AIAA JOURNAL

Vol. 45, No. 10, October 2007

Simultaneous Mass, Damping, and Stiffness Updating

for Dynamic Systems

Sau-Lon James Hu∗

University of Rhode Island, Narragansett, Rhode Island 02882-1197
and
Huajun Li†
Ocean University of China, 266100 Qingdao, People’s Republic of China
DOI: 10.2514/1.28605
This paper presents an efficient and systematic approach to simultaneously update the mass, damping, and
stiffness matrices of linear dynamic systems, given few (say two) measured complex vibration modes (complex
eigenvalues and eigenvectors). The method is termed the cross-model cross-mode method because it involves solving a
set of linear simultaneous equations in which each equation is formulated based on the product terms from two same/
Downloaded by PURDUE UNIVERSITY on August 29, 2014 | [Link] | DOI: 10.2514/1.28605

different modes associated with the mathematical and experimental models, respectively. Two numerical examples
are demonstrated: a 4-degree-of-freedom mass–spring–damper system and a 30-degree-of-freedom finite element
model for a cantilever beam. The numerical updating by the cross-model cross-mode method is excellent for all
system matrices when the measured modes are spatially complete and noise free. The cross-model cross-mode
method, together with the Guyan reduction scheme, also performs reasonably well under a spatial incompleteness
situation.

Nomenclature j , 0j , 00j = jth eigenvalue of baseline, updated,

Scales and true models, respectively
r = damping ratio of the rth mode of the
Cij† =

i T C
0j dynamic system
†
Cn;ij =

i T Cn
!r = natural frequency of the rth mode of the
† † dynamic system
Cn;m = same as Cn;ij with new index m to replace ij
cn = nth damping coefficient of the mass–spring–
Vectors and Matrices
damper model
fm† = see Eq. (21) C, C0 , C00 = damping matrix of the baseline, updated, and
Kij† =

i T K
0j true models, respectively
† =

i T Kn
0j
Kn;ij Cn = preselected damping submatrix of the
† = † baseline model
Kn;m same as Kn;ij with new index m to replace ij
kn = nth stiffness coefficient of the mass–spring– C† = Nm  NC complex matrix
damper model cn = element damping matrix of the nth element
Mij† =

i T M
0j of the analytical model
†
Mn;ij =

i T Mn
0j F† = real column vector of size 2Nm , Eq. (25)
†
Mn;m = †
same as Mn;ij with new index m to replace ij f† = complex column vector of size Nm
mn = nth mass coefficient of the mass–spring– G† = see Eq. (24)
damper model G] = pseudoinverse of G† , Eq. (27)
NC = number of correction coefficients for the K, K0 K00 = stiffness matrix of the baseline, updated, and
damping matrix true models, respectively
Ni = number of modes from the baseline model Kn = preselected stiffness submatrix of the
Nj = number of modes from the updated model baseline model
NK = number of correction coefficients for the K† = Nm  NK complex matrix
stiffness matrix kn = element stiffness matrix of the nth element of
NM = number of correction coefficients for the the mass–spring–damper model
mass matrix M, M0 , M00 = mass matrix of the baseline, updated, and
Nm = Ni  Nj true models, respectively

n = nth element of
Mn = preselected mass submatrix of the baseline
n = nth element of  model
n = nth element of  M† = Nm  NM complex matrix
x, x,
_ x
= displacement, velocity, and acceleration
vectors, respectively
Received 29 October 2006; revision received 15 May 2007; accepted for
= vector of correction coefficients for the
publication 18 May 2007. Copyright © 2007 by the American Institute of stiffness matrix
Aeronautics and Astronautics, Inc. All rights reserved. Copies of this paper  = vector of correction coefficients for the mass
may be made for personal or internal use, on condition that the copier pay the matrix
$10.00 per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood
Drive, Danvers, MA 01923; include the code 0001-1452/07 $10.00 in
 = 
 T
correspondence with the CCC.  = vector of correction coefficients for the
∗
Professor, Department of Ocean Engineering; hu@[Link]. damping matrix
†
Professor, Department of Ocean Engineering, 238 Songling Road;
j ,
0j ,
00j = jth mode shape of the baseline, updated, and
huajun@[Link]. true models, respectively
2529
 2530 HU AND LI

