.

© All Rights Reserved

2 tayangan

.

© All Rights Reserved

- Using Excel to Fit a Titration Curve
- Numerical Methods for Unconstrained Optimization and Nonlinear Equations Classics in Applied Mathematics
- Econometrics
- Testing of Survey Network Adjustments
- lec10
- Valladolid Research Presentation.
- Reading Stories to Enhance English Grammar Intake-Correlational Analysis
- lavaan
- hsc-commerce-2015-march-maths2.pdf
- 1506.06099 finite element
- FullReport IBBS
- 10.Estimation - Introduction
- ESTIMATION OF REGRESSION COEFFICIENTS USING GEOMETRIC MEAN OF SQUARED ERROR FOR SINGLE INDEX LINEAR REGRESSION MOD
- CDB 3093 Data Handling, Statistic and Errors
- Cost and Decision Making
- Pearsons.pdf
- Samper 2006 Physics and Chemistry of the Earth, Parts ABC
- Simple Regression
- On the Least-squares Residual Anomaly Determination - Abdelrahman1985
- Vaz, Brendon

Anda di halaman 1dari 14

www.elsevier.com/locate/compstruc

using correlation and sensitivity of modal data

B.H. Koh

a

a,*,1

, S.J. Dyke

Department of Mechanical Engineering, Dongguk University, 3-26 Pil-dong, Chung-gu, Seoul 100-715, South Korea

b

Department of Civil Engineering, Washington University in St. Louis, St. Louis, MO 63130-4899, USA

Received 5 August 2005; accepted 16 September 2006

Available online 28 November 2006

Abstract

This study investigates the use of correlation-based damage detection methods for long-span, cable-stayed bridges. The proposed

approach is based on the multiple damage location assurance criterion (MDLAC), which combines a correlation-based technique with

a forward-type estimation of damage-sensitive structural parameters. Observing the level of correlation between the variations in the

measured and analytically synthesized natural frequencies enables damage localization. The sensitivity matrix, developed from the nite

element model, further accommodates multiple damage detection. The locations of damage are determined by iteratively searching for

the combination of structural parameters that maximizes the correlation coecient through the application of genetic algorithms. It is

demonstrated that correlation-based modal analysis is successful for damage detection and localization using a numerical model of a fullscale cable-stayed bridge.

2006 Elsevier Ltd. All rights reserved.

Keywords: Structural health monitoring (SHM); Cable-stayed bridge; Multiple damage location assurance criterion (MDLAC); Sensitivity matrix;

Genetic algorithm (GA)

1. Introduction

Large civil structures such as long-span bridges must be

periodically inspected to ensure structural integrity [1]. The

conventional approach, typically visual inspection, tends to

be labor-intensive, tedious, expensive, inconsistent and

subjective. The last few decades have seen enormous

resources dedicated to the development of reliable structural health monitoring (SHM) techniques [2]. Among

the variety of damage detection methods, modal-based

techniques have been the most widely investigated due to

their global nature and simplicity. Modal-based methods

exploit the observable variation in modal parameters

*

Corresponding author. Tel.: +82 2 2260 8591; fax: +82 2 2263 9379.

E-mail addresses: bkoh@dongguk.edu (B.H. Koh), sdyke@seas.wustl.

edu (S.J. Dyke).

1

The work performed while a postdoctoral researcher at Washington

University in St. Louis.

0045-7949/$ - see front matter 2006 Elsevier Ltd. All rights reserved.

doi:10.1016/j.compstruc.2006.09.005

Unlike local methods such as visual inspection, ultrasonic,

and acoustic techniques, modal-based SHM approaches

can readily accommodate automated damage localization

and result in consistent damage assessment. The nature

of damage-sensitive modal properties, i.e., natural frequencies and mode shapes, can be used to further categorize the

damage detection algorithms methodologies.

Although complete information regarding the mode

shapes and natural frequencies of a structure facilitates reliable judgments regarding the location and extent of damage, obtaining complete modal data is far from a reality

due to restrictions regarding the characteristics of the excitation, as well as the presence of unreachable DOFs

(degrees-of-freedom), measurement noise, modeling errors

and environmental inuences. Thus, a challenge encountered by modal-based SHM techniques is how to reliably

extract a sucient amount of damage-sensitive parameters

given limited and incomplete modal data measured from a

118

real structure. Sohn et al. provides a comprehensive literature review regarding recent research eorts on vibrationbased damage detection method [2]. Palacz and Krawczuk

also present a comparative investigation toward various

damage detection methods employing natural frequencies,

mode shapes, and frequency response functions [3].

