

Q fixing

frequency components, it is possible to establish Q


xm simultaneous equations that can be solved for
unknown coefficient matrices.

[A 0 ] T 0(w 1) + [AP]Tp(wl) = [DR(wt)]


[A 0 ] T0(w 2 )

+ [AP] Tp(wz) = [ DR(wz)]

UJ 1

(14)

where [ lf is a diagonal matrix of scalars (1/ Oi)


and if

0; is less than a tolerance limit

z, 1/ CJ,

is

set to zero.
With this approach, the reduced
dynamic stiffness matrix represents the spatial
properties of the original system regardless of the
"cut off "

(J)G.

(11)

3.

IDENTIFICATION FORMULA

Identification means construction of a mathematical

676

model which explains the dynamic characteristics of


a mechanical structure. A method is presented for
directly using the experimental measurements of
some column of FRF matrix to identify the dynamic
stiffness matrix of the system. Implementation of
the identification procedure involves two steps :
First, to consider the frequency dependency in
reduced coordinates, the dynamic stiffness matrix is
assumed
to
be
represented
by
Chebyshev
polynomials. Then, the coefficient matrices are
determined from receptance equations. For a single
point hannonic force at lth dof, m responses are
assumed to be measured.
When the dynamic
stiffness matrix associated with the measurement
points is represented by orthogonal polynomials,
measured FRF matrix is defined as equation (15).
[H( cv)J = [DR( cvlr 1

= [ ~ 0 [A 1 ]Ti(w)]

(15)
-I

possible to establish Qxm simultaneous equations


that can be solved for unknown vector {Y}.
[A] {Y} = {I}

(18)

where

[A]=

[ T 0 (w 1)[B(w 1)]:
[ T 0 (w 2)[B(w 2)]:

:
::

Tp(w 1)[B(wu]]
Tpcw 2)[B(w 2J]]

[ T 0 (wQ)[ B(wQ)] : : Tp(WQ)[B(wQ)]]

The numbers of the columns and the rows of the


matrix [A] are
(p+l)x(m+1)xm/2 and m:xQ,
respectively. It should be noted that equation (21)
can be solved by least squares method only when
the number of lows of [A] is larger than that of
columns of [A], so the number of fixing frequency
points ( Q ) must satisfy the following condition

Q
According to
the orthogonality between
the
measured FRF and dynamic stiffness matrix, the
receptance equation are
(16)

>

(m+ 1)x(p+ 1)/2

If the rank of [A]

is equal to the number of


unknown coefficients, the solution vector {Y} can be
determined by mmimizmg the residual vector
(II[A]{Y} - {1}11). The least square solution is

where { It } is the /th column of mxm unit matrix


and {l-L( cv)} is the lth column of FRF matrix. From
equation ( 16), it is possible to identify the terms
within [Ai] matrices. Assuming that [Ai]' s are
symmetric matrices, equation ( 16) can be rewritten
as
[To( cv)[B( cv)]

(19)

(20)

4. NUMERICAL SIMULATIONS
4.1 Reduction from finite element model

: : Tp( cv) [B( cv)]]{Y}={It} (17)

The beam supported at each end is modelled with


16 dofs, as shown in Fig. 2. In order to consider
four modes in the frequency range of interest, four
coordinates are selected as master set of dofs and
reduction is performed by the proposed method.
[B( cv)]

0
Here, Ti ( cv) is the orthogonal polynomial of order
i and {Y} is the coefficient vector. [B( cv)] is the
matrix composed of measured FRFs and the
numbers of rows and columns of [B( cv)] are m
and mx(m+1)/2, respectively. As all of [Ai]'s are
symmetric matrices, the number of unknown
coefficients within {Y} is (p+ 1)x(m+ 1)xm/2. For

Case 1) CVG is inside the frequency range of


interest.
For considering four modes in the frequency range
of interest (0-600Hz), master dofs are selected as
{xi xs x 11 XIs}, so that the "cut off" (t)G is
determined to be 420Hz according to equation (8).
Dynamic stiffness matrix associated with the
selected coordinates is assumed to be represented
by fifth order Chebyshev polynomial. After
selecting equally spaced fixing frequencies( .Q=60,

