Professional Documents
Culture Documents
Isma2010 0103
Isma2010 0103
Abstract
Since its original development in the early eighties, the modal assurance criterion (MAC) has become a
major tool in modal analysis. This measure of the degree of linearity (consistency) between two modal
vectors provides an invaluable means of comparing and contrasting modal vectors from different origins.
However, to the authors’ knowledge, its extension to complex vectors has not been addressed thoroughly.
Such situations occur when the system under study does not comply with the classical assumptions of struc-
tural dynamics. This paper proposes a generalization of the MAC criterion that is applicable to any types
of vectors. It also introduces an enhanced version of this criterion that tremendously improves its ability to
discriminate modes when few measurements are available.
1 Introduction
The Modal Assurance Criterion (MAC) is an essential tool in modal analysis. Its use to perform the pairing
between two sets of modal vectors is now widespread. The great success of the MAC has fostered the
emergence of numerous derivative criteria designed to deal with more specific situations.
However a rigorous treatment of the complex case still seems to be lacking. Such situations are relatively
frequent and occur for systems that do not fully comply with the classical assumptions of structural dynamics.
The analysis of the aeroelastic behaviour of an aircraft is a typical example of such a situation. The other
drawback of the MAC criterion resides in its limitation to discriminate efficiently the modes when only a
few measurements are available. Again, the in-flight analysis of an aircraft structure is an illustration of this
problem since the number of sensors is sometimes limited.
The development of criterions that are both appropriate to complex modes and to situations with few mea-
surements constitutes the topic of this paper. It is composed of four main sections. The first one briefly
reviews the classical assumptions in structural dynamics and the definition of the MAC criterion. It also
introduces notations used in the paper. The following section deals with the definition of a degree of com-
plexity which quantifies how much a vector differs from a real-valued one. This analysis is important because
it establishes the basic principle which shows why the MAC criterion is not quite appropriate to process com-
plex vectors. The third section of the paper is devoted to an extension dubbed MACX of the MAC criterion
to complex vectors. This extension is constructed from the physical interpretation of the modal contributions
on the measurements of a system. Finally the last section addresses the enhancement of the MACX criterion
for situations with a limited number of measurements. This improved version named MACXP integrates
the information about the modal frequencies and dampings in order to increase the discrimination between
modes.
2713
2714 P ROCEEDINGS OF ISMA2010 INCLUDING USD2010
Mechanical models are based on the fundamental equations of dynamics. They are composed of a set of
second order differential equations on the generalized displacements of a structure. These displacements are
gathered in a vector denoted q. A modeling of the damping which quantifies the dissipation of the energy in
the structure is also necessary. The most common practice is to use a viscous damping model which leads to
linear equations which have the following form:
M q̈ + D q̇ + K q = B U
(1)
Y = C0 q + C1 q̇ + C2 q̈
M, D and K designate the matrices of mass, damping and stiffness. The last equation of the model describes
how the vector of measurements Y is connected to the displacements q. The form adopted here encompasses
the various types of measurements used in structural analysis: displacements, velocities, accelerations.
Two additional hypotheses are formulated for modeling the dynamics of a structure:
• the matrices M, D et K are positive symmetric matrices
• they are diagonalizable in the same basis.
With such conditions, the eigenvectors of the system 1 are identical to those of the associated undamped
system (D = 0). If moreover the damping matrix D is composed of small components, the poles of the
systems λk are complex and stable.
With these hypotheses, each pair of conjugate poles (λk , λk ) shares a common eigenvector ψk and this
vector is real-valued. The mode defined by the quantities λk , λk and ψk is said to be real. When the above
hypotheses are not complied with, each pole λk is associated to a complex eigenvector ψk . The mode is
said to be complex. It is characterized by the pair of conjugate poles (λk , λk ) and the pair of conjugate
eigenvectors (ψk , ψ k ).
A mode shape is a vector µk that characterizes how a pole λk affects the measurements Y . It is given by the
following expression:
µk = C0 + C1 λk + C2 λ2k ψk (2)
In this general formulation, the mode shape µk is not real-valued even if ψk is a real vector. However, if the
measurements performed on the system are all of the same type, the vector µk is proportional to a real-vector.
In this case, only one matrix amongst C0 , C1 and C2 in the model is non-zero. It is designated Ci and we
have
µk = λik (Ci ψk ) (3)
Complex vectors which are equal to the product of a complex scalar by a real-valued vector are named
“monophase” because the phases of their components are all equal modulo π.
