Damage Identification in a Truss Tower by Regularized

Model Updating
Benedikt Weber1 and Patrick Paultre2
Downloaded from [Link] by Guru Nanak Dev Engineering College on 11/07/22. Copyright ASCE. For personal use only; all rights reserved.

Abstract: This paper presents a sensitivity-based damage identification of a three-dimensional truss tower tested in the laboratory. A
finite-element model is updated by modal parameters obtained from ambient vibration measurements. The paper focuses on details of
modeling and model updating. To get a realistic model, it was necessary to include eccentric connections, bending stiffness in truss
members, and foundation flexibility. For model updating, a number of mathematical techniques are combined in a consistent way,
including regularization of the nonlinear updating problem and its linearization. Considering all these details correctly, damage in the truss
has been successfully identified. However, it is also demonstrated how neglecting some of these algorithmic details can lead to incorrect
results. It is also shown how static condensation can lead to a model that is valid for the undamaged case but cannot adequately represent
the damaged case.
DOI: 10.1061/共ASCE兲ST.1943-541X.0000105
CE Database subject headings: Damage; Trusses; Parameters; Vibration; Measurement; Sensitivity analysis; Laboratory tests.
Author keywords: Model update; Regularization; Damage identification; Sensitivity-based; Truss.

Introduction for damage identification is the benchmark problem developed by

the joint IASC-ASCE Task Group on Structural Health Monitor-
This paper describes damage identification in a truss tower. ing. The benchmark structure is a four-storey rectangular steel
Trusses have a long tradition in damage identification, with a building. Damage has been simulated by removing bracing within
large body of literature on the subject. Only a few selected refer- the structure at different stories and by loosening bolts on beams.
ences which include experimental data are considered here. A The benchmark consisted of an analytical phase with simulated
major experimental effort was the National Aeronautics and data 共Johnson et al. 2004兲 and an experimental phase with mea-
Space Administration 共NASA兲 eight-bay three-dimensional truss sured data 共Dyke et al. 2003兲. Most publications related to this
in the 1990s. Data from these experiments were used extensively benchmark determine first the modal parameters and then con-
to verify a whole series of new damage detection algorithms, such struct or update a structural model 共Lus et al. 2004; Caicedo et al.
2004; Yuen et al. 2004; Lam et al. 2004; Giraldo et al. 2003;
as the eigenstructure assignment 共Zimmerman and Kaouk 1992兲,
Ching and Beck 2004兲. These publications focus more on system
best achievable eigenvectors 共Lim and Kashangaki 1994兲, and the
identification than on the structural model. Most writers based the
minimum rank update 共Zimmerman and Kaouk 1994; Doebling
damage identification on a three-dimensional shear-type building
1996兲. Classical sensitivity-based update methods have also been
with 12 degrees of freedom 共DOF兲. In this case, the stiffness
used for trusses. Ricles and Kosmatka 共1992兲 used a two-stage
matrix has a regular structure and can be determined relatively
procedure: they first determined the location of damage by the
easily. Other methods are based on the flexibility matrix 共Bernal
residual force method and then performed a sensitivity-based up-
and Gunes 2004兲 or the damage index 共Barroso and Rodriguez
date of parameters in the damaged region. Weighting and regular- 2004兲.
ization have been taken into account by statistical concepts. Amin In this paper, we present damage identification in a three-
et al. 共2002兲 used a sensitivity-based update of frequencies using dimensional truss tower with a sensitivity-based algorithm pre-
constrained optimization. The NASA experiments as well as the sented earlier 共Weber et al. 2007, 2009兲. In contrast to the papers
other experiments cited here use measured accelerations in three cited above for the ASCE benchmark problem, the emphasis is
directions at each node. not so much on system identification but more on model updating
A more recent effort to compare commonly used techniques and parameter estimation of the structural model. To be useful in
practical applications, the algorithm has to work with a limited
1
Senior Scientist, Empa, Swiss Federal Laboratories for Materials number of sensors and a few modes from output-only system
Testing and Research, Structural Engineering Research Laboratory, CH- identification. The update algorithm also has to cope with inevi-
8600 Duebendorf, Switzerland. E-mail: [Link]@[Link] table errors in the modal parameters and in the finite-element
2
Professor, Univ. of Sherbrooke, Dept. of Civil Engineering, Sher- model. Since the problem is generally ill-conditioned, inconsis-
brooke, Canada J1K 2R1 共corresponding author兲. E-mail: patrick. tencies between the structural model and the modal parameters
paultre@[Link]
can completely invalidate the results if no special measures are
Note. This manuscript was submitted on July 16, 2008; approved on
September 4, 2009; published online on September 10, 2009. Discussion taken. As presented in Weber et al. 共2007, 2009兲, the problem of
period open until August 1, 2010; separate discussions must be submitted ill-conditioning can be solved with regularization, which is a stan-
for individual papers. This paper is part of the Journal of Structural dard tool in the mathematical literature. However, when applied
Engineering, Vol. 136, No. 3, March 1, 2010. ©ASCE, ISSN 0733-9445/ consistently, regularization affects all parts of the update algo-
2010/3-307–316/$25.00. rithm including the Gauss-Newton iteration as well as line search

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / MARCH 2010 / 307

J. Struct. Eng., 2010, 136(3): 307-316

 and stopping criterion. What has been found in previous work by rk+1 = rk − Sk共ak+1 − ak兲 共6兲
numerical simulations is verified in this paper for the experimen-
tal analysis: the consistent algorithm yields excellent damage where rk = r共ak兲 and Sk = sensitivity matrix at ak. Minimizing the
identification results, while some of the more intuitive versions norm of the linearized residual leads to the following Gauss-
greatly degrade procedure performance. Besides algorithmic as- Newton iteration step:
pects, this paper also demonstrates the importance of constructing ak+1 = ak + 共STk Sk兲−1STk rk 共7兲
a meaningful model that captures the essential behavior of the
undamaged and the damaged structure.
Regularization

Theoretical Background The update problem is generally ill-posed and small errors in
measurements can lead to unreasonably large errors in the param-
Downloaded from [Link] by Guru Nanak Dev Engineering College on 11/07/22. Copyright ASCE. For personal use only; all rights reserved.