Q fixing frequency points of measurements, it is

677

fl. f= 10Hz) in the frequency range of interest, the

"exact" reduced dynamic stiffness matrix is


calculated for each frequency by use of equation
(6). In particular, S.V.D technique is used for
frequency
components close to (()G. Then, the
coefficient matrix [Ai]'s are obtained by solving an
overdetermined set of equations constructed from
equation (11').
To prove the accuracy of the
reduced model, FRFs calculated from the reduced
model are compared with those from the complete
finite element model, and they are also compared
with those obtained from static reduction. Fig. 3
shows the results. FRFs from the reduced model
are almost consistent with those from the complete
finite element model. On the other hand, FRFs
obtained from static reduction differ greatly from
those from the complete finite element model
especially in high frequency region.
Case 2) (.(JG is outside the frequency range of
interest.
As in the previous case, four modes are considered
in the frequency of interest.
Master dofs are
selected as {x3 xs xu xd, so the "cut off" (.()G is
determined to be 612Hz. Selecting the same
frequencies as in Case1, the proposed method is
applied. The results are summarized in Table. 1 for
comparison with those of identification. A typical
result is shown in Fig. 4. FRFs from the reduced
model coincide exactly with those from the complete
finite element model.
However, FRFs which are
obtained by using static reduction represent only
three modes in the frequency range of interest
although four coordinates are selected for reduction.
4.2 Identification from measured data
Four responses are assumed to be measured at X3,
x 5, xu and Xi3 The dynamic stiffness matrix
associated with measurement points is assumed to
be represented by fifth order Chebyshev polynomial.
After selecting 60 fixing frequencies ( Q=60, fl.
f= 10Hz)
in
the
frequency
range
of
interest(0-600Hz), an overdetermined equation of
motion is constructed as equation ( 18). Since the
mathematical model of a structure is overdetermined
, system matrix [A] is often rank deficient. By use
of S.V.D technique, equation (18) is solved for
unknown coefficient vector {Y} and the coefficient
matrices [Ai] 's are summarized in Table. 2. From
the Table 1 and Table 2, it can be noted that the
matrices [Ai]'s derived from measured FRFs are

consistent with those reduced from complete finite


element model.
The identified dynamic stiffness
matrix is used to produce FRFs, which are
compared with measured data in Fig. 5. From
Fig.5, the identified model represents the dynamic
characteristics of a structure in the frequency range
of interest.

5. CONCLUSIONS
From the results of numerical simulations, the
proposed reduced model is consistent with that from
FRFs and the valid frequency range is not limited
by the "cut-off" (.()G, which is not the case for
conventional static reduction. In addition, while the
coefficient matrices of the polynomials do not have
physical meanings such as mass and stiffness
matrices, they approximate accurately the dynamic
behavior of complete finite element model or
measured data. Thus, it is possible to compare
directly theoretical and experimental models in
physical coordinates, which enables the method to be
applicable to validation of finite element model and
structural dynamics modification.
A point to be addressed is the order of the
orthogonal polynomials required in the identification
process.
Although higher order polynomials may
better approximate the actual structure in the
frequency range of interest, system matrix becomes
rank-deficient, so that S.V.D should be used. Since
the rank of system matrix is related to the number
of modes affecting the frequency range of interest,
care should be given in the selection of the order of
the orthogonal polynomials.
This aspect will be
investigated in a future investigation.

6.ACKNOWLEDGEMENTS
The authors would like to express gratitude to
Korea Institute of Construction and Structural
Safety for providing the financial support.

7. REFERENCES
[1] Robert j. Guyan, Reduction of Stiffness and
Mass Matrices, AIAA Journal, Vol.3, pp.380, 1971.

[2] Shah, V. N. and Raymond, M., Analytical


Selection of Masters for Reduced Eigenvalue
Problems, Inter. Jour. of Num. Meth. in Engineering,

678

Vol.l8, pp.89-98, 1982.


-1.0499E+5

[3] N. Bouhaddi, S. Cogan and R. Fillod, Dynamic


Substructuring by Guyan Condensation ; Selection
of the Master dof, Proceedings of the lOth
International Modal Analysis Conference, pp.328 333, 1992.

Symmetry

[A

[4] A. J Fricker, A ~ew Approach to the Dynamic


Analysis of Structures Using Fixed Frequency
Dynamic Stiffness Matrices, Inter. Jour. of Num.
Meth. in Engineering, Vol.l9, pp.llll-1129, 1983.
[5] H. G. Lee and B. j. Dobson, The Direct
Measurement of Structural Mass, Stiffness and
Damping Properties, Journal of Sound and Vibration,
Vol.l45, pp.61-81, 1991.