The Modal Assurance Criterion (MAC) is a measure of the degree of linearity between two vectors. Given
two vectors µ1 , µ2 , it is defined by
2
| µ∗1 µ2 |
MAC(µ1 , µ2 ) = = cos2 (µ1 , µ2 ) (4)
k µ1 k k µ2 k
M ODEL UPDATING AND CORRELATION 2715
where ∗ designate the conjugate transpose of a complex vector. The product µ∗1 µ2 is called the Hermitian
inner product between two vectors.
This definition can be interpreted geometrically since the MAC criterion depends on the angle between two
vectors. This criterion can be applied to both real-valued and complex-valued vectors. It is insensitive to the
modulus and the phase of the vectors µ1 and µ2 . For this reason, it is well-suited to the analysis of monophase
vectors. On the other hand, we can notice the value of the MAC is sensitive to conjugate operations on the
vectors µ1 and µ2 .
As recommended by Allemang [1], this criterion is appropriate to compare two sets of modes and perform
pairing between these sets. One should however be aware of a few limitations. First the value of the MAC
is sensitive to the large components in the vectors µ1 and µ2 . If this difference on the amplitude of the
associated measurements is systematic, one could possibly applied scalings on the vectors in order to equalize
their contributions in the computation of the MAC. For instance, on aircraft, sensors installed on the wing
tips produce larger values than those at the center of the fuselage.
The other defect of the MAC criterion is its reduced sensitivity when the number of components in the
vectors µ1 and µ2 is small. Few spatial information is then available to differentiate mode shapes. This
phenomenon is illustrated by the figures 1 and 2 where the modes of two aeroelastic models are compared
with the MAC criterion. In the first case, only 18 measurements are considered. We can notice that modes
quite apart exhibit significant MAC values. The situation is much clearer in figure 2 with 123 measurements.
In practice [2], two vectors are considered correlated when the MAC is greater than 0.9 which correspond to
an angle lower than 18 degrees. They are judged uncorrelated when the MAC is lower than 0.6 which means
that they are separated by an angle greater than 39 degrees.
18 measurements
1
0.9
1 0.8
0.8 0.7
0.6 0.6
0.4 0.5
0.2 0.4
0.3
0
30 0.2
20 30
20 0.1
10 10
0
Model 1 0 0
Model 2
3 Complexity criterion
In practical situations, the actual mode shapes of a physical system are not exactly monophase vectors even
if the aforementioned hypotheses for structural modeling are supposed fulfilled. It is therefore necessary to
define a criterion to quantify how distant a complex vector is from a monophase one.
2716 P ROCEEDINGS OF ISMA2010 INCLUDING USD2010
123 measurements
1
0.9
1 0.8
0.8 0.7
0.6 0.6
0.4 0.5
0.2 0.4
0.3
0
30 0.2
20 30
20 0.1
10 10
0
Model 1 0 0
Model 2
In the literature, several indicators of the complexity of a vector µ are available. The thesis by Purekar
[3] presents several of them. For illustrative purposes, we begin this section by a graphical approach of
complexity. We then concentrate on a criterion called “Modal Phase Collinearity” (MPC) developed by
Pappa, Elliott, and Schenk [4]. A geometrical interpretation of the complexity of a vector is also presented
and its connection with the MAC criterion is established.
A simple approach to analyze the complexity of a vector consists in plotting its components in the complex
plane as illustrated in figure 3. One can then easily determine if these components tend to be aligned accord-
ing to a specific direction. One can also visually assess the dispersion of the components about this mean
direction. This gives an indication of the degree of complexity of the vector µ. For instance, in figure 3, the
first vector is close to a monophase vector whereas the second one is really a complex vector.
Mode 1 Mode 6
Imag. part
Imag. part
In order to quantify the complexity of a vector µ, Pappa et al. [4] defined a criterion that makes use of the
real part µr and the imaginary part µi of µ. The variance and the covariance of these parts are computed
according to
Sxx = µ>r µr
Syy = µ>
i µi (5)
Sxy = µ>
r µi
This phase normalization process is illustrated in figure 4 which depicts the rotation by the angle βe applied to
the components of a vector in the complex plane. The angle βe maximizes the alignment of the components
with the real axis.
