A COkRELATION COEFFICIENT FOR MODAL VECTOR ANALYSIS

B. J. Allemang D. L. Drown
Assistant Professor Research Associate Professor
Mechanical and Industrial Engineering Mechanical and Industrial Engineering
University of Cincinnati University of Cincinnati
Cincinnati, .Qhio 45221 Cincinnati, Ohio 45221

ABSTRACT

Multiple, independent modal vector estimates may be response function matrix. Specifically,
generated whenever multiple rows or columns of the information in the residue matrix corresponding to
frequency response function matrix are available. each Pole of the system is evaluated in a least
These independent estimates of the same modal squares error approach to determine separate
vector need to be processed into a single best estimates of the same modal vector. This
estimate of that particular modal vector. The evaluation consists of the calculation of a complex
development of the concept of consistency of modal modal scale factor (relating two modal vectors) and
vectors, .evaluated through the use of the modal a scalar modal assurance criterion (measuring the
assurance and’modal scale factor, is useful in consistency between two modal vectors).
computing a best estimate of the modal vector and
useful in understanding the errors among separate The function of the modal scale factor (MSF) is to
estimates of the same modal vector. provide a means of normalizing all estimates of the
same modal vector. When two modal vectors are
INTRODUCTION scaled similarly, elements of each vector can be
averaged (with or without weighting), differenced,
The common approach to estimation of modal vectors or sorted to provide a best estimate of the modal
from the frequency response function method is to vector or to provide an indication of the type of
measure a complete row or column of the frequency error vector superimposed on the modal vector.
response function matrix. This will give
reasonable definition to those modal vectors that The function of the modal assurance criterion (MAC)
have a non-zero modal coefficient at the excitation is to provide a measure of consistency between
location and can be completely unooupled with the estimates of a modal vector. This provides an
forced normal mode excitation method. When the additional confidence factor in the evaluation of a
modal coefficient at the excitation looation of a modal vector from different excitation locations.
modal vector is zero (very small with respect to The modal assurance criterion also provides a
the dynamic range of the modal vector) or when the method of determining the degree of causality
modal vectors cannot be uncoupled, the estimation between estimates of different modal vectors from
of the modal vector will contain potential bias and the same system.
variance errors. In such oases additional rows
and/or columns of the frequency response function The modal scale factor and the modal assurance
matrix are measured to detect such potential criterion also provide a method of easily comparing
problems. estimates of modal vectors originating from
different sources. The modal vectors from a finite
Richardson and Kniskern [ll have suggested that a element analysis can be compared and contrasted
simple procedure of averaging two columns will with those determined experimentally as well as
reduce the variance error on the resulting estimate modal vectors determined by way of different
of the modal vector. Identification and proper experimental or modal parameter estimation methods.
weighting of rows or columns containing poor In this approach, methods can be compared and
estimates of a particular modal vector should contrasted in order to evaluate the mutual
greatly improve this process. Additionally, consistency of different procedures rather than
though, much more information concerning each modal estimating the modal vectors specifically.
vector used in such a procedure as well as an
indication of the presence of bias errors in the FREQUENCY RESPONSE FUNCTION THEORY
estimate of the modal vector is desirable. MODAL VECTOR ESTIMATION

In general, this paper centers on developing The formulation of the frequency response function
practical techniques for the utilization of the matrix can be made in terms of the more general
redundant modal vector data in the frequency case of the transfer function. Therefore:
110
IX(s)1 q CH(s)l IF(s)] (I) [A(r)1 =

where :
U(2,r)U(l,r) U(2,r)U(2,r) U(2,r)U(3,r) .*.
IX(s)1 = Response vector U(3,r)U(l.r) U(3,r)U(2,r) U(3,r)U(3,r) *w*
lx(s)1 q m x 1 column vector U(4,r)U(l,r) U(4,r)U(2,r) U(4,r)U(3,r) ...
[H(s)] = Transfer function matrix U(5,r)U(l,r) U(5,r)U(2,r) U(5,r)U(3,r) l -o
m x q rectangular matrix k(r) .. . . . . . . . . . . . .. . . . . . . . . . .... ........
[H(s)]

{F(s)1
=

= Excitation vector .. . . . . . . . . . . ......... ... .. . . . . . . .. . .

