Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

Journal of Sound and Vibration (1996) 198(5), 527545

CRACK IDENTIFICATION IN A CANTILEVER


BEAM FROM MODAL RESPONSE
K. D. H S. S
Department of Civil Engineering, University of Illinois, Urbana, Illinois 61801, U.S.A.
(Received 23 February 1995, and in final form 24 April 1996)
A damage detection and assessment algorithm is developed based on system
identification using a finite element model and the measured modal response of a structure.
The measurements are assumed to be sparse and polluted with noise. A change in an
element constitutive property from a baseline value is taken as indicative of damage. An
adaptive parameter grouping updating scheme is proposed to localize the damage zones
in the structure and a Monte Carlo method is used with a data perturbation scheme to
provide a statistical basis for assessing damage. Damage indices computed from the Monte
Carlo sample of data perturbations are used to assess damage. The threshold values, which
distinguish damage from measurement noise, are established through Monte Carlo
simulation on the baseline structure. The proposed algorithm is applied to the problem of
locating a crack in a cantilever beam. A BernoulliEuler beam model and a plane stress
model are employed to illustrate the use of the method and compare the efficacy of the
two models for crack detection.
7 1996 Academic Press Limited

1. INTRODUCTION

Inspection of structural components for damage is vital to making decisions about their
repair or retirement. The consequences of failing to detect damage vary greatly according
to the application and the importance of the component, but can be considerable from
an economic or safety point of view. Visual inspection is costly and tedious and often does
not yield a quantifiable result. For some components, visual inspection is virtually
impossible. The importance and difficulty of the damage detection problem has
precipitated a great deal of research on quantitative methods of damage detection based
upon physical testing. Among the many possible physical tests, the use of the modal test
has emerged as a particularly promising tool for use in damage detection [12].
In a modal test, one excites the structure either in free vibration or in forced resonance
and extracts the natural frequencies and mode shapes. There are many algorithms available
to extract the modal data [13]. The main idea behind damage detection schemes that use
modal data is that a change in the system due to damage will manifest itself as changes
in the natural frequency spectrum and the associated mode shapes. Early attempts to detect
damage purely from changes in the frequency spectrum met with modest success. More
recent work using mode shapes in addition to the natural frequencies has demonstrated
far more potential.
Damage to a structure can result from any of a number of well-known causes. In the
present paper we are particularly interested in damage associated with a breach in the
material; i.e., a crack. A crack in a solid body manifests itself as a change in the geometry
of the body with respect to the uncracked solid (i.e., the traction-free crack surface of the
damaged solid was an internal, stress-transmitting surface in the uncracked solid), and
527
0022460X/96/500527 + 19 $25.00/0

7 1996 Academic Press Limited

528

. . .

generally occurs without a change in the properties of the material adjacent to the crack.
There is a substantial literature on the topic of cracked rotors that treats the change in
stiffness associated with the crack explicitly, in the context of beam theory, by introducing
a crack element with compliances determined from fracture mechanics concepts [14] that
treat the crack as a geometric defect. Studies concerned with identification of cracks in
beams using this approach have been quite successful [3, 4, 10], but it is not entirely clear
that this approach can be generalized to other types of damage or to more complex bodies;
e.g., structures that cannot be modelled as beams.
The purpose of the present paper is to present and examine an algorithm with a very
general framework that can identify damage from modal data. The algorithm looks for
changes in parametric constitutive properties of a finite element model of the structure (a
model without geometric defects), and takes those changes to be indicative of damage.
Hence, we model the change in stiffness as a smeared crack rather than a discrete
geometric crack. The treatment of damage as a reduction in a constitutive property lends
the algorithm its generality, but also raises the question of its applicability to the
identification of a discrete geometric defect such as a crack. The focus of the present paper
is on the examination of the efficacy of this smeared crack approach to crack identification.
We adopt the cracked cantilever beam as a model problem for this study primarily because
this structure has been the subject of several other investigations related to crack
identifiation.
In addition to the modelling problem, damage detection from measured data has two
practical obstacles. First, measured data are polluted with random measurement errors.
(Unrecognized or unresolved systematic testing errors will not be directly treated here).
The damage detection algorithm must be able to distinguish damage from measurement
error. Second, the measurements will generally be discrete and sparsely distributed over
the spatial domain of the structure. Furthermore, only a few of the natural modes (usually
those associated with the lowest natural frequencies) will be accessible through testing. The
damage detection algorithm must be robust in the face of sparse data. Of the many damage
detection and assessment algorithms that have been proposed, few seem to be well suited
to noisy and sparse data.
In this paper we present an algorithm for detecting damage based on modal data. The
damage detection algorithm has three key features: (a) parameter estimation, (b)
damage localization and (c) damage assessment. We infer damage from changes in the
constitutive properties of elements in a finite element model of the structure. The
values of the parameters are estimated from measured modal data using the modal
displacement error method described by Hjelmstad et al. [15]. The discretization of a
continuous structural system by finite elements introduces a systematic modelling error
that can spoil the damage detection results. To avoid these systematic errors one must use
a suitably refined finite element model. However, mesh refinement generally implies far
more elements, and hence element properties, than the available data can reliably
estimate. To resolve this problem we introduce an adaptive parameter grouping
algorithm for the damage localization phase of the algorithm. The parameter values
generated by the parameter estimation scheme will differ from the baseline values because
of noise in the measurements, even when no damage has occurred. Consequently, the
algorithm must be able to distinguish measurement noise from damage in the damage
assessment phase. We propose a data perturbation scheme, based on the Monte Carlo
method, to generate statistical indices for damage assessment. We use Monte Carlo
simulation on the baseline structure to determine the limits for the damage indices, above
which they indicate damage and below which changes are probably due to measurement
noise.

