Journal of Sound and Vibration (1999) 219(1), 522

Article No. jsvi.1998.1821, available online at http://www.idealibrary.com.on

DEVELOPMENT OF THE EXPERIMENTAL

SPATIAL MATRIX IDENTIFICATION METHOD
(THEORY AND BASIC VERIFICATION WITH A
FRAME STRUCTURE)
M. O, Q. S T. O
Department of Mechanical Engineering and Science,
Tokyo Institute of Technology, 2-12-1, O-okayama, Meguro-ku, Tokyo, Japan
(Received 16 April 1997, and in final form 28 May 1998)
In the present paper, the authors propose a new experimental method of
identifying the set of spatial matrices. This method identifies the spatial matrices,
namely mass matrix, damping matrix and stiffness matrix, which describe the
properties of dynamic structures, using experimentally measured FRFs
(Frequency Response Functions) within the frequency range of interest. In this
paper, the theory developed for this method is introduced. The validity of the
proposed method is verified by its application to an actual frame structure using
experimental FRFs measured by hammering tests.
7 1999 Academic Press

1. INTRODUCTION

In general, mechanical vibration analysis begins with the formulation of the

equations of motion in the physical domain. For example, the equations of motion
are expressed as equation (1) by assuming viscous damping:
[M]{x (t)} + [C]{x (t)} + [K]{x(t)} = { f (t)},

(1)

where [M] is the mass matrix of the objective structure, [C] is the viscous damping
matrix, [K] is the stiffness matrix, {x(t)} is the displacement response vector with
{x } and {x } representing its velocity and acceleration, and { f (t)} is the applied
external force vector.
The mass matrix, the damping matrix and the stiffness matrix are called spatial
matrices because they are formulated in the physical domain. If spatial matrices
are accurately obtained, it is possible to perform any kind of vibration and design
analysis successfully. For example, it is quite easy to obtain the modal parameters
from the spatial matrices by the numerical methods presently available for
eigenvalue problems.
The spatial matrices can be theoretically formulated. The most primitive method
for this formulation would be the manual formulation for massspring systems.
A more sophisticated example would be the finite element method. Such
theoretical approaches have the advantage that they can be carried out without
0022460X/99/010005 + 18 $30.00/0

7 1999 Academic Press

. .

using actual objective structures. However, it is often impractical to theoretically

formulate accurate spatial matrices of actual mechanical structures due to
structure complexity. Therefore, experimental approaches are desirable from a
practical viewpoint. If spatial matrices can be experimentally as well as
theoretically formulated for any complex structure, it becomes much easier to link
experimental approaches to theoretical ones in MCAE (Mechanical Computer
Aided Engineering). To the authors knowledge, however, it is generally
considered difficult to identify spatial matrices representing the high frequency
dynamic characteristics of structures from experimental FRFs (Frequency
Response Functions). Instead of spatial matrices, modal parameters, such as
natural frequencies and natural modes, are the parameters generally identified
from experimental FRFs. Many methods [1] have already been proposed to
identify modal parameters, and are widely used as basic tools for vibration
analysis. In order to link the experimental modal parameters to spatial matrices
for MCAE vibration analysis, a number of additional mathematical operations,
some possibly complex, may be required. The methods of experimentally
identifying spatial matrices have been proposed by Leuridan [2], Roemer [3],
Minas [4], Peterson [5], Okuma [6] and others. However, unfortunately, many of
these methods have not been developed to the point that they can be
experimentally applied. The authors believe that the spatial matrices to be
identified experimentally must be subject to the following conditions from the
practical viewpoint of structural vibration analysis:
(1) It must be possible to set the number of degrees of freedom of spatial
matrices to a value much larger than the number of resonant frequencies located
inside the frequency range of interest.
(2) The spatial matrices identified must be able to represent the dynamic
characteristics of the structure under arbitrary boundary conditions, even
conditions that differ from those in place at the time of the identification.
The first condition is introduced in order to meet the requirement for the
adequate number of degrees of spatial freedom necessary for structural vibration
analysis. The second condition implies the practical use of the spatial matrices for
such processes as structural modification, optimum design, and vibration control.
In this paper, a method is presented for identifying spatial matrices that
satisfactorily meets both the above conditions.
In the following section, the theory and the procedure of the method are
explained. In the third section, as an application, spatial matrices are identified
from the FRFs experimentally measured at 22 points to represent an actual frame
structure under the freefree boundary condition. Finally, some basic
investigations of spatial matrices are discussed in order to demonstrate the validity
and the usefulness of the method for practical applications.
2. THEORY