IF(s)1 q q x 1 column vector .. . . . . . . . . . . . ........ ... .. . . . . . . . .. .
m = Number of response stations .. . . . . . . . . . . . ........ ... .. . . . . . . . . . .
q = Number of excitation stations ......... ... . ........ ... .. . . . . . . . . . .
s = LaPlace variable
s = +jw
Each transfer function can now be written as a
partial fraction expansion in terms of the system (5)
poles and residues. Assuming that the poles are Equations 4 and 5 state that each element of the
global properties of the system yields: residue matrix for modal vector r consists of the
product of a scaling constant. the modal
coefficient of the excitation location, and the
n modal coefficient of the response location.
a(i,j,r) &i,j,r)
H(i,j,s) = + (2) Equation 5 can be restated in a more concise matrix
II s - p(r) s - h-1
form as follows:
ra I

[A(r)1 = k(r) {U(r)] IU(r)jr (6)

where:

a(i,j,r) = Residue From Equation 6 it is easy to see that each row or

p(r) q System pole for moder
l = Complexconjugate column is simply the samemodal veator multiplied
i = Rowindex times the modal coefficient of the common response
= Responselocation location or excitation location respectively.
5 q Columnindex Therefore, the proportionality constant between
j
q Excitation location rows C and D or columns C and D can be referred to
r = Modenumber as a modal scale factor equal to:
n = Numberof degrees of freedom
Rewriting Equation 2 in matrix form to correspond U(c,r)
XG?(c,d,r) q - (7)
to Equation 1 gives: U(d,r)
n
[A(r)1 [Ay(r)l Yhile the formulation of Equation 7 is madewithout
[H(s)] = + (3)
c s - p(r) s - p?r) designation of row or column, the application of
r.t the concept of modal scale factor will differ
according to whether the modal scale factor is
where: determined from row or column relationships. This
will be discussed further in a later section,
[A(r)] q Residue matrix
[A(r)] = m x q rectangular matrix
ORTHCCONALITY
OF MODALVECTORS
Furthermore, each element of the residue matrix of
&u&Lion 3 C&i1be written:
Theoretically, for the case of proportional
damping, each modal vector of a system will be
a(i, j,r) = k(r) U(i,r) U(j,r) (4) orthogonal to all other modal vectors of that
system when weighted by the mass, stiffness, or
dampingmatrix. In practice, these matrices are
where: madeavailable by way of a finite element analysis
and normally the massmatrix is considered to be
k(r) = Scaling constant for moder the most accurate. For this reason, any further
U(i,r) = modal coefficient for location i discussion of orthogonality will be madewith
of moder respect to massmatrix weighting. As a result, the
U(j,r) = modal coefficient for location j orthogonality relations can be stated as follows:
of moder
In terms of matrix notation: IU(r)J' CM1 {U(q)1 = 0 rfs (8)