The theoretical and numerical aspects of model updating for dam- eters in a. To avoid this problem, we use Tikhonov regularization
age identification were presented in Weber et al. 共2009兲 and are 共Hansen 1998; Vogel 2002兲. Applying regularization to the update
summarized here for the reader’s convenience. step in Eq. 共7兲 leads to
ak+1 = ak + 共STk Sk + ␥I兲−1STk rk 共8兲
Model Updating This update is frequently found in the literature but is inconsis-
The first step is the parametrization of the global stiffness matrix tent. The problem is that the effect of regularization diminishes as
the residual decreases and is essentially lost after a large number
p
of iterations. A consistent formulation is found by applying the
K = K共0兲 + ⌬K = K共0兲 + 兺
j=1
␣ jK共j兲 共1兲 regularization before linearization 共Vogel 1987兲. This is achieved
by minimizing the functional
The first matrix K共0兲 = stiffness matrix of the reference model 共un-
J共ak+1兲 = 储r共ak+1兲储22 + ␥储ak+1储22 共9兲
damaged state兲 and the matrices K共j兲 = stiffness matrices of the
elements or element groups being updated. The partial stiffness This functional consist of two parts: the first part–the residual
changes are expressed by the update parameters ␣ j, which are functional–minimizes the difference between measured and cal-
collected in vector a = 关␣1 , . . . , ␣ p兴. Model updating is performed culated quantities; the second part–the penalty functional–restricts
by minimizing the weighted residual the size of the update parameters. The tradeoff between the two

冋 册
objectives is controlled by the regularization parameter ␥. After
W␭r␭共a兲 linearization 关Eq. 共6兲兴, the functional is
r共a兲 = 共2兲
W␾r␾共a兲
J共ak+1兲 = 储rk − Sk共ak+1 − ak兲储22 + ␥储ak+1储22 共10兲
This residual contains the eigenvalue residual r␭共a兲 and the eigen-
Minimizing Eq. 共10兲 leads to the update equation
vector residual r␾共a兲. The weighting matrices W␭ and W␾ allow
scaling of the eigenvalues and eigenvectors to similar magnitudes. ak+1 = ak + 共STk Sk + ␥I兲−1关STk rk − ␥ak兴 共11兲
For each updated mode i, the eigenvalue residual has a com-
ponent Note the extra term compared to Eq. 共8兲. The regularized solution
is less accurate in matching measurements, but it avoids exces-
␭i共a兲 sively large changes in the model due to noise. The optimal
关r␭共a兲兴i = 1 − 共3兲 tradeoff, that is, the optimal regularization parameter ␥, can be
␭meas
i
found by minimizing the generalized cross-validation function
Likewise, for each mode i, the eigenvector residual has a subvec- 共O’Sullivan and Wahba 1985兲
tor
1
储r␥储22
␾meas ˜ i共a兲
␾ n

冋 册
关r␾共a兲兴i =
i
− 共4兲 GCV共␥兲 = 2 共12兲
z̃Ti ␾meas 1
i z̃Ti ␾
˜ i共a兲 trace共In − A␥兲
n
with z̃i = any vector such that the denominators are not zero. The
tilde is used here to indicate that only measured DOF are consid- where A␥ = matrix
ered. Scaling both the measured and the numerical mode shapes A␥ = 共S␥TS␥ + ␥I兲−1S␥T 共13兲
the same way ensures that the residual is zero for an exact match.
The simplest procedure would be to scale a particular DOF to and r␥ and S␥ = residual and the sensitivity matrix at the con-
one. This condition would be expressed by a vector z̃i with one at verged regularized solution, respectively. The regularized residual
the particular DOF and zero elsewhere. However, we prefer to use decreases with decreasing regularization parameter, but the de-
a mass-weighted scaling of all measured DOF and select z̃i as nominator in Eq. 共12兲 also decreases at the same time. Some-
where, the quotient takes its minimal value, indicating the optimal
z̃i = M̃␾meas 共5兲 regularization parameter.
i
In the nonlinear scenario, the update procedure starts with a
where M̃ = mass matrix taken at the measured DOF only. high value of ␥ and evaluates the update parameters ak+1 itera-
The residual depends nonlinearly on the update parameters a. tively. After convergence, the generalized cross-validation func-
The nonlinear least-squares problem is solved by the Gauss- tion is calculated. Then, the value of ␥ is decreased and the
Newton method, which solves the linearized least-squares prob- update is repeated using the results from the previous calculation
lem iteratively. The linearization at iteration k is given by as starting values. In this way, the generalized cross-validation

308 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / MARCH 2010

J. Struct. Eng., 2010, 136(3): 307-316

 function is calculated for a whole range of regularization param- derivatives can be calculated simply by solving a system of linear
eters and the optimal value can be determined. Starting with a equations. The numerical complexity is similar to Nelson’s
large regularization parameter rather than a small one makes the method but the implementation is much simpler.
procedure more stable.
Line Search and Stopping Criterion
Derivatives of Eigenvalues and Eigenvectors
There are a few other important details to consider when imple-
To calculate the sensitivity matrix, we need the partial derivatives menting a successful update algorithm. If the nonlinearity of the
of the eigenvalues and eigenvectors with respect to the update residual with respect to the update parameters is very strong, the
parameters. Although, for the axisymmetric truss analyzed in this update steps resulting from the Gauss-Newton iteration can be-
paper, repeated eigenvalues do occur theoretically, we can avoid come too large such that the functional Eq. 共9兲 is increasing in-
Downloaded from [Link] by Guru Nanak Dev Engineering College on 11/07/22. Copyright ASCE. For personal use only; all rights reserved.