Figure 1 shows a flow-chart of the proposed identification method. It is assumed

here that a structure to be identified is under the freefree boundary condition for

Start

< The second improvement >

< The first improvement >

< Vibration measurement >

Measurement of FRF's, coherence

functions and co-ordinates of
measurement points

Improvement of the mass

matrix as positive definite
Improvement of the stiffness
matrix as a (semi-)
positive definite

< Physical modelling >

Definition of physical connectivity

Creation of constraint equations

< Set up of targets >

Frequency range

Improvement of correspondence
of undamped natural
frequencies

A scalar multiplication of spatial

matrices for the best fitting of
gain level of FRF's
Improvement of the set of
spatial matrices for better fitting
of FRF's

Output of resultant spatial

matrices

Improvement of correspondence
of normal natural modes
of interest

End

Natural frequencies

Degree of
correspondence
of modes

Modal damping ratios

Unsatisfied

Satisfied

Natural modes
Substitution of target
natural modes for discordant
modes [K] and [M]
full matrices

< Initial matrices >

Creation of initial matrices to be

improved as spatial matrices

Creation of viscous damping

matrix by copying the stiffness
matrix
Improvement of correspondence
of modal damping ratios
of interest

Degree of
correspondence
of modal damping
ratios
Unsatisfied
Satisfied

Substitution of modal
dampings calculated with
targets for discordant
modal dampings [C]
full matrix

Figure 1. Flow-chart of the theory.

vibration testing. Rubber strings or other elastic devices of small mass are expected
to be the most useful for suspending the structure.
First, the FRFs and their associating coherence functions are measured at the
chosen measurement points, and the co-ordinates of the measurement points must
be expressed in terms of a defined co-ordinate system. These two sets of data, the
experimental FRFs and the measurement point co-ordinates, are the essential
inputs for the identification method. The number of measurement points should
be adequate to depict the shapes of the resonant vibration modes existing in the
identification frequency range. External force should be applied at an appropriate

. .

point of the test structure such that all feasible rigid motions may be excited
simultaneously.
The process of physical modelling consists of creating physical connectivity
among the measurement points in order to define the structural direct connectivity
among them. This preparation method is similar to the method of defining
connectivity using elements of the finite element method. It is also very similar to
the preparation method of setting contour lines for displaying mode animation by
experimental modal parameter estimation programs. According to the
connectivity, this last method can automatically determine the physical
connection, such as which elements are necessarily set as zero-elements in the
spatial matrices. Then, the constraint equations among the other elements can be
formulated based on the following principle. The mass matrix of any system
having multi degrees of freedom must be defined by
[c]T[M][c] = [Mrigid ],

(2)

where [c] is the matrix of mutually independent rigid motion modes, [M] is the
mass matrix to be identified, and [Mrigid ] is a rigid body mass matrix.
It is well known that [c] can be formulated with the co-ordinates of the
measurement points. The proper formation of any rigid body mass matrix is well
known.

K m
m
sym.
G 0
G 0
0
m
[Mrigid ] = G
0 C
B Ixx
G
0 A Iyx
Iyy
G C
A
0 Izx
Izy
Izz
k B

L
G
G
G,
G
G
l

(3)

where m is the mass of the structure; Ixx , Iyy , Izz are inertia of moment around
x-axis, y-axis and z-axis respectively; Iyx , Izx , Izy are products of the moment
inertia; (xg , yg , zg ) is the co-ordinate of the center of gravity of the structure; and
A, B, C are parameters based on the relations: A = mxg , B = myg , C = mzg .
As a result, even without knowing the rigid body properties of the structure,
several constraint equations among the elements of the mass matrix [M] can be
created by equation (2). Namely, some elements are expressed as dependent
variables by the linear combination of the other elements, which are dealt with
as independent variables, by the constraint equations. Furthermore, if some of the
rigid body properties are known, for example the mass of the structure, it is
possible to use the known values to create the constraint equations.
With respect to the stiffness matrix, the following equation can be formulated
according to the principle that no stress appears at any point of the structure for
any feasible rigid body motion:
[K][c] = [0],

(4)

where [K] is the stiffness matrix to be identified, and [0] is a zero element matrix.