111
lU(1-11~ [Ml {U(q)] = M(r) r=q (9)
where:
M(r) = Generalized mass of mode r
Experimentally, the result of zero in Equation 8
can rarely be achieved but values up to one tenth
of the magnitude of the generalized mass of each
mode are considered to be acceptable. It is common
procedure to,form the modal vectors into a
normalized set of mode shape vectors with respect
to the mass matrix weighting. The accepted
criterion in the aerospace industry, where this
confidence check is made most often, is for all of
the generalized mass terms to be unity and all
terms resulting from Equation 8 to be less than
0.1. Often, even under this criteria. an attempt
is madeto adjust the modal vectors so that
Equations 8 and 9 are satisfied [2,3,4].
In Equations 8 and 9 the massmatrix must be an m x
m matrix corresponding to the measurementlocations
on the structure. This meansthat the finite
element massmatrix must be modified from whatever
size and distribution of grid locations required in
the finite element analysis to the m x m square
matrix corresponding to the measurementlocations.
This normally involves somesort of reduction
algorithm as well as interpolation of grid
locations to match the measurementsituation [5,6].
WhenEquation 8 is not sufficiently satisfied, one
(or more) of three situations may exist. First,
the modal vectors can be invalid. This can be due
to measurementerror or problems with the modal
parameter estimation algorithms. This is a very
common assumption and many times contributes to the
problem. Second, the massmatrix can be invalid.
Since the massmatrix is not easily related to the
physical properties of the system, this probably
contributes significantly to the problem. Third,
the reduction of the massmatrix can be invalid
ISI. This can certainly be a realistic problem and
cause severe errors. The most obvious example of
this.situation would be when a relatively large
amount of massis reduced to a measurementlocation
that is highly flexible, such as the center of an
unsupported panel. In such a situation the
measurementlocation is weighted very heavily in
the orthogonality calculation of Equation 8 but may
represent only incidental motion of the overall
modal vector.
In all probability, all three situations contribute
to the failure of Equation 8 to be satisfied on
occasion. WhenEquation 8 is not satisfied, this
result does not indicate where the problem
originates. From an experimental point of view, it
is important to try to develop methods that
indicate confidence that the modal vector is or is
not part of the problem.
Since the residue matrix contains redundant
information with respect to a modal vector, the
consistency of the estimate of the modal vector
under varying conditions such as excitation
location or modal parameter estimation algorithms
112
can be a valuable confidence factor to be utilized
in the process of evaluation of the experimental
modal vectors.
CONSISTENCY
OF MODALVECTORS
Although it has long been known that the frequency
response matrix (and therefore the residue matrix)
contains redundant information, little use of this
fact has been made. In fact, the primary use of
this information has been to justify measuring the
minimumnumberof rows or columns to obtain a
complete set of modal vectors. The primary reason
for measuring the minimumamount of the matrix has
been the time required to obtain and process the
data. With the increasing availability of digital
test equipment, parallel data processing and
parallel data analysis, the time constraint is no
longer as severe. Since it is often not possible
to excite all modal vectors of interest from a
single excitation location, a minimumof two or
three rows or columnswill have to be measured.
Presently, the redundant information obtained in
this manneris examinedvisually to evaluate
discrepancies and determine which modal vector
represents the best estimate. As more rows and/or
columnsare involved, this approach cannot fully
take advantage of the redundant information
efficiently.
Ideally, all data in each residue matrix should be
evaluated, weighted as to importance, and utilized
in somemanner, manual or automatic, in order to
determine the best estimate of the modal vector or
the source of the contamination of the modal
vector. In order to proceed along these lines, a
correlation coefficient between modal vector
estimates would be very beneficial.
MODALASSURANCE
CRITERION
If Equation 7 is used as the basis for a model to
calculate a least squares error estimate of the
proportionality constant between rows or columns of
the residue matrix, the model is linear as follows:
a(c,j,r) = MSF(c,d) a(d,j,r) ( 10)
In vector notation this would be:
IA(c,r)1 q MSF(c,d) IA(d,r)1 (11)
All of the elements of the modal vectors in
rows/columns c and d exhibit the relationship
stated in Equation 11. The value of the modal
scale factor is to be calculated so as to minimize
the sumof the squared errors between corresponding
elements of each modal vector [9]. All or part of
each modal vector can be used in such a
calculation. Obviously, if someelements are
consciously excluded a form of weighted least
squares error estimation is involved. The modal
scale factor is defined, according to this
approach, as follows:
 1 MOM(c,d) I2
MSF(c,d) MAC(c,d) = (18)
MOM(c,c) MOM(d,d)