Correlation-based methods have been widely employed

for damage detection. One of the earliest attempts in using

the correlation of the natural frequencies as a damage metric was a study by Cawley and Adams [4]. Lew later proposed a coherence-based method that exploits parameter

changes in the transfer functions for determining the damage type and location [5]. Geometrically, the correlation

coecient represents the angle between two vectors, one

of which is the vector of measured frequencies and transfer

function coecients of a potentially damaged structure,

and the other is the parameter vector predicted for various

damage scenarios. The coecient value becomes unity if

the measured frequency change vector precisely matches

the paired hypothesis vector, indicating the hypothesis vector corresponds to the correct location of damage. The correlation coecient becomes zero if the two vectors do not

have a linear relationship. Koh and Ray used coherencebased methods to localize damage in a cantilevered beam

through enhanced sensitivity of the closed-loop natural frequencies in a smart structure framework [6]. Armon presents a rank-ordering of modal frequency changes as a

discrete signature for characterizing damage in a cantilevered beam [7]. The study shows that the rank-ordering

of frequency shifts is only a function of damage location

and is insensitive to the severity of damage. Messina

et al. proposed the DLAC (damage location assurance criterion) which is similar to the concept of a MAC (modal

assurance criterion) [8]. Later, they further developed the

MDLAC (multiple damage location assurance criterion)

that accommodates multiple damage locations. The advantage that the MDLAC oers over the DLAC is that the

sensitivity matrix is incorporated into the correlation equation [9]. By including the derivatives of the natural frequencies with respect to the damaged parameters, the

correlation problem evolves into an optimization problem.

However, the main disadvantage of using only natural frequencies for damage detection is that natural frequency

changes are not unique, i.e., there could exist a number

of damage combinations that generate exactly the same

natural frequency variations [10].

Mode shapes retain the spatial prole of exibility at

each measured DOF, which could be directly or indirectly

used for locating damage. However, in practice, identied

mode shape information tends to be less accurate than natural frequency identication [11]. Extracting high-delity

mode shape information requires a large number of sensors

combined with signicant post-measurement analysis. The

simplest way to compare mode shapes between the damaged and healthy structures is by calculating the MAC

value of the paired modes [12]. The MAC value (ranging

from 0 to 1) quanties the extent of linear correlation

MAC value that is less than one for paired modes indicates

the existence of damage-induced disparity. MAC-based

damage detection techniques are discussed by Kim et al.

[13] who proposed several dierent versions of a MAC in

solving a damage localization problem. Ndambi et al.

showed that the MAC and Coordinate MAC value can

be used as an indication of damage through a study of a

reinforced concrete beam structure [14]. The study found

MAC values to be less sensitive to structural defects than

natural frequencies. Similar discussions regarding the low

sensitivity of MAC values to structural damage can be also

found in other studies [15]. Shi et al. proposed an extended

method of MDLAC, which exploits mode shape data

instead of natural frequencies. The study investigates a

multiple damage localization problem in planar truss structure using incomplete mode shape information [16].

Genetic algorithms (GA) has been one of the most successful search algorithms since its advent in 1970s [17].

Essentially, GA follows the theories of Darwin governing

the biological world to solve optimization problems.

Recently, GA has received increased attention in structural

health monitoring research [18,19]. The nature of damage

detection and optimization problems are closely related

in the sense that detecting damage is equivalent to nding

a set of design (damage) variables that best represent the

current state of damaged structure. Moslem and Nafaspour presented a two-stage damage detection process

[20]. The rst stage focuses on screening possible damage

locations using a modal residual vector. The next stage

determines the extent of damage through GA. This numerical study deals with an 8-bay planar truss structure given

measured natural frequency and mode shape data. Mares

and Surace also presented a GA-based damage detection

method using the modal force error [21] while Krawczuk

and Ostachowicz investigated a damage detection method

using only changes in the natural frequencies [22]. The latter study exploits the DLAC value as an objective function,

which varies from 0 to 1 depending on the level of correlation. Friswell et al. considered GA-based optimization

problem that minimizes the weighted sum of the squares

of the dierences in natural frequencies [23].

There is particular interest in advancing SHM methods

that are specically applicable to exible structures such as

long-span bridges [24]. In this paper, we investigate the feasibility of SHM techniques for long-span civil engineering

structures using a numerical model of the Bill Emerson

Memorial Bridge located in Cape Girardeau, Missouri.

The bridge spans the Mississippi river connecting Missouri

and Illinois and was the subject of two phases of a benchmark problem on the control of cable-stayed bridges [25

27]. The bridge has been open to the public since its completion in December 2003. The seismic risk to the bridge

was a critical design consideration due to its proximity to

the New Madrid seismic zone. Previously, the Emerson

Bridge has been studied by several researchers. Gattulli

investigated the dynamic interactions between the cables,

implemented a damage detection technique based on a

least-squares solution of the characteristic equation [29].

Caicedo and Dyke also constructed a scaled laboratory test

specimen representing half of the bridge and successfully

implemented a least-squares damage identication method

[27]. That study shows that the elemental stiness matrix

can be systematically parameterized for estimating Youngs

modulus of each of the elements, given modal frequencies

and mode shapes.

This paper intends to illustrate the practical aspects of

damage detection methods that exploit the correlation of

natural frequencies and mode shapes. Several approaches

are discussed such as least-squares, frequency-error, and

sensitivity-based correlation method (MDLAC). Instead

of using inverse-type algorithms, this study demonstrates the performance of forward-type, sensitivity-based

MDLAC methods when applied to a model of a cablestayed bridge. To avoid performing an exhaustive search

of the parameter space, GA is adopted to rapidly and eciently identify multiple damage locations using only natural frequencies. Local damage is modeled as a reduction of

the Youngs modulus of particular combinations of elements in the structural model. Although damage is a nonlinear phenomenon, this approach is adopted because the

structural behavior before and after damage occurs, for relatively small disturbances, is linear. Stacked mode shape

correlation is also suggested as a viable tool for damage

localization.