Consequently, several constraint equations can also be created from equation

(4). And the constraint equations regarding the viscous damping matrix can be
created based on the identical concept of the stiffness matrix as well.
In the process of setting up the targets, one first sets up the frequency range for
identification. The lower boundary frequency must be set at a value below the first
resonant frequency of the structure based on observing the accuracy of the phases,
coherence functions, and other attributes of the FRFs. The lower the frequency
is set, the better will be the expected accuracy. The upper boundary frequency may
be set at a high value, provided that the shapes of all natural modes existing in
the identification frequency range can be distinguished clearly by depiction with
measurement points. After that, one identifies the natural frequencies, the natural
modes and the associating modal damping ratios in the identification frequency
range. Any modal parameter estimation method [1, 7] can be used for this
procedure. Hereafter, the modal parameters are referred to as targets.
In the process of defining the initial matrices, the initial matrices have to be
set to begin the identification because of the iterative nature of the method. The
initial elements of the mass matrix and the stiffness matrix are set up by
substituting random numbers into the independent variables of the constraint
equations. This method is fast and simple, and no better way to set up the initial
matrices has yet been found.
In the process of defining the first improvement, the initial mass matrix is
improved to become a positive definite matrix. The mass matrix is normalized by
the biggest absolute value of the elements. Then, the eigenvalues are calculated
[M]{f} = l{f}.

(5)

If any of the eigenvalues of equation (5) is negative, the matrix is judged as

definitively negative. Then, the first order differentials of the negative eigenvalues
with respect to the independent variables are computed by (reference [8]):
1l/1mij = {f}T(1[M]/1mij ){f}/{f}T{f},

(6)

where mij is the element of mass matrix of row i and column j. The sensitivity
analysis using only these differentials cannot transform the negative eigenvalues
into positive ones rapidly, and thus, some other sensitivity equations must be
added to them. The additional sensitivity equations are formulated with the
differentials of the diagonal elements and the off-diagonal elements in order to
move them toward the value of one and zero respectively. These additional
equations work effectively to accelerate the movement of the initially negative
eigenvalues to positive eigenvalues. The differentials of both diagonal and
off-diagonal elements can be obtained as follows. Since all diagonal elements are
always dealt with as dependent variables in the constraint equations, the
coefficients of independent variables in the linear combinations can be taken as
simply the differentials of the diagonal elements with respect to the independent
variables. The differentials of the off-diagonal elements as dependent variables can
also be expressed by the coefficients of independent variables in the constraint
equations. It is clear that the differentials of off-diagonal elements as independent

. .

10

variables all have the unit value of one. The aforementioned normalization is
carried out at every iterative step of the sensitivity analysis.
The initial stiffness matrix is improved to become a semi-positive definite matrix
by the above method. The eigenvalues calculated by equation (7) are controlled.
It is noted here that the constraint equations of the physical modelling have
necessarily set the differentials of the eigenvalues corresponding to the feasible
rigid body natural motion modes to a value zero in advance.
[K]{f} = l{f}.

(7)

After this operation, some undamped first order natural frequencies obtained by
the use of equation (8) are suited to fit well into the natural frequencies actually
located in the identified frequency range by the sensitivity based analysis.
([K] V2[M]){f} = {0}.

