See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228977533

Comparison of FRF and modal methods for combining experimental and

analytical substructures

Article · January 2007

CITATIONS READS

47 845

2 authors:

Mathew S. Allen Randall L. Mayes

Brigham Young University - Provo Main Campus Sandia National Laboratories
215 PUBLICATIONS   2,659 CITATIONS    65 PUBLICATIONS   776 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Substructuring Methods for Nonlinear Structural Dynamics View project

Structural Health Monitoring of wind turbine blades View project

All content following this page was uploaded by Randall L. Mayes on 24 September 2014.

The user has requested enhancement of the downloaded file.

 Comparison of FRF and Modal Methods for Combining
Experimental and Analytical Substructures
Matthew S. Allen1 and Randall L. Mayes2
1
University of Wisconsin-Madison, Corresponding Author: [email protected]
2
Sandia National Laboratories*, [email protected]

Abstract:
This paper investigates methods for coupling analytical dynamic models of
subcomponents with experimentally derived models in order to predict the response of the
combined system, focusing on modal substructuring or Component Mode Synthesis (CMS), the
experimental analog to the ubiquitous Craig-Bampton method. While the basic methods for
combining experimental and analytical models have been around for many years, it appears that
these are not often applied successfully. The CMS theory is presented along with a new strategy,
dubbed the Maximum Rank Coordinate Choice (MRCC), that ensures that the constrained
degrees of freedom can be found from the unconstrained without encountering numerical ill
conditioning. The experimental modal substructuring approach is also compared with frequency
response function coupling, sometimes called admittance or impedance coupling. These
methods are used both to analytically remove models of a test fixture (required to include
rotational degrees of freedom) and to predict the response of the coupled beams. Both rigid and
elastic models for the fixture are considered. Similar results are obtained using either method
although the modal substructuring method yields a more compact database and allows one to
more easily interrogate the resulting system model to assure that physically meaningful results
have been obtained. A method for coupling the fixture model to experimental measurements,
dubbed the Modal Constraint for Fixture and Subsystem (MCFS) is presented that greatly
improves the result and robustness when an elastic fixture model is used.

1. Introduction
Modal substructuring or Component Mode Synthesis (CMS) has been standard practice
for many decades in the analytical realm. Countless works have treated the subject with regards
to Finite Element model reduction, where component models are used to create reduced order
models of complex subsystems. A wide variety of flavors of Component Mode Synthesis are
available, the most common apparently being that by Craig and Bampton [1, 2]. Many standard
texts on vibratory systems or modal analysis treat this subject [1, 3, 4]. Using these methods,
accurate models of complex systems have been created with orders of magnitude fewer degrees
of freedom than would have been required if model reduction had not been employed. A number
of works have also investigated the possibility of applying the methods to nonlinear systems with
promising results [5, 6].
There has also been considerable interest in creating models from experimental
measurements that can be combined with analytical models. Subcomponents are often designed
by a number of independent groups that do not have the information or the resources to model
the macro system. In other applications some particular components may be too difficult to
model analytically with the required precision. A number of researchers have investigated the
possibility of combining analytical and experimental models. Martinez, Miller and Carne used
*
Sandia is a multiprogram laboratory operated by Sandia Corporation, a Lockheed Martin Company, for the
United States Department of Energy’s National Nuclear Security Administration under Contract DE-AC04-94AL85000.
these methods to couple experimental models for a beam like system and a shell-payload system
[7]. Urgueira compared the modal substructuring approach with the impedance (or admittance)
coupling approach treating issues such as the need for residual compliances when free modes
are used and the difficulty in measuring rotations [8]. The more recent doctoral work by Ind [9] is
most relevant to this work; he investigated the possibility of determining the response of a
delicate subcomponent by attaching a fixture, testing the combined system and then using a
variety of methods to remove the effects of the test fixture from the measured responses. All of
these works showed promising results, while also indicating significant difficulties in obtaining the
required responses with sufficient precision, such as rotational and residual responses, in dealing
with uncertainties, and in avoiding numerical ill conditioning and non-physicality in the responses.
Very large errors were sometimes obtained due to small errors in the subcomponent models [10].
This paper presents a subset of work performed at Sandia National Laboratories, aimed
at revisiting these issues on a simple system in an effort to better establish the procedures and
methodologies so that these methods can be used with more confidence. A system similar to
that discussed here was previously investigated by Simmons, Smith, Mayes and Epp in [11],
demonstrating how sensitive the admittance or impedance coupling method can be to
experimental errors when dealing with lightly damped structures. This paper compares the modal
substructuring or Component Mode Synthesis (CMS) approach with the FRF coupling admittance
(or impedance) approach. The CMS method is employed using both free-free and mass-loaded-
interface modes in an analytical study and using only the latter in the experimental study. A
companion paper discusses the experimental issues addressed and presents more detailed
admittance coupling results [12].
This paper is organized as follows. The modal substructuring theory is briefly presented
in Section 2, along with the Maximum Rank Coordinate Choice (MRCC) method that avoids ill
conditioning when enforcing constraints between structures. A method is also presented that
joins the modal degrees of freedom of a fixture to the measured responses on a subcomponent
that has been dubbed the Modal Constraint for Fixture and Subsystem (MCFS). The system of
interest is discussed in Section 3, followed by an analytical study using finite element models in
Section 4. The experimental results are also presented, and finally some conclusions in Section
5.