Superscripts which are likely to be spatially incomplete and complex valued [1].
Although complex modes occur in real dynamic systems for a variety
T = superscript used as transpose operator of reasons, vibration modes of conventional (i.e., nonrotating) linear
† = superscript used for “cross” terms between structures can be complex only if the damping is distributed in a
baseline and updated models nonproportional way. In real structures, nonproportional damping
 = superscript used for complex conjugates can arise readily because the majority of the damping is found to be
concentrated at the joints between components of a structural
assembly that does not result in a proportional distribution in
Introduction damping. This paper further develops the CMCM method to render it
suitable to damped systems, where the damping matrix can be either
D ISAGREEMENTS often exist between analytical models and
experimentally obtained data. Model updating, defined as the
process of correcting the numerical values of individual parameters
proportional or nonproportional. The precise objective is the
reconstruction of mass, damping, and stiffness matrices using few
in a mathematical model using data obtained from an associated measured complex modes, while maintaining the physical
experimental model such that the updated model more correctly connectivity of the mathematical model. Two particular structural
describes the dynamic properties of the subject structure, has become models—a 4-degree-of-freedom (DOF) mass–spring–damper
a common method to improve the correlation between finite element system and a 30-DOF cantilever beam structure—are to be chosen
models and measured data [1–3]. A number of approaches to the for the numerical examples, where the measured modal information
problem exist, based on the type of parameters that are updated and will be synthesized from using a finite element model that is similar
the measured data that are used. This study belongs to the category of to the analytical model, but with different sets of system coefficients.
updating the complete structural matrices, including mass, damping, In the numerical studies, scenarios with both spatially complete and
Downloaded by PURDUE UNIVERSITY on August 29, 2014 | [Link] | DOI: 10.2514/1.28605

and stiffness matrices, from modal data. incomplete modes will be investigated.
Most papers in the area of finite element model updating (FEMU)
addressed undamped systems. Traditionally, the modal-based
FEMU for undamped systems can be classified into two major Preliminaries: Eigenanalysis of Linear Damped
groups: direct matrix methods [4–6] and indirect physical property Dynamic Systems
adjustment methods [1]. The first of these two groups is generally of Consider the homogeneous equations of motion for a linear
noniterative methods, all of which were based on computing changes structure
made directly to the mass and stiffness matrices. Such changes may
have succeeded in generating modified models which had properties M x
 Cx_  Kx  0 (1)
close to those measured in the tests, but these resulting models
become abstract “representation” models and cannot be interpreted where M, C, and K are mass, damping, and stiffness matrices,
in a physical way. The second group is of those methods that are in respectively; and x, x, _ and x
are displacement, velocity, and
many ways more acceptable in that the parameters which they adjust acceleration vectors, respectively. Matrices M, C, and K are
are much closer to physically realizable quantities. Methods in this restricted to be real and symmetric. Let the solution to Eq. (1) have
second group seek to find correction factors for each individual finite the form:
element or for each design parameter relating to each finite element
and are generally viewed as the main hope for updating technology x 
et (2)
even though they require a much greater computation effort. They are
all iterative, in contrast to the direct formulas of the earlier methods. Substituting Eq. (2) and its time derivatives into Eq. (1) yields
Taking a completely different approach from the traditional methods,
Hu et al. [7] recently developed the cross-model cross-mode
2 M  C  K
 0 (3)
(CMCM) model updating method for the simultaneous updating of
the stiffness and mass matrices. The method was so named because it The solution of Eq. (3) constitutes a complex eigenproblem. When
involves solving a set of linear simultaneous equations for the the structure is with N degrees of freedom, there are 2N eigenvalues
physically meaningful correction factors, in which each equation is occurring in complex conjugate pairs, as a result of the fact that all the
formulated based on the product terms from two same/different coefficients in the matrices are real and thus any eigenvalues (or
modes associated with the mathematical and experimental models, roots) must either be real or occur in complex conjugate pairs. The
respectively. The CMCM method is a noniterative method and corresponding eigenvectors to these eigenvalues also occur as
therefore very cost effective in computational time. It also has the complex conjugates. Denote the eigensolution as r , r ,
r , and
r ,
advantage of preserving the initial model configuration and physical r  1; . . . ; N. In vibration analyses, it is customary to express each
connectivity of the updated model. All the FEMU methods eigenvalue r in the form:
mentioned previously used real-valued mode shapes and natural
frequencies. p













r  !r
r  i 1 r2  (4)
Applying FEMU to damped systems was first conducted by
Friswell et al. [8], who extended the traditional direct methods to p