(8)

The sensitivities appearing in reference [8] of the rth order eigenvalue with respect
to independent variables of the stiffness matrix and the mass matrix can be
computed by equations (9) and (10), which are derived from equation (8),
respectively:
1Vr2 /1kij = {fr }T

1[K]
{fr }/{fr }T[M]{fr },
1kij

1Vr2 /1mij = lr {fr }T

1[M]
{fr }/{fr }T[M]{fr },
1mij

(9, 10)

where {fr } is the rth eigenvector, Vr is the rth natural frequency, and lr is the rth
eigenvalue. The natural frequencies located in the identification frequency range
must be satisfactorily controlled. If they are not, the modelling is judged to be
unacceptable and an improvement of the physical modelling, i.e., of the
connectivity definition among the measurement points, is required. When the
correspondence of the natural frequencies is satisfied, their associating natural
modes are improved to correspond with the target modes by the sensitivity analysis
of natural modes with respect to the independent variables of the stiffness and
mass matrices. The sensitivities of the rth natural mode with respect to an element
of the stiffness and the mass matrices are computed by equations (11) and (12),
respectively:
1{fr }/1kij = [F]{h},

(11)

where kij is the element of stiffness matrix of row i and column j, [F] is the
eigenvector matrix, and {h} is the linear combinatioin coefficient vector whose
elements can be calculated by,
hp = {fp }T(1[K]/1kij ){fr }/Vp2 Vr2 (p $ r),
hr = 0,

(12)

1{fr }/1mij = [F]{h},

(13)

11

where mij is the element of mass matrix of row i and column j, while the elements
of {h} can be calculated as
hp = lr {fp }T(1[M]/1mij ){fr }/(Vp2 Vr2 )(p $ r),
hr = 12 {fr }T(1[M]/1mij ){fr },

(14)

When both the natural frequencies and the natural modes of interest are
controlled satisfactorily, one proceeds to the process for determining the viscous
damping matrix. Otherwise, the unsatisfactory natural modes calculated by
equation (8) are coercively replaced by their target vectors normalized with respect
to the mass matrix. One denotes the resultant modified natural mode matrix by
[F'] here. Then, the mass matrix and the stiffness matrix are modified by;
[M] = ([F']T)1[I][F']1,

[K] = ([F']T)1[V2][F']1

(15)

where [I] is an identity matrix, and [V2] is the diagonal matrix of eigenvalues. Note
that both matrices become full element matrices.
The process for determining the viscous damping matrix is considered. At first,
the initial viscous damping matrix is created by copying the resultant stiffness
matrix. Then, the matrix is multiplied by a scalar value. The value is denoted as
a, which is the proportional coefficient of the viscous damping matrix to the
stiffness matrix. The modal damping ratio of the rth natural mode can be
computed from the resultant spatial matrices. Under the normalization of the
natural modes with respect to the mass matrix, the rth modal damping ratio zr
is defined with the undamped natural frequency Vr as expressed by:
zr = 12 aVr .

(16)

The target modal damping ratios of the natural modes located in the identification
frequency range have already been estimated in the process of setting up the
targets. Therefore, by substituting the target modal damping ratios into equation
(16), the most suitable value for the proportional coefficient a can be determined
by the least mean squares method.
Furthermore, by conducting the sensitivity analysis of the eigenvalues with
respect to the independent variables of the viscous damping, the viscous damping
matrix is improved to make the modal damping ratios correspond better with the
target values, provided that the natural modes obtained by equation (17) continue
to correspond well with the natural modes already obtained by equation (8).
([C] lr [M]){fr } = {0}.

(17)

Equation (17) implies that eigenvalues beginning with the smallest positive one
should be controlled to correspond with the values defined by:
lr = 2zr Vr (r = 1n),

(17)

where lr is the rth eigenvalue to be derived from equation (17), zr is the target
modal damping ratio of the rth natural mode, Vr is the target natural frequency
of the rth order, and n is the number of resonant vibration modes actually located
in the identified frequency range.

. .