this problem by making the analytical model slightly asymmetric. stead of decreasing. This can be avoided by reducing the step
Only simple eigenvalues are thus considered in the following. length to ␩⌬ak with ␩ ⬍ 1. The functional Eq. 共9兲 in the direction
Classical formulas for derivatives of simple eigenvalues and re- of the update step is
lated eigenvectors can be found in the literature 共Fox and Kapoor
1968; Nelson 1976兲. Here, we use a less known formulation by ␺共␩兲 = 储r共ak + ␩⌬ak兲储22 + ␥储ak + ␩⌬ak储22 共21兲
Andrew et al. 共1993兲. Since we are only interested in stiffness The requirement that the functional has to decrease is formally
changes, the mass is assumed to be constant in the following stated by the sufficient decrease condition, also known as the
derivation. second Wolfe condition or Armijo condition 共Nocedal and Wright
The derivative of the eigenvectors by the classical formulas is 1999兲
based on the constraint that the eigenvectors remain mass-
normalized ␺共␩兲 ⬍ ␺共0兲 + c1␩␺⬘共0兲 共22兲
⳵ ␾i where c1 = small positive number, typically c1 = 10−4. If the suffi-
␾Ti M =0 共14兲 cient decrease condition is not satisfied, the step length is re-
⳵ ␣j
duced. A simple way to determine a value for ␩ is to approximate
However, for the scaling used in Eq. 共4兲 it is advantageous to the function ␺共␩兲 by a parabola defined by ␺共0兲, ␺⬘共0兲 and ␺共1兲
enforce the following normalization condition 关orthogonally con- and to minimize this parabola. To avoid very small steps, an ad-
strained eigenvectors according to Andrew et al. 共1993兲兴: ditional condition that ␩ ⱖ 0.5 is applied. If the sufficient decrease
condition is still not satisfied for the reduced step, further reduc-
⳵ ␾i
zTi =0 共15兲 tions are required until the condition is satisfied. More details are
⳵ ␣j given in Weber et al. 共2009兲.
where zi = constant nonzero vector. If we take zi to be the same as A second detail of implementation is the stopping criterion.
z̃i at the measured DOF and zero elsewhere, Eq. 共15兲 implies The update has converged when the functional Eq. 共9兲 is close to
zero. This criterion is not practical, however, since the possible
⳵␾˜i minimum is unknown. A more suitable criterion based on the
z̃Ti =0 共16兲 gradient is given in Dennis and Schnabel 共1996兲. To apply it, the
⳵ ␣j
linearized functional Eq. 共10兲 is expressed as
This normalization involves only measured DOF and fits the scal-
ing used in the residual. The derivative of the scaled eigenvectors J共ak+1兲 = 储r̂k − Ŝk共ak+1 − ak兲储22 共23兲
appearing in the sensitivity matrix is

冋 册
with

冋冑 册 冋冑 册
⳵ ␾i共a兲 1 ⳵ ␾i共a兲
= 共17兲 Sk rk
⳵ ␣ j z̃Ti ␾
˜ i共a兲 z̃i ␾i共a兲 ⳵ ␣ j
T˜
Ŝk = and r̂k = 共24兲
␥I − ␥ak
Because of the condition in Eq. 共16兲, there is no additional term
from the denominator in the left-hand side. Now consider the The proposed criterion is then cos ␪ ⬍ tol, where tol = tolerance
eigenvalue problem 共tol = 0.002 in the examples兲 and

共K − ␭iM兲␾i = 0 共18兲 r̂Tk Ŝk共ŜTk Ŝk兲−1ŜTk r̂k

cos ␪ = 共25兲
where K depends on ␣ j as given in Eq. 共1兲 and M = constant. 储r̂k储2 · 储Ŝk共ŜTk Ŝk兲−1ŜTk r̂k储2
Differentiating Eq. 共18兲 with respect to ␣ j leads to
⳵ ␾i共a兲 ⳵ ␭i共a兲
共K − ␭iM兲 = − K共j兲␾i + M␾i 共19兲
⳵ ␣j ⳵ ␣j Application to an Aluminum Truss Tower
This equation can be combined with the normalization condition
关Eq. 共15兲兴 using bordered matrices Experiments

冋 K − ␭iM − M␾i
zT 0
册冋 册冋
⳵ ␾ i/ ⳵ ␣ j
⳵ ␭ i/ ⳵ ␣ j
=
− K共j兲␾i
0
册 共20兲
The proposed method has been applied to the truss tower shown
in Fig. 1共a兲. It consists of L-shaped aluminum members con-
nected with single steel bolts and has a total height of 2.8 m. For
As shown in Andrew et al. 共1993兲, the resulting system of linear easy fabrication, members were bolted directly together without
equations is nonsingular and gives the desired derivatives of the using gusset plates 关Fig. 1共b兲兴. While this design made it easy to
eigenvector and the eigenvalue. The solution with bordered ma- assemble the structure, it turned out that the eccentric connections
trices is very attractive because the eigenvalue and eigenvector complicated the finite-element model considerably. The truss is

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / MARCH 2010 / 309

J. Struct. Eng., 2010, 136(3): 307-316

 Reference
35

36 34

27
28
26
25

Damage

19
20 22.09 Hz 23.44 Hz 53.47 Hz
Downloaded from [Link] by Guru Nanak Dev Engineering College on 11/07/22. Copyright ASCE. For personal use only; all rights reserved.