529

We apply the damage detection algorithm to the problem of detecting a crack in a


cantilever beam, using data obtained from an experiment by Rizos et al. [10]. We compare
two different finite element models of the beam, a BernoulliEuler beam model and a plane
stress model. Through the example, we demonstrate how the algorithm is applied and how
to interpret the results. We also show the importance of appropriate model selection in
quantitative damage detection schemes.
2. SYSTEM MODELLING AND PARAMETER ESTIMATION

Let us assume that we can characterize our structure with a finite element model with
Nd degrees of freedom. For the damage detection problem it is reasonable, even if not
necessary to the development of the algorithm, to assume that the damage processes do
not change the mass M of the structure. As such, we view it as constant and known from
the baseline properties of the structure. The linear stiffness K of the structure, on the other
hand, can change as a result of damage. Although we seek to locate and assess damage
that could be considered to be a change in the topology or geometry of the structure (i.e.,
a crack) we shall model the damage as a change in constitutive properties of a structure
with fixed topology and geometry. Accordingly, the stiffness matrix K(x) is parameterized
by Np constitutive parameters x.
The selection of an appropriate parameter estimation algorithm is inextricably linked
to the parameterization of the model. In the literature one finds many modal parameter
estimation algorithms based upon a nodal representation of the properties of the structure
(e.g., the elements of the stiffness matrix K are taken as parameters). While this approach
is useful for improving a finite element model to better represent measured data, it is not
particularly well suited to damage detection because damage is most naturally associated
with element behavior. An alternative approach to structural modelling, the one we adopt
here, is to take element constitutive properties as model parameters (see, for example,
Hjelmstad et al. [16]). This approach is attractive because it preserves structure topology,
and hence preserves the essential features of the flow of forces through the structure. The
element-based model of the structure is also useful for the damage detection problem
because the estimated parameters are often directly indicative of damage.
Let us assume, for the sake of discussion, that a group of elements Vn can be
characterized by a single constitutive parameter xn , a multiplier of a nominal constitutive
property value, for example. Each element in the model belongs to one of the groups
{V1 , V2 , . . . , VNp }. Thus, the parameter associated with element m $ Vn is xn . With these
conventions, the linear stiffness matrix of the model has the explicit form
Np

K(x) = K0 + s s fm (xn )Gm ,

(1)

n = 1 m$Vn

where K0 is the fixed part of the stiffness matrix. In the parameterized part of the stiffness
matrix, fm (xn ) is the constitutive function and Gm is the Nd Nd globalized kernel matrix
of element m. When the stiffness matrix is linear with respect to the parameters,
fm (xn ) = xn . The simplest possible parameterization is to take xn as a scalar multiplier of
the nominal element stiffness matrix, then Gm = km [5]. Some structural models have
element stiffness matrices that depend upon two essentially different phenomena. For
example, a Timoshenko beam possesses both flexural and shear modes of behavior. It may
be important to parameterize those different modes independently. The structrure of
equation (1) can be easily generalized to accomodate elements with multiple parameters
by adding an additional summation over parameter types [17]. The mechanical theory may

530

. . .

suggest that an element has more than one constitutive parameter, but one of those
parameters might be far less sensitive to damage than the other. For example, in a plane
stress problem Youngs modulus is more sensitive to damage than is the Poisson ratio. One
can view the insensitive constitutive parameter as fixed at some nominal value, thereby
eliminating it from the damage detection scheme. The part of the stiffness associated with
the fixed parameters are assembled into K0 .
The process of assembling the stiffness matrix in equation (1) requires only a slight
modification of the usual assembly process for a finite element model. Consequently, many
of the computations associated with parameter estimation can be organized in a manner
similar to direct finite element computations. We will show in a subsequent section that
damage localization can be viewed as a search for an optimal parameterization over a fixed
finite element mesh. The model implicit in equation (1) is well suited to such a search.
The undamped, free vibration of a structural system gives rise to the matrix eigenvalue
problem
K(x)f = lMf,
(1)
where K(x) is the parameterized stiffness matrix, M is the mass matrix, l is the eigenvalue
(the square of the natural frequency) and f is the eigenvector. The natural frequencies and
natural modes can be measured in a modal test [13]. Typically, the modal test would extract
information about Nm modes. For each mode, we assume that the frequency can be
measured accurately and that the mode shape would be sampled at N
d discrete locations.
The finite element model, with specific values of the parameters x, gives rise to Nd modes
sampled at Nd locations in space. For simplicity, we assume that the measurement locations
correspond with degrees of freedom of the finite element model, so that the N
d
measurement locations are a subset of the Nd locations on the model. We call the
measurements sparse if the measurement locations are few, N
d Nd , if the number of
modes sampled is few, Nm Nd , or if both are few Nm N
d Nd2 . Let $
be the set of degrees
of freedom of the finite element model associated with measurement locations on the test
piece, and let $ be the set of remaining degrees of freedom. We can partition the
eigenvectors and system matrices along these same lines. For example, we shall call f
the
part of the eigenvector f associated with measurements values, and f
 the part of f
associated with no measurements. We shall assume that the measurement locations are the
same for each of the Nm modes.
The damage detection and assessment algorithm proposed herein requires that the
parameters of the model be estimated from the measured modal properties
{(li , f
 i ), i = 1, . . . , Nm } for a fixed grouping V 0 {V1 , V2 , . . . , VNp } at each stage of the
localization algorithm. We adopt the output error estimator of Hjelmstad et al. [15]. Since
the performance of the damage detection algorithm will be a function of the performance
of the parameter estimator on which it is based, let us briefly summarize the estimator that
will be used in the study of the cantilever beam.
Let us partition the mass matrices along the lines drawn above. In particular, let M
be
the part of the mass matrix associated with the degrees of freedom $
, and let M
 the portion
associated with the degrees of freedom $. Define the matrix Bi (x) as
Bi (x) 0 K(x) li [O=M
 ],

(2)

where O is an Nd N
d zero matrix. The output error for the modal response can be defined
from equation (2) as
ei (x) 0 f
i li QB+

f
i ,
(3)
i (x)M
where B+ indicates the generalized inverse of the matrix B and the N
d Nd Boolean matrix
Q extracts the components of the eigenvector associated with degrees of freedom $
from

531

the complete eigenvectors as f


i = Qfi . The parameters are given as the solution to the
following constrained least-squares optimization problem:
Nm

minimize
J(x) = 12 s ai >ei (x)>2,
N
x$R p

subject to x E x E x ,

(4)

i=1

where x and x are the lower and upper bound vectors, and ai is the weight factor for the
ith mode.
There is a lower limit on the amount of information required to make an estimate with
the output error estimator. Let IC denote the ratio of data to unknowns as
IC = Nm N
d /Np .