12

When this control can be achieved, one advances to the last process. When
control cannot be achieved, the unsatisfactory eigenvalues to be controlled are
replaced by the values themselves as calculated by equation (18). Therefore, the
viscous damping matrix is formulated by
[C] = ([F']T)1[L][F']1,

(19)

where [L] is a diagonal matrix the elements of which are solved by equation (18).
The matrix becomes a full element matrix and the resultant damping matrix
becomes non-proportional to both the stiffness and mass matrices.
Finally, one moves to the last process referred to as the second improvement
in Figure 1. Since the spatial matrices are computed basically under the treatment
of the normalization mentioned above, there is no guarantee that the gain of the
FRFs will correspond to that of the experimental FRFs after the previous
processes. Consequently, the best scalar value should be determined by the least
mean squares method in order that the gain of the FRFs fits well with the
experimental FRFs. The value should be multiplied by the stiffness, mass, and
damping matrices. Furthermore, the spatial matrices are improved in order to
better fit the FRFs computed from the equations of motion to the experimental
FRFs by a mathematical optimization method. The objective function to be
minimized, subject to the physical modelling constraint equations mentioned
above is:
p

J = s {Dh(vi )}H[W(vi )]{Dh(vi )}

(20)

i=1

where p is the number of sampling frequencies in the identification range, {Dh(vi )}

is the difference vector between experimental FRFs and the FRFs calculated with
spatial matrices, H is symbol of the conjugate transpose, [W(vi )] is the weighting
function (a diagonal matrix), which are derived from coherence functions and
FRFs based on the theory of the statistical likelihood estimation [9], and vi is the
frequency.
A modified steepest descent method, explained below, is used. According to the
authors examinations of the identification by its application to some kinds of
actual structures, quite large variances are observed among the sensitivities of the
objective function, equation (20), with respect to the elements of the stiffness, mass
and damping matrices. The spatial matrices having comparatively very small
sensitivities cannot change at all by a steepest descent method that handles the
independent variables of all spatial matrices simultaneously. Even after many
iterations, the objective function does not decrease very substantially by this
orthodox method. Nevertheless, the objective function may reduce more by an
optimization method that partially selects the elements of the matrices as variables
in every iterative process. That is to say, only the elements of the stiffness matrix
are dealt with as variables in an iterative process. In the next iterative process, only
the elements of the mass matrix are treated as variables. And in the still next
iterative process, only the elements of the damping matrix are dealt with as
variables. Finally, in the next process, all spatial matrices are handled

13

simultaneously. The various ways of handling the variables are repeated one after
another in turns until the objective function converges. The results of this process
are demonstrated by the practical application in section 3.
It should also be noted that the resultant spatial matrices are not the unique
solution of the structure to be identified because this is a system identification
which basically uses only experimental FRFs of a single point excitation and a
limited identification frequency range. However, the resultant spatial matrices can
represent the dynamic characteristics of the structure in the identification
frequency range even in the case of changed boundary conditions and/or
connection of some additional masses and stiffeners to this structure. Therefore,
the spatial matrices can be used for many kinds of practical analyses.
3. BASIC VERIFICATION: AN EXPERIMENTAL APPLICATION

To show the practical validity of this method, a set of spatial matrices of an

actual frame structure is identified, along with its FRFs, which are obtained by
hammering tests.
Figure 2 illustrates a schematic view of the frame structure, which is constituted
with L-shaped cross-sectional steel components. The structure is suspended by
four rubber strings to simulate the freefree boundary condition on vibration
testing. The origin of the co-ordinate system is placed at measurement point 12.
The x-axis and y-axis are set vertically and horizontally, respectively. The z-axis
is set with the right hand thumb rule as shown in Figure 2. Only the z-axial FRFs
between the single point excitation at measurement point 1 and the multi-point
responses at all 22 measurement points (122) are measured by hammering tests.
The vibration responses are measured with a small accelerometer.
The dotted line in Figure 3 shows one of the experimental FRFs used as the
input data for identification. The lower boundary frequency of the identification
frequency range is set at 10 Hz due to the very low accuracy of the experimental
FRFs below it. On the other hand, the upper boundary frequency is set at 180 Hz.
The first four resonant vibration modes below 180 Hz are observed for the first
torsional mode, the first bending mode, the second torsional mode and the second
bending mode in ascending order of frequency as shown in Figure 4. The resonant
frequencies of these modes are about 14, 86, 115 and 177 Hz, respectively. The
number of measurement points lining up along the direction of the x-axis is 11
and it can be judged that these four resonant modes can be well distinguished by
these 22 measurement points. This is the reason why 180 Hz was chosen as the
upper boundary frequency here, although it should be noted that there are many
other frequencies that might feasibly have been used. According to the actual
frame structure, connectivity data among those 22 measurement points can be
created to form a feasible physical modelling, and spatial matrices with 22 degrees
of freedom are to be identified.
The accuracy of fitting FRFs of one of the measurement points is illustrated in
Figure 3, in which the solid line denotes the FRFs calculated from the identified
spatial matrices, and the results show that the identified FRFs are consistent with
the experimental data. The FRFs of other measurement points calculated from the