2. Theory
2.1. Component Mode Synthesis
The basic modal substructuring or Component Mode Synthesis theory will be presented
briefly for convenience, following the notation used in Ginsberg [3]. Modal substructuring is best
understood in the context of the Ritz series. The modal models for each subcomponent define
N×N mass, stiffness and (sometimes) damping matrices for the subcomponent, where N is the
number of modes in the model of the substructure. As in the Ritz method, the number of physical
degrees of freedom or node points Np associated with the modal model may not be the same as
the number of modes N. The model for each subcomponent is uniquely defined by its modal
parameters where each mode is defined by its natural frequency ωr and mode vector {φr}. The
damping matrix can be defined by augmenting this set with the modal damping ratios ζr or loss
factors for viscous modal damping or structural damping models respectively. A non-proportional
damping model requires the modal damping ratios and a set of complex mode vectors in place of
the real mode vectors {φr} to complete the state space description [3]. The discussion here will
be confined to real mode vectors and modal damping for simplicity, although the extension to
state space is relatively straightforward.
The modal parameters define a model for each subsystem, whose equations of motion
are the following if the mode vectors {φr} are mass normalized
 [ I ]{q} + ⎡⎣ \ 2ζ rωr \ ⎤⎦ {q} + ⎡⎣ \ ωr 2 \ ⎤⎦ {q} = [Φ ] {F }
T

{ y} = [Φ ]{q} (1),

[Φ ] = ⎡⎣{φ1} " {φN }⎤⎦

where [I] is an N×N identity matrix and ⎡⎣ 2ζ rωr \ ⎤⎦ and ⎡⎣ ωr \ ⎤⎦ are diagonal matrices containing
\ \ 2

the modal damping constants and modal natural frequencies squared, respectively. The mode
matrix [Φ] is NpxN and forces {F} are applied at any of the Np physical coordinates.
Once the equations of motion for the subsystems are defined, they can be easily joined
simply by defining constraints between the subsystems. For example, consider joining two
subsystems A and B whose individual models are defined in the form of Eq. (1), having NA and NB
modes and NpA and NpB physical coordinates respectively. We shall follow the notation in
Ginsberg [3] for simplicity, although it is more efficient to implement this using Finite Element
assembly methods. The total set of equations of motion are

⎡[ I A ] 0 ⎤ ⎧qA ⎫ ⎡ ⎣⎡ 2ζ rωr \ ⎦⎤ A ⎤ ⎧q ⎫ ⎡ ⎡ \ ωr 2 \ ⎤ ⎤ q

\
0 0
⎢ ⎥⎨ ⎬+ ⎢ ⎥ ⎨ A ⎬ + ⎢⎣ ⎦A ⎥ ⎧⎨ A ⎫⎬ =
⎣ 0 [ I B ]⎦ ⎩qB ⎭ ⎢⎣ 0 ⎡⎣ 2ζ rωr \ ⎤⎦ ⎥ ⎩q B ⎭ ⎢ 0
\
B⎦ ⎣
⎡⎣ \ ωr 2 \ ⎤⎦ ⎥ ⎩qB ⎭
B⎦

⎡[ Φ A ]T
0 ⎤ ⎧ FA ⎫
=⎢ ⎥
T ⎨ ⎬
⎢⎣ 0 [Φ B ] ⎦⎥ ⎩ FB ⎭
⎧ y A ⎫ ⎡[ Φ A ] 0 ⎤ ⎧q A ⎫
⎨ ⎬=⎢
⎩ yB ⎭ ⎣ 0 [Φ B ]⎥⎦ ⎨⎩qB ⎬⎭
(2),
where 0 denotes a matrix of zeros of the appropriate dimensions.
These systems are to be joined in some way. The connections can usually be described
by linear constraints between the physical degrees of freedom in the following form
⎧y ⎫
⎡⎣ a p ⎤⎦ ⎨ A ⎬ = {0} (3).
⎩ yB ⎭
These in turn couple the modal degrees of freedom {q} as follows
⎡[ Φ ] 0 ⎤ ⎧q A ⎫ ⎧q A ⎫
⎡⎣ a p ⎤⎦ ⎢ A ⎥ ⎨ ⎬ = [ a ] ⎨ ⎬ = {0} (4),
⎣ 0 [ Φ B ]⎦ ⎩ q B ⎭ ⎩ qB ⎭
where [a] is an Nc×(NA+NB) matrix of constraints and Nc is the number of constraints.
Equations (2) and (4) define a correct set of equations of motion and could be used to simulate
the response of the system to a variety of inputs. However, not all of the generalized coordinates
in Eq. (2) are truly free, the system only has (NA+NB) - Nc degrees of freedom, so it is beneficial to
condense the equations of motion to an unconstrained set. To do this, one must first define
which modal coordinates {qc} will be constrained and which will remain unconstrained {qu}. The
procedure is simpler if these are placed at the top of the generalized coordinate vector.
⎧q A ⎫ ⎧ qc ⎫
⎨ ⎬ = [ P] ⎨ ⎬ (5)
⎩ qB ⎭ ⎩qu ⎭
This choice, if done acceptably, does not affect the final results since only the physical degrees of
freedom are meaningful in the combined system. Ginsberg suggests that this be done manually
with a sorting matrix [P] that moves some of the modal coordinates, perhaps the last few modes
for each subsystem, to the top of the generalized coordinate vector and assigns them as
constrained generalized coordinates. The previous equation is combined with Eq. (4) resulting in
the following
⎧q ⎫ ⎧q ⎫ ⎧q ⎫
[ a ][ P ] ⎨qc ⎬ = [ aˆ ] ⎨qc ⎬ = [ aˆc aˆu ] ⎨ c ⎬ = {0} (6)
⎩ u⎭ ⎩ u⎭ ⎩ qu ⎭
that is solved as follows to express the constrained generalized coordinates in terms of the
unconstrained.

{qc } = − [ aˆc ] [ aˆu ]{qu }

−1
(7)
Unfortunately, an incorrect choice in constrained generalized coordinates can result in a
[ ] meaning that Eq. (6) cannot be solved for the particular choice of
singular matrix aˆc
constrained generalized coordinates. This is best illustrated by an example:

Example: Choosing Constrained Generalized Coordinates

Consider joining two one dimensional systems at a single point. The constraint matrix [ap]
consists of a value of 1 and -1 at the two coordinates that are to be joined. The resulting [a] matrix
is then
⎧q ⎫ ⎧q ⎫
⎡ ⎡Φ y 1 " Φ y N ⎤ − ⎡⎣Φ yc 1 " Φ yc N B ⎤⎦ ⎤ ⎨ A ⎬ = [ a ] ⎨ A ⎬ = {0} (8),
⎣⎣ c c A ⎦A B⎦ q
⎩ B⎭ ⎩ qB ⎭
where yc indexes the row of the mode matrices corresponding to the physical coordinates that are
being coupled. One generalized coordinate must now be defined as the constrained coordinate
and the preceding equation will be solved for that coordinate in terms of the others as done in Eq.
(7). Suppose that we assign the first modal coordinate of the A system (q1)A as constrained. It is
found in terms of the other generalized coordinates as

( q1 ) A = − ( Φ y 1 )
−1
⎡ ⎡Φ y 2 " Φ y N ⎤ − ⎡⎣Φ yc 1 " Φ yc N B ⎤⎦ ⎤ ⎨
(
⎧⎪ q2:N
A
) A
⎫⎪
⎬
c ⎣⎣ c c A ⎦A B⎦
⎪⎩ qB ⎭⎪
(9).
It is now clear that this is not possible if the mode coefficient for this mode at the
connection point ( Φ ) is zero. When structures are joined at multiple points the analogous
yc 1

situation occurs when the matrix [ aˆc ] is singular.

