Modal Correlation
Modal Correlation
Modal Correlation
OF
ANALYTICAL
AND
EXPERIMENTAL
MODAL
MODEL
CORRELATION
TECHNIQUES
Peter Avitabile
Copyright 1998
ABSTRACT
Correlation of analytical and experimental modal models for dynamic applications as well as the correction of
analytical models to better reflect the actual structural dynamic system has become increasingly important in
many engineering analyses for a wide variety of applications. Much work has been expended in the area of
model reduction and model expansion to further this cause. Also, correlation tools have been developed to
assist in the process. However, clear identification of where discrepancies exist is not always possible with
existing tools.
The thrust of this work is to further define correlation tools that help better identify where discrepancies exist
between the analytical and experimental data bases. In addition, some work is also presented to provide
additional tools that assist in the evaluation of test measurement locations that may be critical to the success of
the correlation process.
TABLE OF CONTENTS
1
2
3
4
5
6
7
8
9
Abstract
Introduction
Review of Basic Theory for Analytical and Experimental Modal Analysis
General Model Reduction Techniques
Mode Shape Expansion
Experimental Modal Test Considerations for Finite Element Correlation
Correlation Techniques
Coordinate Orthogonality Check
Pre-Test Evaluation Techniques
Conclusions
References
Bibliography
Appendices
CHAPTER 3
GENERAL MODEL REDUCTION
TECHNIQUES
PREFACING REMARKS
This chapter presents some of the basic approaches to the reduction of finite element models
from a general standpoint. Model reduction is typically performed to obtained a reduced model
for efficiency purposes for other structural dynamic applications such as forced response
analysis and component model synthesis techniques. However, as used for this work, model
reduction is specifically used to form a mapping between the very large set of finite element
degrees of freedom and the relatively small set of tested degrees of freedom.
In general we can write a relationship between the full set of analytical or finite element dof and
the reduced set of active or condensed dof as
{x n } = x a = [T]{ x a }
x
or
{ x1} = [ T12 ]{ x2 }
(3-1)
The 'n' subscript denotes the full set of analytical dofs, the 'a' subscript denotes the active set of
dof (sometimes referred to as master dof and for correlation studies referred to as test dof) and
the subscript 'd' denotes the deleted dof (sometimes referred to as embedded or omitted dof);
MODAL MODEL CORRELATION TECHNIQUES
Chapter 3 - Model Reduction Techniques
Rev 031198
Peter Avitabile
the [T] transformation relates the transformation between these two sets of dofs. (Note: that the
subscript 1 and 2 are also sometimes ued to denote 'state 1' and 'state 2')
Since the energy of the system needs to be perserved then we can write a balance between the
energy at state 1 and state 2 as
U=
1
1
T
T
x1} [ K1 ]{ x1} = { x 2 } [ K2 ]{ x2 }
{
2
2
(3-2)
1
T
1
T
T12 ]{ x2 } } [ K1 ]{ [ T12 ] { x 2 } } = { x 2 } [ K2 ]{ x 2 }
{
[
2
2
(3-3)
1
1
T
T
T
x 2 } [ T12 ] [ K1][ T12 ]{ x2 } = { x 2 } [ K2 ]{ x 2 }
{
2
2
(3-4)
Therefore, we can see that the reduced stiffness is related to the original stiffness as
or
[ Ka ] = [T]T[ K n ][ T]
(3-5)
[ M a ] = [T]T[ M n ][ T]
(3-6)
The [T] transformation can be a variety of different matrices depending on the transformation
technique utilized.
Once these new mass and stiffness matrices are available in 'a' space then the equation of
motion becomes
[ M a ]{&&xa } + [ Ka ]{ xa } = { Fa ( t)}
(3-7)
[[ Ka ] [ Ma ]]{x a } = {0}
(3-8)
Depending on the reduction scheme utilized, the eigenvalues of the reduced system will generally
MODAL MODEL CORRELATION TECHNIQUES
Chapter 3 - Model Reduction Techniques
Rev 031198
Peter Avitabile
Guyan Condensation
For a static system, the equation of motion can be written in partitioned form as
[ K aa ]
[ Kda ]
[ Kad ] x a = Fa
[ Kdd ] xd Fd
(3-9)
again the 'a' subscript denotes the master or active set of dof and the 'd' subscript denotes the
embedded or deleted dof.
Assuming that the forces on the deleted dof are zero, then the second equation can be
expanded as
(3-10)
(3-11)
(3-12)
(3-13)
Therefore, a relationship is available relating the the active dof to the full set of dof as
{x n } = [ K
[I ]
1
{ xa } = [ Ts ] { x a }
dd ] [ K da ]
(3-14)
Using this transformation matrix, we can write the reduced system stiffness matrix as
[ K ] = [ T ] [ K ][ T ]
G
a
(3-15)
Rev 031198
Peter Avitabile
Both Guyan and Irons proposed that the same system transformation matrix used to modify the
stiffness matrix be used to modify the system mass matrix
[M ] = [T ] [M
G
a
][ Ts ]
(3-16)
This transformation attempts to convert the system mass to the set of active dofs. However,
since this technique is based soley on the static stiffness of the system, there is no guarantee that
the reduced matrix will be accurate for dynamic applications.
The solution of the reduced problem will contain eigenvalues and eigenvectors that are similar to
the eigenvalues and eigenvectors of the full system model. The degree of sililarity is heavily
dependent on the selection of the set of adof - both the total number of dof as well as the
distribution of the dof. In general the relative difference increases as the mode number increases
with the lower order modes generally having less discrepancy than the higher order modes.
(3-17)
where
[ t s ] = [ Kdd ]1 [Kda ]
[ 0]
[ ti ] = [ 0]
[ 0]
[ ]
Kdd1
1
[ M n ][ Ts ][ M a ] [ Ka ]
(3-18)
The IRS technique generally produces better approximations of the reduced eigensystem when
compared to the Guyan/Irons approach since an estimate of the inertia associated with the
deleted dof is developed as part of the reduction process. While IRS is useful as a model
MODAL MODEL CORRELATION TECHNIQUES
Chapter 3 - Model Reduction Techniques
Rev 031198
Peter Avitabile
reduction technique, expansion and correlation studies usually do not employ this method due to
other much more accurate techniques.
Dynamic Condensation
A dynamic implementation of the Guyan reduction process is the Dynamic Condensation
process which is often used in correlation studies, in particular for expansion of mode shapes.
Let's introduce a shift value, f, into the set of equation describing the dynamic system as
[[ Kn ] ( f )[ M n ]]{x n } = {0}
(3-19)
and rearrange terms to group the constant term f times the mass matrix with the stiffness matrix
to yield
(3-20)
[ D n ] = [ Kn ] + f [ M n ]
(3-21)
Using the same approach as done with Guyan condensation, these equations can be written in
partitioned form as
[ D aa ]
[ Dda ]
[ Dad ] x a = Fa
[ Ddd ] xd Fd
(3-22)
Assuming that the forces on the deleted dof are zero, then the second equation of the partitioned
set can be expanded as
(3-23)
(3-24)
(3-25)
5
Rev 031198
Peter Avitabile
Upon substitution
(3-26)
Therefore, the relationship between the active dofs and the full set of analytical dofs can be
written as
{x n } = [ D
[I]
1
{ x a } = [ Tf ] { x a }
dd ] [ D da ]
(3-27)
[ K ] = [T ] [ K ][ T ]
[ M ] = [T ] [ M ][ T ]
T
f
a
f
a
(3-28)
(3-29)
Due to the formulation of the dydnamic condensation process, the eigensolution of the reduced
matrices will result in on eigenvalue which will correspond to the shift value used for the
reduction process. If the shift value happens to correspond exactly to one of the eigenvalues of
the system, then this eigenvalue will be preserved accurately in the reduced mode and will also
produce and expanded eigenvector which will be exactly the same as the corresponding
eigenvector from the full finite element model relating to the shifted eigenvalue. None of the
other eigenvalues will correspond to any of the eigenvalues of the full system.