. .

14

identified spatial matrices also fit those of the experiment, as shown in Figure 3.
The fitness of FRFs may be the most basic indicator for evaluation of the validity
of the identification. It is also verified that the mode shapes calculated using the
identified spatial matrices from the first to the fourth correspond very well with
those obtained from the experimental FRFs by modal identification algorithm.
Table 1 lists all natural frequencies and the modal damping ratios calculated
from the identified spatial matrices. The modelling of the vibration when
considered only in the z-axial direction involves three feasible natural rigid motion
modes. Therefore, the first three natural frequencies are 0 Hz, and the associating
damping ratios are also zero. The next four natural frequencies show the
resonances of the FRFs in the identified frequency range. The results show that
the proposed method can control all residual natural frequencies located at higher
frequencies than the identified frequency range. This control thus constitutes an
essential function of the method, as explained in the previous section.
As one step in the verification of this method, the FRFs based prediction of
locations that are not used in the identification is carried out with the identified
spatial matrices. Figure 5 shows the comparison of FRFs between the
measurement points 11 and 14 as one of the results. In the figure, the solid line
denotes the predicted FRFs from the identified spatial matrices, and the dotted
line indicates the experimental FRFs measured later for verification. The above
results indicate that the prediction has an acceptable accuracy for this real
application.
The next part of the investigation of the validity concerns the prediction
accuracy of the structure under changing boundary conditions. FRFs of the
structure defined by clamping four measurement points, 10, 11, 21 and 22, are
calculated from the identified spatial matrices under the freefree boundary
condition by simply deleting the four degrees of freedom corresponding to those
four measurement points in the spatial matrices. Figure 6 shows the prediction

A
6

10

11
0.003 m

A'

A'

0.03 m

0.003 m

0.5 m

A'

y
0.03 m
Cross section AA'
12
z

13

14
x

15

16

17

18

19

20

21

22

1m

Figure 2. A frame structure as specimen.

15

Phase (deg)

180
0
180

Compliance (m/N)

1 102

1 104

1 106

1 108

40

80
120
Frequency (Hz)

160

Figure 3. Fitting of FRFs by the identification: ----, identified; , experimental.

results. In the figure, the solid line denotes one of the predicted FRFs of the
structure under the clamping, and the dotted line represents the counterpart of
the experimentally measured FRFs. The resonant frequencies located in the
identification frequency range are observed at about 12, 20, 90 and 112 Hz on
the experimental FRFs. On the other hand, Table 2 lists all natural frequencies
calculated with the spatial matrices. It is successfully demonstrated that rigid body
natural motion modes disappear under clamped boundary condition, and the
lowest four natural frequencies are consistent with the experimental ones. All
residual natural frequencies successfully move to higher frequencies above the
(a)

(b)

(c)

(d)

Figure 4. Mode shape orders: (a) first; (b) second; (c) third; (d) fourth.

. .