estimating both the damping and stiffness matrices of a damaged where i  1, !r is the natural frequency, and r is the damping
cantilever beam while assuming that its mass matrix was known. ratio for that mode. The significance of complex eigenvectors is that
Their algorithm has the drawback that it does not guarantee the the mode shapes are complex. In effect, a complex mode is one where
connectivity of the original finite element model. Kuo et al. [9] each part of the structure has not only its own amplitude of vibration
extended the direct method to a more general problem where the but also its own phase. As a result, each part of a structure which is
analytical mass, damping, and stiffness matrices were all allowed to vibrating in a complex mode will reach its own maximum deflection
be updated. Model updating for damped systems also appeared in the at a different instant in the vibration cycle to that of its neighbors
literature under the mathematical term quadratic inverse eigenvalue which all have different phases.
problems (QIEP). There are many articles in the area of QIEP. Often numerical libraries only provide first order eigenvalue
Particularly, Starek and Inman [10] studied the QIEP associated with solvers, thus a transformation from Eq. (3) to its first order form is
nonproportional underdamped systems. The latest progress in needed. For keeping symmetric square matrices for the generalized
solving QIEP has been detailed in a recent book by Chu and Golub eigenvalue analysis, a common 2N state space representation of
[11]. Eq. (1) can be expressed as
A common practice in the industry to measure the modes of a real      
dynamic system has been the experimental modal analysis (EMA). C M x_ K 0 x
 (5)
The EMA data are usually characterized by a small number of modes, M 0 x
0 M x_
 HU AND LI 2531

Solving the generalized eigenproblem of Eq. (5) yields the Cij† 

i T C
0j (14)
eigenvalue r and its corresponding eigenvector in the form as
     
C M
r K 0
r
r  (6)
M 0 r
r 0 M r
r Kij† 

i T K
0j (15)

One thus can immediately obtain the eigenvector
r associated with

Eq. (3) from extracting the first half of the eigenvector shown in †
Eq. (6). Kn;ij 

i T Kn
0j (16)

Cross-Model Cross-Mode Method †

Using Complex Modes Mn;ij 

i T Mn
0j (17)
Throughout this paper, to distinguish symbols associated with the and
models before and after updating, without a superscript is used for the
original (or baseline) model, and the superscript “0” is used for the †
Cn;ij 

i T Cn
0j (18)
updated model. For instance, M0 and M represent the mass matrix of
the updated and baseline models, respectively. For clarity, symbols with superscript “†” throughout this paper are
First, the stiffness, damping, and mass matrices of the baseline “cross” terms calculated from both baseline and updated models.
model, denoted as K, C, and M, respectively, have been initially
Downloaded by PURDUE UNIVERSITY on August 29, 2014 | [Link] | DOI: 10.2514/1.28605

Using a new index m to replace ij, Eq. (12) becomes

modeled. In the proposed cross-model cross-mode approach, the
stiffness matrix K0 of the updated model is a modification of K via X
NK  XNC 
Km†  †

n Kn;m †
 0j Cm  †
n Cn;m
X
NK n1 n1
K0 K 
n Kn (7)  X
NM 
† †
n1  02
j Mm  n Mn;m 0 (19)
n1
where an individual Kn is a preselected stiffness submatrix of the
baseline model;
n are unknown stiffness correction factors to be Rearranging Eq. (19), one obtains
determined; and NK is the number of stiffness correction terms for the
stiffness matrix. Often Kn could be chosen to be the stiffness matrix X
NM
†
X
NC
†
X
NK
†
of each element, and then NK the number of elements. Likewise, one n
02
j Mn;m   n
0j Cn;m 
n Kn;m  fm† (20)
writes the corresponding expression for the mass matrix M0 and n1 n1 n1

viscous damping matrix C0 , respectively, as where

X
NM
fm† 
02 † 0 † †
M0  M n Mn (8) j Mm  j Cm  Km  (21)
n1
When Ni and Nj modes are taken from the analytical model and the
and measured model, respectively, totally Nm  Ni  Nj complex
equations can be formed from Eq. (20). Those equations are named
X
NC CMCM equations in view of the fact that they are formed by crossing
C0 C  n Cn (9) over two models, for two arbitrary modes. Equation (20) can be
n1 rewritten in a matrix form:
where the individual Cn and Mn are preselected submatrices of the K †
 M†   C†   f† (22)
baseline model; n and n are correction coefficients to be
determined; and NM and NC are the number of mass and damping where K† , M† , and C† are Nm -by-NK , Nm -by-NM , and Nm -by-NC
correction coefficients, respectively. complex matrices, respectively.
, , and  are real column vectors
Denoting the jth eigenvalue 0j and eigenvector
0j associated with of size NK , NM , and NC ; and f† is a complex column vector of size
K , C0 , and M0 , one writes
0
Nm . Denoting Re
z and Im
z as the real and imaginary parts of a
complex number z, respectively, one can rewrite Eq. (22) as
K 0
0j  0j C0
0j  02
j M
j  0
0 0
(10)
G †   F† (23)
In the following development, it is assumed that few and 0j
0j
are
known from measurements. Denoting superscript “T” as a transpose where
operator, and premultiplying Eq. (10) by

i T , which can be either a  
Re
K†  Re
M†  Re
C† 
real or a complex mode, yields G†  (24)
Im
K†  Im
M†  Im
C† 


i T K0
0j  0j

i T C0
0j  02
j

i  M
j  0
T 0 0
(11)
 