It can be show that the optimum value βe is equal to
∠(µ> µ)
βe = mod π (8)
2
where ∠(z) designates the phase of the complex scalar z and > is the transpose operation. If we denote µ e
the phase normalized vector µ
e=µe −j βe
and µ
er et µ
ei its real and imaginary part respectively, we have the
following equalities
1
kµ er k2 = k µ k2 + µ> µ
2
(9)
2 1 2
>
kµei k = kµk − µ µ
2
2718 P ROCEEDINGS OF ISMA2010 INCLUDING USD2010
imag. axis
imag. axis
~
β
real axis real axis
This quantity lies between 0 and 1. It is equal to 1 in when µ is monophase and to 0 when k µ
ei k = k µ er k.
By substituting the relations 9 in the expression of C(µ), we come upon a very simple form of this criterion
which directly depends on the vector µ
> !2 > !2
µ µ µ µ
C(µ) = = (11)
µ∗ µ k µ k2
We are now in position to establish two remarkable equalities. First, by considering the quadratic form
associated with the matrix defined in 3.2, we can show that the criterion C(µ) is strictly equivalent to the
MPC criterion defined in 7. However, the expression 11 provides a much more concise way of computing
this quantity. Moreover, it is based on simple geometric considerations.
Expression 11 can also be recast in the following form
2
| µ∗ µ |
MPC(µ) = C(µ) = (12)
kµk kµk
where µ is the conjugate of µ. Comparing with the definition 4 of the MAC criterion, we deduce that
MPC(µ) is nothing mere than the MAC computed for the vector µ and its conjugate µ:
This result shows that, when the vector µ is not monophase, the angle between a complex vector and its
conjugate is non-zero.
M ODEL UPDATING AND CORRELATION 2719
In this section, we first consider the direct application of the MAC criterion to complex vectors which reveals
the limitations of the classical formulation of the MAC. We then proceed to the definition of an extension of
this criterion to complex vectors that we call MACX.
When complex modes are considered, each mode is characterized by a pair of complex-conjugate mode
shapes. If we want to compute a degree of similarity between two modes using the MAC criterion, we might
wonder, as illustrated by the diagram 5, which vector in each pair of conjugate vectors, we might select
for computing the MAC criterion. When one of the mode shapes µ1 or µ2 is monophase, this choice has
absolutely no incidence on the value of the MAC. But if the two vectors are complex then two values are in
fact possible since the associations of the same color in the diagram 5 produce identical results.
µ1 µ2
?
_ _
µ1 µ2
This selection can lead to significant differences for the values of the criterion. This is illustrated by figure 6
where the differences MAC(µ1 , µ2 ) − MAC(µ1 , µ2 ) between the two possibles values for the MAC are
plotted for the two models considered in figure 2. We notice that these differences range from -0.64 to 0.92.
123 measurements
0.8
1 0.6
0.5 0.4
0 0.2
−0.5 0
−1 −0.2
30
20 30 −0.4
20
10 10 −0.6
0 0
Which of the two cases in figure 5 should we adopt to compute the MAC Criterion?
A first solution would be to make use of the sign of the imaginary part of the poles λ1 and λ2 . These
imaginary parts represent the angular frequencies of the oscillations of the modes. The solution1 would be
1
Figures 1 and 2 were calculated in this way.
2720 P ROCEEDINGS OF ISMA2010 INCLUDING USD2010
to select the two vectors which are associated with poles λ1 et λ2 with imaginary parts of the same sign.
However logic this solution may be, it bears no relation with the underlying physical phenomenons. The
contribution of a complex mode on the measurements is in fact the combination of the two complex-conjugate
mode shapes associated to this mode. The value of the output vector Y (t) at time t is given by
nm
X nm
X
Y (t) = µk xk (t) + µk xk (t) = 2 Re µk xk (t) (14)
k=1 k=1
for some value of the angle β. In order to compare two complex modes, we have then to compare the vectors
of their contributions on the system outputs which are given by the real vectors υ1 (β1 ) et υ2 (β2 ). For given
phases β1 et β2 , the similarity between these vectors depends on their correlation which is equal to the scalar
product
υ1>(β1 ) υ2 (β2 ) (16)
If this product is equal to zero, then the contributions of the two modes are orthogonal.
If one of the mode shapes µ1 or µ2 is monophase, MAC(µ1 , µ2 ) = 0 implies υ1>(β1 ) υ2 (β2 ) = 0 for any
angles β1 et β2 . But if both mode shapes are complex, it can be shown that the scalar product will be equal
to zero only if the following equality between β1 and β2 holds
β1 + β2 = ∠(µ>
1 µ2 ) ± π/2 (17)
In this situation, the MAC criterion clearly gives an underestimated value of the similarity between the two
mode shapes because it indicates completely dissimilar vectors whereas the contributions of these modes will
be orthogonal only for two specific values of the sum β1 + β2 on the interval [ 0 2π [. Finally, for complex
vectors, the fact that υ1>(β1 ) υ2 (β2 ) = 0 for all values of β1 and β2 is equivalent to MAC(µ1 , µ2 ) = 0 and
MAC(µ1 , µ2 ) = 0.