2.1.1. Maximum Rank Coordinate Choice

When the choice of constrained generalized coordinates results in a singular matrix aˆc , [ ]
one recourse is to simply choose a different set of constrained generalized coordinates [3]. The
authors instead devised the following scheme, which has been named the Maximum Rank
Coordinate Choice (MRCC), that defines a set of constrained coordinates that ensure that
[ aˆc ] will be nonsingular unless such a choice is impossible. The authors are not aware of
another method for assuring that the constrained generalized coordinates are well chosen. The
MRCC method begins by finding the singular value decomposition (SVD) of the constraint matrix
[a] as follows
[ a ] = [U a ][ Sa ][Va ]T (10).
(According to the definition of the SVD, [Ua] and [Va] are orthogonal matrices and [Sa] is a
diagonal matrix of singular values in order of decreasing magnitude.) Then let [P]=[Va] and by
Eq. (6) and because [Va] is an orthogonal matrix,
 [ aˆ ] = [U a ][ Sa ] (11).
The choice of [P] has resulted in the largest singular values of [a] and their associated singular
[ ]
vectors being placed in the leading columns of â . As a result, each column of aˆc is linearly [ ]
independent, and in fact orthogonal to every other column, so it can always be inverted and the
constrained generalized coordinates can always be found in terms of the unconstrained via Eq.
(7), so long as the constraint matrix is of rank Nc.
One can now find a reduced set of equations of motion for the unconstrained generalized
coordinates by noting that

⎧q A ⎫ ⎡ − [ aˆc ]−1 [ aˆu ] ⎤

⎨ ⎬ = [ P] ⎢ ⎥ {qu } = [ B ]{qu } (12).
⎩ qB ⎭ ⎢ ⎡⎣ I ( N A + N B − Nc ) ⎤⎦ ⎥
⎣ ⎦
The reduced equations of motion then become
[ M ]{qu } + [C ]{qu } + [ K ]{qu } = {Q}
⎧ y A ⎫ ⎡[ Φ A ] 0 ⎤ (13),
⎨ ⎬=⎢ [ B ]{qu }
⎩ yB ⎭ ⎣ 0 [Φ B ]⎥⎦
where

⎡IN 0 ⎤ ⎡ ⎡ \ 2ζ rωr \ ⎤ 0 ⎤
⎣ ⎦A
[ M ] = [ B ] ⎢ 0A ⎥ [ B] , [C ] = [ B ] ⎢⎢ ⎥ [ B]
T T

⎣⎢ I N B ⎦⎥ 0 ⎡⎣ 2ζ rωr \ ⎤⎦
\ ⎥
⎣ B⎦

⎡ ⎡ \ ωr 2 \ ⎤ 0 ⎤ ⎡[ Φ A ]T 0 ⎤ ⎧F ⎫
⎣ ⎦A
[ K ] = [ B ] ⎢⎢ ⎥ [ B] , {Q} = [ B ] ⎢
T T
⎥
T ⎨
A
⎬
⎣
0 ⎡⎣ \ ωr 2 \ ⎤⎦ ⎥
B⎦
⎢⎣ 0 [Φ B ] ⎥⎦ ⎩ FB ⎭
(14).
One can now find the modes of the combined system by solving the eigenproblem [K-ω2M]{φr}=0.
Note that all physical coordinates of the A and B systems have been retained, even though the
presence of constraint equations implies that some of the physical coordinates might now be
redundant. Typically one would remove any duplicate coordinates from {yA} and {yB} either
before or after coupling the systems. All of the operations discussed in this section are easily
automated. The authors can supply a Matlab function that implements this to interested
researchers.
It is important to note that one could also use these procedures to remove a substructure.
The only difference would be that the structure to be removed would have negative mass,
stiffness and damping. In the context of finite element models this seems very reasonable since
assembly of finite elements into a master system is clearly reversibly by simply changing the
signs on the elements’ contributions to the system matrices and repeating the assembly process.
The more general case of coupling Ritz approximations to substructures that is considered here
is similar in that one is removing mass, stiffness and damping from a system, yet the substructure
models are global Ritz models rather than local finite elements.
2.2. Estimating Connection Point Responses
The problem investigated in the following sections involves coupling two beams together
in bending and extension. The beams must be coupled in both displacement and rotation to yield
a physically meaningful model, yet rotations are difficult to measure directly. This work addresses
this issue by attaching a fixture to the end of the beam and measuring the translations at multiple
points on the fixture. The effects of the fixture are removed using the substructuring procedure
just outlined with negative signs on the subsystem matrices that are being removed. The
measurement points on the fixture are shown with black arrows in Figure 1. These can be used
to estimate the connection point response using the modal filter described in the companion
paper [12]. Unfortunately, this process was found to be very sensitive to measurement errors
and/or errors in the fixture model, as described in Section 4.2.

F1
2 0
A
F2
B
F3
C
Figure 1: System under test consisting of Beam B and Fixture A. The
measurement locations and directions on Fixture A are denoted with black
arrows, the force application locations and directions with blue arrows.