2.1. Single damage detection using natural frequencies

Determining the level of correlation between the measured and predicted (hypothesis) modal frequencies provides a simple statistical tool for locating damage. The

parameter vectors used for evaluating correlation coecients consist of the ratios of the rst n modal frequency

changes due to damage to the modal frequencies, i.e.,

Dx = (xh xd)/xh. Here, xh and xd denote the natural

frequency vectors of the healthy and damaged structures,

respectively. Likewise the corresponding hypothesis vector,

predicted from an analytic model is denoted dx. Given a

pair of parameter vectors, one can estimate the level of correlation in several ways. The easiest way to estimate correlation is to calculate the angle between the two parameter

vectors. A damage localization method using the pair comparison tries to nd linear correlation of modal frequency

variation vectors, as in

Cj

DxT dxj

:

jDxjjdxj j

damage (j = 1,2, . . . , r). Another correlation-based metric

119

expressed in the following form [8]:

T

Dx dxj 2

:

2

DLACj

DxT Dxj dxTj dxj

Similar to Eq. (1), the DLAC compares two frequency

change vectors, one based on measurements obtained from

the test structure, the other from the jth hypothesis of an

analytical model of the structure [30]. Both Eqs. (1) and

(2) evaluate the level of correlation between the two parameter vectors. Another correlation-based technique called

the rank-ordering method also exploits the relationship between the modal frequency shift and the location of damage [7]. However, instead of applying the actual value of

modal frequency change, the rank of this ratio is used.

2.2. Multiple damage detection using natural frequencies

All the aforementioned methods employing natural frequencies are in general only capable of detecting a single

damage location. Because a unique pattern of modal frequency changes only holds for a single damage case, the

algorithm becomes impractical for structures that have

multiple defects, or an unknown number of defects. To

accommodate multiple damage locations, a sensitivity

matrix derived from the analytic model of the structure

can be incorporated into Eq. (1). The sensitivity matrix is

composed of the rst-order derivatives of the modal frequencies with respect to each damage variable. An example

of a damage variable could be the ratio of stiness reduction in each element. Typically, the sensitivity matrix is

derived either from nite-dierences or modal-based methods [31]. In this regard, Messina et al. introduced the

MDLAC (multiple damage localizing assurance criterion)

[9].

Because dx = Sdz,

T

Dx Sdzj 2

;

MDLACj

3

DxT Dxj Sdzj T Sdzj

where the sensitivity matrix S contains the rst-order derivatives of m natural frequencies (x) with respect to r damage

variables (z), as in

3

2

ox1 ox1

ox1

6 oz1

oz2

ozr 7

7

6

7

6 ox

6 2 ox2 ox2 7

7

6

oz

oz2

ozr 7:

4

S6

7

6 1

6 ..

.. 7

..

..

6 .

. 7

.

.

7

6

4 oxm oxm

oxm 5

oz1

oz2

ozr

The only dierence between the previously discussed correlation methods and the MDLAC is the presence of the sensitivity matrix. Although incorporating the sensitivity matrix

into the correlation expression enables the identication of

120

nding the appropriate combination of damage variables

dzi. The objective is to nd a damaged variable vector dzi that

makes the MDLAC equal to one. However, evaluating all

possible combinations of damage variables that maximizes

the MDLAC is prohibitive even for a simple problem. For

example, considering an identication model having 10 damage locations (elements) with ve possible levels of damage

severity produces 510 potential combinations of damage

variables. Thus, an ecient searching algorithm such as

GA is needed to determine the correct set of damage

variables.

2.3. Genetic algorithms

GA exploits the mechanisms of biological evolution to

perform optimization without information regarding the

derivative of the objective function. The search strategy

in GA is based on survival of the ttest; the global optimum can be sought by evolution in a series of generations.

The initial population or trial set of solutions (chromosomes) repeatedly evolve throughout generations until convergence to a desired solution occurs. The key actions in

GA are encoding, selection, crossover, and mutation. First,

a set of solutions or a chromosome (a string of DNA) is

encoded to a string of binary code, which is an ecient

way of representing design variables in the search space.

The selection process determines a pair of superior parents

(best solutions) and reproduces their child as a member of

the next generation. The members of the chromosomes that

give the best value of the objective function are chosen as

the ttest. A portion of the parents chromosome is also

combined and inherited into the next generation through

crossover. Typically, the range of the chromosome that will

be combined is randomly chosen. Although there are several dierent techniques available for the selection process,

elitism has good results in many problems. Elitism allows

the parents to survive through the generations as long as

they are the best among the population. This technique

eliminates the possibility of extinction of dominant chromosomes. Mutation is another key action in GA to reduce

inbreeding and prevent premature convergence to a local

minimum. Crossover and mutation are obvious advantages

GA have over conventional algorithms such as the conjugate gradient method. Thus, GA is particularly ecient

for highly complex optimization problems with many

nested local minima. Subsequently, GA is used herein for

searching for the damage variable vector that maximizes

the correlation coecient of natural frequency change

ratios.