{x n } = x a = [T] { xa }
x
(3-30)
{x n } = x a = Ua { p}
x
(3-31)
Notice that the modal matrix is also partitioned into the 'a' active and 'd' deleted set of degrees
of freedom. Looking at just the relationship for the 'a' set of degrees of freedom, we can write
MODAL MODEL CORRELATION TECHNIQUES
Chapter 3 - Model Reduction Techniques
Rev 031198
Peter Avitabile
{ x a } = [ U a ] { p}
(3-32)
The inverse specification of this equation involves a generalized inverse since the number of
unkowns is not equal to the number of equations need to be solved. There are two possible
solutions to this situation
when the number of equations 'a' are greater than or equal to the number of solution
variables 'm' (an overspecification or equivalence of the system)
when the number of equations 'a' are less than the number of solution variables 'm' (an
underspecification of the system)
Least Squares Solution - a m
{x a } = [ Ua ]{ p}
[ U a ] T { x a } = [ U a ] T [ U a ] { p}
{ p} = [ Ua ]T [ U a ] [ Ua ]T { xa } = [ Ua ]g { xa }
Average Solution - a m
{ p} = [ Ua ]T [ Ua ]T [ Ua ] { x a } = [ Ua ]g { xa }
For most structural dynamic applications in dynamic testing, the least sqaures solution is used
since the number of master dof (or tested dof) is far greater than the number of modes in the
system, then the generalized inverse is
{ p} = [ Ua ]T [ Ua ] [ Ua ]T { x a } = [ Ua ]g { xa }
(3-33)
This equation for the modal displacement can be substituted into the modal transformation
eqaution to give
{x n } = [ U n ][ Ua ]g { xa } = [Tu ]{ xa }
(3-34)
where
Rev 031198
Peter Avitabile
[ TU ] = [ Un ][ Ua ]g
(3-35)
This is the SEREP transformation matrix that is used for either the reduction of the finite element
mass and stiffness matrices or for the expansion of the measured experimental modal vectors.
The System Equivalent Reduction Expansion Process (SEREP) relies on a finite element model
or analytical model from which an eigensolution is performed to develop the mapping between
the full set of finite element dof and the reduced set of 'a' degress of freedom. The eigensolution
of the full set of system matrices yields a set of modal vectors which can be partitioned into
those degrees of freedom that correspond to the active set of 'a' dof and the inactive set of 'd'
dof.
Un =
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
xxxxxxxxxxxxxxxxxxx
Ua =
[ M ] = [ T ] [ M ] [T ]
[ K ] = [ T ] [ K ] [T ]
T
S
a
(3-36)
S
a
The equation of motion for the 'a' set of degrees of freedom can be written as
[ M ]{&&x } + [ K ]{x } = {F (t )}
S
a
S
a
(3-37)
Substituting in the SEREP transformation matrix in equation (3-35) into the reduced mass
matrix in equation (17) gives
Rev 031198
Peter Avitabile
[M ] = [U ]
S
a
gT
[ Un ]T [ M n ][ Un ][ Ua ]g
(3-38)
From mass orthogonality in equation (2-10), equation (3-38) can be written for the reduced
mass as
[M ] = [U ]
S
a
gT
[ Ua ] g
(3-39)
Note that the original system mass matrix is not needed in order to compute the reduced mass
matrix.
Similarly for the reduced stiffness matrix, the SEREP transformation matrix can be substituted
into the reduced stiffness matrix in equation (3-16) to give
[K ] = [U ]
S
a
gT
[ Un ]T[ K n ][ Un ][ Ua ]g
(3-40)
From stiffness orthogonality in equation (2-11), equation (3-40) can be written for the reduced
stiffness matrix as
[ Ka ] = [ Ua ] g
[ ][ U ]
2
(3-41)
Note that the original system stiffness matrix is not needed in order to compute the reduced
stiffness matrix.
While the size the these reduced mass and stiffness matrices is a by a, the rank of the reduced
matrices is only m. Therefore, use of these matrices must be done so with caution. Due to this
rank deficiency, an alternate form of the SEREP reduction process which invokes an exact
solution can be obtained by using a=m for the reduction. This technique is referred to as
SEREPa.
In order to better understand this rank deficiency problem, the following section reviews the
reduced eigenproblem using the SEREP reduced matrices. Singular valued decomposition is
used to illustrate some key features of the SEREP reduced matrices. (SVD is a procedure
which allows for the inversion of coefficient matrices.)
Any matrix [A] can be decomposed into its orthogonal matrices and singular values as
[ A] = [ L][S][ R]T
where
MODAL MODEL CORRELATION TECHNIQUES
Chapter 3 - Model Reduction Techniques
Rev 031198
Peter Avitabile
[ 0]
[S] = [ 0r ] 0
[ ] [ ]
Reduced Modal Matrices
Using SVD, the active set of dof in the modal matrix can be decomposed into
[ Ua ] = [ L][S][ R]T
where
[S] = [ 0m ]
[ ]
Then the generalized inverse can be formed as
[ Ua ]g = [ R][S]g [ L]T
where
[ S ] g = [ m ] 1 [ 0]
[ S]
gT
[ ]
m2
g
S
=
[ ]
[ 0]
[ 0]
[ 0]
10
Rev 031198
Peter Avitabile
([S]
where
gT
[ 0]
[S1] = [ 0m ] 0
[ ] [ ]
[[ Ka ] [ Ma ]]{x a } = [0]
which can be expressed as
([ ] [ ] [ ] [ ])
1 [ R] T 2 [ R ] 1 2
m
m
[ L] m
[ 0]
[ L] T { x a } = { 0}
([0] [ 0])
[ 0]
which is
[ Ua ]
gT
] ([ ]
2 [ I ]
0
[ ]
[0]
U g
[ a ] { xa } = { 0}
([ 0] [ 0]) [0]
[ 0]
and finally
MODAL MODEL CORRELATION TECHNIQUES
Chapter 3 - Model Reduction Techniques
11
Rev 031198
Peter Avitabile
([ ] [ I]){ p} = { 0}
2
Hybrid Reduction
Another method for reduction of the system matrices utilizes the exactness of the SEREP
process and overcomes the rank deficiency by incorporating the effects of Guyan condensation
into the process and is referred to as the Hybrid Reduction [9] technique and is formulated as
(3-42)
While this reduction technique overcomes some of the rank problems associated with SEREP
when using the reduced model for forced response and other studies, there is no inherent
advantage in using this technique as a model reduction scheme for correlation studies.
CONCLUDING REMARKS
Several of the more popular and commonly used reduction schemes were presented. Any of
the transformations described above may be used for the development of the reduced mass and
stiffness for processing such as needed for correlation studies. However, there will be
inaccuracies introduced in the reduced matrices depending on the technique employed and the
set of degrees of freedom chosen for the master set of degrees of freedom.
12
Rev 031198
Peter Avitabile
CHAPTER 4
MODE SHAPE EXPANSION
PREFACING REMARKS
This chapter presents some of the basic approaches for the expansion of measured experimental
modal vectors. Expansion of experimental vectors is often needed for correlation studies and
also used in the model updating procedures typically implemented utilizing current technology.
{x n } = x a = [T] { xa }
x
(4-1)
Rev 031198
Peter Avitabile
Using this expansion concept along with measured experimental modal data, we can write
E
[ E n ] = E a = [ T] [ E a ]
(4-2)
So we see that the measured experimental modal vectors at ADOF are expanded over all the
finite element NDOF using the transformation matrix [T]. This transformation matrix will take
on various forms depending on which technique is utilized.
Guyan Expansion
The Guyan Expansion technique [4] uses the static condensation transformation matrix to
expand the measured DOF over all the finite element DOF. The transformation is given by
[I]
[I ]
[ Ts ] = t = [K ]1 [K ]
da
dd
[ s ]
(4-3)
[ E n ] = E a = [ Ts ][ E a ] = [K
d
[I ]
1
[ E a ]
dd ] [ K da ]
(4-4)
Notice that the ADOF remain unchanged as seen by the upper partition of this equation
[ E a ] = [I ] [ E a ]
(4-5)
[ E ] = [ [K
d
dd
] 1 [ Kda ] [ E a ]
(4-6)
Of course, the Guyan condensation process will not produce acceptable results unless there are
sufficient DOF to describe the mass inertia of the system (as previously discussed in the
reduction section). If sufficient DOF are available, then the Guyan process will produce
reasonably good results but will never produce exact results since the inherent formulation of the
reduction matrix is approximate. The Guyan reduction process is still widely used for model
reduction applications due to it's long historical background but is not widely used for expansion
of mode shapes due to other more accurate techniques that have been developed.