One alternative is to constrain every measured point on the fixture of the C system to
each corresponding point on the fixture model A, yet this assumes that all measurements are
error-free, and was found to give erratic results. One would prefer a method that allows the
fixture model to be joined to the measured substructure model in a least squares sense. The
authors devised the following scheme to apply a least squares constraint to a pair of structures
that does just this.
2.3. Modal Constraint for Fixture and Subsystem (MCFS)
One approach to joining subsystems A and C is to require equal displacement between
all NM of the measurement points as follows
{xC}M = {xA}M (15),
where the vectors denote the NM x1 set of measurement points on C and A respectively. Some of
the constraints above are redundant if the fixture model has fewer than NM (active) degrees of
freedom, yet eq. (15) implies that each equation is important. This can serve to amplify
measurement errors since slight inconsistencies in the measurements enforce additional
nonphysical constraints. One would prefer that eq. (15) instead be satisfied in a least squares
sense so that errors could be averaged out. This can be done by instead satisfying the following
weighted set of constraint equations,
W{xC}M = W{xA}M (16),
where the weighting matrix W has fewer rows than columns. To obtain meaningful results, the
weighting matrix must be chosen to satisfy the essential constraints. If the fixture A is well
described by a modal model with NA < NM modes, then only NA constraints are really needed to
satisfy eq. (15) to within experimental precision. One can then make the following choice for W
that couples the modal degrees of freedom of A to their approximation on C
pinv(ΦA){xC}M = pinv(ΦA){xA}M (17),
where pinv(Φ ) denotes the pseudoinverse of the NM x NA mode matrix for system A. This
A

represents a least squares solution to the constraint equation (15). This approach was found to
give dramatically better results than attempting to use a modal filter to determine the connection
point responses, as will be demonstrated in Section 4.2. This will be referred to as the Modal
Constraint for Fixture and Subsystem (MCFS) method.
 2.4. FRF Based Admittance Coupling
The Frequency Response Function (FRF) based coupling method is based on the same
principles of balancing forces and ensuring compatibility at the interfaces of the substructures, but
the operations are performed on FRFs rather than on models in the form of differential equations.
The derivation for this system is presented in the companion paper [12] so only the final result will
be repeated here:
HC00-1=HB00-1+HA00-1
HB20=HC20*HA00-1*(HA00+HB00) (18).
-1
HE21=HB20*(HB00+HD00) *HD01
The subscripts 0, 1 and 2 represent the connection point, force application point, and
response point respectively. Hence, the FRF matrices HC00 , HB00, HA00 are all three by three (x
and y displacement and rotation about z). Equations (18) are applied at every frequency line of
interest yielding the desired set of FRFs HE21. The computation at each frequency line is
independent and assumes nothing about the mathematical form of the systems except that they
are linear. The admittance predictions here were obtained by reconstructing the FRFs for the C
system from the measured modal parameters and condensing them to the connection point using
the modal filtering procedure in [12].

3. System of Interest
The problem of interest, illustrated in Figure 2, consists of joining two beams B and D at
the connection point labeled 0 to form the combined system E. System C represents physical
hardware for which a test based model will be derived. Beam B is found by performing
measurements on system C and then removing an analytical model of the fixture A attached to
beam B from the measurements. The fixture serves a number of purposes. First, a small
number of accelerometers are placed on the fixture and used to determine the rotation of the
connection point 0. This is necessary since the beams must be coupled in both rotation and
translation to yield a physically correct master structure. The fixture also serves to exercise the
beam near the connection point so that the modes of the experimental system serve as a better
basis to represent the behavior of the combined system. It is generally accepted that the free
modes of a structure do not form an ideal basis for substructure coupling since free modes have
zero moment and shear force at the interface. Furthermore, the additional mass on the end of
the beam serves to bring more modes, and hence more information into the frequency band of
interest. The admittance approach may also benefit if the mass of the fixture is on the same
order as the effective mass of the final structure at the interface, since the net change produced
by the admittance coupling would be smaller.

A
D 0 + 0
B - A =
C
1 2

E
Figure 2: Schematic of substructure coupling problem. Experimental model for
structure B is to be joined to an analytical (FE) model for structure D after the test
fixture A attached to structure B is analytically removed.

Each beam is made of steel with nominal weight density 0.28 lb/in3, modulus of elasticity
30e6 lb/in2, height 0.75 in and width 1.0 in. The fixture is a 1.0 in by 1.0 in by 4.6 in long block of
steel. The fixture was chosen to be long enough to minimize the errors in estimating the locations
of the excitation and measurement points. The fixture is bolted and adhered to beam B using a
US ¼-28, grade 5 bolt and dental cement [12].

4. Results
4.1. Simulation Example
The CMS method was first applied to join simulated experimental and analytical models
for two beams having the same properties as the experimental ones. This allows one to evaluate
the performance of the CMS method in the absence of experimental error. Structures C and D
were simulated using a finite element (FE) model with 40 beam elements per foot and the
nominal dimensions of the beams. The fixture attached to beam B was modeled as a lumped
mass, its mass and inertial properties computed from its density and dimensions and
approximating the block as a solid rectangular bar. This neglects the accelerometers, their
mounting blocks and the excitation pads shown in the figures in [12].
Two cases will be considered here. The first follows the schematic in Figure 2 where
measurements are simulated on the mass loaded system C, the fixture A is analytically removed,
and the resulting model for beam B is coupled to the FE model for D. This will be denoted the
mass loaded (ML) case. The second case simulates measurements taken on the free-free
structure B, when no fixture is attached at yc, and then couples the result to the finite element
model for D. This case will be denoted the free-free (FF) case.
Only axial motion of the beams will be considered for brevity. Similar results were
obtained when the component models were joined in the lateral direction. Measurements are
simulated in both cases by assuming that the modes of the experimental structure (C for ML case
and B for FF case) are only known up to 10 kHz. In each case the modal natural frequency and
the necessary mode shapes are assumed to have been measured perfectly. The data in the
frequency band of interest consists of one rigid body mode and one elastic mode on the
simulated experimental substructure (B or C) for both the FF and ML cases. Table 1 displays the
natural frequencies of system C that are below 10kHz with 1.) the lumped-mass fixture (C-ML),
2.) the natural frequencies of C-ML after analytically removing the fixture (B-MR), which is an
estimate for the free-free B system, and 3.) the natural frequencies of the system without the
fixture (B-FF). The analytical procedure used to remove the lumped mass is not entirely
accurate, as evidenced by the fact that the second B-MR natural frequency is not identical to the
second B-FF natural frequency. This is to be expected, since the model for C from which the
mass was removed consisted of only the rigid body mode and one elastic mode. The error
decreases as more modes are included for beam C, with zero error being obtained when all 41 of
beam C’s axial modes are used to compute the free-free mode, in which case the modal model
for the B-MR structure is complete and identical to the finite element model for the free-free
beam.