2.4. Multiple damage detection using mode shapes

Most of the existing correlation-based methods use the

vector of damage-inuenced modal frequency changes. In

methods employing mode shape data in the damage detection problem, the use of MAC values between paired

most widely studied. More specically, a set of structural

parameters or damage variables is determined to make

the MAC value for paired modes equal to one. Thus, the

number of available MAC values equals the number of

measured modes. Instead of individually evaluating the

MAC value for each mode shape, a single vector of stacked

mode shapes can be used for evaluating linear correlation.

Originally, the idea of using mode shape data as a parameter of MDLAC is proposed by Shi et al. [16]. Unlike natural frequencies, mode shapes retain spatial information

regarding exibility at each DOF. Thus, correlation

between stacked mode shapes can directly indicate damage

locations. Because we are only interested in identifying the

locations of damage, relative change in each mode shape

between healthy and damaged states can be used for calculating correlation coecients. The change of mode shapes

DU(m r) due to damage is transformed to a single vector,

vec[DU](mr 1) by stacking r columns (or modes) of the

matrix DU.

T

SMSCj

vecDU vecdUj

:

jvecDUjjvecdUj j

variation of the identied and predicted mode shapes,

respectively. The correlation coecient between vec[DU]

and vec[dU] is denoted herein as SMSC (stacked mode

shape correlation). The obvious benet of using stacked

mode shapes instead of natural frequencies is the capability

of detecting and localizing multiple damage locations

without incorporating the sensitivity information. Thus, a

searching algorithm is not required for identifying multiple

damage locations in SMSC method. However, the prole

of this correlation coecient does not reveal information

regarding the extent of damage.

3. Illustrative example: cantilevered beam

In this section we demonstrate the performance of the

damage detection algorithms using modal sensitivity and

correlation through a simple numerical example. A nite

element (FE) model of a cantilevered beam having 15

elements is the example structure as shown in Fig. 1.

The length, thickness and width are 2.74, 0.00635, and

0.0760 m, respectively. For comparative investigations,

three dierent damage detection algorithms, including iterative least-squares, frequency-error, and MDLAC methods

are individually tested. Each method exploits the natural

frequency data obtained from the same cantilevered beam

model. And, later in the section, another damage detection

algorithm that uses the correlation of stacked mode shape

is investigated for comparison. The penalty function for the

iterative least-squares and frequency-error methods is the

change in the natural frequencies due to the variation of

damage variables, while the objective function for the

MDLAC method is the coecient in Eq. (2). Both the fre-

121

12

Fig. 1. Cantilevered beam model having 15 elements. The length, thickness and width are 2.74, 0.00635, and 0.0760 m, respectively. Two damages are

located at element numbers 4 and 12.

optimal solution.

3.1. Iterative least-squares method

The sensitivity-based, least-squares or penalty function

method exploits a truncated Taylor series expansion of natural frequencies for iteratively updating the damage variables that minimize a penalty function [32]. The change

in natural frequency dx = xh xd is expressed as a function of the dierence in the damage variable vector dz =

zh zd. Here, the damage variable vector zh represents

the healthy state of the structural parameters, whose elements should each be one (or 100%). First, the sensitivity,

or Jacobian matrix S, is developed using the nite-dierence method. The dierences between the predicted and

the identied natural frequencies, dx, are computed and

pre-multiplied by the pseudo-inverse of the sensitivity

matrix S+ yielding the updated damage variable vector

z(new) as in

dz S dx;

znew zold dz:

6

7

the current state of the sensitivity matrix S

new and to obtain

the natural frequencies from the FE model. Likewise,

damage variables are iteratively updated by projecting

the dierence between the current and identied natural

frequencies. The strength of the iterative least-squares

method is its convergence speed. However, depending on

the available modes and parameter uncertainties, the sensitivity matrix can easily be ill-conditioned, and small errors

can be amplied by the use of the pseudo-inverse of the

sensitivity matrix. Moreover, the resulting solution may

end up with values for the damage variables that do not

have any physical meaning. Thus, extra constraints are

commonly imposed to regulate the stability of convergence

[33].

3.2. Frequency error method

The goal of GA in damage detection problem is to nd

an optimal set of damage parameters that accurately reproduces the identied modal frequencies from the damaged

structure. Typically, an FE model provides predicted

modal frequencies and the sum of the square of the natural

frequency dierences between the identied and predicted

[23] as in

2

m

X

xj xaj

J

Wj

:

8

xj

j1

Here, Wj denotes a weight applied to the jth natural frequency, which is often considered as unity. xj and xaj are

the jth identied and predicted natural frequencies of a

structure, respectively. However, the eigenvalues of the

FE model need to be computed to evaluate the objective

function. This computation could be prohibitive if the

dimension of the FE model is relatively large.

3.3. Multiple damage location assurance criterion

(MDLAC)

The coecient of the MDLAC in Eq. (3) can be used as

an objective function. Since the value of the MDLAC

varies between 0 and 1, the set of damage variables that

produces the maximum objective function, i.e., 1, will

indicate the correct location of damage. Unlike the frequency-error method, the sensitivity matrix is already

incorporated into the objective function. Thus, evaluating

the eigenvalues of the FE model is not required during

the search process. The question arises as to how to obtain

an accurate sensitivity matrix for calculating a reliable

MDLAC. Because all elements in the variable vector in

Eq. (3) are expressed in a ratio, the MDLAC should have

the same value, regardless of the damage severity, as long

as the ratios of the damage variables are identical. However, the absolute severity of damage can be estimated

through the use of a scaling coecient, as described in reference [9].