The nodal assurance OritePiOn iS a Scalar COrdant

relating the causal relationship between two modal
The numerator of Equation 12 can be defined as the vectors. The constant will take on values from
cross moment of the modal vectors. This will be zero, representing no consistent correspondence. to
represented by: one, representing a consistent correspondence. In
this manner. if the modal vectors under
consideration truly exhibit a consistent
n relationship, the modal assurance criterion should
MOM(c,d) = a(c,j,r) itd, j,r) (13) approach unity and the value of the modal scale
c factor can be considered to be reasonable.
j-1

FREQUENCY
RESPONSE
FUNCTION
ANALOGY
Similarly, the denominator of Equation 12 can be
defined as the auto momentof the modal vectors.
This will be represented by: The formulation of the frequency response function
and coherence at a specific frequency is a direct
parallel to the concept of the modal scale factor
and modal assurance criterion. The commonapproach
MOM(d,d) = a(d,j,r) ad,J,r) (14) to the formulation of the frequency response and
f coherence functions is based upon the following
jsl
linear model:
Therefore, Equation 12 can be restated in a more
concise manner
Y = HX+N (19)

MOM(c,d)
MSF(c,d) = (15)
MOM(d,d)
Equation 15 implies that the modal vector of
row/column D is the reference to which the modal
vector of row/column C is compared. In the general
case, modal vector C can be considered to be made
of two parts. The first part will be the part
correlated with modal vector D. The second part
will be the part that is not correlated with modal
vector D and will be madeup of contamination from
other modal vectors and of any randomcontribution.
This error vector will be considered to be noise.
If the modal assurance criterion is defined as a
scalar constant relating the portion of the auto
momentof the modal vector that is linearly related
to the reference modal vector. then the following
equation is applicable:
MAC(c,d) MOM(c,c) = 1 MSF(c,d) I2 MOM(c,d)
This model is used in a least squares error
formulation in order to estimate the frequency
response and coherence functions just as the least
squares approach is used to define the modal scale
factor and the modal assurance criterion. Instead
of momentsbetween modal vectors, the momentsare
calculated between input and output spectrums.
Assumingthat the noise is not correlated with the
input, the estimation of the frequency response
function H(y,x) for an input X and an output Y is
G(y,x)
H(Y,x) = (20 1
'3(x,x)
where*.
G(y,x) = Cross spectrum between output
aild input spectrum
G(x,x) = Auto spectrum of input

(16) The coherctce function is a scalar quantity

indicating the causal relationship between the
output Y biiti ,the input X. It is likewise
Therefore. solving for the modal assurance calculated in a manner corresponding to tlic modal
criterion: assurance criterion.

1 MSF(c,d) te MOM(d,d) I G(Y,x) I2

MAC(c,d) = (17) COH(y,x) = (21)
MOM(c,c) G(x,x) G(Y,Y)

Substituting Equation 15 into Equation 17 gives the where:

final form of the equation defining the modal
assurance criterion. G(Y,Y) = Auto spectrum of output
113
COH(y,x) = Coherence function are consistent. This does not necessarily mean
that they are correct. The modal vectors can be
The reason for noting this parallel analogy is that consistent for any of the following reasons:
much analysis has been done in the area of
frequency response and coherence functions with I) The modal vectors have been incompletely
respect to noise contamination and error measured. This situation can occur whenever
evaluation. Therefore. these equivalent concepts too few response stations have been included
will provide understanding of the applioations and in the experimental determination of the modal
limitations of the modal scale factor and the modal vector. For example, two symmetric modes of
assurance criterion. the wings of an airplane will appear to be
identical if only response stations at the
wing tips are used to define the modal
MODAL ASSURANCE CRITERION PROPERTIES vectors.