Re
f† 
Substituting Eqs. (7–9) into the above equation yields F†  (25)
Im
f† 
X
NK  X
NC 
Kij†  †

n Kn;ij  0j Cij†  †
n Cn;ij and
n1 n1 8 9
 X
NM  <
=
 02 M †
  †
n n;ij  0
M (12)   (26)
j ij : ;
n1 

where Note that there are 2Nm real-valued CMCM equations in Eq. (23).
Analytically, one can solve  in Eq. (23) by a standard inverse
Mij† 

i T M
0j (13) operation,   G† F† , if G† is a nonsingular square matrix. For a
1
 2532 HU AND LI

nonsquare matrix G† , where the number of equations does not equal problems. It is mathematically equivalent to a shear building model,
the number of unknowns, the equivalent operator is the to a lumped-mass finite element model of a rod in longitudinal
pseudoinverse. If G† has more rows than columns, an vibration, to a set of point masses vibrating transversely on a taut
overdetermined case where there are more equations than unknowns, string, and to a finite difference or finite element approximation to a
that is 2Nm >
NM  NC  NK , the pseudoinverse is defined as Sturm–Liouville problem [13].
Following the concept of the finite element method, the element
G ] 
G† G†  1 G†
T T
(27) stiffness and damping matrices of the nth element that connects the
(n 1)th and nth masses are given as [14]
for nonsingular
G† G† . The resulting solution,   G] F† is
T
 
optimal in a least-squares sense. kn kn
kn  (28)
When all nonzero stiffness, damping, and mass coefficients of the kn kn
baseline model are allowed to change, this particular situation can be
termed a complete-updating case, in contrast to a partial-updating and
case where one or more nonzero coefficients of M, C, and K are not  
allowed to vary. If the updated matrices K0 , M0 , and C0 in Eqs. (7–9) cn cn
cn  (29)
are replaced by aK0 , aM0 , and aC0 , where a is an arbitrary positive cn cn
constant not equal to 1, it would correspond to a different set of
correction factors. Although these two dynamic systems, The system stiffness matrix K for a 4-DOF mass–spring–damper
characterized by
M0 ; C0 ; K0  and
aM0 ; aC0 ; aK0 , respectively, system can be assembled as
are different in spatial domain, they are identical in the modal space, 2 3
k1  k2
Downloaded by PURDUE UNIVERSITY on August 29, 2014 | [Link] | DOI: 10.2514/1.28605