18
17
Fig. 3. Modes of damaged truss from output-only test

11
12 the first two modes clearly separate due to damage and the fre-
10
9 quency of the third mode drops down considerably. The mode
shapes of the undamaged and the damaged trusses are more dif-
ficult to compare visually.
y
x
Finite-Element Model
(a) (b) (c) A first finite-element model was constructed with simple truss
elements neglecting the eccentricity of the connections. The hori-
Fig. 1. Aluminum truss tower: 共a兲 model in laboratory; 共b兲 eccentric zontal bracing elements were modeled by two diagonal elements
connections; and 共c兲 sensor locations for output-only tests at every second level. These elements were necessary to retain the
quadratic shape of the truss cross section. Pinned boundary con-
ditions were used to attach the four legs to the ground. This model
attached to a one-inch-thick aluminum plate mounted on a strong showed frequencies much higher than measured and was not con-
floor. Although the original intention was to provide rigid bound- sidered realistic. A closer look at the measurements revealed that
ary conditions, the foundation turned out to have significant flex- the foundation was more flexible than anticipated. Therefore, a
ibility, which has to be included in the finite-element model. rigid base plate was introduced that is connected to the ground
Output-only vibration tests were performed for determining with two rotational springs. These springs allow the model to be
frequencies and mode shapes. Four accelerometers were used: tuned to the first two lateral frequencies. The rotational flexibility
two fixed accelerometers were placed at the top to measure ref- introduced at the base corresponds approximately to two extra
erence values and two moving accelerometers were used to mea- levels below the surface. The frequency of the torsional mode is
sure values at every second level, as shown in Fig. 1共c兲. Since not affected by the rotational flexibility of the base and was still
there was not enough ambient input in the laboratory, the truss much too high 共88.3 Hz instead of 58.8 Hz兲.
was excited by hitting it with a hammer. A revised model was built with beam elements for the vertical
Experimental frequencies and mode shapes were determined members and truss elements for other members, as shown in Fig.
with the Frequency Domain Decomposition algorithm in ARTE- 4共a兲. Three nodes were introduced at each joint to model eccentric
MIS 共2005兲. The first three modes of the undamaged truss are connections. The angle sections of the vertical members have
shown in Fig. 2. Numerical values are provided below in Table 2. their principal axes for bending at 45° with respect to the coordi-
Note the almost identical frequencies for the first two modes. Also nate system of the global geometry, with the strong axis passing
note that these modes move diagonally to the x and y axes. To through the central vertical axis of the truss. Although the dis-
simulate damage, a diagonal member was removed 关Fig. 1共c兲兴.
The corresponding modes are shown in Fig. 3. Numerical values
are provided below in Table 4. The almost identical frequencies of (a) (c)
29
(b)
13 14

25

9 10

21

5 6

17

1 2

23.07 Hz 23.68 Hz 58.84 Hz

Fig. 4. Finite-element model of truss: 共a兲 model; 共b兲 detail deforma-
Fig. 2. Modes of undamaged truss from output-only test tion; and 共c兲 parameterization

310 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / MARCH 2010

J. Struct. Eng., 2010, 136(3): 307-316

 Table 1. Frequencies for the Symmetric and Asymmetric Models Com-
pared to Measured Frequencies of Undamaged Truss 共Hz兲
Mode 1 2 3
Symmetric model 23.19 23.19 57.97
Asymmetric model 23.19 23.21 57.99
Symmetric model with mass change 23.24 23.24 58.15
Measured 23.07 23.68 58.84

diagonal element removed to simulate damage corresponds to pa-

Downloaded from [Link] by Guru Nanak Dev Engineering College on 11/07/22. Copyright ASCE. For personal use only; all rights reserved.

23.19 Hz 23.19 Hz 57.97 Hz rameter 21. The parametrization will be discussed in more detail
in the section on damage identification.
Fig. 5. Modes from symmetric finite-element model Frequencies and mode shapes of the finite-element model are
shown in Fig. 5. Although the frequencies of the undamaged truss
are modeled quite accurately, the analytical mode shapes differ
tance between these nodes is small, bending about the weak prin- significantly from those measured. While the analytical modes
cipal axis of the vertical beams introduces significant additional move basically in the x- and y-directions, the measured modes are
flexibility. This unexpected behavior is mainly visible in torsional rotated by 45°. To see this fact more clearly, the mode shapes are
modes, as shown in Fig. 4共b兲. There are two other important shown again in a different format in Fig. 6. In this figure, the
points to mention concerning this model: 共1兲 the mass of the horizontal planes where accelerations were measured are plotted
horizontal elements needs to be computed as a consistent mass one above the other. For the symmetric model, there is indeed a
because using a lumped mass would yield a wrong inertial mass difference between the experimental mode shapes and those de-
for the torsional modes and 共2兲 rotational DOF about the vertical termined from the finite-element model. The problem here is that
axis are only relevant for torsion of the vertical beam element as the first two modes have very close frequencies and are thus not
truss elements do not participate. This last point leads to local very well determined. This fact is inherent to the eigenvalue prob-
torsional modes of the vertical beams, which are not physically lem and well-documented in the literature 共Golub and Van Loan
present. These local modes are suppressed by fixing the corre- 1996兲. Mathematically, the bordered matrix in Eq. 共20兲 is ill con-
sponding rotational DOF. As before, foundation flexibility is mod- ditioned if some frequencies are very close and singular if two
eled by a rigid base plate connected to the ground by rotational frequencies are identical. The formulas by Fox and Kapoor 共1968兲
springs. The stiffness is adjusted such that the frequencies of the and those by Nelson 共1976兲 have similar problems.
first two modes match the measured frequencies. These springs As an easy way out, we can modify the model and make it
are kept constant for subsequent model updating. To reduce the slightly asymmetrical. If we increase parameter 1 corresponding
computational cost, the additional DOF for the eccentric connec- to a vertical member at the base by 1%, the analytical modes
tions have been condensed out with static condensation. The pa- indeed rotate and move close to the measured modes as shown in
rametrization of the front vertical plane is shown in Fig. 4共c兲. The Fig. 6 for the asymmetrical model. The comparison of the corre-

Fig. 6. Comparison of measured and analytic modes for undamaged truss

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / MARCH 2010 / 311

J. Struct. Eng., 2010, 136(3): 307-316

 Table 2. Mode Shapes of Asymmetric FEM Model Compared with Measured Values of Undamaged Truss
Mode 1 Mode 2 Mode 3
DOF FEM Measurement FEM Measurement FEM Measurement
10x ⫺0.108 ⫺0.096 0.114 0.126 ⫺0.146 ⫺0.140
18x ⫺0.249 ⫺0.237 0.266 0.293 ⫺0.271 ⫺0.265
26x ⫺0.403 ⫺0.324 0.431 0.505 ⫺0.356 ⫺0.362
34x ⫺0.554 ⫺0.437 0.594 0.734 ⫺0.393 ⫺0.404
12x ⫺0.108 ⫺0.094 0.114 0.118 0.146 0.156
20x ⫺0.249 ⫺0.233 0.266 0.279 0.271 0.280
28x ⫺0.403 ⫺0.321 0.431 0.490 0.356 0.364
Downloaded from [Link] by Guru Nanak Dev Engineering College on 11/07/22. Copyright ASCE. For personal use only; all rights reserved.