(5)

If IC Q 1, the parameter estimates will be completely unreliable. If IC e 1, then parameter


estimation is possible. Clearly, more information will increase the reliability of the
estimates. The amount of information redundancy required depends upon the particular
application. With sparse data in a damage detection scheme one often pushes the limits
of identifiability.

3. ADAPTIVE PARAMETER GROUPING ALGORITHM

To localize damage in a systematic manner with sparse data, we propose an adaptive


parameter grouping algorithm. The main idea of the scheme is to isolate damaged parts
in the finite element model by sequentially subdividing parameter groups, starting from
a baseline grouping. Each subdivision stage corresponds to a specific parameter grouping.
Let us designate the parameter grouping at the ith stage as Vi = {V1 , V2 , . . . , VNpi }, where
Npi is the number of parameter groups at subdivision stage i. At each stage the parameters
associated with the grouping are determined by solving the constrained least-squares
problem (4). The grouping Vi is determined from Vi 1 by subdividing one of the groups
in accord with a specified criterion. The baseline grouping is V0 and contains at least one
group. Because several damaged regions with different levels of severity may coexist in a
structural system, the subdivision must be continued until all of the damaged elements are
isolated.
At each stage of the algorithm one must determine which group is the best candidate
for subdivision. Natke [6] and Natke and Cempel [7] recommend comparing the difference
between the estimated parameter and the baseline value. The group with the largest
absolute difference would be selected for subdivision. Simulation studies have
demonstrated that, when noise is present in the measured data, the largest deviation from
baseline is often associated with noise rather than damage. Thus, the deviation from
baseline is not a suitable measure for subdivision when noise is not negligible.
One can also observe from simulations that the value of the error function J(x) always
decreases with subdivision, and decreases precipitously when undamaged members are
separated from damaged ones. Therefore, the best group to subdivide is the one that results
in the greatest decrease in, and hence the smallest value of, J(x). Clearly, one must
investigate the effect of subdividing all candidate groups to find the one that gives the
greatest decrease in error. However, multiple subdivision cases can be computed
concurrently, and the subdivision can be organized as a depth-first search so that there
are only a few candidate groups at each stage. This approach to subdivision is
computationally intensive, but quite reliable for localizing damage. Generally, the

. . .

532

candidate subset is divided in half, but one could just as easily subdivide into a larger
number of subgroups if the data are sufficient. The reduction in estimation error is
considerable when a damaged member is finally isolated.
The adaptive parameter grouping algorithm is best organized as a depth-first search
wherein subdivision continues along a certain line until the penultimate group has only
two elements, until subdivision gives a reduction in J(x) smaller that some predefined
tolerance, or until the change in estimated parameters is smaller than a predefined
tolerance. The parameters associated with the converged group are fixed at their currently
estimated values and removed from the set of unknown parameters. The algorithm then
backs up one level and investigates the branch not taken in the last search. Upon
exhausting the alternate branch, the algorithm backs up one more level and continues until
it returns to the top level. The algorithm terminates when all branches have been
investigated. This algorithm makes optimal use of the sparse data, but it is possible to run
afoul of the identifiability criterion. It at any stage IC Q 1 then damage detection is not
possible.
The subdivision of parameter groups implies a variation in the number of groups Np ,
and hence a change in the complexity of the parameterized model. As the number of
parameters increases the value of the error function J(x) decreases. The theory of
parametric modelling suggests that the most complex model may not be the best when
there is noise in the data. Beyond a certain level of complexity the estimator will start to
represent, rather than filter, the noise in the model. As one approaches an interpolatory
model (IC = 1) the sensitivity of the parameter estimates to noise increases rapidly.
Therefore, it is important to deter the algorithm from seeking models that are too complex.
There are many approaches to limiting model complexity in the literature on system
identification. One common approach is to penalize the error objective J(x) with terms that
grow with the number of parameters. Accordingly, let us modify the original parameter
estimation problem to minimize the squared model error (SME) rather than the objective
function J(x) itself as follows:
minimize
SME(x, Np ) = J(x)r1 (Np ) + r2 (Np ),
N
Np ,x $ R p

subject to

x E x E x ,

(6)

where r1 (Np ) and r2 (Np ) are monotonically increasing functions of the number of
parameters. Contained as a special case of SME is the predicted squared error of Barron
[18] with r1 (Np ) = 1 and r2 (Np ) = 2s 2Np /N*, where s 2 is an estimate of the variance of the
parameter estimates and N* is the maximum number of parameters. Also contained as a
special case of SME is the finial prediction error of Akaike [19] with r2 = 1 and
r1 = (N* + Np )/(N* Np ). For the present study, we select the hybrid penalty functions

r1 (Np ) = 1,

r2 (Np ) = 12 s 2N0

Np
,
Ns Np

(7)

where s 2 is an estimate of the variance in the parameter estimates, N0 = Nm N


p is the
maximum number of parameters for identifiability, and Ns is the maximum number of
parameters possible in the finite element model (e.g., if there is one parameter per element,
then Ns is equal to the number of elements). Other choices are clearly possible.

533

h
b

Figure 1. The cracked cantilever beam specimen (L = 300 mm, l = 140 mm, b = h = 20 mm, a = 10 mm).