k2 0 0
namely, they possess the same eigenvalues 0j and eigenvectors
0j . X4
6 k2 k2  k3 k3 0 7
Because the CMCM equations in Eq. (22) or Eq. (23) are derived K Kn  64 0
7 (30)
based on the usage of 0j and
0j , the corresponding solutions for the n1
k 3 k3  k4 k4 5
correction factors should apply to both dynamic systems. From the 0 0 k4 k4
above statements, one concludes that multiple sets of solution for the
correction factors must exist for a complete-updating case. In theory, where Kn denotes the corresponding kn in the global coordinates,
to gain a unique solution for the correction factors, at least an and the system damping matrix C:
additional constraint equation must be imposed. For instance, a 2 3
c1  c2 c2 0 0
particular mass or stiffness term is predetermined, or the total mass of X4
6 c2 c2  c3 c3 0 7
the system is known. C Cn  6
4 0
7 (31)
n1
c 3 c3  c4 c4 5
0 0 c4 c4
Numerical Studies
Numerical examples are given to illustrate the procedure of where Cn denotes the corresponding cn in the global coordinates.
applying the CMCM method and to demonstrate the accuracy of the The corresponding system mass matrix M is written as
method for correcting the stiffness, mass, and damping coefficients 2 3
m1 0 0 0
of an analytical finite element model based on measured complex X4
6 0 m2 0 0 7
modes. In this paper, both the analytical model and the measured M Mn  64 0
7 (32)
modal information are generated from using finite element models, n1
0 m 3 0 5
with different sets of system coefficients. It is assumed that the 0 0 0 m4
analytical model is with a “wrong” set of coefficients, and the goal is
to correct those wrong coefficients from the measured modal where Mn denotes the corresponding mn in the global coordinates.
information which is simulated from the true model with the “right” Numerical values for the system coefficients of the analytical
coefficients. In the following presentation, the term analytical model model are chosen to be identical to those of Pilkey [12]:
refers to the model with K, M, and C, and the term true model refers m1  m4  5, m2  m3  10; kn  1 and cn  0:01 for n
to the model with K00 , M00 , and C00 . Throughout this paper, symbols 1; . . . ; 4. The system matrices K, C, and M of the analytical model
with superscript “00” always denote quantities associated with the true are numerically determined from using Eqs. (30–32). Because the
model. Although the true model and the updated model share the eigenvalues and eigenvectors occur in complex conjugate pairs, only
same symbolic expression such as Eqs. (7–9), they might be different one root from each pair is needed. Performing the eigenanalysis
numerically. Two particular structural models—a 4-DOF lumped based on Eq. (6), one obtains the four eigenvalues as 1 
mass–spring–damper system, and a 30-DOF cantilever beam 0:0001  0:1256i, 2  0:0007  0:3864i, 3  0:0018
structure—are chosen for the numerical examples. Those structural 0:5922i, and 4  0:0024  0:6959i, and the corresponding
models were investigated previously by Pilkey [12] and Friswell eigenvectors are
et al. [8], respectively. 8 9 8 9
>
> 1:0000 >
> >
> 1:0000 > >
< = < =
1:9211 1:2535
4-DOF Mass–Spring–Damper System
1  ;
2 
>
> 2:5392 >
> >
> 0:3646 >>
Consider the 4-DOF mass–spring–damper system shown in : ; : ;
2:7566 1:4383
Fig. 1. This 4-DOF structure was studied by Pilkey [12] who focused 8 9 8 9
on estimating the damping matrix while assuming that mass and >
> 1:0000 >> >
> 1:0000 > >
< = < =
stiffness matrices were known. Using this simple example allows the 0:2465 0:4211

3  ;
4 
detailed numerical results being presented, and also provides the >
> 1:3715 >
> >
> 0:1969 > >
reader an opportunity to verify the correctness of the CMCM method : ; : ;
1:8202 0:1386
independently. The mass–spring–damper model also has been
investigated by many researchers in the area of inverse eigenvalue Those eigenvectors are obtained to be real vectors because the
system damping of the analytical model has been taken to be a
proportional damping, noting C  0:01K.
The true system stiffness matrix K00 is synthesized by using Eq. (7)
with the above K and assigned quantities
1  0:4,
2  0:1,

3  0:2, and
4  0:1. Similarly, the way to produce M00 and
C00 is according to Eqs. (8) and (9) with 1  0:1, 2  0:2,
Fig. 1 Sketch of a 4-DOF mass–spring–damper system. 3  0:3, 4  0:35, 1  0:1, 2  0:3, 3  0:2, and
 HU AND LI 2533

4  0:15. Sequentially, performing the eigenanalysis associated

with M00 , K00 , and C00 , one obtains the four eigenvalues as 001 
0:0001  0:1329i, 002  0:0006  0:3408i, 003  0:0020
0:5901i, and 004  0:0028  0:7798i, and the corresponding
complex eigenvectors:
8 9 8 9
>
> 1:0000 0:0000i > > >
> 1:0000 0:0000i > > Fig. 3 Sketch of a 10-element cantilever beam.
< = < =
2:2004 0:0009i 1:7977 0:0021i

001  ;
002 
>
> 3:2676 0:0015i > > >
> 0:2365 0:0010i > >
: ; : ; inertia are 1:25  10 3 m2 and 2:604  10 7 m4 , respectively.
3:7670 0:0019i 1:8311  0:0024i
8 9 8 9 Following the standard formulation for a uniform beam element in a
>
> 1:0000 0:0000i > > >
> 1:0000 0:0000i > > plane, one computes the element stiffness matrix kn and element
< = < =
0:8482 0:0039i 0:2149 0:0006i (consistent) mass matrix mn of a size 6  6. There are 10 nodal points

003  ;
004  and each nodal point has 3 DOFs, thus matrices K and M both are
>
> 3:7909  0:0066i > > >
> 0:0746  0:0004i > >
: ; : ; with a size 30  30. With 10 structural elements, the global stiffness
2:3523 0:0017i 0:021 0:0002i
matrix of the analytical model is obtained as
In the current implementation of the CMCM method, all four
modes from the analytical model and the first two complex modes X
10

from the true model are employed. Thus, total 16 real-valued CMCM K Kn
equations can be formed. An additional constraint equation (for n1
Downloaded by PURDUE UNIVERSITY on August 29, 2014 | [Link] | DOI: 10.2514/1.28605

scaling purpose) is that the change of the total mass of the system is
presumably known. In the present numerical example, one has and that of the true model as