Natural Frequencies (Hz)

Mode C-ML B-MR B-FF
(mass-loaded) (mass removed) (free-free)
1 0.0 0.0 0.0
2 5128.7 9425.2 8472.5
Table 1: Natural frequencies of mass-loaded system C (C-ML), free-free system
B (B-FF) and natural frequencies of mass-loaded system C after analytically
removing the fixture (B-MR).

Both the B-FF and the B-MR structures were then coupled to the FE model for the D
structure resulting in the E system natural frequencies shown in Table 2. The result of coupling
the complete finite element model for system B is also shown (FEA), which is taken to be the
truth model, and the errors of the FF and ML cases are shown relative to the FEA case. The ML
case shows very small errors for modes 2 and 3, yet larger errors for modes 4 and 5 than the FF
case. Either method estimates all of the natural frequencies in the 10kHz frequency band with
less than 4% error. It is interesting that the ML modes predict the natural frequencies of the
combined structure well, even though they did not accurately predict the first elastic natural
frequency of the B system (after the end mass was removed.)

Natural Frequencies (Hz)

Mode E-FEA E-FF E-ML Error (E-FF) Error (E-ML)
(free-free) (mass-loaded) (free-free) (mass-loaded)
1 0.0 0.0 0.0 0.0% 0.0%
2 2824.8 2919.4 2829.7 3.3% 0.2%
3 5649.1 5860.6 5651.3 3.7% 0.0%
4 8472.5 8472.5 8678.5 0.0% 2.4%
5 11294.3 11758.9 12171.5 4.1% 7.8%
Table 2: Natural frequencies of combined system, E, using the full FEA model
for system C (FEA), using free-free modes for C (FF) and using mass-loaded
modes for system C (ML). The errors for the FF and ML cases are also shown
relative to the full FEA model.

Figure 3 shows the first three mode shapes obtained for the combined system E using
the full FEA model, and using the free-free modes and mass-loaded modes for system B. The
mode shapes obtained using the ML modes for system B overlay the true FE modes almost
exactly. On the other hand, when the FF modes are used, the mode shapes show noticeable
errors, especially near the connection point. The errors in the mode shapes for the FF case are
about 20% at most, whereas the natural frequency errors for the FF case showed only 4%
maximum error. The FF mode shapes have zero slope at the connection point (24 inches), which
is as one would expect since the shapes for the combined system are linear combinations of the
FF mode shapes, and each FF mode shapes has zero axial force at its free end. The ML mode
shapes do not suffer from this limitation since the mass on the connection point of the C system
results in its mode shapes, and hence the B-MR mode shapes, having nonzero slope at the free
end.
 Mode Shapes - 3 ft Beam - (-) True FE (-.) FF (o:) ML
15

Axial Displacement 10

-5 1
2
3
-10

-15
0 4 8 12 16 20 24 28 32 36
Position (in)

D B

E
Figure 3: First three mode shapes for combined system E using full FEA model
for beam B (solid lines), using free-free modes for beam B (dash-dot lines) and
using mass loaded modes for beam B (dotted lines and circles).

4.1.1. Discussion
The analysis just described assumed that all of the modes of both structures were
available below 10 kHz. In the problem of interest the measurable frequency range is actually 0
to 6400 Hz, as discussed in [12]. If this were used in the previous analysis, the FF model for
beam B would not have any elastic modes in the frequency band of interest while the ML model
would still have one mode in the frequency band. This was found to increase the error in the FF
natural frequencies for the E system to 9.3% and 21.4% for the first two elastic modes. The 10
kHz band was used in the preceding in order to obtain a more uniform comparison between the
performance of mass-loaded and free-free modes. However, in practice, one advantage of mass
loaded interfaces is that they can reduce the frequencies of a system’s modes, bringing more
modes into the measurable frequency range.

4.2. Experimental Application

The modal substructuring method was also applied to an experimental system. The
details of the experiment and data acquisition are provided in [12]. The basic problem is identical
to that presented in the previous section; an experimental model of beam B is to be coupled to an
analytical model for beam D. The combined equations of motion are
 ⎡ ⎡\ ⎤
0 ⎤ ⎧ qA ⎫ ⎢ − ⎣ 2ζ rωr \ ⎦ A
⎤ 0 0
⎡− [ I A ] 0 ⎥ ⎧ q A ⎫
⎢ ⎥⎪ ⎪ ⎢ ⎥ ⎪⎨ q ⎪⎬ + "
⎢ 0 [ IC ] 0 ⎥ ⎨ qC ⎬ + ⎢ 0 ⎣⎡ 2ζ rωr \ ⎦⎤ C
\
0
⎥⎪ C⎪
⎪
0 [ I D ]⎦ ⎩qD ⎭ ⎢⎪
⎣⎢ 0 ⎥
0 0 ⎡⎣ \ 2ζ rωr \ ⎤⎦ ⎥ ⎩ q D ⎭
⎣ D⎦