36x ⫺0.554 ⫺0.549 0.594 0.558 0.393 0.389

9y 0.105 0.113 0.118 0.083 0.146 0.139
17y 0.244 0.265 0.271 0.235 0.271 0.260
25y 0.396 0.440 0.438 0.373 0.356 0.347
36y 0.545 0.587 0.602 0.485 0.393 0.338
11y 0.105 0.114 0.118 0.095 ⫺0.146 ⫺0.151
19y 0.244 0.269 0.271 0.248 ⫺0.271 ⫺0.279
27y 0.396 0.442 0.438 0.367 ⫺0.356 ⫺0.366
35y 0.546 0.616 0.602 0.527 ⫺0.393 ⫺0.405

sponding frequencies shown in Table 1 confirms that this change Damage Identification
is negligible. The numerical values of the mode shapes of the
asymmetrical finite-element model and the measured mode For damage identification, the asymmetric finite-element model
shapes are given in Table 2. with rotational springs for the foundation flexibility is used. The
Table 1 also shows the influence of the mass change when rotational springs are kept constant and are not updated since the
removing a diagonal to simulate damage. Typical damage in main interest is in damage in the truss itself. Also, it would be
structures such as cracking or loose connections only changes the difficult, if not impossible, to distinguish between stiffness
stiffness but not the mass. Therefore, the algorithm considers only changes in the foundation and in the lower part of the truss with
stiffness changes. However, in the experiment, the mass changes the current sensor configuration. As will be shown, this model
when removing a member. This change is, however, quite small. matches the undamaged physical model quite well and is, there-
In terms of measured weights, the mass of a diagonal member is fore, used as base model for damage identification of both the
only 0.85% of the total weight of the structure without base plate undamaged and the damaged structure. The update with the un-
and 0.57% with base plate. This mass change does not affect the damaged structure could be used as a new base model, but the
modal parameters significantly as demonstrated by the frequen- improvement would be minimal.
cies of a model with a massless diagonal shown in row “Symmet- An important step in damage identification is model parametri-
ric model with mass change.” zation. For the truss presented here, the number of parameters
A considerable effort has been undertaken to get a good finite- becomes quite large if every element is considered separately.
element model. Starting from a simple model with truss elements Since measurements are only available at every other level, we do
and coincident nodes, the model has been further refined to obtain not expect to be able to distinguish damage in one member from
the final model with beam elements for the vertical members and damage in the member just above it. It seems, therefore, adequate
truss elements for the horizontal and diagonal members with ec-
centric connections. Material properties of the aluminum profiles
have been verified by axial tension tests and by weighing. Nev- −0.7
10
ertheless, there remain several modeling uncertainties. For in-
Residual

stance, the connections are assumed to be pinned, whereas in

reality there is considerable friction. However, since bending in −0.8
10
the horizontal and diagonal members seems to be negligible, so is
the friction in the connections, at least for the lower modes used 10
−6 −4
10 10
−2
10
0
10
2

for damage identification. Another uncertainty is the boundary γ

−2
condition on the ground. The rigid plate with rotational springs 10
works fine in this application, but the reality is, of course, more
GCV

complex. The practical experience gained with this project is that 10

−3

constructing a reliable model takes time and engineering experi-

ence. A simple variation of the material properties often does not −4
10
give the full picture. For instance, using a distributed mass instead 10
−6 −4
10 10
−2
10
0
10
2

of a lumped mass in the horizontal members turned out to be γ

much more important than using the exact weight of the mem-
bers. Fig. 7. Residual and generalized cross-validation of undamaged truss

312 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / MARCH 2010

J. Struct. Eng., 2010, 136(3): 307-316

 1 1

α 0 0

α
−1 −1
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Parameter Parameter

Fig. 8. Update parameters for undamaged truss Fig. 10. Update parameters for undamaged truss, determined with an
inconsistent algorithm
Downloaded from [Link] by Guru Nanak Dev Engineering College on 11/07/22. Copyright ASCE. For personal use only; all rights reserved.