4. DATA PERTURBATION SCHEME

After element properties have been estimated by the adaptive parameter grouping
algorithm, we must determine which elements are damaged and to what extent.
Philosophically, this assessment is quite simple. If the estimated parameter is different from
the baseline value, then the element is damaged. The presence of measurement noise
complicates the issue considerably. Even if there was no damage at all, the noise in the
measurements and any errors in modelling would cause the estimated parameters to be
different from the baseline values. Presumably, if the errors are small, the estimated
parameters would be close to the baseline values. It therefore seems reasonable to allow
absolute differences between estimated and baseline values that are less than some
predefined tolerance to be interpreted as not damaged. The essential problem is to find
an index of comparison and a tolerance that reliably distinguishes measurement nosie from
damage. We propose a data perturbation scheme for this purpose.
Let us assume that we have at our disposal a single set of measurements. The natural
frequencies usually can be measured with negligible error compared with the mode shapes.
Let us further assume that we know the variance s2 of the measurement errors. (The
variance of the errors can be determined from a repetitive test.) If the measurement noise
is random with zero mean, the measurements are not made worse, in a statistical sense,
by adding noise of the same character to them. Thus, we can create an artificial set of
perturbed data by adding random noise to the measured data. To wit, let the kth
perturbation of the jth component of the ith eigenvector be computed from the jth
component of the ith measured eigenvector as
f ijk = f
ij (1 + zijk ),

(8)

T 1
Normalized measured modal displacements of the cracked cantilever
beam
x/L

f
1

f
2

f
3

01
02
03
04
05
06
07
09
10

0027
0063
0109
0156
0236
0354
0487
0778
0940

0069
0149
0276
0444
0516
0400
0200
0455
0899

0033
0103
0221
0288
0011
0387
0498
0203
0435

x is the distance from the clamped end.

534

. . .

where zijk is a random variate with zero mean and variance s 2. By carrying out the adaptive
parameter grouping algorithm M times, we generate a Monte Carlo sample of parameter
estimates for each of the elements. We can take the mean value of the sample as the
estimated parameter value, for each element. The standard deviation of the sample, for
each element, gives a measure of the sensitivity of that element to the measurement noise.
Let x*
m
m represent the baseline value of the property associated with element m. Let x
be the mean and sm the standard deviation of the Monte Carlo sample of estimates for
element m. Two damage indices can be defined to assist the damage detection and
assessment process as follows:
bias cxm =

=xm x*
m=
,
x*
m

bias sdm =

=xm x*
m=
.
sm

(9, 10)

The first index, bias cxm , indicates how close the averaged element parameter xm is to the
baseline value x*
m . If the measurement noise is small and if the parameter value of element
m is reliably estimated, then bias cxm will be a good indicator of damage. However, when
the noise level is high or when the parameter value of element m is not well represented
in the data, it may be difficult to reliably infer damage from bias cxm . An insensitive
element will generally exhibit a high variance in the Monte Carlo sample. Thus, bias sdm
can be used to reduce the risk of identifying an actually undamaged element as a damaged
one. It is possible for bias sdm to become a large value when sm is small compared with
=xm x*
m = due to an excellent estimation. Therefore, the two damage indices must be used
together to detect damage in a structural system.
Let cxlmt and sdlmt be the threshold values of bias cx and bias sd, respectively. We
will consider an element damaged if
bias cxm q cxlmt

AND

bias sdm q sdlmt.

(11)

The threshold values of the two damage indices must be determined to optimally
distinguish measurement noise from damage. These values can be estimated from a Monte
Carlo simulation of the baseline structure. At the given noise level, cxlmt and sdlmt are
those values that predict no damage in all elements, say, 97% of the time. Damage can
often change the behavior of the structure a great deal. One may wish to repeat the Monte
Carlo simulation on the identified damaged structure, as an a posteriori estimate, to assess
the efficacy of the values chosen from the baseline study. In the following case study we
will demonstrate the overall procedure of damage detection.
5. CRACK IDENTIFICATION CASE STUDY

In this section our damage assessment algorithm is used to locate and assess the severity
of the crack in the cantilever beam shown in Figure 1. The modal responses were obtained
experimentally by Rizos et al. [10]. They measured natural frequencies and modal
displacements of a cantilever structure with a transverse surface crack extending uniformly
along the width. The 300 mm cantilever with square cross-section 20 mm 20 mm was
clamped to a vibrating table. Resonant harmonic excitation was applied and each of three
modes was investigated independently. Even though they tested various cases with
different crack locations and different crack depths, they documented the measurements
appropriate to our damage detection scheme for the case with the crack depth a = 10 mm
and the crack located at l = 140 mm from the clamped end, as shown in Figure 1. The
crack was initiated with a thin saw cut and propagated to the desired depth by fatigue
loading. The crack depth was measured directly and verified with an ultrasonic detector
for uniformity and actual depth of the crack.

535

The vibration amplitude was measured at nine locations, 30 mm, 60 mm, 90 mm,
120 mm, 150 mm, 180 mm, 210 mm, 270 mm and 300 mm from the clamped end. The
amplitude of motion at 240 mm from the clamped end was not measured. The measured
modal displacements were normalized with respect to the maximum vibration amplitude.
Mode shapes were measured by using two calibrated accelerometers mounted on the
cantilever. One accelerometer was kept at the clamped end to give the reference input and
the other accelerometer was moved along the cantilever to measure the mode amplitude.
The mass of the accelerometer was negligibly small compared with the mass of the
cantilever. The measured natural frequencies of the cracked cantilever were 171 Hz,
987 Hz, and 3034 Hz for the first, second and third modes, respectively. The measured
modal displacements at nine locations for each mode were extracted from the mode shape
figures in the paper and are summarized in Table 1. The digitized measured modal
displacements therefore contain errors associated with extracting the measured modal
displacement from the published paper in addition to the measurement error.
In what follows, we shall consider three different structural models of the cantilever
beam. For validation of the experiments we will consider a plane stress model with the
crack present in the geometry of the body, as shown in Figure 2. The finite element model
has three levels of mesh refinement: 60 elements (the coarse mesh), 240 elements and 960
elements. Each refinement is accomplished by quartering the mesh of the previous model.
For damage detection we shall consider two models: (1) the structure modelled by
BernoulliEuler beam elements as shown in Figure 3 and (2) the structure modelled by
two-dimensional plane stress elements as shown in Figure 2, except without the geometric
modelling of the crack. The flexural stiffness EI of the beam is the only constitutive
parameter associated with the BernoulliEuler beam model, with E typically taken equal
to Youngs modulus and I taken equal to the second moment of the cross-sectional area.
The isotropic, linearly elastic plane stress model can be described in terms of Youngs
modulus E and the Poisson ratio n. The two selected structural models of the cantilever
can vary according to the number of finite elements used in the discretization. For plane
stress models, square four-node elements are used.
The structural model needs to be accurate to prevent bias from systematic numerical
modelling errors. The accuracy of the model depends upon the state of damage, which is
not known in advance. To start the damage detection algorithm, the best model is the one
that gives a uniformly small discretization error for the baseline structure (i.e., a uniform
mesh for the present case study). To resolve local damage, the mesh may need subsequent
refinement. Ideally, one would adaptively refine the mesh to concentrate the elements in
the neighbourhood of the damage zones. For the present study we simplify the mesh
refinement question by starting with a uniform mesh fine enough to resolve the damage.
Clearly, for more complicated structures this strategy would not be practical due to the
extensive computations required. We avoid the adaptive mesh refinement question here,
so that we can focus on the adaptive parameter grouping algorithm that forms the basis
of our damage detection scheme. More efficient implementations of the damage detection
algorithm are certainly of great practical interest.
5.1.
The geometry of the cantilever structure and the modulus of elasticity of the material
are given by Rizos et al. [10]. The paper does not name the material used, but gives the
modulus of elasticity as E = 206 105 MPa, from which one can reasonably assume that
the material is steel, and would thus have a density of 7800 kg/m3 and a Poisson ratio of
n = 03. To validate these assumptions the cracked beam was analyzed using a sequence
of successively refined plane stress models. The plane stress model can accurately model