IRS Expansion
Rev 031198
Peter Avitabile
The IRS Expansion technique [6] uses the static condensation transformation matrix along with
adjustment terms to compensate for the inertia associated with the deleted DOF to expand the
measured DOF over all the finite element DOF. The transformation is given by
I
[ Ti ] = [K []]1 [K
dd
[0]
+ [0]
da ]
[0]
1
[ M n ][Ts ][ M a ] [ K a ]
[0]
1
[ M n ][Ts ][ M a ] [ Ka ][ E a ]
[K ]
1
dd
(4-7)
[ E n ] = E a = [Ti ]{E a }
[0]
[I ]
=
+
1
[ Kdd ] [ Kda ] [0]
[K ]
1
dd
(4-8)
Notice that the ADOF remain unchanged as seen by the upper partition of this equation
[ E a ] = [I ] [ E a ]
(4-9)
{E d } =
[ [ [ K
dd
] [[ ]
[ K ]]][ E ]
a
(4-10)
Of course, the IRS technique will improve on the Guyan expansion process but will not produce
acceptable results unless there are sufficient DOF to describe the mass inertia of the system (as
previously discussed in the reduction section). If sufficient DOF are available, then the IRS
process will produce reasonably good results (which are improved over the Guyan results) but
will never produce exact results since the inherent formulation of the reduction matrices is
approximate. While the use of IRS is popular for reduction of matrices for reduced model
processing such as for structural response studies, the technique is not widely used for
expansion of mode shapes due to other techniques more appropriate for this type of process.
Dynamic Expansion
The Dynamic Expansion technique [7,22] is very similar in technique to the static expansion
process except that the stiffness matrix is modified to include the effects of the mass of the
system at a particular frequency; this is accomplished by adding an adjustment term of the
reference frequency times the system mass to the stiffness matrix as shown in the development
MODAL MODEL CORRELATION TECHNIQUES
Chapter 4 - Expansion Techniques
Rev 031198
Peter Avitabile
of the model reduction equations. In essence, this matrix is exact for this one particular
frequency and the transformation matrix will be exact in regards to expanding a mode shape at
that particular frequency; of course, the shift frequency must correspond to one of the
eigenvalues of the system. The transformation is given by
[ I]
[I]
[ Tf ] = t = [D ]1 [D ]
dd
da
[ f ]
(4-11)
[ E n ] = E a = [ Ts ][ E a ] = [D
d
[I ]
1
[ E a ]
dd ] [ D da ]
(4-12)
Notice that the ADOF remain unchanged as seen by the upper partition of this equation
[ E ] = [ I][ E ]
a
(4-13)
[ E ] = [ [D
d
dd
] 1 [ Dda ] [ E a ]
(4-14)
The Dynamic Expansion process will produce exact results for one frequency and only one
frequency. Providing that the shift frequency corresponds exactly to one of the eigenvalues of
the system, then the expansion will produce an exact mode shape for this one eigenvalue. If
additional eigenvectors need to be expanded, then separate shift values need to be processed.
While many matrices need to be processed for each eigenvector that needs to be expanded, the
exactness of the process warrants the additional processing.
SEREP Expansion
The SEREP Expansion technique [8] uses the SEREP transformation matrix to expand the
measured DOF over all the finite element DOF. The transformation is given by
Rev 031198
Peter Avitabile
[ Tu ] = [ Un ][ Ua ]
(
(
)
)
[ U a ] [ U a ] [ Ua ]
=
[ Ud ] [ Ua ]T [ Ua ]
[ Ua ]T
[U a ]
(4-15)
(4-16)
Notice that the ADOF may be changed as seen by the upper partition of this equation
[ E a ] = [ Ua ][ Ua ]g [ E a ]
(4-17)
[ E d ] = [ Ud ][ Ua ]g [ Ea ]
(4-18)
When the ADOF are expanded, there is the possibility that the initial measured DOF may be
modified by the expansion process; this is referred to as smoothing of the measured DOF. This
occurs since the SEREP process is based on a generalized inverse using a least squares error
minimization. Therefore, the measured data is smoothed as part of the process. While much
controversy exists over where or not to smooth the actual measured data, this is the most
proper way to process the data, from a mathematical standpoint.
The SEREP expansion technique is extremely accurate in regards to the expanded mode shapes
- actually it is exact due to it's inherent formulation. However, if the experimental mode shapes
are not correlated well with regards to the analytical mode shapes, then the results can produce
very poor expanded mode shapes. The SEREP process is very unforgiving of small errors that
exist in the measured experimental data base. While the SEREP process is often looked at as
being too harsh in the evaluation of modal vectors, this is exactly what is needed in order to
more clearly identify where errors exist in the measured and/or analytical model.
SEREPa Expansion
The SEREPa Expansion technique [23] extends the SEREP expansion technique such that the
generalized inverse is formulated as a standard inverse. This is accomplished by assuring that
a=m; that is, the number of measurement DOF is equal to the number of modes in the system.
The transformation is given by
MODAL MODEL CORRELATION TECHNIQUES
Chapter 4 - Expansion Techniques
Rev 031198
Peter Avitabile
[ Tu ] = [ U n ][ U a ]
[ U ][ U ] 1
a
= a
1
[ Ud ][ Ua ]
(4-119)
This assures that the measured DOF remain unchanged in the expansion process since
[ E a ] = [ U a ][ Ua ] 1[ E a ]
(4-20)
However, due to the large number of modes that are possibly needed, this technique is not
widely used for most expansion applications.
Modal Expansion
Since the SEREP process generally smoothes the data on the measured DOF in the expansion
process, another technique sometimes used is the Modal Expansion technique [24]. Actually,
this technique is the same as the SEREP process except that the upper partition of the
transformation matrix is forced to be identity which then forces the measured DOF to remain
unchanged in the expansion process. The transformation is written as
[ I]
[ TM ] = U U g
[ d ][ a ]
(4-21)
[I ]
E
[ E n ] = E a = [ TM ][ Ea ] = U U g [ E a ]
d
[ d ][ a ]
(4-22)
The measured DOF remain unchanged in the expansion process and the deleted DOF are
expanded in the same manner as in the typical SEREP expansion process. While this approach
attempts to retain the measured DOF as seen from test, there is a mathematical mismatch
between the ADOF and the deleted DOF which can cause some errors in any further
processing using the expanded mode shapes.
Hybrid Expansion
The Hybrid Expansion technique [9] uses the Hybrid condensation transformation matrix to
expand the measured DOF over all the finite element DOF. The transformation is given by
Rev 031198
Peter Avitabile
(4-23)
This transformation was developed to address some of the rank deficiencies inherent in the
SEREP reduction process when the number of DOF is greater than the number of modes in the
system. While Hybrid is a very useful technique for model reduction applications, the extra
computation necessary does not warrant it use as an expansion process.
CONCLUDING REMARKS
Several of the more popular and commonly used expansion schemes were presented. Any of
the transformations described above may be used for the development of the expansion matrix
for processing such as needed for correlation studies. However, there will be inaccuracies
introduced in the expansion process depending on the technique employed and the set of
degrees of freedom chosen for the master set of degrees of freedom.
Rev 031198
Peter Avitabile
CHAPTER 6
CORRELATION TECHNIQUES
PREFACING REMARKS
Many different correlation techniques exist for the comparison of analytical and experimental
modal vectors. The correlation techniques can be broken down into techniques relating to the
modal correlation in a vector sense and those relating to the correlation in a DOF sense; both of
these categories can be further broken down into those techniques which do not use any mass
scaling and those techniques that do use mass scaling.
In general, the techniques which do not employ any mass scaling are easier to implement and
use form a practical standpoint but usually are not as discerning since they lack any scaling of
the system mass properties. Techniques that do employ mass scaling are more difficult to use
since the mass matrix needs to be condensed but usually offer more robust evaluation of the
data evaluated.
Vector based correlation techniques are the MAC, POC and RVAC whereas DOF based
techniques are the COMAC, ECOMAC, CORTHOG and FRAC. All of these techniques are
described in the following sections with the exception of CORTHOG which is discussed in
Chapter 7.
CORRELATION TECHNIQUES
Modal Assurance Criteria (MAC)
The Modal Assurance Criteria (MAC) [1] is an extremely useful technique which gives a first
indication as to the level of correlation that exists between the analytical and experimental modal
vectors and is given by
MAC ij =
{ u i} T {e j}
[{ u } { u }] {
T
e
j
(6-1)
}{ }
T
e j
In this formulation, the values of MAC range between 0 and 1, where zero indicates that there is
little or no correlation between the vectors and one indicates that there is a high degree of
similarity between the modal vectors. MAC is ideal for identify which analytical modes
correspond to which experimental modes and is very useful when identifying mode switching.
MODAL MODEL CORRELATION TECHNIQUES
Chapter 6 - Correlation Techniques
Rev 031198
Peter Avitabile
MAC is very sensitive to the DOF that are largest in value and is very insensitive to very small
DOF in the mode shape vector. Mass weighting is not used in this formulation which has an
advantage in that mass reduction is not needed but the scaling effects of the mass matrix are
useful in weighting various dofs for a correlation study.