to use only one parameter for two vertical members or for two
diagonal members situated one above the other between two mea-
update formula. Another important aspect is model condensation.
suring levels. Horizontal members and bracing elements are not
In this paper, static condensation of the eccentric nodes has been
updated, since damage in them is almost impossible to detect.
used to reduce computation time. However, translational and ro-
This is due to the fact that reducing the stiffness in these elements
tational DOF 共except torsion of vertical members兲 have been re-
has almost no influence on the frequencies and mode shapes. The
tained for the principal nodes at each level. In another context,
proposed grouping leads to 32 parameters: 16 for vertical mem-
model condensation is sometimes proposed to cope with incom-
bers and 16 for diagonal members. The association of the mem-
plete measurements. To see if this is a valid approach, the model
bers with parameters is shown in Fig. 4共c兲 for the front vertical
has been reduced to the measured DOF. This reduced model
plane. Using 3 modes and 16 sensors yields 48 equations for the
works fine for the undamaged case, as can be seen from the up-
mode shapes and 3 equations for the frequencies. The total of 51
date parameters presented in Fig. 11. However, it does not ad-
equations is used to determine 32 unknown parameters. Consid-
equately represent the damaged truss, as will be shown below.
ering only two modes would only give 34 equations, which is not
The frequencies corresponding to the different updating proce-
enough for the generalized cross-validation to work.
dures are compared in Table 3. It should be noted that the regu-
As described in the previous section, a considerable effort has
larized update of the asymmetric model gives the correct update
been made to construct a good finite-element model closely
parameters close to zero but not the closest match of the frequen-
matching the measured modes. An update of the undamaged
cies to the measured frequencies. This is exactly the point of
model is thus not strictly necessary. However, to be useful for
regularization. Using a very small regularization parameter or the
monitoring, a damage identification algorithm has also to be able
inconsistent algorithm improves the frequency match but gives
to identify an intact structure as undamaged. A first step is thus to
unreliable update parameters.
run the identification algorithm for the undamaged structure. Fig.
The damaged case is presented next. As stated at the beginning
7 shows the residual and the generalized cross-validation function
of this section, the same base model is used as for the undamaged
versus the regularization parameter ␥. The calculations have been
case, since this model is already very close to the undamaged
performed for decreasing ␥. For decreasing values of ␥, the re-
structure. Fig. 12 shows the mode shapes before and after model
sidual norm decreases while the generalized cross-validation
updating. The numerical values are provided in Table 4. The cor-
function first decreases and then increases again. The optimal
responding frequencies are shown in Table 5. The mode shapes
value is where the generalized cross-validation functions takes its
and frequencies after the update are reasonably close to the mea-
minimum, indicated by a circle. The corresponding update param-
sured values. Note, however, that a very close match is no guar-
eters are shown in Fig. 8. They are close to zero, indicating an
antee for a good solution. In fact, a better frequency match is
undamaged structure.
The following shows what happens if some algorithmic details
are not correctly implemented. A first point is that regularization
is essential here since the results without 共or with only very small兲 1
regularization are unusable as shown in Fig. 9. These results are
completely unrealistic: several vertical members 共parameters
0
α

1–16兲 would require a stiffness increase of around a factor of 10.

It is also important to apply regularization consistently in a non-
linear update algorithm. Fig. 10 shows the update parameters ob- −1
tained with the inconsistent update formula in Eq. 共8兲. These 0 5 10 15 20 25 30
Parameter
results are clearly inferior to the results obtained by the consistent
Fig. 11. Update parameters for undamaged truss, determined with a
condensed model
15
10 Table 3. Frequencies for Different Updating Procedures of Undamaged
Truss Compared to Measured Values 共Hz兲
5
α

Mode 1 2 3
0
−5
Update with regularization 共Fig. 8兲 23.20 23.56 58.66
0 5 10 15 20 25 30 Update with ␥ = 10−6 共Fig. 9兲 23.07 23.68 58.82
Parameter
Inconsistent algorithm 共Fig. 10兲 23.07 23.68 58.82
Fig. 9. Update parameters for undamaged truss, determined with a Condensed model 共Fig. 11兲 23.09 23.66 58.82
small regularization parameter 共␥ = 10−6兲 Measured 23.07 23.68 58.84

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / MARCH 2010 / 313

J. Struct. Eng., 2010, 136(3): 307-316

Downloaded from [Link] by Guru Nanak Dev Engineering College on 11/07/22. Copyright ASCE. For personal use only; all rights reserved.

Fig. 12. Comparison of measured and analytic modes for damaged truss

achieved by several less reliable procedures and models discussed tact with update parameters close to zero. Parameter 21 refers to
below. two diagonals, one situated above the other without a measuring
Fig. 13 shows the residual and the generalized cross-validation point in between. From an identification point of view, these two
function versus the regularization parameter ␥. Here as well, the diagonals thus act in series 共like a chain with two links兲 and an
calculations proceed from high values to low values of ␥. Con- update parameter of ␣ = −1 implies the loss of either one or both
trary to the undamaged case, the residual and the generalized diagonals.
cross-validation function drop suddenly at ␥ = 3 ⫻ 10−2. A sudden As for the undamaged case, reliable results are only obtained
drop indicates a large improvement in the residual without a large with a consistent algorithm and an appropriate model. A first ob-
increase of the update parameter, which is the exact outcome servation is that regularization is also essential for the damaged
sought. The generalized cross-validation clearly has its minimum case. Without or with only very small regularization, the update
there, indicating the optimal regularization parameter. The update parameters are not usable, as shown in Fig. 15. Other algorithm
parameters shown in Fig. 14 indicate a total loss of stiffness 共␣ components are important as well. Not applying regularization
= −1兲 for update parameter 21 with other members remaining in- consistently yields much less useful results. Fig. 16 shows the

Table 4. Mode Shapes of Updated FEM Model Compared with Measured Values of Damaged Truss
Mode 1 Mode 2 Mode 3
DOF FEM Measurement FEM Measurement FEM Measurement
10x 0.164 0.150 0.053 0.051 0.099 0.077
18x 0.457 0.466 0.147 0.127 0.366 0.366
26x 0.691 0.706 0.218 0.267 0.384 0.399
34x 0.918 0.932 0.287 0.415 0.347 0.343
12x 0.159 0.141 0.049 0.041 ⫺0.178 ⫺0.196
20x 0.394 0.385 0.122 0.100 ⫺0.284 ⫺0.283
28x 0.624 0.623 0.191 0.230 ⫺0.432 ⫺0.436
36x 0.849 0.846 0.259 0.142 ⫺0.538 ⫺0.534
9y ⫺0.053 ⫺0.058 0.177 0.147 ⫺0.139 ⫺0.120
17y ⫺0.153 ⫺0.157 0.401 0.385 ⫺0.325 ⫺0.308
25y ⫺0.234 ⫺0.217 0.656 0.657 ⫺0.410 ⫺0.402
36y ⫺0.312 ⫺0.308 0.907 0.861 ⫺0.448 ⫺0.387
11y ⫺0.048 ⫺0.049 0.180 0.160 0.139 0.147
19y ⫺0.090 ⫺0.079 0.426 0.415 0.324 0.342
27y ⫺0.166 ⫺0.149 0.684 0.678 0.405 0.425
35y ⫺0.243 ⫺0.200 0.936 0.953 0.437 0.443