. . .

536
6

11

16

21

26

31

36

46

51

Measurement locations
Figure 2. The structural model with 50 BernoulliEuler beam elements.

the geometry of the cracked beam, unlike the BernoulliEuler beam model. The three
lowest computed natural frequencies from 60-, 240- and 960-element meshes are
summarized in Table 2. From the table, we can observe that the computed frequencies
differ from the measured data by less than 2% when 960 elements are used. The close
correspondence between measurements and model suggest that the properties and
geometry used in the model closely reflect the features of the experiment.
The results in Table 2 also show that the less refined models predict natural frequencies
higher than the measured values for all the three modes. Despite the fact that these models
are not completely accurate, we shall demonstrate that they are still useful for the problem
of damage detection. The error in frequency will be distributed across the structure as a
reduction in element stiffness by the parameter estimation algorithm, rather than as local
changes in element properties, and will not necessarily mask the real damage. The question
of how accurate the finite element model needs to be to minimize the affects of these
systematic modelling errors is problem dependent, but can always be addressed by a mesh
refinement study.
The amplitude of noise in the measured modal displacements is needed in the damage
assessment algorithm. Typically, this information would be available from a baseline
evaluation of the undamaged structures. No baseline measurements are available for the
current problem. However, since the exact crack location and depth are given from the
experiment by Rizos et al. [10], the missing information may be deduced from the
analytical study of the cracked beam. The error in the modal displacement is the
combination of the actual sensor error and the error that we committed in digitizing the
data from the original paper. To quantify the error between the analytical mode shape and
the measured mode shape, we employ the well-known modal assurance index.
bi 0

(c
i f
i )2
,
>c
i >2>f
i >2

(12)

where c
i and f
i are the ith analytical and measured mode shapes, respectively, with only
the measurement locations included. Clearly, when the analytical mode shape is close to
the measured shape, the value of bi is close to unity. The discrepancy between the computed
and measured modes for the three different plane stress models is summarized in Table 3.
From the table, one can observe that the error increases with the mode number, but the
errors do not change with refinement of the finite element mesh. There is a considerable
difference between the third mode shape predicted by the finite element model and the one
measured.
The amplitude of measurement noise can be deduced from the modal assurance index
if one assumes that errors are random and proportional to the measurement value, in
L = 300 mm
140 mm

Figure 3. The structural model with plane elements (coarse mesh).

537

T 2
Natural frequencies of the cracked beam, plane stress model
Number of
elements

Frequency (Hz)

ZxxxxxxxCxxxxxxxV
v1

v2

v3

60
240
960

188
178
175

1079
1012
983

3263
3106
3062

Measured value

171

987

3034

accord with equation (8). If the random variable zijk has a uniform distribution between
the values of 2oi , then one can show that
oi 1 z6(1 zbi ).

(13)

Equation (13) suggests that the noise amplitudes for the three modes are 10%, 39% and
94%, respectively, if averaged over all components in the mode shape vector. The
estimated noise in the higher modes seems unreasonably high. They may be dominated
by the components with smallest amplitude. When the error is investigated for the modal
displacement component the amplitude of which is the largest in each mode, we obtain
values of 5%, 17% and 54%, respectively. The values of bi may also be influenced by
systematic modelling errors, which equation (13) does not take into account. For the
subsequent crack identification studies we shall generally assume the value of 10%
proportional error for the average amplitude of measurement noise. However, for some
case studies, 20% amplitude of proportional noise is applied to investigate the effects of
noise amplitude on the identification results. We repeat that one should evaluate oi from
repetitive testing, a luxury not available in the present case.
The damage assessment algorithm assumes that the geometry of the structure does not
change and that damage manifests itself as a reduction in certain constitutive properties.
Before we attempt to identify the crack in the beam from the measurements, let us examine
the validity of using a model parameterized by the constitutive properties to characterize
and assess the damage associated with a crack. A model with 200 BernoulliEuler beam
elements and a model with 960 plane stress elements are investigated to compare their
modal behaviors with the measured data. For the BernoulliEuler beam, we examine the
response of the cantilever as the single element closest to the actual crack location varies
in flexural stiffness EI. For the plane stress element model, we examine the response as
Youngs modulus E varies for the band of elements at the crack location, all the way
through the thickness. The constitutive properties of the other elements are at the nominal
baseline values for both models. The three lowest computed natural frequencies are shown
in Figure 4, in which frequency is plotted against the ratio of the reduced element property
to the nominal element property, EI/EI0 for the beam model and E/E0 for the plane stress
model. The figures suggest that both models are generally capable of representing the
spectrum of the cracked beam. Even though the computed natural frequencies of the third
mode from the BernoulliEuler beam model cannot match the measured frequency when
a single beam element is simulated as damaged, the error in the third natural frequency
can be diminished by allowing stiffness of some other elements also to change.

. . .