X
4
X
10
n mn  0:25 K 00 
1 
n Kn
n1 n1

to be consistent with the simulated data. There are 12 unknown

correction coefficients to be estimated. With 17 equations and 12 where Kn is the corresponding kn in global coordinates. Likewise,
unknowns, one solves it as an overdetermined case. Shown in Fig. 2 for the global mass matrix of the analytical and true models, one has
is resulting estimates of the stiffness correction coefficients
n , the
mass correction coefficients n , and the damping correction X
10
M Mn
coefficients n for n  1; . . . ; 4, plotted against those coefficients n1
originally used to generate K00 , M00 , and C00 . It is found that the
correction factors estimated from applying the CMCM method
and
match perfectly with the preset values.
X
10
30-DOF Cantilever Beam System M 00 
1  n Mn
The CMCM method will be tested on a 10-element steel cantilever n1
beam shown in Fig. 3, where each element is modeled as a uniform
beam element. The beam is of length 1 m, breadth 25 mm, and respectively, where Mn is the corresponding mn in global
thickness 50 mm. Young’s modulus is taken as 2:05  1011 Pa, and coordinates. The analytical model is considered to be an undamped
mass density 7860 kg=m3 (or mass density per unit length system, that is C  0, and the global damping matrix of the true
9:825 Kg=m). The cross-section area and the associated moment of model is assumed to have the form

Fig. 2 Comparison of the preset and estimated correction coefficients
n , n , and n for the 4-DOF mass–spring–damper system.
 2534 HU AND LI
Downloaded by PURDUE UNIVERSITY on August 29, 2014 | [Link] | DOI: 10.2514/1.28605

Fig. 4 Comparison of the preset and estimated correction coefficients
n , n , and n for the beam model when two spatially complete modes are used.

X
10 In the present CMCM implementation, 10 bending modes from
C 00  n Cn the analytical model and two complex bending modes from the true
n1 model are used, thus 40 real-valued CMCM equations can be
formed. Those CMCM equations, together with a scaling constraint
where Cn is considered to have a scaled form of Kn . Precisely, it is equation that the first stiffness term is unchanged, that is,
1  0, are
assumed Cn  10 5 Kn . Note that the interpretation of n differs sufficient to solve for the 30 correction coefficients. Shown in the top
from that of
n or n . The term n Cn is an extra damping term from panel of Fig. 4 are the resulting estimations of the stiffness correction
zero damping, rather than a modification of an existent Cn . Although coefficients
n , plotted against the preset coefficients which were
n must be positive to justify a positive energy dissipation, the value used to generate K00 . Likewise, the estimated and preset mass
of
n or n must be greater than 1 to justify a positive stiffness or correction coefficients n and damping correction coefficients n are
mass term. shown in the middle and bottom panels of Fig. 4, respectively. For
Similar to the case study in Friswell et al. [8], the true model is providing quantitative detail of those correction coefficients shown
considered to be a damaged cantilever beam. The damage is modeled in Fig. 4, Table 1 lists all the preset and estimated values as well as
as a reduction in the stiffness of element 4 by 25% from the analytical their relative errors. At the damaged element, the relative errors of
4
model, that is
4  0:25. The damping in element 4 is assumed to and 4 are 8:73  10 7 and 4:05  10 7 , respectively, which are
be 5  10 5 times the stiffness of element 4, that is 4  5. This practically negligible. The worst updating occurs at 1 , which has a
damaged model is motivated by the realization that damage often relative error 0.12. Overall, the numerical results indicate that the
reduces stiffness and adds damping locally. In addition to
4  correction coefficients estimated from applying the CMCM method
0:25 and 4  5, the parameters of the true model are considered to match nicely with the target values.
be slightly different from those of the analytical model for other In the above calculation, only two complex modes of the damaged
elements as well. Specifically, the true model is produced with the beam are used, but all DOFs of the measured mode are assumed
quantities
n generated by using a Gaussian random number known. If only a subset of the DOFs is measured (a spatially
generator with the mean equal to 0 and standard deviation equal to incomplete situation), then the analytical model must be reduced or
0.02, and n being generated based on mean 0 and standard deviation the measured mode shapes must be expanded. When only the
0.01. Physically, damping terms must be positive. The quantities for translational DOFs in the transverse direction are measured for the
n are generated by taking the absolute value of the Gaussian random first two complex modes and the Guyan reduction scheme [15] is
numbers based on mean 0 and standard deviation 0.25. used to reduce the analytical model, implementing the CMCM