-3.8250E+4

-2.4760E+5
-1.4587E +6

Symmetry

-1.2321E+6

8.7675E+5
-2.3879E-6

Symmetry

Symmetry

l.0879E+6
-2.8179E+6
-1.5212E+6

8.9270E-5
-2.3231E+6

Symmetry

4.8704E-S
-l.6451E+6
-2.3230E+6

l.0880E+6
-2.8740E+5
4.6919E+4

-4.4633E+5

4.1973E-5
-1.1622E+6

4.1703E+5
-1.1632E+6
Symmetry

-1.4638E-4
6.6622E+4
7.8222E+4
-3.8251E+4

1.1131E+6
-2.5611E+6
-1.5135E+6

-4.2109E-5
1.1450E+6
-2.7943E+5
4.5343E+4

4.9321E+5
-1.6317E -6
-2.4120E+6

-1.3561E+5
4.8699E+5
8.9002E+5
-1.2212E+6

2.5992E5
-9.6390E+5
-1.1543E+6

-6.5914E+4

l.2288E+5
-5.4200E+5
-5.3184E+5

-2.8319E+4

2.5976E+5
4.2103E+5
-4.4599E+5

4.8707E+5
8.9269E~5

-1.0389E+5

-l.2403E-6

2.6307E+5
-9.6635E+5
-1.1631E+6

1.2680E~5

-i.3499E-5

1.5328E-5
-5.2851E+5

Symmetry

-4.4659E~5

1.6479E+5
-1.0499E -5

-4.1256E+5

Symmetry

-1.2403E+6

-5.3067E-5

6.6617E+4
-2.7885E+5
-2.8478E+5

7.8223E+4
-2.8480E+5

4.5633E+04

1 Reduced spatial properties from finite


element model.

-2.8719E~5

-2.8528E+4

Table. 2 Identified spatial properties from measured


FRFs.

[7] W. H. Press, B. P. Flannery and S. A.


Teukolsky, Numerical Recipes
The Art of
Scientific Computing, Cambridge Cniversity Press,
1986.

-l.5216E+6

1.2678E+5
-5.1019E+5

matrix

Symmetry

[6] R. L. Kidder, Reduction of Structural Frequency


Equations, AIAA Journal, Vol.ll, pp.892, 1973.

Table.

l.6480E+5
-5.3071E+5

-6.6074E+4

[A

4
]

l.2490E+5
1.6981E+5
-1.1489E+5

matrix

2.6309E+5
4.1703E+5
-4.4659E+5

-3.7410E+4

7.7298E+4
-2.8541E-5

Symmetry

679

6.7150E+4
-2.7003E5
-2.7957E+5

-1.4224E+4
6.6941E+4
7.7911E+4
-3.8058E+4

Measured FRF

~(response

model)\

10'
Cll

c
0

a_

Cll

[M], [C], [K]

CJr,iPr, (r
(modal model)

10

(spatial model)

~ _____

10'~

L __ _ _ _ _ _~_ _ _J__~

200

Fig.l Representation of mathematical models.

400

600

frequency (Hz)

(a) FRF ( H(i,j), i=ll, j=ll


- - - full model
- - - - - - - static reduction
- - proposed method

Kyy

Kyy
5

= 2.2 x 10 N/m
Kee = 20 N m/rad

Kyy

E
A

20,68E+6 N/m
1.25E-4 m 2

L
I

,,

''"

0.65 m
2.6E-10 m 4

L------~-----~----~

200

____

600

(b) FRF ( H(i,j), i=l3, j=ll )

Fig.2 Finite element model of a beam structure.


--

400

frequency (Hz)

Fig.4

___ _____

Comparison of FRFs ( case2).

10'
Q)

1o'r
Q_
Q)

10 ~
7

L______[__._________]____

200

400

600

200

frequency (Hz)

(a) FRF ( H(i,j),i=ll,j=ll


- - - f u l l model
- - - - - - - staf1c reduction
proposed method

400

600

frequency (Hz)

(a) FRF ( H(i,j),i=ll,j=ll


I

true
identified

II
II
II
II
I\
I
\

'"
10

10

10'
0

200

400

600

L_------~------~------~~

200

frequency (Hz)

(b) FRF ( H(i,j),i=15,j=ll

Fig.3

400

frequency (Hz)

(b) FRF ( H(i,j},i=13,j=ll

Comparison of FRFs (case 1 ).

Fig.5

680

True and Identified FRFs.

600

Вам также может понравиться