3.4. Simulation results

In this study, the rst ve natural frequencies are used

for damage detection. Damage is induced at elements 4

and 12 of the cantilevered beam as shown in Fig. 1. Three

dierent damage detection methods are tested and the identied damage variables are expressed in ratios of Youngs

modulus (E) reduction as shown in Fig. 2. The vertical

solid bars indicate the true location and the extent of damage induced in the model, while the hollow bars show the

identied damage variables for each element. In each

graph, identied damage ratios are given (e.g. 0.3 indicates

a 30% reduction of E of the beam element). Although the

122

generation is set to p = 20, the crossover probability is

Pc = 0.1, and the mutation probability is Pm = 0.01. The

best solution of the current generation is replicated to the

next generation, adopting the elitism strategy. Although

the same GA parameters are used with both methods, the

MDLAC (Fig. 2c) has better results in terms of accuracy

and computing time.

0.4

Damage Ratio

0.3

0.2

0.1

0.1

1

10

11

12

13

14

15

10

11

12

13

14

15

10

11

12

13

14

15

Damage Ratio

0.4

0.3

0.2

0.1

0.1

0.4

Damage Ratio

0.3

0.2

0.1

0.1

Element Number

Fig. 2. Damage localization results: (a) iterative least squares, (b)

frequency-error, (c) MDLAC. Solid bar indicates location and amount

of damage induced. Hollow bar shows the value of identied damage

variable.

20 iterations (Fig. 2a), signicant leakage is observed in

some of the undamaged elements. The other two methods

produced relatively reliable damage localization results.

The same GA parameters are used for both the frequency-error and the MDLAC methods as shown in

This section demonstrates the performance of correlation-based method, SMSC, as an eective tool for damage

localization which is introduced in the study of Shi et al.

[16]. Rather than using a vector of natural frequency ratios,

stacked mode shape vectors are used for computing the

correlation coecients. As mentioned previously, correlation-based methods that do not incorporate a sensitivity

matrix are not generally capable of localizing multiple

damage locations. This limitation is due to the fact that

natural frequencies are global properties of the structure,

and they do not directly provide information on the location of structural defects. However, mode shapes exhibit

spatial proles of exibility between each DOF in a structure. The dierence in the mode shapes between the healthy

and damaged structures reveals points of discontinuity or

damage. For numerical demonstration, the cantilevered

beam used previously is employed (Fig. 1). However, the

FE model here is rened such that a total of 50 beam elements are used rather than the original 15 elements to

increase the complexity of the problem and demonstrate

the ecacy of the technique.

Reducing Youngs modulus of elements 15 and 25 by

10% simulates the eects of damage in the beam. Fig. 3a

shows a stacked vector of mode shape dierences for the

rst six modes. Here, due to practical diculties involved

in accurately measuring rotational components, the mode

shapes include only translational DOFs of the beam model.

Thus, the stacked mode shape has a total of 300 data

points. The easiest way of locating damage using mode

shape dierences is by graphically nding discontinuities

caused by structural defects. However, noise-corrupted

mode shape data (2% standard deviation) as shown in

Fig. 3b make it dicult to visually distinguish damageinduced peaks from noise. Interestingly, even with noisecorrupted mode shape dierences, the correlation

coecient, i.e., Eq. (5), successfully detects multiple damage locations as shown in Fig. 3c. In the gure, the vertical

bar indicates the true locations of damage while the peaks

of SMSC curve indicate potential locations of damage.

Note that uncorrelated random values have been individually imposed on the mode shapes of both the healthy and

damaged state to simulate measurement noise. Fig. 4 also

illustrates similar localization result for a dierent scenario,

i.e., with elements 8 and 35 damaged. Because Eq. (5) represents the level of linear correlation between two sets of

discrete measurements, it provides a robust parameter for

123

x 10

5

0

Mode 1

Mode 5

10

0

3

x 10

50

100

150

200

250

300

50

100

150

200

250

300

1

0.5

0

0.5

1

0

0.8

0.6

SMSC

0.4

0.2

0

0.2

0.4

0

10

15

20

25

30

35

40

45

50

Element Number

Fig. 3. Damage localization results from stacked mode shape correlation: (a) noise-free stacked mode shape dierence, (b) noise-polluted stacked mode

shape dierence, (c) correlation value of SMSC. Vertical bars indicate the correct locations of damage induced.

have more data points than natural frequencies, making

them eective for uniquely locating damage as compared

to the previous frequency-based correlation methods. In

summary, one can detect multiple damage locations using

the correlation coecient of stacked mode shape dierences, although the extent of damage is not readily identied in this study.