538

5.2.
Let us turn our attention to the problem of identifying the location of the crack using
a BernoulliEuler beam model. Since we assume no knowledge of the crack location in
advance, it is most logical to discretize the cantilever beam with equal sized elements. The
finite element model does not change during the damage localization process, so it is
important to begin with a suitably fine mesh at the outset. For the current study, we use
a model with 50 BernoulliEuler beam elements. Although 50 elements is far more than
required to reduce the finite element discretization error, this refinement is necessary to
accurately resolve the damage. The data perturbation iterations are executed 50 times,
which is adequate to establish statistical significance of the estimates.
The BernoulliEuler beam element has a single constitutive parameter, the flexural
stiffness EI. Damage is inferred as a reduction in the flexural stiffness of an element or
group of elements. Since the number of measured modal displacement for each node N
d
is equal to nine and the number of measured modes Nm is equal to three, the limit to the
number of unknown parameters in the model N0 is 27, according to equation (5). Since
the member properties of the uncracked baseline cantilever structure can be assumed to
be uniform over the whole length of the structure, the number of parameter groups starts
with Np = 1. Subdivision can continue up to a maximum of 27 groups.
Before evaluating the cracked cantilever structure, it is necessary to determine the upper
limit values for the damage indices from a simulation study of the baseline structure. The
determined upper limit values cxlmt and sdlmt for damage indices bias cx and bias sd,
respectively, will be used to discern damage from noise in each element. In Figure 5(a) are
shown the mean and standard deviation of the parameter estimates of the baseline
structure (no crack) over 50 Monte Carlo trials with 10% imposed noise. One can see that
the elements at the tip of the cantilever beam are the most influenced by noise. The
probability of detecting each member as damaged as a function of the sdlmt value is shown
in Figure 5(b). There is a curve plotted for each element, with the ordinate giving the
percentage of times the member was identified as damaged in the Monte Carlo sample for
the given sdlmt value. One can see that for values of sdlmt less than 1, many of the members
are detected as damaged, some of them quite often, even though there is no damage.
Therefore, an appropriate value of sdlmt for damage detection is 11. For this value of
sdlmt no member in the baseline structure is detected as damaged in more than 3% of the
Monte Carlo trials. One can do the same exercise to determine cxlmt, but in this case the
structure is not sensitive to cxlmt. Shin and Hjelmstad [17] have suggested setting cxlmt
equal to the noise amplitude, 10% in the present case.
Now let us try to locate the crack in the damaged structure. In Figure 6(a) are shown
the computed mean and standard deviation values of flexural stiffnesses of the damaged
structure, as determined from the 50 data perturbation iterations. The damage indices are

T 3
Computed modal displacement errors, plane stress model
Number of
elements
60
240
960

Modal assurance index

ZxxxxxxxCxxxxxxxV
b1

b2

b3

0996
0996
0997

0956
0953
0951

0729
0729
0730


200
(a)
Frequency (Hz)

539

1200

3300
(b)

190

1100

180

1000

170

900

160

800

(c)
3200
3100

150
0.0

0.1

0.2

700
0.0

3000

0.1
0.2
Stiffness ratio

2900
0.0

0.2

0.1

Figure 4. Variation of the natural frequencies with respect to the stiffness ratio for (a) the first mode; (b) the
second mode and (c) the third mode. , Measured value; W, BernoulliEuler beam model; R, plane stress
model.

shown in Figure 6(b). The bias sd values of elements 125 are rather large, but the bias cx
values are less than cxlmt. Hence, these elements are not detected as damaged. Element
26 is clearly identified as damaged. It is quite near the actual crack. With cxlmt = 10%
and sdlmt = 11, elements 31 and 46 are also detected as damaged. This anomalous result
is worth noting, and is indicative of the problem of identifying a crack with a
BernoulliEuler beam model.
The fact that we very nearly detected all of members 125 as damaged deserves further
examination. In Figure 7 are shown the computed natural frequencies of the damaged
cantilever beam from three different estimated models. In case (i) we set EI of element 26
equal to its estimated values but all others at their nominal values. In case (ii) we set EI
of elements 26, 31 and 46 equal to their estimated values with all others at their nominal
values. In case (iii) we set EI of all elements equal to their estimated values. Case (i) matches
the first frequency very well, but has difficulty with the second and third. Thus, changing
element 26 gets us fairly close to the measured response. The changes in parameters in the
other elements are needed to match the higher modes more closely. This problem is an
artifact of the inability of the BernoulliEuler beam to capture the spectrum of the cracked
beam noted earlier. The estimated parameters will attempt to make up for shortcomings
in the model, often to the detriment of damage detection. The mode shapes for case (i)

3000

Flexural stiffness

Damage probability (%)

(a)

2500
2000
1500
1000
500
0

40
(b)
30
20
10

0
1

15 22 29 36 43
Element ID number

1
sdlmt

50

Figure 5. (a) The mean and standard deviation of the estimated flexural stiffnesses from the Monte Carlo
sample: , baseline; W, mean; R, standard deviation. (b) The variation of the damage probability with respect
to sdlmt for the baseline structure for cxlmt = 10%.

. . .

540

100
biassd

80
3000

40
20
0

2000

1 8 15 22 29 36 43 50

1500
biassd

Flexural stiffness

2500

60

1000

0
1

15 22 29 36 43
Element ID number

50

biassd

500

60
50
40
30
20
10
0
15
10
5
0
1 8 15 22 29 36 43 50
Element ID number

(a)

(b)

Figure 6. Damage localization and assessment with the 50 BernoulliEuler beam element model. (a) Parameter
estimates and standard deviations from perturbation iterations: , base line; W, mean; R, standard deviation.
(b) Damage indices for each element.

are compared with the measured data in Figure 8. The mode shapes from the other two
cases also yielded roughly the same results.
One of the primary interests of the current study is to investigate how well a
BernoulliEuler beam model can locate the crack and assess damage by a reduced flexural
stiffness. In Figure 9, the results from three different BernoulliEuler beam models with
20, 30 and 50 beam elements are compared. For the purpose of comparison, the location
of a crack is defined as the distance from the fixed end to the middle of the element which
was detected as the most highly damaged. The estimated flexural stiffness of the detected
element is also compared in the same figure. We can observe that the identified crack
location approaches the actual cracked section as a more refined beam model is applied.
As more elements are used, flexural stiffness of the element identified as damaged reduces
rapidly because the length of the damaged element decreases.