This result shows that, to decide on the full orthogonality of the contribution of two complex modes, i.e. for
all values of β1 and β2 , we must consider simultaneously the two associations in diagram 5. This explains
why the conventional MAC formulation is not sufficient for complex vectors.
In this section, we propose an extension of the MAC that is valid for any type of vectors. The approach
adopted is quite similar to the phase normalization of complex vectors developed in subsection 3.3. The
definition of the new criterion is based on the correlation 16 computed for angles βb1 and βb2 which maximize
the absolute value of this quantity. These angles solve the following optimization problem which is illustrated
in figure 7
υ1 (β1 ) = Re µ1 e−j β1
(βb1 , βb2 ) = arg max υ1>(β1 ) υ2 (β2 ) with
β1 , β 2
υ2 (β2 ) = Re µ2 e−j β2
M ODEL UPDATING AND CORRELATION 2721
~
β
imag. part
10
^β
1
5
0
real part real part real part
imag. part
10
~β 5
2
0
real part real part real part
b1 = µ1 e−j β1 = µ
µ br1 + j µ
bi1 e1 = µ1 e−j β1 = µ
µ er1 + j µ
ei1
b e
b2 = µ2 e−j β2 = µ
µ br2 + j µ
bi2 e2 = µ2 e−j β2 = µ
µ er2 + j µ
ei2
b e
By using equations 9 and 19, we can derive a simpler expression that can be computed directly from the
vectors µ1 and µ2
2
| µ∗1 µ2 | + µ>1 µ2
MACX(µ1 , µ2 ) = ∗ (20)
µ1 µ1 + µ >
1 µ1
µ∗2 µ2 + µ>
2 µ2
This formulation is quite similar to the classical MAC since we only need to perform the following substitu-
tion in the expression 4 of the MAC:
1
| µ∗1 µ2 | −→ | µ∗1 µ2 | + µ>
1 µ 2 (21)
2
123 measurements
0.8
0.6
0.4
0.2
−0.2
30
20 30
25
20
10 15
10
5
0 0
Model 1 Model 2
In subsection 2.3, we mention that the MAC is of little use when a limited number of measurements is avail-
able. This is illustrated by figure 1 where two aeroelastic models are compared using only 18 measurements.
The MACX criterion has the same drawback. Anyhow, in some instances, it would be very useful to be able
to discriminate modes with few measurements.
In this section, we present an improvement of the MACX criterion that takes into account the poles associated
to the mode shapes. We first present the concept of pole weighting that has already been used in the literature.
Then we define a criterion named MACXP based on this concept.
Pole weighting consist in integrating, in the MACX criterion, the information about the poles associated
to the mode shapes for a better mode differentiation. Instead of computing the criterion on the sole mode
shapes, the idea is to use the free-decay responses realted to each pole.
Given the pole λ and the associated mode shape µ, the expression of the free-decay response at time t is
given by
d(t) = µ x(t) = x0 µ eλ t (22)
were x0 is the initial complex state associated to this pole. For each pole we can form a vector V by
concatenating the free-decay response at several equispaced instants (0, ∆t, 2 ∆t, · · · , nr ∆t) where ∆t
is a constant period. Such a vector is defined by
µ
µ e λ ∆t
V = . (23)
.
.
µ enr λ ∆t
The pole weighted version of the MACX criterion is nothing mere than the application of the MACX formu-
lation to two of these vectors
We can notice incidentally that, even if the mode shape is monophase, the vector V will not be monophase
as soon as nr > 1. Thus, even for real modes, the use of a complex version of the MAC is unavoidable.
Pole-weighted versions of the MAC have already been worked out in the literature. This approach was used
by Juang [5] at NASA to evaluate the quality of modes identified with the ERA algorithm (Eigensystem
Realization Algorithm). This concept was also studied more recently by Phillips and Allemang [6] who
named it pwMAC (Pole-Weighted MAC). Anyhow the complex nature of the vectors V in equation 23 was
not addressed.