FINITE ELEMENT
MAC
MODAL
ASSURANCE
CRITERIA
MATRIX
MODE
SWITCHING
VECTOR CORRELATION
1
FEM 5
0.8
FEM 4
0.6
1
FEM 5
0.8
0.4
FEM 3
0.2
FEM 4
0.6
0.4
FEM 2
FEM 3
0.2
FEM 2
EXP1 EXP 2
EXP 3 EXP 4
EXP 5
FEM 1
FEM 1
EXP1
EXP 2
EXP 3
EXP 4
EXP 5
EXPERIMENTAL
Orthogonality Checks
Two orthogonality checks are often made in evaluating vector correlation. The Cross
Orthogonality Check or the Pseudo Orthogonality Check. The obvious hurdle to overcome is
whether to reduce the finite element mass and stiffness matrices to the set of tested DOF or to
expand the measured experimental modal vectors to the full space of the finite element model;
alternately, there could also be some combination of both reduction and expansion in order to
compute the orthogonality.
The Cross Orthogonality Check is an orthogonality check where the modal vector matrices are
obtained from the experimentally measured modal data.
CROSS = [ E] [ M ][ E] = [I ]
?
(6-2)
The Pseudo Orthogonality Check [10] is essentially an orthogonality check where one of the
modal vector matrices is replaced with the experimentally measured modal data.
POC = [ E]
[ M] [ U] = [I]
(6-3)
2
Rev 031198
Peter Avitabile
As mentioned above, in order to accomplish this triple product, the size of the matrices must be
consistent. Therefore, either the system mass matrix must be reduced to the set of tested DOF
(and corresponding ADOF from the modal matrix) or the experimentally measured modal
vectors must be expanded to the full space of the finite element model. Of course, the results of
either of these checks will be dependent on the type of reduction or expansion utilized with the
exception of the SEREP process which preserves the dynamics of the system in the reduced
model.
In reviewing the results of the Cross Orthogonality Check and the Pseudo Orthogonality Check,
there are similarities that exist since one check basically uses the experimentally measured
vectors twice in the computation of orthogonality and therefore, the resulting terms are larger.
However, there is no benefit of using these vectors twice in the Cross Orthogonality Check; no
additional insight is gained in this process. Thus, there is no benefit in using the Cross
Orthogonality Check rather than the Pseudo Orthogonality Check for general orthogonality
using the usual model reduction and expansion processes such as Guyan and IRS. However,
when using the SEREP process, there are other significant computational benefits that are
obtained when using the Pseudo Orthogonality Check that are described next.
(6-4a)
POC = [ E n ] [ M n ][ Un ]
Post-multiplying by the experimental modal vectors
T
(6-4b)
Rev 031198
Peter Avitabile
POC = [ Ua ] [ M a ][ E a ]
T
(6-4c)
POC = [ Un ] [ M n ][ En ]
T
(6-4d)
Now let's consider one of the forms above and shown that whether the POC is performed at
the reduced model size or at the full model size produces exactly the same results.
POC = [ E a ] [ M a ][ Ua ]
T
(6-5)
Substituting the reduced mass matrix in equation (3-36) into equation (6-5) gives
(6-6)
[ E a ]T [ M a ][ Ua ] = [ Ea ]T [ U a ]g [ U n ]T [ M n ][ Un ][ Ua ]g [ Ua ]
T
(6-7)
The expanded experimental modal vector in equation (4-16) can be transposed to give
[ E n ] T = [ Ea ] T [ U a ] g [ U n ] T
T
(6-8)
and it can be seen that the first three terms of equation (6-7) are equation (6-8). Now also
realize that the last two terms of equation (6-7) can be regrouped
[ Ua ]g [ Ua ] = ([ Ua ]T[ U a ]) [ Ua ]T [ Ua ]
(6-9)
[ Ua ]g [ Ua ] = ([ U a ]T[ Ua ])
[ U a ] [ Ua ] = [ I]
(6-10)
Substituting equation (6-8) and equation (6-10) into equation (6-7) it can easily be seen that
[ E a ]T [ M a ][ Ua ] = [ E n ]T[ M n ][U n ]
(6-11)
This implies that the POC will produce exactly the same results at either the full set of 'n' finite
element degrees of freedom or at the reduced set of 'a' tested degrees of freedom. Obviously
MODAL MODEL CORRELATION TECHNIQUES
Chapter 6 - Correlation Techniques
Rev 031198
Peter Avitabile
there are significant computational benefits to be obtained if the computations are performed at
the set of 'a' tested degrees of freedom rather than at the set of 'n' finite element degrees of
freedom.
Now let's show that the POC can be written such that neither the full or reduced analytical mass
matrix is needed. Recall the POC at the full set of finite element degrees of freedom in equation
(6-4) and substitute the relationship for the expanded experimental modal vectors in equation
(23)
[ E n ]T[ M n ][ U n ] = ([ Un ][ Ua ]g [ Ea ]) [ M n ][ U n ]
T
(6-12)
Taking the transpose of the relationship for the expanded experimental modal vectors in
equation (4-16) gives
(6-13)
[ E n ]T[ M n ][ Un ] = [ Ea ]T[ Ua ]g
(6-14)
Since the POC was shown to be identical at either the full set of finite element degrees of
freedom or at the reduced set of tested degrees of freedom in equation (6-11), we see that the
POC is no more than
[ E a ]T [ M a ][ Ua ] = [ E n ]T[ M n ][ Un ] = ([ Ua ]g [ Ea ])
(6-15)
[ Ua ]T[ M a ][ E a ] = [ U n ]T[ M n ][ E n ] = ([ Ua ]g [ Ea ])
(6-16)
[ E a ]T [ Ka ][ Ua ] = [ E n ]T[ Kn ][ Un ] = ([ Ua ]g [ E a ])
[ ]
2
(6-17)
(6-18)
Rev 031198
Peter Avitabile
Two important items can be noted from the development presented above. First, the Pseudo
Orthogonality Check will provide exactly the same results whether the check is done at the full
space of the finite element model or at the reduced space of the tes model. Therefore, the
computation is most efficiently performed in the reduced space of the test model. Second,
tremendous computational and procedural benefits are obtained for the Pseudo Orthogonality
Check using the SEREP process since the mass matrix is not needed for this computation.
L#
M
#
M
M
#
M
M
#
M
#
M
M
#
M
M
#
M
#
M
M
#
M
M
#
M
#
M
M
#
M
M
#
M
#
M
M
N#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#O
P
#P
#P
P
#P
P
#P
#P
P
#P
P
#P
#P
P
#P
P
#P
#P
P
#P
P
#P
#P
Q
L#
M
#
M
M
#
M
M
#
M
#
M
M
#
M
M
#
M
#
M
M
#
M
M
#
M
#
M
M
#
M
M.
M.
M
M.
M
M.
M.
M
M.
M
M.
M.
M
M.
M
M.
M.
M
M.
M
M
N.
#
#
#
#
#
#
#
#
#
#
#
#
.
.
.
.
.
.
.
.
.
.
.
.
.
#
#
#
#
#
#
#
#
#
#
#
#
.
.
.
.
.
.
.
.
.
.
.
.
.
#
#
#
#
#
#
#
#
#
#
#
#
.
.
.
.
.
.
.
.
.
.
.
.
.
#
#
#
#
#
#
#
#
#
#
#
#
.
.
.
.
.
.
.
.
.
.
.
.
.
#
#
#
#
#
#
#
#
#
#
#
#
.
.
.
.
.
.
.
.
.
.
.
.
.
#
#
#
#
#
#
#
#
#
#
#
#
.
.
.
.
.
.
.
.
.
.
.
.
.
#
#
#
#
#
#
#
#
#
#
#
#
.
.
.
.
.
.
.
.
.
.
.
.
.
#
#
#
#
#
#
#
#
#
#
#
#
.
.
.
.
.
.
.
.
.
.
.
.
.
#O
P
#P
#P
P
#P
P
#P
#P
P
#P
P
#P
#P
P
#P
P
#P
#P
P
.P
P
.P
.P
P
.P
P
.P
.P
P
.P
P
.P
.P
P
.P
P
.P
.P
P
.P
Q
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
FEM 5
0.1
FEM 4
FEM 3
FEM 2
EXP1
EXP 2
EXP 3
FEM 1
EXP 4
EXP 5
#
O
P
# #
P
P
# # #
P
# # #
P
P
# # #
P
P
# # #
P
# # #
P
P
# # #
P
P
# # #
P
# # #
P
P
# # #
P
P
# # #
P
P
# # #
P
# # #
P
P
# # #
P
# # #
P
P
# # #
P
P
# # #
P
P
# # #
P
# # #
P
P
# # #
P
# # #
P
# # # P
P
# # #P
P
# #P
Q
L#
M
#
M
M
#
M
M
#
M
#
M
M
#
M
M
N#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#
#O
P
#P
#P
P
#P
#P
P
#P
P
#P
Q
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
FEM 5
0.1
FEM 4
FEM 3
FEM 2
EXP1
EXP 2
EXP 3
FEM 1
EXP 4
EXP 5
Rev 031198
Peter Avitabile
CoMAC( k) =
m ( c ) (c )
uk ek
c=1
(
m
c= 1
2
u(kc )
c= 1
(6-19)
( )
2
e (kc )
However, without mass scaling to properly weight the dofs, at times it is difficult to determine
the degree of correlation that exists. Another drawback of the CoMAC is that it can only be
used for correlated mode pairs which implies that only the diagonal related terms of the MAC
correlation matrix can be assessed (once the vectors are arranged in propoer correlated order if
need be).