⎡ − ⎡ \ ωr 2 \ ⎤ ⎤
⎥ ⎧ q A ⎫ ⎢[ A ]
0 0 ⎡Φ T 0 0 ⎤ ⎧F ⎫
⎢ ⎣ ⎦A
⎥⎪ A⎪
⎪ ⎪ ⎢
⎢ ⎥ [ΦC ]
T
0 ⎡⎣ ωr \ ⎤⎦
\ 2
0 ⎨q ⎬ = 0 0 ⎥ ⎨ FC ⎬
⎢ C ⎥⎪ C⎪ ⎢ ⎥
T ⎪ ⎪
⎢
⎣
0 0 ⎡⎣ \ ωr 2 \ ⎤⎦ ⎥ ⎩qD ⎭ ⎢⎣ 0
D⎦
0 [ Φ D ] ⎥ ⎩ FD ⎭
⎦
(19).
These are supplemented with six constraint equations enforcing equal axial, lateral and rotational
displacement at the connection point. The constraint equations are used to reduce the equations
of motion to an unconstrained set using the MRCC method.
The analytical model for beam D is an Euler-Bernoulli solution for a beam in bending and
(a rod in) extension. The Euler-Bernoulli beam equation was solved for the first 10 elastic modes
and these were supplemented with the first 9 axial modes found by solving the rod equation.
These 19 modes were augmented with the three rigid body modes resulting in a 22 mode model
for beam D. The natural frequencies of this model were adjusted to match those found by testing
a 24-in beam in the laboratory.
System C consists of beam B with the fixture A attached. A modal test was performed on
system C with free-free boundary conditions and all of the bending and axial modes below 6400
Hz were extracted. An additional axial mode at 8293 Hz was also extracted. Measurements on
the fixture A attached to beam B were used to deduce the displacement and rotation of the
connection point (yc in Figure 2) as discussed in [12]. Models for beam B were obtained by
subtracting a model for the fixture A from the modal model for system C. Both rigid and elastic
models for the fixture were considered, as described in the following two subsections.
4.2.1. Rigid Fixture
The first model for the fixture treated it as a rigid body with a mass of 3.33e-3 lb-s2/in and
moment of inertia 6.99e-3 lb-in-s2 about the axis perpendicular to the plane containing the beam
and the fixture. The mass normalized rigid body mode shapes for the fixture were computed at
the measurement points using the known mass. This information was used to generate a 3-mode
model for the fixture as in Eq. (1) with ωn = 0. The mass matrix for this system was a negative
identity matrix, since the intent is to subtract the fixture A from the C system. The negative fixture
model was coupled to the Beam C model and the Beam D model according to Eq. (2) with six
constraint equations to enforce equal axial, lateral and rotational displacement at the connection
point.
The resulting model for the combined system E was then used to compute the FRFs
(HE21) between response at point 2 due to force at point 1 (see Figure 2) in both the axial and
lateral directions. Figure 4 shows the lateral force to lateral response HE21 FRF. The result of
using the Admittance procedure from [12] is also shown, as well as the analytical prediction
computed by solving the Euler-Bernoulli beam equation for a 36 in beam. Both predictions agree
with the analytical result almost perfectly, although there are some very small frequency
differences between the predictions and the analytical result. The modal substructuring (MS)
result also shows a few small amplitude spikes at 2400 and 5200 Hz, which correspond to the
frequencies of the axial modes of the beam.
 Lateral FRFs

Phase [deg]
0
-180
-360

5
10
FRF magnitude []

4
10

3
10
Analytical
CMS
Admit

0 1000 2000 3000 4000 5000 6000

Frequency (Hz)
Figure 4: Lateral FRF predicted by modal substructuring (MS) and admittance
using a rigid fixture model and that predicted analytically.

The axial force to axial response HE21 FRFs predicted by CMS and admittance are
shown in Figure 5 and compared with the analytical prediction. The CMS and Admittance
predictions are very similar, yet both underestimate the natural frequencies of both axial modes.
There were some differences in the way these methods were implemented [12], otherwise one
would expect identical results so long as numerical ill-conditioning is not encountered.
 Axial FRFs

Phase [deg]
0
-180
-360
6
10

5
10
FRF magnitude []

4
10

3
10

2
10 Analytical
CMS
Admit
1
10
0 1000 2000 3000 4000 5000 6000
Frequency (Hz)
Figure 5: Axial FRF predicted by modal substructuring and admittance using a
rigid fixture model and that predicted analytically.

It was noted that the modal substructuring approach returned purely imaginary natural
frequencies for two of the modes of the E system. The absolute value of the first was very small
suggesting that it should have been a rigid body mode. This eigenvalue was assigned a natural
frequency of zero and retained. The other had a very large natural frequency, well beyond the
frequency band of interest, so it was discarded along with its mode vector. Further investigation
found that one of the eigenvalues of the E system mass matrix was -0.0115 (in modal
coordinates). Most of the eigenvalues of the mass matrix were unity, so this value was relatively
small. This is clearly non-physical since it would imply that the system could have negative
kinetic energy for a certain combination of the modal amplitudes. Other than these two features,
the model obtained by MS was physical.
It is much more difficult to evaluate the physicality of the admittance prediction. There
are properties that one could test, such as whether the imaginary part of the FRF matrices are
always positive at the driving point [13], yet this must be done at every frequency line.
4.2.2. Elastic Fixture Joined at Connection Point
A 3D finite element analysis of the beam-fixture subsystem C revealed that the fixture
was significantly flexible in the system’s axial direction. This can be accounted for by treating the
fixture as elastic and subtracting an elastic fixture model from the measured model for C. A
simple elastic model for the fixture was created by modeling the fixture with a single bending
mode. In doing so, it was thought that since beam B is coupled to the fixture over a 0.75-in by 1-
in area, the connection might restrain bending in fixture over that area. Hence, the connection
was approximated by modeling the fixture as a guided free beam with length (4.6-0.75)/2-in with a
mass at the guided end equal to the mass of the 0.75/2 in3 segment of the fixture that was fixed
to the end of the beam. Analytical solution of the Euler-Bernoulli beam equation for this guided-
free model resulted in a first natural frequency of 17.3 kHz. This model is not expected to be
highly accurate since the beam is not long and slender as required by the Euler-Bernoulli theory,
although it might be sufficiently accurate to account for the flexibility of the fixture adequately.
Figure 6 shows the lateral FRFs predicted for the combined system using the flexible
beam model in both the admittance and modal substructuring approaches, compared with the
analytical result. Both results agree very well with the analytical model, as they did in Figure 4 for
the rigid fixture.

Lateral FRFs
Phase [deg] 0
-180
-360

5
10
FRF magnitude []

4
10

3
10
Analytical
CMS
Admit

0 1000 2000 3000 4000 5000 6000

Frequency (Hz)
Figure 6: Lateral FRF predicted by modal substructuring and admittance using
an elastic fixture model and that predicted analytically.

The elastic fixture model has a much more significant effect on the axial modes, as
illustrated in Figure 7, which compares the predicted FRFs using the one-elastic mode fixture with
the analytical prediction. Comparison to rigid fixture result in Figure 5 reveals that the natural
frequency of the first axial mode is predicted much more accurately with this elastic fixture model
for both the modal substructuring and admittance approaches. On the other hand, the second
axial mode is even more severely underestimated when the elastic fixture model is used. Both
the MS and Admittance models predict an anti-resonance near 5500 Hz. This is clearly non-
physical since Beam E is a free-free structure. An anti-resonance would imply that forces internal
to the beam can cause zero displacement at the response point (a vibration absorber), and this is
not possible at a free end since the vibration absorber would have to be located beyond the end
of the beam.
 Axial FRFs

Phase [deg]
0
-180
-360
5
10
Analytical
CMS
4 Admit
10
FRF magnitude []

3
10

2
10

1
10

0 1000 2000 3000 4000 5000 6000

Frequency (Hz)
Figure 7: Axial FRF predicted by modal substructuring and admittance using an
elastic fixture model and that predicted analytically.