The MACXP criterion proposed in this paper is simply the limit of the pole-weighted version of the MACX
criterion defined in 24 when ∆t tends towards 0 and nr towards infinity:
It can be shown that the Hermitian and scalar products between vectors V1 and V2 in the expression of the
MACX criterion have the following limits
Z ∞ Z ∞
lim V1∗ V2 ∆t = d1 (t)∗ d2 (t) dt lim V1> V2 ∆t = d1 (t)> d2 (t) dt (26)
∆t −→ 0 0 ∆t −→ 0 0
nr −→ ∞ nr −→ ∞
where d1 (t) and d2 (t) are the free-decay responses of the poles λ1 and λ2 as defined in 22. Under the
assumption that these poles are stable, the integrals in these equations are given by
Z ∞ Z ∞
∗ −1 ∗ −1
d1 (t) d2 (t) dt = µ1 µ2 d1 (t)> d2 (t) dt = µ∗ µ 2 (27)
0 λ1 + λ2 0 λ1 + λ2 1
Substituting for these results in the definition 25 we obtain the following formulation of the MACXP crite-
rion:
2
| µ∗1 µ2 | | µ>
1 µ2 |
+
| λ1 +λ2 | | λ1 +λ2 |
MACXP(µ1 , µ2 ) = (28)
µ∗1 µ1 | µ>1 µ1 | µ∗2 µ2 | µ>
2 µ2 |
2 | Re λ1 | + 2 | λ1 | 2 | Re λ2 | + 2 | λ2 |
As compared to the basic MACX formulation 20, we clearly see in the above expression the influence of
pole weighting. The MACXP criterion can be interpreted in terms of the correlation function between the
real decay responses associated to two modes: it can be shown that the square root of this criterion is quite
close to the maximum value of the modulus of this correlation function normalized by the L2 -norm of these
free-decay responses.
The application of this criterion for the comparison of the aeroelastic models with few measurements appears
in figure 9. We can see that the situation is tremendously improved as compared to figure 1. For instance,
this latter criterion clearly reveals an exchange between modes 24 and 26.
18 measurements
1
0.9
1
0.8
0.5 0.7
0.6
0
0.5
30
0.4
20 0.3
0.2
10 30
25 0.1
20
15
Model 1 10 0
0 5
0 Model 2
6 Conclusion
This paper is devoted to the study of criterions so as to compare complex modes when a small number of
measurements are available. This study explains why the utilization of the MAC criterion is limited to real
modes. Two extensions of this criterion are proposed.
The first one called MACX enlarges the application of the MAC criterion to complex modes. Moreover this
criterion is strictly equivalent to the MAC when real modes are considered. The second criterion named
MACXP dramatically betters the analysis of modes when few measurements are available.
One can think of several applications in modal analysis that can take advantage of these criterions:
• mode matching in stabilization diagrams
• association of identified modes in various experimental situations (for instance, different flight condi-
tions for the aeroelastic testing of an aircraft)
• comparison of results between several identification algorithms
• mode pairing between a model of the physical system and the identification results for this system.
• ...
Acknowledgements
The authors greatly thank Stéphane Leroy, Aurélien Cordeau, Adrien Pavie and Adrien Bérard at Airbus for
their technical support on the aeroelastic models.
References
[1] Randall J. Allemang. The modal assurance criterion: Twenty years of use and abuse. Sound and vibration
Magazine, 37(8):14–23, 2003.
[2] Etienne Balmès, Jean-Philippe Bianchi, and Jean-Michel Leclère. Structural Dynamics Toolbox - FEM-
Link, User’s Guide. SDTools, Vibration Software and Consulting, version 6.1 edition, Septembre 2009.
[3] Dhanesh M. Purekar. A Study of Modal Testing Measurement Errors, Sensor Placement and Modal
Complexity on the Process of FE Correlation. Master of science, University of Cincinnati, Department
of Mechanical Engineering of the College of Engineerring, 2005.
[4] Richard S. Pappa, Kenny B. Elliott, and Axel Schenk. Consistent-mode indicator for the eigensystem
realization algorithm. Journal of Guidance, Control, and Dynamics, 16(5):852–858, Spetember–October
1993.
[5] Jer-Nan Juang. Applied System Identification. PTR Prentice Hall, 1994.
[6] Allyn W. Phillips and Randall J. Allemang. Data presentation schemes for selection and identification
of modal parameters. In 23th International Modal Analysis Conference, Orlando, 2005.
2726 P ROCEEDINGS OF ISMA2010 INCLUDING USD2010