300

1150

Frequency (Hz)

3300
(b)

(a)

(c)

200

1050

3200

100

950

3100

(i)

(ii) (iii)

850

(i)

(ii) (iii)

3000

(i)

(ii) (iii)

Figure 7. Computed natural frequencies with reduced flexural stiffness in different elements; cases (i), (ii) and
(iii) for (a) the first mode, (b) the second mode and (c) the third mode. The solid horizontal line represents the
measured value.

541

1
(a)

Amplitude

0
1
(b)
0
1
1
(c)
0
1
0.0

0.2

0.4

0.6

0.8

1.0

x/L
Figure 8. The computed lowest three mode shapes of the BernoulliEuler beam model with 50 elements when
the flexural stiffness of element 26 is reduced with (a) the first mode, (b) the second mode and (c) the third mode.
, Identified; W, measured.

We have repeated the analysis of the cracked beam using Timoshenko beam elements.
The results, presented in detail in reference [17], are omitted here for the sake of brevity.
The Timoshenko beam element has twice as many constitutive parameters as the
BernoulliEuler beam. Consequently, the advantages gained by accounting for shear
deformation in the vibration frequencies and mode shapes are counteracted by the relative
loss of information content in the data. That observation notwithstanding, the results of
the Timoshenko beam model are qualitatively similar to those of the BernoulliEuler beam
model, leading us to conclude that the performance of these two models is dominated by
the one-dimensional nature of the approximation rather than the importance of shear.

300
250
200
150
100
50
0

Flexural stiffness

Crack location

5.3.
For a plane stress model, there are two constitutive parameters in each element, Youngs
modulus E and the Poisson ratio n. By modelling the cantilever with plane stress elements,
one can capture many physical features of the cracked beam that the BernoulliEuler
model precluded. For example, shear deformation becomes important in studying the
modes of vibration of higher frequencies [20]. Also, the stress concentration effects of a
crack can be represented by a plane stress model.

(a)

10

20

30

40

50

60

1200
(b)
900
600
300
0

10

20

30

40

50

60

Number of beam elements


Figure 9. The estimated crack location (a) and the flexural stifiness of the cracked element (b). , Actual
crack location; W, estimated crack location.

. . .

Eb

542
5000
4000
3000
2000
1000
0
5000
4000
3000
2000
1000

(b)

10

Damage probability (%)

(a)

20
30
40
Element ID number

10

50

60

(c)

0.5
sdlmt

1.0

Figure 10. The mean and standard deviation of E from the Monte Carlo sample for (a) top elements and (b)
bottom elements: , baseline; W, mean; R, standard deviation. (c) Variation of the damage probability with
respect to sdlmt for the baseline structure for cxlmt = 10%.

When there are more than two parameters per element, it may be that they are not
equally represented in the data. In the plane stress problem, the Poisson ratio is typically
difficult to estimate from the given (transverse displacement) measurements because it is
not sensitive to them. In a simulation study of the cantilever beam, we observed that
estimated values of the Poisson ratio hit the upper or lower bounds frequently during the
identification processes, while Youngs modulus was more robustly estimated. To generate
an alternate measurement set, we used a 960-element plane stress model. Those data were
then used with a 60-element plane stress model for damage detection. Estimating two
parameters in each element, we found that the estimates of the Poisson ratio were not
reliable for locating damage in the structure, while Youngs modulus estimates were. We
therefore consider the Poisson ratio for all elements to be fixed at the nominal value for
all of the subsequent studies. A model with 60 plane stress elements is selected to carry
out damage detection on the cantilever structure. The initial number of parameter groups
starts from one because the stiffness properties of the uncracked baseline cantilever
structure can be assumed to be uniform over the whole structure.
The upper limit values for the damage indices are determined by the damage probability
curves obtained from a simulation study of the baseline structue as shown in Figure 10.
Mean and standard deviations from the Monte Carlo sample for the baseline structure are
also shown. From the figure, one can observe that most of the elements are sensitive
enough with respect to the test conditions with their small standard deviations, even
though the elements near the free end show a little higher standard deviation values. From
the damage probability figure, we can determine the upper limit values as sdlmt = 048 and
cxlmt = 10% (equal to the noise amplitude).
The computed mean and standard deviation values and the damage indices for each
element compiled in Figure 11, from 50 data perturbation iterations. From the figure, the

(a)

(b)

biascx (%)
biascx (%)

(c)

(e)

20
30
40
Element ID number
800
(d)
600
400
200
0
800
(f)
600
400
200

50

60

biassd

10

0
80
60
40
20
0
80
60
40
20

543

biassd

Eb


5000
4000
3000
2000
1000
0
5000
4000
3000
2000
1000

10 20 30 40 50 60
Element ID number

10 20 30 40 50 60
Element ID number

Figure 11. Damage detection and assessment with the 60 plane stress element model for (a) top elements and
(b) bottom elements. Damage measures for (c, d) top elements and (e, f) bottom elements. , baseline; W,
mean; R, standard deviation.

range of elements 2430 (three on the top row and three on the bottom row) are detected
as damaged. The crack actually lies between elements 28 and 30 at the top of the cantilever.
The estimated Youngs moduli in the elements near the fixed end differ from the baseline
value, even though those elements are sensitive with respect to the test conditions as
demonstrated from the simulation study for the baseline structure. This phenomenon is
again attributable to the error in modelling (recall that we needed 960 elements to really
get the spectrum right in the cracked cantilever beam model). When 20% noise amplitude
was applied instead of 10% in the data perturbations with the same model structure,
almost the same estimates of the non-damaged elements were obtained. The modes shapes
computed with a plane stress model with the moduli of elements 2430 reduced to their
estimated values with all others at their baseline values are nearly identical to those shown
in Figure 8. Like the mode shapes from the BernoulliEuler beam model, the first and
second mode shapes are close to the measured data but the identified third mode does not
match as well. The modal displacement errors computed for the beam model and the plane
stress model are nearly the same, suggesting that those errors are not attributable to
modelling errors.