314 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / MARCH 2010

J. Struct. Eng., 2010, 136(3): 307-316

 Table 5. Frequencies before and after Update Compared to Measured 1
Values of Damaged Truss 共Hz兲
0
Mode 1 2 3

α
Before update 共asymmetric model兲 23.19 23.21 57.99 −1
After update 共Fig. 14兲 22.18 23.38 53.27 −2
After update, ␥ = 10−6 共Fig. 15兲 22.10 23.44 53.43 0 5 10 15 20 25 30
Parameter
Inconsistent algorithm 共Fig. 16兲 22.11 23.42 53.44
Condensed model 共Fig. 17兲 22.10 23.42 53.46 Fig. 16. Update parameters for damaged truss, determined with an
Measured 22.09 23.44 53.47 inconsistent algorithm
Downloaded from [Link] by Guru Nanak Dev Engineering College on 11/07/22. Copyright ASCE. For personal use only; all rights reserved.

update parameters using the inconsistent formulation in Eq. 共8兲.

yields surprising results and is not recommended. The corre-
Although the missing member corresponding to parameter 21 is
sponding update parameters are presented in Fig. 17. The stiffness
clearly indicated, other parameters also indicate substantial
corresponding to the removed diagonal member 共parameter 21兲
changes. The inconsistent formulation shows much less favorable
drops down more than 100% and the stiffness for the adjacent
convergence properties. While the consistent formulation con-
vertical member 共parameter 6兲 is reduced to zero. The reduction
verges in a few iterations to a certain value of the functional, the
of more than 100% is, of course, not physically possible. The
functional continues to decrease in the inconsistent formulation,
problem with the condensed model is that it is able to reproduce
even after a large number of iterations. This behavior reflects
the undamaged structure but not the damaged one. The physically
again the fact that the penalty functional decreases as the update
impossible result could be prevented by additional constraints, but
increment becomes smaller with the number of iterations until the
this would just hide the problem of an inappropriate model. Note
regularization effect is essentially lost. Model condensation has
that all these inappropriate algorithms and models result in a bet-
been demonstrated to work well for the undamaged case. How-
ter match between calculated and measured frequencies as shown
ever, for the damaged case, reducing the model to measured DOF
in Table 5. Therefore, a close frequency match is not an indication
of reliable parameters.
−0.3
10
We also note here that the updating problem is highly nonlin-
ear due to the close frequencies of the first two modes. Line
Residual

search is thus essential here for convergence. A typical case where

the step length has to be reduced is plotted in Fig. 18. The figure
−0.7
10 shows the calculated values of the function ␺共␩兲 given in Eq.
10
−6 −4
10 10
−2
10
0
10
2 共21兲. Note that only the values at ␩ = 0 and at ␩ = 1 are required in
γ the algorithm; intermediate values were only calculated for dem-
−2.3 onstration. Also shown is the tangent at ␩ = 0 determined from the
10
derivative. The sufficient decrease condition 共Wolfe兲 is given by
the straight dotted line with slope c1␺⬘共0兲. Clearly, the function
GCV

value ␺共1兲 is above this line and the step has to be reduced. The
−2.9 approximating parabola is also shown with a minimum near ␩
10
10
−6 −4
10 10
−2
10
0
10
2 = 0.4. In the case shown, the additional restriction that ␩ ⱖ 0.5 is
γ controlling. At ␩ = 0.5 the sufficient decrease condition is satis-
fied.
Fig. 13. Residual and generalized cross-validation of damaged truss

1 1

0 0
α

−1 −1

−2 −2
0 5 10 15 20 25 30 0 5 10 15 20 25 30
Parameter Parameter

Fig. 14. Update parameters for damaged truss Fig. 17. Update parameters for damaged truss, determined with con-
densed model
15
10
Calculated Parabola Wolfe
5
α

0
−5
0 5 10 15 20 25 30
Parameter

Fig. 15. Update parameters for damaged truss, determined with a

small regularization parameter 共␥ = 10−6兲 Fig. 18. Line search

JOURNAL OF STRUCTURAL ENGINEERING © ASCE / MARCH 2010 / 315

J. Struct. Eng., 2010, 136(3): 307-316

 Conclusions Dennis, J. E., and Schnabel, R. B. 共1996兲. Numerical methods for uncon-
strained optimization and nonlinear equations (classics in applied
A sensitivity-based damage identification algorithm previously mathematics), SIAM, Philadelphia.
Doebling, S. W. 共1996兲. “Minimum-rank optimal update of elemental
proposed by the writers has been applied to an aluminum truss
stiffness parameters for structural damage identification.” AIAA J.,
tower. It has been shown that for a successful damage identifica- 34共12兲, 2615–2621.
tion many modeling and algorithmic details have to be considered Dyke, S. J., Bernal, D., and Ventura, J. B. C. 共2003兲. “Experimental phase
and applied correctly. II of the structural health monitoring benchmark problem.” Proc.,
An important part of model-based damage identification is a 16th ASCE Engineering Mechanics Conf., 具[Link]
correct model of the structure. The modeling part requires engi- [Link]/em2003/proceedings典.
neering skills and can take more effort than the actual damage Fox, R. L., and Kapoor, M. P. 共1968兲. “Rates of change of eigenvalues
identification. For the truss investigated here, the important mod- and eigenvectors.” AIAA J., 6共12兲, 2426–2429.
Downloaded from [Link] by Guru Nanak Dev Engineering College on 11/07/22. Copyright ASCE. For personal use only; all rights reserved.