CoMAC
COORDINATE
MODAL
ASSURANCE
CRITERIA
Experimental
Analytical
DOF CORRELATION
FINITE ELEMENT
EXPERIMENTAL
Modulus Difference
The modulus difference is another tool that helps to identify the discrepancy that a given mode
pair may have.
MODAL MODEL CORRELATION TECHNIQUES
Chapter 6 - Correlation Techniques
Rev 031198
Peter Avitabile
(6-20)
This is a subset of COMAC since it evaluates the difference between DOF for a given
correlated mode pair rather that over a set of correlated modes.
ECoMAC( k ) = c =1
2m
(6-21)
The ECOMAC is very good for identifying gross problems in the measured modal vectors such
as polarity of a transducer from a test. Again no system mass matrix is used to weight the dofs
in this evaluation.
Frequency Response Assurance Criteria (FRAC)
The Frequency Response Assurance Criteria [15] evaluates each DOF based on the FRF
comparison of the analytical and experimentally derived functions. The formulation is very
similar to the MAC function and has similar interpretation. The FRAC is given by
FRAC( ) =
)(
(6-22)
The FRAC is a useful tool for evaluating FRFs but the main drawback is that the analytical
model may have similar shape characteristics but differ slightly in frequency which can cause
significanly low FRAC values - so the function is constructed with a shifting function to allow for
some frequency adjustment due to global stiffness differences. The FRAC is mainly used for
correlation that leads into frequency response based model updating studies. The majority of
studies in this research are directed towards modal vector based correlation; FRAC is included
here for completeness of the presentation of different correlation tools.
Rev 031198
Peter Avitabile
FRAC
FREQUENCY
RESPONSE
ASSURANCE
CRITERIA
FINITE ELEMENT
EXPERIMENTAL
DOF CORRELATION
(6-23)
Again this technique is more applicable to correlation for frequency response based model
updating and is included here for completeness.
RVAC
RESPONSE
VECTOR
ASSURANCE
CRITERIA
FINITE ELEMENT
EXPERIMENTAL
VECTOR CORRELATION
Rev 031198
Peter Avitabile
CHAPTER 7
COORDINATE ORTHOGONALITY CHECK
PREFACING REMARKS
The correlation tools discussed thus far consisted of vector correlation techniques (both with
and without mass scaling) and degree of freedom correlation techniques (that use no mass
scaling). The vector correlation techniques assess the vector in a global sense and the
correlation of the vector is stated in terms of a scalar quantity that provides a measure of the
level of correlation achieved. The Modal Assurance Crieria was very easy to compute and
provides a first level of correlation of the vector sets. However, mass scaling can hamper the
technique. Orthogonality Checks such as the Cross Orthogonality Check and Pseudo
Orthogonality Check provide a better indicator of the level of correlation achieved since mass
scaling is included but again, the correlation of the vector is stated in terms of a scalar quantity
that provides a measure of the level of correlation achieved. The degree of freedom correlation
techniques such as CoMAC and ECoMAC provide a better assessment of the correlation on a
degree of freedom basis but also are hampered by the fact that mass scaling is not included in
the formulation. A new technique referred to as the Coordinate Orthogonality Check [12,14]
has been developed to address the degree of freedom correlation of two modal vectors on a
mass scaled basis. The Coordinate Orthogonality Check was developed to more clearly
identify the discrepancy that exists between analytical and experimental modal vectors on a
degree of freedom basis and is formulated with the mass as a weighting function on the DOF.
The simplest statement concerning the Coordniate Orthogonality Check is as follows:
The Coordinate Orthogonality Check (CORTHOG) is simply the comparison of what
should have been obtained analytically for each degree of freedom in an orthogonality
check to what was actually obtained for each degree of freedom in a POC from test.
The CORTHOG is shown schematically in Figure 7-1. The formulation of the CORTHOG is
presented herein.
CORTHOG
COORDINATE
ORTHOGONALITY
CRITERIA
Experimental
Analytical
DOF CORRELATION
FINITE ELEMENT
EXPERIMENTAL
Rev 031198
Peter Avitabile
[ U] T [ M ] [ U] [ I]
(7-1)
This is the condition for which the analytical vectors are orthogonal with respect to the system
mass matrix.
The experimental vectors can be used as one of the set of vectors in equation (7-1) to obtain an
indication as to how well the measured experimental modal vectors are related to (or are
orthogonal to) the analytical mass matrix and the analytical modal vectors. The Pseudo
Orthogonality Check relating the correlation between the analytical modal vectors, [U], and the
experimental modal vectors, [E], with an analytical mass matrix, [M], can be performed using
POC = [ E]
[ M ] [ U] = [ I ]
(7-2)
and can be done at either the set of 'a' tested DOF or at the set of 'n' finite element DOF.
(Note that unit modal mass scaling is assumed throughout this theoretical development.) At the
reduced model size, the mass matrix can be obtained through the SEREP reduction scheme. At
the full model size, the measured experimental vectors can be expanded using the SEREP
expansion process.
Each term of the POC matrix can be described as
POC ij = e ki m kp u pj
k
(7-3)
To easily describe the concept of CORTHOG, let's consider a simple 3 DOF system with a
lumped mass matrix. The POC can be written as
e11
POC = e 21
e 31
e12
e 22
e 32
e 13
e 23
e 33
m11
m 22
m 33
2
u11
u
21
u 31
u 12
u 22
u 32
u 13
u 23
u 33
(7-4)
Rev 031198
Peter Avitabile
Considering only the 'i'th experimental 'e' mode with the 'j'th analytical 'u' mode,
POC ij = e1i
e2i
m11
e 3i
m 22
u1j
u
2j
m 33 u 3 j
(7-5)
Equation (7-5) can be written in expanded form for the 3 DOF system as
? 1 ( for i = j)
POC ij = ( e1i m11 u1 j + e 2 i m 22 u 2 j + e 3i m 33 u 3 j ) =
0 ( for i j)
(7-6)
Clearly, all DOFs have a contribution to one particular diagonal or off-diagonal term of the
POC matrix. It is very important to note that for an off-diagonal term to become zero, the
vectors need not be correlated; also that for off-diagonal terms, each of the individual
multiplications are not zero themselves but rather the summation of all the multiplications should
produce a value of zero. The individual multiplications that make up a POC term can be
inspected, but there is no way to assess whether a given value is too high or too low if just the
individual 'emu' terms are evaluated.
If we recall the statement of mass orthogonality for this simple 3 DOF system then Equation (71) can be written in expanded form as
1 ( for i = j)
ORTHOG ij = ( u1i m11u 1 j + u 2 i m 22 u 2 j + u 3i m 33 u 3 j )
0 ( for i j)
(7-7)
'umu' term
DOF 1
e 1i m11 u1 j
u1i m11u 1 j
DOF 2
e 2 i m 22 u 2 j
u 2i m 22 u 2 j
DOF 3
e 3i m 33 u 3 j
u 3i m 33u 3 j
_________
POC ij
___________
ORTHOGij
(7-8)
Figure 7-2 shows a plot of the comparison of the actual individual 'emu' values from POC and
MODAL MODEL CORRELATION TECHNIQUES
Chapter 7 - Coordinate Orthogonality Check
Rev 031198
Peter Avitabile
expected individual 'umu' values from one orthogonality term for each degree of freedom for a
given 'ij' mode pair; this is a plot of the individual values shown in Equation (7-8). In Figure 72, it is clear that both the POC and orthogonality terms all sum to zero, but that the difference
between them on a degree of freedom by degree of freedom basis is not the same. From this
plot, this discrepancy becomes apparent and an assessment can be made as to the correlation
that exists on a degree of freedom basis.
The Coordinate Orthogonality Check (CORTHOG) is simply the comparison of what should
have been obtained analytically for each degree of freedom in an orthogonality check using
Equation (7-7) to what was actually obtained experimentally for each degree of freedom in a
POC using Equation (7-8).