6. CONCLUSIONS

In this paper we have proposed a new algorithm for damage detection and assessment.
The algorithm considers changes in estimated constitutive properties of a finite element
model of the structure, with fixed topology and geometry, to be indicative of damage. An
output-error parameter estimation algorithm is used in conjunction with an adaptive
parameter grouping scheme to localize damage from sparse, noisy measurements.
Parameter changes due to damage are distinguished from changes due to measurement

544

. . .

noise with a data perturbation scheme. The limiting values for the two statistical damage
indices are established through a Monte Carlo simulation of the baseline structure. One
of the advantages of the algorithm presented here is that the sensitivity of each element
parameter can be evaluated during the process of the damage detection. The sensitivity
of the element parameters need not be determined separately before trying to detect
damage in the structure.
The damage detection algorithm was applied to the problem of identifying a crack in
a cantilever beam. The measured data for the cracked beam were taken from the literature.
All aspects of the damage detection scheme were examined except for establishing the
baseline parameters from measurements, which were not available among the reported
results. (We considered the baseline model to be adequately described by the nominal
dimensions and properties of the beam.) The BernoulliEuler beam model identified the
crack, but also suggested damage at other locations along the beam. Analysis showed that
the kinematic constraints associated with the beam theory caused difficulties in matching
the response of the higher frequency modes when only the beam modulus EI was available
to adapt to the measurements. A reduced bending modulus is only a marginally effective
way of modelling a crack in a beam. (Beam theory with a crack element as in reference
[10] that explicitly models the geometric defect does not suffer from this problem.) The
plane stress model gave much better results, giving a reliable and accurate assessment of
the crack location. Thus, one can conclude that using a change in element constitutive
properties is a reasonable approach to the problem of detecting topological damage like
cracks. Furthermore, the mesh refinement required to detect damage is relatively modest.
The 60-element model proved adequate to detect damage even though the 60-element
model with the crack explicitly modelled showed errors in frequency of the order of 10%.
It seems intuitively clear that the smeared crack approach cannot do better than an
approach that geometrically models the crack because the singularity in the strain field
associated with the crack is very important. In particular, all other things being equal, one
should expect the estimates from the crack element approach [10] to be sharper than those
obtained here. However, identification of geometric features of a solid body in general is
inherently more difficult than identification of constitutive parameters. In the case study
examined here there was a single vertical crack in the beam. In a practical application we
may not have such knowledge. The strength of the present approach lies in its generality.
We need know nothing about the damage (e.g., that it even manifests as a crack) before
we begin our assessment of it. The method can adapt to a variety of defects without
requiring a reformulation. We have shown that this general approach has merit in the
context of the difficult problem of crack identification.

ACKNOWLEDGMENT

The research reported here was supported by the National Science Foundation under
a Presidential Young Investigator Award. The results, opinions and conclusions expressed
in the paper are solely those of the writers and do not necessarily represent those of the
sponsors.

REFERENCES
1. M. S. A, S. F. M, R. K. M and T. K. C 1991 Journal of Engineering
Mechanics, American Society of Civil Engineers 117, 370390. System identification approach
to detection of structural changes.

545

2. P. H and F. J. S 1990 American Institute of Aeronautics and Astronautics Journal 28,


11101115. Structural damage detection based on static and modal analysis.
3. F. I, A. I and H. R. M 1990 Journal of Sound and Vibration 140, 305317.
Identification of fatigue cracks from vibration testing.
4. T. Y. K and T. Y. L 1994 International Journal of Solids and Structures 31, 925940. Crack
size identification using an expanded mode method.
5. T. W. L and T. A. L. K 1994 American Institute of Aeronautics and Astronautics
Journal 32, 10491057. Structural damage detection of space truss structures using best
achievable eigenvectors.
6. H. G. N 1989 89, Politechnike Pozuanska, 99110. Identification approaches
in damage detection and diagnosis.
7. H. G. N and C. C 1991Mathematical Systems and Signal Processing 5, 345356. Fault
detection and localization in structures: A discussion.
8. A. K. P, M. B and M. M. S 1991 Journal of Sound and Vibration 145,
321332. Damage detection from changes in curvature mode shape.
9. J. M. R and J. B. K 1992 American Institute of Aeronautics and Astronautics
Journal 30, 23102316. Damage detection in elastic structures using vibratory residual forces and
weighted sensitivity.
10. P. F. R, N. A and A. D. D 1990 Journal of Sound and Vibration 138,
381388. Identification of crack location and magnitude in a cantilever beam from the vibration
modes.
11. N. S, T. H. B and R. O 1990 International Journal of Analytical and
Experimental Modal Analysis 5, 6779. Global non-destructive damage evaluation in solids.
12. D. C. Z and M. K 1992 American Institute of Aeronautics and Astronautics
Journal 30, 18481855. Eigenstructure assignment approach for structural damage detection.
13. D. J. E 1984 Modal Testing: Theory and Practice. New York: John Wiley.
14. J. W 1990 Applied Mechanics Reviews 43, 1317. On the dynamics of cracked rotors: a
literature survey.
15. K. D. H, M. R. B and M. R. B 1995 Earthquake Engineering and Structural
Dynamics 24, 5367. On building finite element models of structures from modal response.
16. K. D. H, S. L. W and S. J. C 1992 Journal of Structural Engineering, American
Society of Civil Engineers 118, 223242. Mutual residual energy method for parameter estimation
in linear structures.
17. S. S and K. D. H 1994 Structural Research Series Report No. 593,
UILU-ENG-94-2013, Department of Civil Engineering, University of Illinois at UrbanaChampaign. Damage detection and assessment of structural systems from measured response.
18. A. R. B 1984 in Self-organizing Methods in Modelling (S. J. Falow, editor). New York:
Marcel Dekker. Predicted squared error: a criterion for automatic model selection.
19. H. A 1972 in Proceedings of the 2nd International Symposium on Information Theory (B. N.
Petrov and F. Csaki, editors), 267281. Budapest: Akademiai Kiado. Information theory and
an extension of the maximum likelihood principle.
20. S. T, D.H. Y and W. W, J. 1974 Vibration Problems in Engineering New
York: John Wiley; fourth edition.

You might also like