The modal substructuring method returned three complex natural frequencies when the
flexible fixture was used. The first had a small absolute value, so it was assumed to be a rigid
body mode and it was assigned a natural frequency of zero. The next two were a complex
conjugate pair at 8951.1+2450.6i Hz. Their accompanying mode vectors were also a complex
conjugate pair whose coefficients were about four times stronger in the axial direction than they
were in the lateral. It was also noted that the mass matrix for the E system had all positive
eigenvalues except one that was equal to -1. The corresponding eigenvector for that eigenvalue
was a vector of zeros in physical coordinates. The eigenvalue of -1 is clearly nonphysical, as it
would imply that a state exists that possesses negative kinetic energy. On the other hand, it is
interesting that the accompanying pattern of motion exists only internally as it corresponds to zero
motion at all of the system’s measurement points.
The fixture mode shapes were used in two ways in these analyses. First, they were used
to solve a least squares problem to determine the connection point rotation and translation from
the responses measured by the five accelerometers on the fixture. This is discussed in more
detail in the companion paper [12]. Second, they were part of the dynamic model for the fixture
that was subtracted from the C system to estimate the B system. The analysis in this section was
repeated using the flexible mode shape of the fixture in the modal filtering step, but omitting it
from the modal model for the fixture that was subtracted from system C. The results obtained
were virtually identical to those presented in Figures 6 and 7. The differences between Figures 4
and 6 and Figures 5 and 7 must then be due to differences in the estimates of the connection
point translation and rotation mode shapes rather than to differences in the subcomponent
models. Interestingly, this alternative analysis (when the elastic mode of the fixture was omitted
from the CMS model) yielded a positive definite mass matrix and only two complex natural
frequencies, both of which had very small magnitudes and hence were assumed to be rigid body
modes.
4.2.2.1. Discussion
The results presented in the previous sections show that the modal substructuring
approach gives excellent results in the lateral direction; the lateral FRFs of the E system agree
very well, implying that the E system’s bending modes are predicted very accurately. The results
are also quite good in the axial direction, except for the last axial mode that is predicted to be at a
10% lower frequency than it actually is. The magnitudes of the errors in this mode in both
Figures 5 and 7 are much larger than one would expect based on the results of the simulation
presented in Section 4.1. It was noted that one could obtain results that were virtually identical to
those in Figures 6 and 7 by using the flexible mode of the fixture model in the modal filtering step
and omitting it from the system model. The additional flexible mode had a significant effect on the
modal filter that was used to estimate the connection point responses. The estimate of the first
axial frequency improved when this mode was added to the modal filter, although perhaps at the
expense of the second. These observations seem to suggest that the errors encountered in
estimating the natural frequencies of the axial modes were not due to deficiency in the size of the
modal models for the A and C systems but in the difficulty associated with accurately measuring
the mode shapes at the connection point.
4.2.3. Elastic Fixture Joined Using MCFS Method
The MCFS method in equation (17) was also employed to couple the negative fixture
model A to the C component at all of the measurement points (in a least-squares sense). The D
component was coupled to the connection point degrees of freedom of the A component of the
combined A-C system. The resulting predicted axial FRFs are shown in Figure 8. The bending
FRFs were similar to those presented previously in Figures 4 and 6. The least squares method
predicts the axial FRF very accurately, with excellent agreement at both axial natural frequencies.
This analysis yielded a combined system mass matrix having one negative eigenvalue and the
overall system had one complex rigid body mode and a pair of complex natural frequencies at 13
kHz. It was also noted that nearly identical results were obtained when the fixture model was
replaced with 1.) an Euler-Bernoulli free-free beam model (4.6-in long), 2.) a 2.3-in long guided-
free Euler-Beronoulli beam model, and 3.) the 1.925-in guided-free model described previously
that was used in Figures 4 through 8. The Admittance result and the CMS result found using a
modal filter to determine connection point responses (Figure 7) both changed significantly in the
vicinity of the second axial mode depending on which fixture model was used.
 Axial FRFs

Phase [deg]
0
-180
-360
5
10

4
10
FRF magnitude []

3
10

2
10

Analytical
CMS
1
10
0 1000 2000 3000 4000 5000 6000
Frequency (Hz)
Figure 8: Axial FRF predicted by CMS using an elastic fixture model and
coupling A and C using the least squares method. The analytical FRF is also
shown.

4.2.3.1. Discussion
Sections 4.2.2 and 4.2.3 presented the results of coupling the analytical elastic fixture
model to the measured C system in two different ways. The first used a modal filter to estimate
the three connection point coordinates from the measurements on the C system. The second
also used a modal filter based on the fixture model, but coupled the four modal coordinates of the
A system to their representation on the C system. The A system is described by a four mode
model, so the amplitude of each modal contribution can be determined in a least squares sense
from the six measurement points using the known mode matrix (i.e. {qA} = pinv(ΦA){xA}M since ΦA
is NmxNA and Nm > NA.) However, the modal amplitudes cannot be determined uniquely from the
instantaneous connection point displacements alone because there are only three connection
point DOF and four modes. If the model for the fixture (A) were known perfectly, then knowledge
of the motion at the connection point and the fixture model would be sufficient to determine the
motion of the rest of the fixture as a function of time. This is related to the frequency domain
inverse method for determining the forces acting on a structure and its variants [14, 15], which
are known to be quite sensitive to errors in the system model. It seems logical that it might be
easier to constrain the motion of C to match that of A using four points on the perimeter of the
fixture rather than using only the three degrees of freedom at the connection point even though
both are theoretically possible. Hence, it is not really surprising that the MCFS method that
employs one constraint per mode to directly constrain the modes of the A system to their
projection onto C is more accurate and much more robust than constraining the structures at the
connection point alone when the fixture (A) has more than three active modes.
This is not so much an issue when one is interested only in combining substructures
since a physically meaningful result is obtained so long as the boundaries of each substructure
are properly joined. On the other hand, when one wishes to subtract a component from a
measured model, one must assure that all of the modes the component model are adequately
constrained to the corresponding system measurements or the result may be highly sensitive to
the component model.
 4.2.4. Remarks
Before concluding, it should be noted that a close inspection of the C system’s mode
shapes revealed that the motion of the fixture in the second axial mode included some rotation
about the connection point. This suggests either that the fixture was not attached exactly in line
with its center of mass or that the mass of the accelerometers, their mounting blocks and the
force pads (which was neglected in the fixture model) was sufficient to cause asymmetry. This
may have contributed to some of the difficulties described in the preceding sections.
It was also noted that complex natural frequencies were encountered and that the
combined system mass matrix was not always positive definite. This is somewhat troubling since
it leads to a non-physical model, yet it should be expected when one subcomponent is removed
from a system because there is always the possibility that the removed mass or stiffness could be
greater than what was originally in the system. This is the analog to high frequency errors in
natural frequency predictions when two finite-length modal models are joined. One would hope
that these errors would only affect modes that are out of the frequency band of interest, as is
always the case when joining structures. This seemed to be the case here since most of the
complex natural frequencies were above the frequency band of interest. The other complex
natural frequency had small magnitude, appearing to be the missing rigid body mode, so it
seemed quite reasonable to set it to zero. Future efforts will be directed at modeling the fixture
more accurately to avoid or minimize these types of errors.