16

T 1
Natural frequencies and modal damping ratios calculated from identified matrices
Order

Natural
freq. (Hz)

Damping
ratios

Order

Natural
freq. (Hz)

Damping
ratios

1
2
3
4
5
6
7
8
9
10
11

00
00
00
138
864
1140
1766
2235
2816
4003
4471

00
00
00
00288
00071
00126
00048
00348
00211
00451
00393

12
13
14
15
16
17
18
19
20
21
22

4773
4949
5230
5351
5665
5868
5947
6181
6268
6748
7371

00453
00416
00532
00475
00609
00643
00605
00676
00706
00704
00696

Phase (deg)

frequency range. Figure 7 shows the natural modes calculated with the spatial
matrices for the frequency range of interest. These are the first bending mode, the
first torsional mode, the second bending mode, and the second torsional mode in
ascending frequency order. It is experimentally verified that this order is correct.
It is noticed here that the change of the orders of the torsional mode and bending
mode of the structure under the clamping condition is correctly predicted.
It is already well known that if spatial matrices are identified experimentally,
various kinds of experimentally based structural analyses, such as analysis of
structural modification, optimum design, simulations integrated with the finite
180
0
180

Compliance (m/N)

1 102

1 104

1 106

1 108

40

80
120
Frequency (Hz)

160

Figure 5. Predicted FRFs of structure excited at another point (driving point: 11, response point:
14). Key as for Figure 3.

Phase (deg)

17

180
0
180

Compliance (m/N)

1 10

1 10

1 106
0

40

80
120
Frequency (Hz)

160

Figure 6. Predicted FRFs of structure under a different boundary condition (driving point: 1,
response point: 7). Key as for Figure 3.

element method, vibration control design, etc., may be performed in a much

simpler manner. However, if one tries to predict the dynamics of the structure
under the clamping boundary condition by using the modal parameters identified
under the freefree boundary condition by experimental modal analysis, one has
to measure FRFs by changing excitation locations at a minimum of three different
points in order to estimate the three feasible natural rigid body motion modes from
three sets of inertia terms identified together with the modal parameters of these
four resonances [10]. The predictions based on the analysis in the modal domain
could not be successfully performed using only modal parameters of these four
resonances in the identification frequency range. In the case of this application,
T 2
Predicted natural frequencies and modal damping ratios of structure under a different
boundary condition
Order

Natural
freq. (Hz)

Damping
ratios

Order

Natural
freq. (Hz)

Damping
ratios

1
2
3
4
5
6
7
8
9

151
175
928
1162
1887
2658
2902
4188
4445

00424
00092
00048
00084
00197
00059
00257
00421
00465

10
11
12
13
14
15
16
17
18

4788
5086
5203
5781
5857
6035
6212
6677
7368

00503
00541
00633
00610
00418
00696
00701
00697
00644

. .

18
(a)

(b)

(c)

(d)

Figure 7. Mode shape orders under clamping four points. Key as for Figure 4.

Phase (deg)

in which only experimental FRFs under a single point excitation at the

measurement point 1 are used, the proposed method obtains successful results, but
the prediction analysis in the modal domain does not. These results are compared
in Figure 8. In this figure, the dark dotted line expresses the experimental FRF,
the light dotted line represents the result of the present method, and the solid line
denotes the FRFs predicted by the analysis in the modal domain.
The rigid body properties can also be calculated from the identified mass
matrices. Table 3 lists the rigid body properties obtained from the identified mass
matrix and the results by other means. The values of the principal inertia of
moments in the Measurement column are obtained by the hammering tests as
180
0
180

Compliance (m/N)

1 102

1 104

1 106

1 108

40

80
120
Frequency (Hz)

160

Figure 8. Predicted FRFs of structure under a different boundary condition by modal analysis.
Key: ----, modal analysis; , present method;
, experiment.

T 3
Rigid body properties

Principal moments of inertia (kgm2)

Measurement

Approximate by
hand calculation

66
Xg = 0466
Yg = 0242
Zg = 00
I1 = 03741
I2 = 07094
I3 = 00

65
0482
0244
00
026 2 007
082 2 015
00
Assumed as
(1, 0, 0)
(0, 1, 0)
(0, 0, 1)

67
0490
0240
00
024
063
00
Assumed as
(1, 0, 0)
(0, 1, 0)
(0, 0, 1)

Principal moments of axes of inertia

I1-axis (099, 0013, 00)
I2-axis (0013, 099, 00)
I3-axis (00, 00, 10)

Mass (kg)
Center of gravity (m)

From the mass matrix

19

. .