4. Case study: cable-stayed bridge

4.1. Description of the long-span bridge

Fig. 5 provides a schematic drawing of the Bill Emerson

Memorial Bridge. Although the total length of the bridge is

1205.8 m, only the span between Bent 1 and Pier 4 (636 m)

is considered in this study because the dynamic eect of the

Illinois approach is known to be negligible. Previously, a

three-dimensional FE model of this bridge was developed

to address control issues of cable-stayed bridges under seismic excitation [25,26]. The FE model developed for that

study is used here. The model is composed of beam, cable,

lumped mass and rigid elements to fully represent the

bridges dynamic characteristics. The bridge model has a

total of 579 nodes, 420 rigid links, 162 beam elements

and 128 cable components. As shown in Fig. 6, cables that

weight of the bridge. Each tower has 50 nodes, 43 beam elements and 74 rigid links. Boundary conditions at Bent 1

and Pier 4 only allow translation in X and rotation in Y

and Z directions. Additionally, Pier 2 and 3 only restrict

translational movement in Y direction.

It is well known that cable-stayed bridges exhibit a signicant amount of nonlinear behavior due to tensioning

and sag in the cables, beam-column interaction, and large

deformation resulting from the bridges self-weight. To

faithfully represent the realistic aspects of cable-stayed

bridges, FE analysis uses the post-deformed geometry of

the bridge based on the results of nonlinear static analysis.

Further details regarding the nonlinear static analysis of

this model can be found in references [25,26]. Representative mode shapes from this model are shown in Fig. 7.

4.2. Results and discussion

In total, 66 beam elements are used to model the deck of

the bridge. Reducing E (Youngs modulus) of an individual

beam element simulates a structural defect in the deck.

Considering the total number of DOFs, the frequencyerror function (Eq. (4)) is not appropriate damage detection metric for the bridge structure. Thus, the MDLAC

combined with GA is used here for locating damage.

124

4

x 10

10

5

0

5

0

50

100

150

200

250

300

250

300

x 10

3

2

Mode 1

Mode 5

1

0

1

0

50

100

150

200

SMSC

0.4

0.2

0

0.2

0.4

0

10

15

20

25

30

35

40

45

50

Element Number

Fig. 4. Damage localization results from stacked mode shape correlation: (a) noise-free stacked mode shape dierence, (b) noise-polluted stacked mode

shape dierence, (c) correlation value of SMSC. Vertical bars indicate the correct locations of damage induced.

Y

X

Deck 15

Pier 4

Pier 3

Deck 52

B: (Deck 15, 52)

Pier 2

Bent 1

Deck 20

Deck 10

Fig. 6. Finite element model of the Bill Emerson Memorial Bridge.

125

evaluate the technique for long-span bridges. Each case has

arbitrarily chosen dual damages. For damage case A, E of

beam elements 10 and 20 were reduced by 50%, as depicted

in Fig. 6. Likewise, case B considers the same extent of

damage at deck elements 15 and 52. Although a real bridge

exhibits a mixture of axial, bending and torsional motions,

only the bending modes are used herein for calculating the

MDLAC. Apparently, the bending modes are considered

to be the most sensitive to this type of damage and thus will

provide the best results.

Fig. 8 shows the damage localization results for case A

using the rst 28 modes (ranging from 0.29 to 3.90 Hz).

In previous studies researchers have readily obtained 23

modes from a similar bridge with a limited number of sensors. Using optimal sensor placement techniques combined

with appropriate modal analysis methods, 28 modes is not

beyond current capabilities [34]. Fig. 8a depicts the search

process to nd the best set of damage variables over 5000

generations. The hollow bars in Fig. 8b indicate the identied damage variable, i.e., ratio of E for the damaged structure to that of the healthy structure for the corresponding

element. On the other hand, the solid bars show the true

in the model (0.5 means 50% reduction of E). The

MDLACGA method successfully locates the damage,

although their true magnitudes are not correctly captured.

Relatively small leakage is present in some of the locations

in which no damage occurs. This behavior is common in

most damage detection techniques. Figs. 813 also illustrate the convergence speed of the GA in terms of number

of generations. The search prole shows that convergence

occurs very fast at the beginning and slows down as the

objective function approaches steady-state value, which is

a typical characteristic of GA.

Fig. 9 shows the localization results for damage case B

demonstrating a level of accuracy similar to the results of

case A. Figs. 10 and 11 indicate the localization results of

cases A and B, respectively, using only the rst 20 natural

frequencies. Interestingly, reducing the number of available

modes from 28 to 20 specically deteriorates the results

when the damage is in a particular region. Damage case

A (Fig. 10) is more signicantly aected by number of

modes than damage case B (Fig. 11). Damage detection

results for case A using the data set of noise-corrupted

(2% standard deviation) natural frequencies are illustrated

126

1

Objective Function

0.95

0.9

0.85

0.8

0.75

0.7

0.65

0

10

10

10

Generations

10

10

Damage Ratio

0.8

0.6

0.4

0.2

0

10

20

30

40

50

60

Element Number

Fig. 8. Damage localization results of MDLACGA using rst 28 modes: (a) performance of GA, (b) identied damage variables (hollow bar), damage

case A (solid line).

Objective Function

0.95

0.9

0.85

0.8

0.75

0.7

0

10

10

10

10

10

Generations

1

Damage Ratio

0.8

0.6

0.4

0.2

0

10

20

30

40

Element Number

50

60

Fig. 9. Damage localization results of MDLACGA using rst 28 modes: (a) performance of GA, (b) identied damage variables (hollow bar), damage

case B (solid line).