Table 1 The preset correction coefficients (denoted by
n , n , and n ) and their estimations (denoted by
^n , ^n , and ^n ) for the beam model when two
spatially complete modes are used

^ n
n ^n n ^ n n
n
n
^ n
n
n ^n n
n ^ n n

1 0 1:22e 17 —— 7:26e 03 8:13e 03 1:20e 01 3:34e 01 3:34e 01 2:06e 07

2 2:51e 03 2:51e 03 7:28e 06 5:88e 03 6:00e 03 1:93e 02 1:79e 01 1:79e 01 2:32e 07
3 5:75e 03 5:75e 03 1:38e 05 2:18e 02 2:19e 02 1:18e 03 4:06e 01 4:06e 01 4:72e 07
4 2:50e 01 2:50e 01 8:73e 07 1:36e 03 1:36e 03 3:04e 05 5:00e  00 5:00e  00 4:05e 07
5 2:38e 02 2:38e 02 2:65e 05 1:14e 03 1:13e 03 9:52e 03 2:15e 01 2:15e 01 1:66e 06
6 2:38e 02 2:38e 02 4:32e 05 1:07e 02 1:07e 02 1:53e 03 3:14e 01 3:14e 01 8:98e 07
7 7:53e 04 7:51e 04 2:61e 03 5:93e 04 5:75e 04 3:07e 02 3:98e 01 3:98e 01 1:08e 06
8 6:55e 03 6:54e 03 7:75e 04 9:56e 04 9:42e 04 1:55e 02 3:60e 01 3:60e 01 2:34e 06
9 3:49e 03 3:51e 03 4:09e 03 8:32e 03 8:33e 03 8:86e 04 1:43e 01 1:43e 01 4:68e 06
10 3:73e 03 3:81e 03 2:04e 02 2:94e 03 2:95e 03 5:51e 04 9:99e 02 9:99e 02 1:00e 04
 HU AND LI 2535
Downloaded by PURDUE UNIVERSITY on August 29, 2014 | [Link] | DOI: 10.2514/1.28605

Fig. 5 Comparison of the preset and estimated correction coefficients
n , n , and n for the beam model when only translational DOFs of two modes are
measured.

method based on the same consideration previously yields the result unchanged masses of the adopted analytical model are different from
shown in Fig. 5. The estimations of the stiffness and damping terms, those of the true model, which has been employed to generate
including those of the damaged element, are reasonably well, except measurements. Figure 7 compares the preset and estimated
the mass estimate of the first element. It is interesting to note that if correction coefficients
n and n when only the 10 translational
two rotational DOFs associated with element 4 are also measured, DOFs of the first mode are measured. The estimations for both
then the result improves significantly as shown in Fig. 6. damping and stiffness correction coefficients seem pretty good.
Because mass terms usually can be modeled properly and are In practice, modal measurements always contain errors. Figure 8
unlikely to change significantly even when damages occur, the shows the updating results based on using a corrupted measured
remaining numerical study considers not updating mass terms. When mode that possesses a 0.1% proportional random error. Identical
only damping and stiffness terms are to be updated, using just one mass modeling error and spatial incompleteness mentioned in Fig. 7
measured complex bending mode, along with 10 bending modes are also included while obtaining Fig. 8. The value of the corrupted
from the analytical model, is sufficient in the implementation of the measured mode at each nodal point is generated by multiplying its
CMCM method. Neglecting the update for the mass terms is true value with a factor (1  ), where  has been simulated from a
equivalent to introducing the modeling error in mass because the Gaussian random number generator with mean 0 and standard

Fig. 6 Comparison of the preset and estimated correction coefficients
n , n , and n for the beam model when 10 translational and two rotational DOFs
of two modes are measured.
 2536 HU AND LI
Downloaded by PURDUE UNIVERSITY on August 29, 2014 | [Link] | DOI: 10.2514/1.28605

Fig. 7 Comparison of the preset and estimated correction coefficients
n and n for the beam model when n are not updated and 10 translational DOFs
of the first mode only are measured.