Now let's extend this to the general case of a fully populated mass matrix for the 3 DOF system.
Considering only the 'i'th experimental 'e' mode with the 'j'th analytical 'u' mode,
POCij = e1i
e 2i
m11
e 3i m 21
m 31
m12
m22
m32
m13 u1 j
m 23 u2 j
m 33 u 3 j
(7-9)
(7-10)
+ e 3i m 31 u1 j + e 3i m 32 u 2 j + e 3i m 33 u 3 j )
As stated earlier, all DOFs have a contribution to one particular diagonal or off-diagonal term of
the POC matrix. In terms of the individual DOFs, the contribution to the POC for the first,
second and third DOF, respectively, is
POCijDOF1 = e1i m11u1 j + e1i m12u 2 j + e1im 13u3 j
(7-11a)
(7-11b)
= e 2i ( m 21u1j + m 22 u2 j + m23 u3 j )
POCijDOF3 = e 3i m 31u 1j + e3i m 32u 2 j + e 3i m 33u 3 j
(7-11c)
= e 3i ( m 31u1 j + m 32 u 2 j + m 33u 3 j )
and the contribution to the orthogonality for the first, second and third DOF, respectively, is
Rev 031198
Peter Avitabile
1
ORTHOG DOF
= u1i m11u1 j + u1i m12 u 2 j + u1i m13u 3 j
ij
(7-12a)
(7-12b)
(7-12c)
Using this notation, the subscript on the POC and ORTHOG refer to the analytical and
experimental mode pair being evaluated whereas the superscript refers to the particular DOF
that is being evaluated. For purposes of general notation used hereafter, equation (7-11) and
(7-12) can be written in general terms as
POCijk = e ki m kpu pj
p
for an individual DOF. The sum over all DOF yields the usual 'ij' term of the POC or
orthogonality.
Rev 031198
Peter Avitabile
experimental values for each individual DOF. All other techniques described below provide
similar information but are presented in some sort of scaled fashion which may accentuate
differences and allow easier interpretation of the data.
Simple Difference
A simple approach is to determine the difference between the summation of the individual
multiplications for each DOF for a POC off-diagonal term and its expected value based on the
analytical vectors given as
SD = CORTHOG ijk =
e ki m kp u pj u ki m kp u pj
(7-13)
An advantage for this formulation is that the magnitude and direction of the error is retained. By
retaining the magnitude of the error, the value may be evaluated but there is no way to assess
whether a given discrepancy is either acceptable or unacceptable. This magnitude may be
considered in the same way as off-diagonal terms are evaluated; an arbitrary limit for a DOF,
such as 0.01, may be used as a criteria for acceptance. By retaining directional information,
differences relative to other degrees of freedom can be assessed. The summation of the plus
values and minus values that make up the POC terms may be inspected lending further insight
into the discrepancies.
Normalized SD Maximum
Another approach is to normalize the simple difference to a certain value such as the maximum
difference given as
e ki m kp u pj u ki m kp u pj
NSD M = CORTHOG kij =
e ki m kp u pj u ki m kp u pj
p
max
(7-14)
which has the advantage of setting the maximum difference equal to 1; all other differences
would be a percentage of the maximum. However, since every mode pair evaluated may have
widely varying levels of correlation, it is difficult to determine which mode pairs of vectors are
the most and least correlated. That is to say that this is a tool to be used in conjunction with one
of the other scaling methods.
Normalized SD Total
Rev 031198
Peter Avitabile
e ki mkp u pj u ki mkp u pj
NSDT = CORTHOG ijk
(7-15)
e ki mkp u pj u ki mkp u pj
k
each difference will be given as a percentage of the POC term. Thus, the maximum DOFs
contributing to the given off-diagonal term can easily be evaluated. Note that if the POC term is
exactly equal to zero then this normalizing scheme cannot be used.
Absolute SD
By taking the absolute difference between the summation of individual multiplications for each
experimental DOF of a POC off-diagonal term and its expected value based on the analytical
vectors, the absolute difference is given as
ASD = CORTHOG kij =
e ki m kp u pj u ki m kp u pj
(7-16)
and can be easily compared. This again highlights which DOFs are most important to the given
off-diagonal term. The advantage here is that the sum of all the terms will not arbitrarily become
zero due to the addition of many different plus and minus contributions. When used in this
fashion with an orthogonality check, this is referred to as an Absolute POC.
Normalized ASD Maximum
The absolute difference when normalized to the maximum difference is given as
e ki m kp u pj u ki mkp u pj
NASD M = CORTHOG kij =
e ki m kp u pj u ki m kp u pj
p
max
(7-17)
and sets the maximum DOF difference to 1. All other differences are percentages of the
maximum difference. This will provide a consistent scale but does not allow for direct
comparison with other mode pairs evaluated.
Normalized ASD Total
MODAL MODEL CORRELATION TECHNIQUES
Chapter 7 - Coordinate Orthogonality Check
Rev 031198
Peter Avitabile
e ki mkp u pj u ki m kp u pj
NASD T = CORTHOG kij =
(7-18)
e ki mkp u pj u ki mkp u pj
k
and will put the DOF discrepancy in terms of percent of the total difference.
The formulations discussed above are plotted in Figures 7-3 through 7-7 for the 3 DOF
example presented in Figure 7-2. Note that the normalized absolute difference could not be
plotted due to division by zero. Again, not all of the normalization schemes are routinely used
but are included here for reference.
(7-19)
(7-20)
g
u12
u g22
M
u gm2
e 11 e12
g
u13
L ug1a
e 21 e 22
ug23 L u g2a
e
e
31 32
M O
M
M
u gm3
ugma
e a1 e a 2
L e1m
L e 2m
L e 3m
eam
(7-21)
Rev 031198
Peter Avitabile
ug
i1
POCij =
u gi2
u gi3
L u gia
e1 j
e2j
e3j
M
e aj
(7-22)
(7-23)
and if we multiply out the corresponding term from the orthogonality relationship, we get
ORTij = u gi1u1 j + ugi 2u 2 j + u gi3 u 3 j + L + u gia u aj = 0
(7-24)
The CORTHOG manipulation is the comparison of each of the terms of the relationships in
Equation 7-23 and 7-24 for corresponding degrees of freedom.
POC ij = ugi1e1j + u gi2 e2 j + ugi 3e 3 j + L + ugia e aj
c
CORTHOG ij
M
dof 2
(7-25)
M
dof 3
Then any one of the comparison techniques (scaling, normalization, etc.) described previously
can be computed. There are tremendous computation benefits to formulating the Coordinate
Orthogonality Check using the SEREP process.
Rev 031198
Peter Avitabile
'emu'
DOF 1
'umu'
DOF 2
'emu'
'umu'
'emu'
DOF 3
'umu'
-8
-6
-4
-2
10
Rev 031198
Peter Avitabile
SD
DOF1
DOF2
DOF3
-1
-0.5
0.5
NSD
DOF1
DOF2
DOF3
-1
-0.5
0.5
ASD
DOF1
DOF2
DOF3
0
0.5
NASDM
DOF1
DOF2
DOF3
0
0.5
NASDT
DOF1
DOF2
DOF3
0
0.5
11
Rev 031198
Peter Avitabile
CHAPTER 8
PRE-TEST EVALUATION TECHNIQUES
PREFACING REMARKS
Selection of DOF for measurement locations can be critical to the success of an experimental
modal survey. It is very important to assure that an adequate number of proper points are
identified for the collection of data. This chapter briefly reviews some of the existing techniques
used for pre-test evaluations and then offers some additional tools based mainly on the
Coordinate Orthoganlity technique.
[A(s)]
= Qk { u k } { u k }
(8-1)
So that the drive point residues (DPR) [16], that is where the input and output are measured at
the same location, can be evaluated for each DOF over a set of modes from
a iik = q k u ik u ik
(8-2)
Thus as different modes, k, are evaluated, different DPRs are computed. Typical summation
methods include an assessment on the minimum value, the maximum value, the average value
and a weighted average (which is the average times the minimum value) for each DOF over a
set of modes. The problems with the summing schemes is that often effects are smeared into
one value which does not adequately depict the proper selection of points for test DOFs.