20

Difference of FRFs

600

400

200

200

400

600

Iteration count

Figure 9. Convergence of fitness of FRFs in the second improvement. Key: ----, proposed
method; , orthodox method.

follows. The structure is suspended from two strings connected to two points on
an assumed principal axis of the moment of inertia. The structure is hit at a point
away from the principal axis with a vibration testing hammer, and the acceleration
at this point is measured in a rotational manner around the axis by a small
accelerometer. The principal moment of inertia is estimated from the applied force
and the acceleration and distance of the stricken point using a primitive calculation
based on Newtons Second Law. The values listed in the right column are obtained
by a simple calculation with the standardized material constants of steel and the
specifications of the structural size. Considering that the FRFs are measured by
hammering tests, one can conclude that the mass matrix is well identified due to
the acceptable accuracy of the rigid body properties regarding the matrix.
(a)

(b)

(c)

(d)

Figure 10. Some of mode shapes of residual orders (order nos. correspond with those of Table 1).
Key: (a) eighth; (b) ninth; (c) 21st; (d) 22nd.

21

Finally, the convergence of the objective function in equation (20) is

investigated. Figure 9 shows the comparison of the modified and orthodox steepest
descent methods. The solid line denotes the convergence of the steepest descent
method modified in this paper as stated in section 2, and the dotted line expresses
the convergence of the orthodox steepest descent method. From the figure, it is
clear that the proposed modification improves the convergence.
One can summarize the experimental application as follows. Spatial matrices
identified by the proposed method are available for such applications as structural
analysis, modification and optimum design, and vibration control design.
However, it should again be pointed out that the identified spatial matrices may
not constitute a unique solution, since they fail to represent realistic natural
vibration modes of the residual orders, as depicted in Figure 10. Further
investigations will be needed to address this problem.

4. CONCLUSIONS

The authors have presented the current theoretical basis for their experimental
spatial matrix identification method. The method can be used to identify the
spatial matrices under two necessary conditions: (1) It must be possible to set the
number of degrees of freedom of spatial matrices at a value much larger than the
number of resonant frequencies located inside the frequency range of interest. (2)
The spatial matrices identified must be able to represent the dynamic
characteristics of the structure under arbitrary boundary conditions, even
conditions that differ from those in place at the time of the identification. The
method is then verified through its application to an actual frame structure, and
relevant investigations are discussed. Finally, a remaining shortcoming of the
method, which should be addressed in future investigations, is described.

REFERENCES
1. H. V, J. K, G. T. R and R. R 1982 SAE Paper, 820194. A
multi-input modal estimation algorithm for multi-computers.
2. J. M. L, D. L. B, and R. J. A 1982 AIAA-82-0767-CP, 497.
Direct system parameter identification of mechanical structure with application to
modal analysis.
3. M. J. R and D. J. M 1990 Proceedings of the Eighth IMAC. Robust
time-domain identification of mass, stiffness and damping matrices.
4. C. M and D. J. I 1991 Transactions of ASME, Journal of Vibration and
Acoustics 133, 219224. Identification of a nonproportional damping matrix from
incomplete modal information.
5. L. D. P 1992 Proceedings of the Tenth IMAC. Efficient computation of the
eigensystem realization algorithm.
6. M. O, T. O, K. N and A. N 1989 SAE paper, 891139,
131138. Application of a new experimental identification method to an engine rigid
mount system.
7. M. O and R. M 1997 Transactions of JSME 63 (616), 41664170
Estimation of modal damping ratios from mode indicator functions (in Japanese).

22

. .

8. R. L. F and M. P. K 1968 AIAA Journal 6, 24262477. Rates of change of

eigenvalues and eigenvectors.
9. M. O, M. Y and A. N 1989 Proceedings of the ASME PVP
Conference 179, 5777. Practical and realistic weighting function for modal curve
fitting.
10. M. O and Q. S 1997 Transactions of ASME, Journal of Vibration and Acoustics
119, 341345. Identification of the principal rigid body modes under freefree
boundary condition.