eling parts are the eccentric connections and the proper orienta- Giraldo, D., Caicedo, J. M., and Dyke, S. J. 共2003兲. “Experimental phase
of the SHM benchmark studies: Damage detection using NExT
tion of the beam sections. Condensation of the model can have a
and ERA.” Proc., Engineering Mechanics Conf., 具[Link]
negative influence.
[Link]/em2003/proceedings典.
The second important part of damage identification is a good
Golub, G. H., and Van Loan, C. 共1996兲. Matrix computations, 3rd Ed.,
and mathematically sound updating algorithm. All the necessary Johns Hopkins University Press, Baltimore.
mathematical tools are known in the literature but have to be Hansen, P. C. 共1998兲. “Rank-deficient and discrete Ill-posed problems:
combined consistently. Intuitive application often leads to incon- Numerical aspect of linear inversion.” SIAM monographs on math-
sistent procedures and less useful results. The necessary tools are ematical modeling and computation, SIAM, Philadelphia.
regularization to reduce the influence of measurement noise, gen- Johnson, E. A., Lam, H. F., Katafygiotis, L. S., and Beck, J. L. 共2004兲.
eralized cross-validation to determine the optimal regularization “Phase I IASC-ASCE structural health monitoring benchmark prob-
parameter, line search to control the step length, and a stopping lem using simulated data.” J. Eng. Mech., 130共1兲, 3–15.
criterion. Lam, H. F., Katafygiotis, L. S., and Mickleborough, N. 共2004兲. “Appli-
The proposed algorithm has been successfully applied to dam- cation of a statistical model updating approach on Phase I of the
IASC-ASCE structural health monitoring benchmark study.” J. Eng.
age detection in an aluminum truss tower. The location and extent
Mech., 130共1兲, 34–48.
of damage are accurately identified. Lim, T. W., and Kashangaki, T. A. L. 共1994兲. “Structural damage detec-
tion of space truss structures using best achievable eigenvectors.”
AIAA J., 32共5兲, 1049–1057.
Lus, H., Betti, R., Yu, J., and De Angelis, M. 共2004兲. “Investigation of a
Acknowledgments system identification methodology in the context of the ASCE bench-
mark problem.” J. Eng. Mech., 130共1兲, 71–84.
Nelson, R. B. 共1976兲. “Simplified calculation of eigenvector derivatives.”
This research was supported by the ISIS Canada Network of Cen-
AIAA J., 14共9兲, 1201–1205.
tres of Excellence, the Natural Sciences and Engineering Re- Nocedal, J., and Wright, S. J. 共1999兲. Numerical optimization, Springer,
search Council of Canada, and Hydro-Québec. New York.
O’Sullivan, F., and Wahba, G. 共1985兲. “A cross validated Bayesian re-
trieval algorithm for nonlinear remote sensing experiments.” J. Com-
References put. Phys., 59, 441–455.
Ricles, J. M., and Kosmatka, J. B. 共1992兲. “Damage detection in elastic
structures using vibratory residual forces and weighted sensitivity.”
Amin, M. S., Humar, J. L., and Soucy, Y. 共2002兲. “Experimental verifi-
AIAA J., 30共9兲, 2310–2316.
cation of a vibration based damage detection technique.” Proc., Int. Vogel, C. R. 共1987兲. “An overview of numerical methods for nonlinear
Modal Analysis Conf., IMAC-XX, SPIE, Bellingham, Wash. ill-posed problems.” Inverse and ill-posed problems, H. W. Engl and
Andrew, A. L., Chu, K.-W. E., and Lancaster, P. 共1993兲. “Derivatives of
C. W. Groetsch, eds., Academic, San Diego, 231–245.
eigenvalues and eigenvectors of matrix functions.” SIAM J. Matrix Vogel, C. R. 共2002兲. “Computational methods for inverse problems.”
Anal. Appl., 14共4兲, 903–926.
Frontiers in applied mathematics, SIAM, Philadelphia.
ARTEMIS. 共2005兲. ARTEMIS Extractor software, Structural Vibration Weber, B., Paultre, P., and Proulx, J. 共2007兲. “Structural damage detection
Solutions, Aalborg, Denmark. using nonlinear parameter identification with Tikhonov regulariza-
Barroso, L. R., and Rodriguez, R. 共2004兲. “Damage detection utilizing the tion.” Struct. Control Health Monit., 14共3兲, 406–427.
damage index method to a benchmark structure.” J. Eng. Mech., Weber, B., Paultre, P., and Proulx, J. 共2009兲. “Consistent regularization of
130共2兲, 142–151. nonlinear model updating for damage identification.” Mech. Syst. Sig-
Bernal, D., and Gunes, B. 共2004兲. “Flexibility based approach for damage nal Process., 23共6兲, 1965–1985.
characterization: Benchmark application.” J. Eng. Mech., 130共1兲, 61– Yuen, K.-V., Au, S. K., and Beck, J. 共2004兲. “Two-stage structural health
70. monitoring approach for Phase I benchmark studies.” J. Eng. Mech.,
Caicedo, J. M., Dyke, S., and Johnson, E. A. 共2004兲. “Natural excitation 130共1兲, 16–33.
technique and eigensystem realization algorithm for Phase I of the Zimmerman, D. C., and Kaouk, M. 共1992兲. “Eigenstructure assignment
IASC-ASCE benchmark problem: Simulated data.” J. Eng. Mech., approach for structural damage detection.” AIAA J., 30共7兲, 1848–
130共1兲, 49–60. 1855.
Ching, J., and Beck, J. 共2004兲. “Bayesian analysis of the Phase II IASC– Zimmerman, D. C., and Kaouk, M. 共1994兲. “Structural damage detection
ASCE structural health monitoring experimental benchmark data.” J. using a minimum rank update theory.” J. Vibr. Acoust., 116共2兲, 222–
Eng. Mech., 130, 1233–1244. 231.

316 / JOURNAL OF STRUCTURAL ENGINEERING © ASCE / MARCH 2010

J. Struct. Eng., 2010, 136(3): 307-316