127

Objective Function

0.95

0.9

0.85

0.8

0

10

10

10

10

10

Generations

1

Damage Ratio

0.8

0.6

0.4

0.2

0

10

20

30

40

50

60

Element Number

Fig. 10. Damage localization results of MDLACGA using rst 20 modes: (a) performance of GA, (b) identied damage variables (hollow bar), damage

case A (solid line).

Objective Function

0.95

0.9

0.85

0.8

0.75

0.7

0.65

0

10

10

10

Generations

10

10

Damage Ratio

0.8

0.6

0.4

0.2

0

10

20

30

40

50

60

Element Number

Fig. 11. Damage localization results of MDLACGA using rst 20 modes: (a) performance of GA, (b) identied damage variables (hollow bar), damage

case B (solid line).

128

0.6

Objective Function

0.55

0.5

0.45

0.4

0.35

0.3

0

10

10

10

10

10

Generations

1

Damage Ratio

0.8

0.6

0.4

0.2

0

10

20

30

40

50

60

Element Number

Fig. 12. Damage localization results of MDLACGA using noise-corrupted rst 28 modes: (a) performance of GA, (b) identied damage variables

(hollow bar), damage case A (solid line).

Objective Function

0.95

0.9

0.85

0.8

0.75

0.7

0.65

0

10

10

10

10

10

Generations

1

Damage Ratio

0.8

0.6

0.4

0.2

0

10

20

30

40

50

60

Element Number

Fig. 13. Damage localization results of MDLACGA using the last 18 modes: (a) performance of GA, (b) identied damage variables (hollow bar),

damage case A (solid line).

the noise-corrupted case only reaches 0.55, while all other

noiseless cases exceed 0.9. Although the results wrongly

indicate false-positive outcomes in undamaged elements

34 and 53 as potential damage locations, no false negative

outcomes occurred.

It is notable that while damage in a cantilevered beam

model has been successfully localized using only ve natural frequencies (Fig. 2c), for the more complex bridge

model, we require up to 28 natural frequencies to achieve

similar results (Fig. 8b). To explain this conclusion, recall

that the number of parameters (potential damage locations) is quite dierent between the two models. The cantilevered beam has 15 variables while the cable-stayed bridge

model has 66 variables. Additionally, the cable-stayed

bridge exhibits a completely dierent pattern of natural frequencies than the cantilevered beam, i.e., the dominant low

frequencies are closely-spaced in a exible structure such as

a long-span bridge. Thus, locating the damage, which is

more sensitive to higher modes, is more challenging.

Fig. 13 illustrates a localization result that uses only the

upper 18 natural frequencies out of the rst 28, i.e., modes

1128 (ranging 1.263.90 Hz). Obviously, Fig. 13 (using 18

modes) shows less accurate damage detection than the

results shown in Fig. 8 which used 28 modes. It is concluded that the set including modes 1128 contains more

unique information regarding damage locations than the

set including modes 120. An important observation can

also be made from the results in Figs. 10 and 13. For the

same damage case the results are signicantly dierent

from the rst 20 modes (Fig. 10) and the last 18 modes

(Fig. 13), showing that the success of the MDLAC-based

method heavily depends on particular set of natural

frequencies rather than simply the number of modes

employed. Thus, for a given application it may be necessary to nd a suitable set of measurable modal frequencies

that is successful for a broad set of damage cases when the

suspected locations of damage are not known in advance.

Alternatively one might weight sets of modes that are more

successful for damage cases with a higher likelihood of

occurrence.

5. Conclusions

This paper discusses the capabilities and limitations of

several correlation-based damage detection algorithms,

and evaluates their performance through two numerical

examples. First, a cantilevered beam model is used for a

comparative study, i.e., iterative least squares, frequencyerror, and MDLAC methods. It is also shown that the correlation of stacked mode shape dierences serves as an

eective tool for locating damage in civil structures, particularly when multiple damage locations are present. Finally,

a cable-stayed bridge model is used for simulating localization of damage, given the MDLAC as the cost function for

the optimization. The damage variable set that best represents the modal properties of the damaged bridge structure

129

is obtained through genetic algorithms. Numerical simulation shows that the MDLACGA approach yields successful localization of multiple damage locations presented in a

cable-stayed bridge structure. Future studies will consider

experimental validation of this approach.

Acknowledgements

This research was conducted with partial support from

the US National Science Foundation (Grant No. CMS

02-45402). The authors would like to acknowledge the support and advise of Prof. Juan M. Caicedo (University of

South Carolina) in providing and using the model of the

Bill Emerson Memorial Bridge.

References

[1] Code of Federal Regulations, Title 23, vol. 1, April 1, 2002.

[2] Sohn H, Farrar CR, Hemez FM, Shunk DD, Stinemates DW, Nadler

BR. A review of structural health monitoring literature: 19962001.

Technical reports LA-13976-MS, Los Alamos National Laboratory,

2003.

[3] Palacz M, Krawczuk M. Vibration parameters for damage detection

in structures. J Sound Vib 2002;249(5):9991010.

[4] Cawley P, Adams RD. The localization of defects in structures from

measurements of natural frequencies. J Strain Anal 1979;14:4957.

[5] Lew J-S. Using transfer function parameter changes for damage

detection of structures. AIAA J 1995;33(11):218993.