Fig. 8 Comparison of the preset and estimated correction coefficients
n and n for the beam model when n are not updated and 10 translational DOFs
of a noisy second mode only are measured.

deviation 0.001. Figure 8 clearly exhibits the fact that the stiffness updating problems for linearly damped vibrating systems, where the
and damping terms at the free end of the beam are more influenced by damping matrix could be either proportional or nonproportional. The
the measurement noise. precise objective was to simultaneously update the mass, damping,
Note that the numerical observations related to the mode and stiffness matrices of the finite element model for a dynamic
incompleteness, modeling error, and measurement noise in this system, when two or more “measured” complex modes were
article are problem and parameterization dependent. A theoretical provided. The system reconstruction would satisfy modal and
investigation of the effect on the CMCM method due to various structural constraints simultaneously, where the modal constraint
sources of error remains to be done in the future. refers to matching the specified modes (eigenvalues and
eigenvectors) and the structural constraint refers to keeping the
desirable features associated with the mathematical model, including
Conclusions the physical connectivity of the finite element model, and the
This paper developed a simple, efficient, and systematic approach, symmetric real-valued mass, damping, and stiffness matrices. The
named the cross-model cross-mode method, for solving model accuracy of the CMCM method was demonstrated numerically by
 HU AND LI 2537

two simulated examples, a 4-DOF mass–spring–damper system and No. 11, 1978, pp. 1208–1210.
a 30-DOF finite element model for a damaged cantilever beam. In [5] Berman, A., and Nagy, E. J., “Improvement of a Large Analytical
both examples, when two spatially complete modes were available, Model Using Test Data,” AIAA Journal, Vol. 21, No. 8, 1983, pp. 1168–
1173.
the reconstruction of all system (mass, damping, and stiffness)
[6] Wei, F.-S., “Structural Dynamic Model Improvement Using Vibration
matrices was found to be excellent. In the second example, when Test Data,” AIAA Journal, Vol. 28, No. 1, 1990, pp. 175–177.
only 10 transverse DOFs associated with the 30-DOF mathematical [7] Hu, S. J., Li, H., and Wang, S., “Cross-Model Cross-Mode Method for
model were measured, the reconstruction by the CMCM method, Model Updating,” Mechanical Systems and Signal Processing, Vol. 21,
together with the classical Guyan reduction scheme, for all system No. 4, 2007, pp. 1690–1703.
matrices was found to be satisfactory also. The proposed CMCM [8] Friswell, M. I., Inman, D. J., and Pilkey, D. F., “Direct Updating of
method has provided a new direction on solving the quadratic inverse Damping and Stiffness Matrices,” AIAA Journal, Vol. 36, No. 3, 1998,
eigenvalue problems. The mathematical procedure of the CMCM pp. 491–493.
method seems to have enough generality that, with some suitable [9] Kuo, Y., Lin, W., and Xu, S., “New Methods for Finite Element Model
modifications, it can be applied to other types of partially described Updating Problems,” AIAA Journal, Vol. 44, No. 6, 2006, pp. 1310–
1316.
inverse eigenvalue problems as well. [10] Starek, L., and Inman, D. J., “Design of Nonproportional Damped
Systems via Symmetric Positive Inverse Problems,” Journal of
Acknowledgments Vibration and Acoustics, Vol. 126, No. 2, 2004, pp. 212–219.
[11] Chu, M. T., and Golub, G. H., Inverse Eigenvalue Problems: Theory,
H. Li acknowledges financial support by the 863-project of China Algorithms, and Applications, Oxford Univ. Press, Oxford, England,
(Program No. 2006AA09Z331), and by the China National Science U.K., 2005.
Fund for Distinguished Young Scholars under the grant [12] Pilkey, D. F., “Computation of a Damping Matrix for Finite Element
Downloaded by PURDUE UNIVERSITY on August 29, 2014 | [Link] | DOI: 10.2514/1.28605

no. 50325927. Model Updating,” Ph.D. Dissertation, Department of Engineering

Mechanics, Virginia Polytechnic Institute and State University,
Blacksburg, VA, 1998.
References [13] Gladwell, G. M. L., Inverse Problems in Vibration, Martinus– Nijhoff,
[1] Ewins, D. J., Modal Testing: Theory, Practice and Application, 2nd ed., Dordrecht, The Netherlands, 1986.
Research Studies Press, Baldock, Hertfordshire, England, U.K., 2000. [14] Craig, R. J., Structural Dynamics: An Introduction to Computer
[2] Friswell, M. I., and Mottershead, J. E., Finite Element Model Updating Methods, Wiley, New York, 1981.
in Structural Dynamics, Kluwer Academic, Norwell, MA, 1995. [15] Guyan, R. J., “Reduction of Stiffness and Mass Matrices,” AIAA
[3] Mottershead, J. E., and Friswell, M. I., “Model Updating in Structural Journal, Vol. 3, No. 2, 1965, p. 380.
Dynamics: A Survey,” Journal of Sound and Vibration, Vol. 167,
No. 2, 1993, pp. 347–376.
[4] Baruch, M., “Optimization Procedure to Correct Stiffness and B. Balachandran
Flexibility Matrices Using Vibration Data,” AIAA Journal, Vol. 16, Associate Editor