5. Conclusions
The CMS method has been compared to the impedance or admittance coupling method
for a simple structure comprised of two beams joined at their ends. The general theory employed
was described as well as a novel method that can be used to avoid ill conditioning when
eliminating constrained degrees of freedom from the combined system model, dubbed the
Maximum Rank Coordinate Choice (MRCC). The modal substructuring method was then
demonstrated on two systems modeled using finite elements, demonstrating the utility and high
accuracy of the methods, especially when mass-loaded-interface modes are employed. The
companion paper [12] showed that poor results were obtained when admittance coupling was
performed on the measured FRFs. On the other hand, good results were obtained using both the
CMS and admittance approaches when the admittance approach employed reconstructed
Frequency Response Functions (FRFs). A familiar rule of thumb that states that one should use
all of the subcomponents’ modes covering a frequency band of 1.5 to 2.0 times the frequency
band of interest for the assembly. Taken in reverse, this suggests that since the subcomponents
in this study were only described from 0 to 6400 Hz by the modal models used, the results should
only be trusted out to 3200 or 4200 Hz. Significantly better results were obtained in the axial
direction when the Modal Constraint for Fixture and Substructure (MCFS) method was employed,
which coupled the modal coordinates of the fixture to their projection onto the C system. The
MCFS method obtained excellent agreement over the entire 6400 Hz measurement band. The
good performance of the MCFS method was attributed to the fact that direct constraint between
all of the fixture modes and the measured component assures that the fixture model and the
component are adequately constrained, even if the fixture model is less than perfect.
Some non-physicality was identified in the final combined system model found by the
CMS method and its character and sources were investigated. It was noted that the methods
available to evaluate the physicality of the admittance coupling results are somewhat more
cumbersome because they must be applied at every frequency line. The modal substructuring
method has the advantages of easy integration into existing finite element codes, a compact and
easy to manipulate database for the experimental structure, and the fact that one can apply
insight gained from the voluminous works on CMS, modal substructuring and the Ritz method.

6. References
[1] R. R. J. Craig, Structural Dynamics. New York: John Wiley and Sons, 1981.
[2] R. R. J. Craig and M. C. C. Bampton, "Coupling of Substructures Using Component
Mode Synthesis," AIAA Journal, vol. 6, pp. 1313-1319, 1968.
 [3] J. H. Ginsberg, Mechanical and Structural Vibrations, First ed. New York: John Wiley and
Sons, 2001.
[4] D. J. Ewins, Modal Testing: Theory, Practice and Application. Baldock, England:
Research Studies Press, 2000.
[5] D. J. Segalman, "Model Reduction of Systems with Localized Nonlinearities," Sandia
National Laboratories, Albuquerque, NM SAND2006-1789, 2006.
[6] D. T. Griffith and D. J. Segalman, "Finite Element Calculations Illustrating a Method of
Model Reduction for the Dynamics of Structures with Localized Nonlinearities," Sandia
National Laboratories, Albuquerque, NM SAND2006-5843, 2006.
[7] D. R. Martinez, A. K. Miller, and T. G. Carne, "Combined Experimental and Analytical
Modeling of Shell/Payload Structures," presented at The Joint ASCE/ASME Mechanics
Conference, Albuquerque, NM, 1985.
[8] A. P. V. Urgueira, "Dynamic Analysis of Coupled Structures Using Experimental Data," in
Imperial College of Science, Technology and Medicine. London: University of London,
1989.
[9] P. Ind, "The Non-Intrusive Modal Testing of Delicate and Critical Structures," in Imperial
College of Science, Technology & Medicine. London: University of London, 2004.
[10] M. Imregun, D. A. Robb, and D. J. Ewins, "Structural Modification and Coupling Dynamic
Analysis Using Measured FRF Data," presented at 5th International Modal Analysis
Conference (IMAC V), London, England, 1987.
[11] L. A. Simmons, G. E. Smith, R. L. Mayes, and D. S. Epp, "Quantifying Uncertainty in an
Admittance Model Due to a Test Fixture," presented at 23rd International Modal Analysis
Conference (IMAC XXIII), Orlando, Florida, 2005.
[12] R. L. Mayes and E. C. Stasiunas, "Lightly Damped Experimental Substructures for
Combining with Analytical Substructures," presented at 25th International Modal Analysis
Conference (IMAC XXV), Orlando, Florida, 2007.
[13] T. G. Carne and C. R. Dohrmann, "Improving Experimental Frequency Response
Function Matrices for Admittance Modeling," presented at 24th International Modal
Analysis Conference (IMAC XXIV), St. Louis, Missouri, 2006.
[14] J. A. Fabunmi, "Effects of Structural Modes on Vibratory Force Determination by the
Pseudoinverse Technique," AIAA Journal, vol. 24, pp. 504-509, 1986.
[15] D. C. Kammer and A. D. Steltzner, "Structural identification using inverse system
dynamics," Journal of Guidance Control and Dynamics, vol. 23, pp. 819-825, 2000.

View publication stats