The value of the modal assurance criterion can give 2) The modal vectors are the result of a
an indication as to the validity of the modal scale forced excitation other than the desired
factor. While certain implications of the modal input. This would be the situation if, during
assurance criterion are dependent upon the the measurement of the frequency response
calculations involving rows or columns, some function, a rotating piece of equipment with
general discussion is applicable to all cases. an unbalance is present in the system being
tested.
The modal assurance criterion can take on value:
between zero and one. If the modal assurance 3) The modal vectors are primarily coherent
noise. Since the reference modal vector may
criterion has a value near zero. this is an
indication that the modal vectors are not be arbitrarily chosen, this modal vector may
consistent. This can be due to any of the not be one of the true modal vectors of the
system. It could simply be a random noise
following reasons: vector or a vector reflecting the bias in the
modal parameter estimation algorithm. In any
1) The system is non-stationary. This can
occur whenever the system is undergoing a case, themodal assurance criterion will only
change in mass or stiffness during the testing reflect a causal relationship to the reference
period. modal vector.

2) The system is nonlinear. System 4) The modal vectors represent the same modal
vector with different arbitrary scaling. If
nonlineari ties will appear differently in
frequency response functions generated from the two modal vectors being compared have the
different exciter positions or excitation same expected value when normalized, the two
The modal parameter estimation modal vectors should differ only by the
signals. complex valued scale factor which is a
algorithms will also not handle the different
nonlinear characteristiics in a- consistent function of the common modal coefficients
manner. between the rows or columns.

3) There is noise on the reference modal Therefore. if the first three reasons can be
vector. This case is the same as noise on the eliminated, the modal assurance criterion indicates
input of a frequency response function that the modal scale factor is the complex constant
measurement. No amount of signal processing relating the modal vectors and that the modal scale
can remove this type of error. factor can be used to average, difference,.or sort
the modal vectors.
4) The modal parameter estimation is invalid.
The frequency response functions measurements It is very important to notice that the modal
may contain no errors but the modal parameter assurance criterion can only indicate consistency,
estimation may not be consistent with the not validity. If the same errors, random or bias,
data. For example, the modal parameter exist in all modal vector estimates, this will not
estimation algorithm may utilize a complex be delineated by the modal assurance criterion.
system pole model when only real valued syster! Invalid assumptions are normally the cause of this
poles exist. sort of potential error. Even though the modal
assurance criterion is unity, the assumptions
5) The modal vectors are from linearly involving the system or the modal parameter
unrelated mode shape vectors. Hopefully, estimation techniques are not necessarily correct.
since the different modal vector estimates are The assumptions may cause consistent errors in all
from different excitation positions this modal vectors under all test conditions verified by
measure of inconsistency will imply that the the modal assurance criterion.
modal vectors are orthogonal.