Effective Independence
Rev 031198
Peter Avitabile
One technique that is used is to assess the rank of the modal matrix for a set of DOF that is
chosen as candidate measurement locations. The rank of the equation can be determined
through the singular value decomposition of
[[ U ][ U ] ] [][ ][ ]
T
(8-3)
The Effective Independence [17] can be used to determine the contribution of each retained
DOF using
EfI = diag i [ Ua ][ ][ ]
[ ]T [ Ua ]T ) = diag i [ Ua ][[U a ]T [ Ua ]] [ U a ]T
1
(8-4)
This process is continued until an acceptable set of DOF are identified as candidate modal
vectors. Due to it's inherent formulation, the DOF that are selected are very good for identifying
the independent vectors for possible measurement locations. Unfortunately, this method will
routinely specify DOF which are close to or actually at nodes of the system. These DOF are
highly suspect in terms of accuracy and are not considered good measurement locations for an
experimental modal test.
MAC Contribution
One fairly simple technique is to use the MAC and evaluate the effect of different DOF to the
values of the off-diagonal terms of the MAC matrix from
[{u } {u }]
=
[{u } {u }] [{u } {u }]
2
MAC ij
( 8-5)
Referred to as the MAC Contribution (MACCO) [18], the MAC calculation can be extended
to include the effect of an additional DOF as
[{u } {u } + u u ][{u } {u } + u u ]
=
[{u } {u } + u u ] [{u } {u } + u u ]
T
MAC
p
ij
pi
pj
pi
pj
pj pj
pi
pi
(8-6)
This requires only a minor calculational effort. The addition of extra DOF tend to make the off
diagonal terms of the MAC more acceptable thereby indicating a better set of points has been
selected. An iterative procedure can be developed to determine the effects of adding additional
candidate DOFs for measurement and to rank them based on the effect on the reduction of the
MODAL MODEL CORRELATION TECHNIQUES
Chapter 8 - PreTest Evaluation Techniques
Rev 031198
Peter Avitabile
off-diagonal terms of the MAC matrix. This can provide some useful information, however, as
in the case of MAC used for correlation, there is no mass matrix available for appropriate
weighting of important DOFs.
Other Analytical Methods (Henshell Method, etc.)
Other analytical techniques [19] exist for automatic selection of active DOF for model reduction
purposes, but generally these techniques also use averaging by one means or another to obtain
some "best" location for the active DOFs. The combination of all the DOFs for all the modes of
the system generally break down in applications where the structure contains directional modes
which is often the case.
Mode Shape Evaluation Techniques
Since the CORTHOG was developed as a correlation tool which inherently contains mass
scaling, then it would be reasonable to assume that it might also be useful for pre-test
evaluations for selection of DOF especially when the resulting experimental modal data base will
be used for correlation studies. Several new graphical tools have been developed, several of
which are based off of the CORTHOG technique [20].
Mode Shape Summation Plot - MSSP
Due to the fact that information is averaged across many modes for the majority of the
techniques, one possible alternative is to use a simple graphical approach to view the mode
shapes. Recall that the modal matrix is nothing more than the mode shapes arranged in column
form with the DOFs arranged in rows of the modal matrix. One technique is simply the
graphical presentation of this data into the Mode Shape Summation Plot (MSSP).
u11
u
21
u 31
[ U] = u 41
u 51
M
u a1
u12
u22
u32
u42
u52
M
ua 2
u13
u23
u33
u43
u53
M
ua 3
L
L
L
L
L
O
L
u1m
u 2 m
u3 m
u4m
u5m
M
u am
(8-7)
Basically, the MSSP is a bar graph of the set of modes over the set of DOFs. Each (DOF) bar
is created by stacking (summing) the corresponding mode shape components for all of the
modes of interest. By creating the mode shape sum in this manner, the contribution of each
mode to each DOF is preserved and more clearly identifies how the energy is distributed on a
mode by mode basis for each DOF. Separate plots can be also be created for each mode. In
MODAL MODEL CORRELATION TECHNIQUES
Chapter 8 - PreTest Evaluation Techniques
Rev 031198
Peter Avitabile
addition, the plots can be scaled to reflect displacement, velocity or acceleration, as desired.
Generally, the high amplitude bars in the MSSP represent good candidates for test response
or reference DOFs. However, typical, high amplitude DOFs for certain modes are at or near
node DOFs for other modes. Therefore, it may be advantageous to use different sets of
measurement DOFs in the correlation of each mode. The advantage of the MSSP over other
approaches previously utilized is that the contribution of each mode is presented and the user
can visually evaluate how well each mode will be represented by or excited at each DOF.
CORTHOG Pretest Plot - MSSP
A variation on the MSSP is the CORTHOG Pretest Plot (CPP). Recall that the Coordinate
Orthogonality Check (CORTHOG) was developed as a tool to aid in determining the
correlation of analytical and experimental modal data on a mass scaled DOF basis. Used as a
Pre-Test Tool, CORTHOG helps to clearly show how important each individual DOF will be
in an orthogonality check. This is a mass weighted preview of the mode shape, therefore, it will
yield additional information that will not be evident when using other Pre-Test tools such as
DPRs and MAC.
Like the MSSP, the CPP is a summed bar-chart of the mass-scaled mode shapes (columns
of [Ua]g). The importance of each individual DOF is mass scaled in the same way it is in an
orthogonality check.
[ U n ] T[ M n ][ E n ] = [ Ua ]T [ M a ][ E a ] = [ U ag ][ E a ]
[U ] [M ] = [U
T
g
a
u11
u
21
u 31
= u 41
u51
M
u
a1
u 12
u13
u 22
u 23
u 32
u 33
u 42
u 43
u 52
u 53
ua2
ua3
L u 1m
L u 2m
L u 3m
L u 4m
L u 5m
O M
L u am
(8-8)
(8-9)
The CPP is useful in identifying and evaluating potential response and reference DOFs on the
basis of modally active masses. The CPP plotted in conjunction with the MSSP helps to identify
small mode shape components that may be difficult to measure but important to an orthogonality
check.
MODAL MODEL CORRELATION TECHNIQUES
Chapter 8 - PreTest Evaluation Techniques
Rev 031198
Peter Avitabile
u r 1
[U]
(79)
u r2
u r3
O
u 11u r1
u 21u r1
u 31u r1
= M
u r1u r1
u rm M
u u
a 1 r1
u 12u r2
u 22 u r2
u 32 u r2
M
u r 2 u r2
M
ua2 ur 2
u1m u rm
u 23u r 3 L u 2 m u rm
u 33u r 3 L u 3 mu rm
M
L
M
u r 3u r 3 L u rm u rm
M
O
M
u a 3u r 3 L u amu rm
u 13u r 3 L
These "drive point scaled" modal vectors can be plotted in bar chart form and used to evaluate
how well each mode will be represented in the set of measured frequency response functions
using the chosen reference DOF.
While these new tools are not expected to replace previously developed techniques, they can
provide additional insight into issues pertaining to correlation of analytical and experimental data.
Rev 031198
Peter Avitabile
REFERENCES
1)
2)
3)
4)
5)
6)
7)
8)
9)
10)
11)
12)
13)
14)
15)
16)
17)
18)
Allemang,R.J. and Brown,D.L., "A Correlation Coefficient for Modal Vector Analysis", First
International Modal Analysis Conference, Orlando, Florida, November 1982
Lieven,N.A.J., Ewins,D.J., "Spatial Correlation of Mode Shapes, the Coordinate Modal
Assurance Criteria (COMAC)", Sixth International Modal Analysis Conference, Orlando,
Florida, February 1988
Hunt,D., "Application of an Enhanced Coordinate Modal Assurance Criterion", Tenth
International Modal Analysis Conference, San Diego, California, February, 1992
Guyan, R.J., "Reduction of Stiffness and Mass Matrices", AIAA Journal, Vol. 3, No 2, 1965
Conti,P., "A Higher Order Approximation of the Guyan Reduction Method", Seventh
International Modal Analysis Conference, Las Vegas, Nevada, February 1989
O'Callahan,J.C., "A Procedure for an Improved Reduced System (IRS)", Seventh
International Modal Analysis Conference, Las Vegas, Nevada, February 1989
Kidder, RL, "Reduction of Structure Frequency Equations", AIAA, Vol 11, No 6, June
1973
O'Callahan,J.C., Avitabile,P., Riemer,R., "System Equivalent Reduction Expansion Process",
Seventh International Modal Analysis Conference, Las Vegas, Nevada, February 1989
Kammer,D.C., "A Hybrid Approach to Test Analysis Model Development for Large
Space Structures", submitted for publication in Journal of Spacecraft and Rockets,
November 1989
Avitabile,P., O'Callahan,J.C., Milani,J. "Model Correlation and Orthogonality Criteria", Sixth
International Modal Analysis Conference, Orlando, Florida, February 1988
Avitabile,P., Li,P., Foster,T., Effects of DOF Selection on Orthogonality using Reduced
System Matrices", Eleventh International Modal Analysis Conference, Orlando, Florida,
February 1992
Pechinsky,F., Avitabile,P., Coordinate Orthogonality Check (CORTHOG)", Twelfth
International Modal Analysis Conference, Honolulu, Hawaii, February 1993
P.Avitabile, T.Foster, J.O'Callahan, "Comparison of Reduction vs. Expansion for
Orthogonality Checks", Twelfth International Modal Analysis Conference, Honolulu,
Hawaii, January 1994.