COvI.ously, if the first four reasons can be APPLICATIONS

c!.‘,.:inated, the modal assurance criterion can be
!.r.t ,rpreted in a similar way as an orthogonality
cc>. .:ulation. Under the constraints mentioned in the previous
section, the modal assurance criterion can be
If the modal assurance criterion has a value near applied in many different ways. Due to space
unity, this is an indicaticn that the modal vectors limitations, these concepts will be discussed
114
briefly but examples of these types of applications
can be found in References 9 and 10. The modal
assurance criterion can be used to verify or The concept of consistency in the estimate of modal
correlate an experimental modal vector with respect vectors from separate testing constraints is
to a theoretical modal vector (eigenvector). This important considering the potential of multiple
can be done by computing the modal assura.nce estimates of the same modal vector from numerous
criterion between q modal vectors estimated from input configurations and modal parameter estimation
experimental data and r modal vectors estimated algorithms. The computation of modal scale factor
from a finite element analysis evaluated at common and modal assurance criterion results in a complex
stations. This process results in a q by r scalar and a correlation coefficient which does not
rectangular modal assurance criterion matrix with depend on weighting information outside the testing
values that approach unity whenever an experimental environment. Since the modal scale factor and
modal vector and an analytical modal vector are modal assurance criterion are computed analagous to
consistently related. the frequency response function and coherence
function, both the advantages and limitations of
Experimental modal vectors can be averaged, the computation procedure are well understood.
differenoed, or sorted to determine the best single These characteristics, as well as others, provide a
estimate or the potential source of contamination. useful tool in the processing of experimental modal
Since the modal scale factor is a complex scalar vectors.
that allows two vectors to be phased the same and
to the same mean value. these vectors can be
subtracted to evaluate whether the error is random REFERENCES
or biased and, if the error appears to be random
and the modal assurance criterion is high, the 1. Richardson, M.; Kniskern, J.
modal vectors can be averaged, using the modal "Identifying Modes of Large Structures from
scale factor, to improve the estimate of a modal Multiple Input and Response Measurements"
vector. If the error appears to be biased or SAE Paper Number 760875
skewed, the error pattern often gives an indication 1976, 12 pp.
that the error originates due to the location of
the excitation or due to an inadequate modal 2. Gravits, S. I.
parameter estimation process. Based upon partial "An Analytical Procedure for Orthogonalization
but overlapping measurement of two columns of the of Experimentally Measured Modes"
frequency response function matrix, modal vectors Journal of the Aero/Space Sciences
can be sorted, assuming the modal assurance Volume 25
function indicates consistency, into a complete 1958, pp. 721-22
estimate of each modal vector at all measurement
stations. 3. McGrew, J.
"Orthogonalization of Measured Modes and
The modal assurance criterion can be used to Calculation of Influence Coefficients"
evaluate modal parameter estimation methods if a AIAA Journal, Volume 7. Number 4
set of analytical frequency response functions with 1969, PP. 774-776
realistic levels of random and bias errors is
generated and used in common to a variety modal 4. Targoff, W. P.
parameter estimation methods. In this way, "Orthogonality Check and Correction of
agreement between existing methods can be Measured Modes"
established and new modal parameter estimation AIAA Journal, Volume 14, Number 2
1976, pp. 164-167
methods can be checked for characteristics that are
consistent with accepted procedures. Additionally,
this approach can be used to evaluate the 5. Guyan, R.J..
characteristics of each modal parameter estimation "Reduction of Stiffness and Mass Matrices"
method in the'presenoe of varying levels of random AIAA Journal, Volume 3. Number 2
and bias error. February 1965, pp. 380

Finally, the modal assurance criterion can be 6. Irons, B.

utilized in an experimental test procedure to Yitructural Eigenvalue Problems: Elimination
determine the modal vectors without knowing the of Unwanted Variables"
explicit details of the input. AIAA Journal, Volume 3. Number 5
This procedure 961-962
requires implicit knowledge of the general May 1965, pp.
characteristics of the input. The ratio of
resPonses between a reference point and a point of 7. Downs, B.
interest on the structure is measured and averaged "Accurate Reduction of Stiffness and Mass
in the presence of single inputs being applied at Matrices for Vibration Analysis and a
different random points on the structure, one point Rationale for Selecting Master Degrees
of Freedom"
per average. This is repeated for all points of ASME Paper Number 79-DET-18
interest on the structure. The resulting frequency 1979, 5 PP.
domain data is not a frequency response function
but can be used to determine modal vectors if no
modal mass computation is required [9,11].

115
0. Sowers, J.D.
"Condensation of Free Body Mass Matrices
Using Flexibility Coefficients"
AIAA Journal, Volume 16. Number 3
March 1978, pp. 272-273

9. Allemang, R.J.
"hVeStiQStiOn Of SOmeMultiDle InDUt/OUtDUt
Frequent; Response Function Experimental
Modal Analysis Techniques"
Doctor of Philosophy Dissertation
University of Cincinnati
Department of Mechanical Engineering
1980, 358 PP.
10. Allemang, R.J.; Brown, D.L.; Olt, R.A.;
Rost, R.W.; Steedman, J.
"Using Dual Input RandomExcitation
Experimental Modal Analysis"
Final Report - Contract F08635-80-C-0166
Eglin Structural DynamicsLaboratory
1981, 171 PP.
11. Allemang, R.J.; Zimmerman,R.D.;Brown,D.L.
"Determining Structural Characteristics
from ResponseMeasurements"
Presentation-Application of Systems
Identification Techniques
Division af Dynamic Systems and Control
ASMEWinter Annual Me&ng
1979, 39 PP.

116