Pechinsky,F., Avitabile,P., "Evaluation of DOF Correlation using the Coordinate
Orthogonality Check", Thirteenth International Modal Analysis Conference, Nashville,
Tennessee, February 1995.
Lammens,S., "Frequency Response Based Validation of Dynamic Structural Finite
Element Methods", PhD Dissertation, Katholieke Universiteit, Leuven, Belgium, 1995
Kientzy,D., Richardson,M., Blakely,K., "Using Finite Element Data to Setup Modal
Tests", Sound & Vibration, June 1989
Kammer, DC, "Effects of Noise on Sensor Placement for On-Orbit Modal Identification
of Large Space Structures:, Journal of Dynamic System, Measurement and Control, Vol
114, September 1992
Carne,T.G., Dohrmann,C.R., "A Modal Test Design Strategy for Model Correlation,
Thirteenth International Modal Analysis Conference, Nashville, Tennessee, February
Rev 031198
Peter Avitabile
1995.
19) O'Callahan, JC, Li, P, "An Automatic Selection of Reduced Degrees of Freedom",
Twelfth International Modal Analysis Conference, Honolulu, Hawaii, February 1994.
20) DeClerck,J.P., Avitabile,P., Development of Several New Tools for Pre-Test Modal
Evaluation", Fourteenth International Modal Analysis Conference, Dearborn, Michigan,
February 1996.
21) Irons, B, "Structural Eigenvalue Problems: Elimination of Unwanted Variables", AIAA
Journal, Vol 3, No 5, 1965
22) Flanigan, CC, Freymiller, JE, "Using Dynamic Expansion to Improve Test-Analysis
Correlation", Twelfth International Modal Analysis Conference, Honolulu, Hawaii,
February 1994.
23) O'Callahan, JC, Li, P, "The Effects of Modal Vector Expansion on Finite Element Model
Updating", Thirteenth International Modal Analysis Conference, Nashville, Tennessee,
February 1995.
24) Kammer,D.C., "Test-Analysis-Model Development Using an Exact Modal Reduction",
International Journal of Analytical and Experimental Modal Analysis, October 1987
25) Avitabile,P., O'Callahan,J., "Mass and Stiffness Orthogonality Checks without a Mass or
Stiffness Matrix", Thirteenth International Modal Analysis Conference, Nashville, Tenn,
February 1995
BIBLIOGRAPHY
Model Reduction References
1) Kammer,D.C., "A Hybrid Approach to Test Analysis Model Development for Large
Space Structures", submitted for publication in Journal of Spacecraft and Rockets,
November 1989
2) Kammer,D.C., "Test-Analysis-Model Development Using an Exact Modal Reduction",
International Journal of Analytical and Experimental Modal Analysis, October 1987
3) Freed,A.M., Flannigan,C.C., "A Comparison of Test Analysis Model Reduction
Methods", Eighth International Modal Analysis Conference, Orlando, Florida, February
1990
4) Avitabile,P., Pechinsky,F., O'Callahan, J.C., "Study of Modal Vector Correlation using
Various Techniques for Model Reduction", Tenth International Modal Analysis Conference,
San Diego, California, February, 1992
5) Guyan, R.J., "Reduction of Stiffness and Mass Matrices", AIAA Journal, Vol. 3, No 2, 1965
6) O'Callahan,J.C., "A Procedure for an Improved Reduced System (IRS)", Seventh
International Modal Analysis Conference, Las Vegas, Nevada, February 1989
7) O'Callahan,J.C., Avitabile,P., Riemer,R., "System Equivalent Reduction Expansion Process",
Seventh International Modal Analysis Conference, Las Vegas, Nevada, February 1989
8) J.O'Callahan, "Comparison of Reduced Model Concepts", Eighth International Modal
Analysis Conference, Orlando, Florida, January 1990.
9) P.Avitabile, J.O'Callahan, J.Milani, "Comparison of System Characteristics Using
Various Model Reduction Techniques", Seventh International Modal Analysis
MODAL MODEL CORRELATION TECHNIQUES
References/Bibliography
Rev 031198
Peter Avitabile
Rev 031198
Peter Avitabile
5)
6)
7)
8)
9)
10)
11)
12)
13)
14)
15)
16)
February 1995.
Flanigan, CC, Freymiller, JE, "Using Dynamic Expansion to Improve Test-Analysis
Correlation", Twelfth International Modal Analysis Conference, Honolulu, Hawaii,
February 1994.
Levine-West, MB, Kissil, A, Milman, M, "Evaluation of Mode Shape Expansion
Techniques on the Micro-precision Interferometer Truss", Twelfth International Modal
Analysis Conference, Honolulu, Hawaii, February 1994.
Hemez, FM, Farhat, C, "Comparing, Mode Shape Expansion Methods for Test-Analysis
Correlation", Twelfth International Modal Analysis Conference, Honolulu, Hawaii,
February 1994.
Roy, N, Girard, A, Bugeat, LP, "Expansion of Experimental Mode Shapes - An
Improvement of the Projection Technique", Eleventh International Modal Analysis
Conference, Orlando, Florida, February 1993.
Imregum, M, Ewins, D, "An Investigation into Mode Shape Expansion Techniques",
Eleventh International Modal Analysis Conference, Orlando, Florida, February 1993.
Lieven, NAJ, Ewins, D, "Expansion of Modal Data for Correlation", Eighth International
Modal Analysis Conference, Orlando, Florida, February 1990.
O'Callahan, JC, Lieu, IW, Chou, CM, "Determination of Rotational Degrees of Freedom
for Moment Transfers in Structural Modifications", Third International Modal Analysis
Conference, Orlando, Florida, February 1985.
Gysin, HP, "Comparison of Expansion Method for FE Modeling Error Localization",
Eighth International Modal Analysis Conference, Orlando, Florida, February 1990
Gysin, HP, "Expansion: The Achilles Heel of Fe Model Updating", Proceedings of the
15th International Seminar on Modal Analysis, September 1990
O'Callahan, JC, Li, P, "A Non-Smoothing SEREP Process for Modal Vector
Expansion", Twelfth International Modal Analysis Conference, Honolulu, Hawaii,
February 1994
Lipkins, J., Vandeurzen, U., "The Use of Smoothing Techniques for Structural
Modification Applications", Proceedings of the 12th International Seminar on Modal
Analysis, 1987
Lang,G.F., "Including Rotational Terms in a Modal Model...An Experimental Prelude",
Sound & Vibration, November 1989
Rev 031198
Peter Avitabile
5)
6)
7)
8)
9)
10)
11)
12)
13)
Rev 031198
Peter Avitabile
Rev 031198
Peter Avitabile
7)
8)
9)
10)
11)
12)
13)
Software References
1) Cada-X Collection of Modal Software Packages, Leuven Measurement Systems,
Belgium
2) LINK Correlation/Updating Software Package, Leuven Measurement Systems, Belgium
3) Dynaview Correlation Software Package, MicroDyne, Inc
4) STAR Modal Analysis Software, Spectral Dynamics, San Jose, California
5) I-DEAS Collection of Software Packages- Structural Dynamics Research Corporation
6) NASTRAN Finite Element Software, McNeal-Schwendler Corp, Los Angeles,
California
7) ANSYS Finite Element Software Package, Swanson Analysis Systems
8) FEMTools Correlation Software, Dynamic Systems Solutions, Belgium
9) FEMAP Pre/Post Processor Software, Enterprise Systems, Pennsylvania
10) MATSAP/MATRIX Toolbox, UMASS Lowell, J.C.O'Callahan
Miscellaneous References
1) Strang, G, "Linear Algebra and Its Applications", Academic Press, 2nd Edition, 1980
2) Cook, RD, Malkus, DS, Plashha, ME, "Concepts and Applications of Finite Element
Analysis", John Wiley and Sons, Third Edition, 1989
3) Craig, RR, "Structural Dynamics - An Introduction to Computer Methods", John Wiley
and Sons, 1981
4) Bathe, KJ, "Finite Element Procedures in Engineering", Prentice-Hall, Englewood Cliffs,
NJ, 1982
5) Ewins, DJ, "Modal Testing: Theory and Practice", John Wiley and Sons, Inc, 1984
Rev 031198
Peter Avitabile