[6] Koh BH, Ray LR. Feedback controller design for sensitivity-based

damage localization. J Sound Vib 2004;273(12):31735.

[7] Armon D, Ben-Haim Y, Braun S. Crack detection in beams by rankordering of eigenfrequency shifts. Mech Systems Signal Process

1994;8(1):8191.

[8] Messina A, Jones IA, Williams EJ. Damage detection and localisation

using natural frequency changes. In: Proc. conf. identication in

engineering systems, Swansea, UK, 1996. p. 6776.

[9] Messina A, Williams EJ, Contursi T. Structural damage detection by

a sensitivity and statistical-based method. J Sound Vib

1998;216(5):791808.

[10] Abdo MA-B, Hori M. A numerical study of structural damage

detection using changes in the rotation of mode shapes. J Sound Vib

2002;251(2):22739.

[11] Friswell MI, Penny JET. The practical limits of damage detection and

location using vibration data. 11th VPI & SU symposium on

structural dynamics and control. Virginia: Blacksburg; 1997. May.

[12] Salawu OS, Williams C. Damage location using vibration mode

shapes. In: Proc. 12th int. modal analysis conf. Honolulu, Hawaii,

February, 1994. p. 9339.

[13] Kim J-H, Jeon H-S, Lee C-W. Application of the modal assurance

criteria for detecting and locating structural faults. In: Proc.10th int.

modal analysis conf. San Diego, CA, February, 1992. p. 53640.

[14] Ndambi J-M, Vantomme J, Harri K. Damage assessment in

reinforced concrete beams using eigenfrequencies and mode shape

derivatives. Eng Struct 2002;24:50115.

[15] Fox CHJ. The location of defects in structures: a comparison of the

use of natural frequency and mode shape data. In: Proc.10th int.

modal analysis conf. San Diego, CA, February, 1992. p. 5228.

[16] Shi ZY, Law SS, Zhang LM. Damage localization by directly using

incomplete mode shapes. J Eng MechASCE 2000;126(6):65660.

[17] Holland JH. Adaptation in natural and articial systems. Ann

Arbor, MI: University of Michigan Press; 1975.

[18] Rao MA, Srinivas J, Murthy BSN. Damage detection in vibrating

bodies using genetic algorithms. Comput Struct 2004;82(1112):

9638.

130

[19] Pawar PM, Ganguli R. Genetic fuzzy system for damage detection in

beams and helicopter rotor blades. Comput Methods Appl Mech Eng

2003;192(1618):203157.

[20] Moslem K, Nafaspour R. Structural damage detection by genetic

algorithm. AIAA J 2002;40(7):1395401.

[21] Mares C, Surace C. An application of genetic algorithms to identify

damage in elastic structures. J Sound Vib 1996;195(2):195215.

[22] Ostachowicz W, Krawczuk M. Identication of delamination in

composite beams by genetic algorithm. Sci Eng Compos Mater

2002;10:14755.

[23] Friswell MI, Penny JET, Garvey SD. A combined genetic and

eigensensitivity algorithm for the location of damage in structures.

Comput Struct 1998;69(5):54756.

[24] Farrar CR, Doebling SW. Lessons learned from applications of

vibration-based damage identication methods to large bridge

structures. In: Proc. int. workshop on structural health monitoring.

Stanford, CA, September, 1997. p. 35170.

[25] Dyke SJ, Caicedo JM, Turan G, Bergman LA, Hague S. Phase I

benchmark control problem for seismic response of cable-stayed

bridge. J Struct EngASCE 2003;129(7):85772.

[26] Caicedo JM, Dyke SJ, Moon SJ, Bergman LA, Turan G, Hague S.

Phase II benchmark control problem for seismic response of cablestayed bridges. J Struct Control 2003;10(July):13768.

monitoring for exible bridge structures. Struct Control Health

Monitor 2005;12(34):42543.

[28] Gattulli V. Dynamic interactions in cable-stayed bridges. In: 4th int.

workshop on structural control. Columbia University, New York,

USA, 2004. p. 27281.

[29] Caicedo JM. Structural health monitoring of exible civil structures,

Ph.D. dissertation, Department of Civil Engineering, Washington

University in St. Louis, August, 2003.

[30] Clayton EH, Koh BH, Xing G, Fok CL, Dyke SJ, Lu C. Damage

detection and correlation-based localization using wireless mote

sensors. In: 2005 int. sympos. intelligent control 13th mediterranean

conf. control and automation, Cyprus, June, 2005.

[31] Fox RL, Kapoor MP. Rates of change of eigenvalues and eigenvectors. AIAA J 1968;6:24269.

[32] Lew SJ, Juang J-N. Structural damage detection using virtual passive

controller. J Guidance, Control, Dynam 2002;25(2):41924.

[33] Ahmadian H, Mottershead JE, Friswell MI. Regularisation methods

for nite element model updating. Mech Systems Signal Process

1998;12(1):4764.

[34] Nagayama T, Abe M, Fujino Y, Ikeda K. Structural Identication of

a nonproportionally damped system and its application to a full-scale

suspension bridge. J Struct Eng 2005;131(10):153645.

- Using Excel to Fit a Titration CurveDiunggah olehMark
- Numerical Methods for Unconstrained Optimization and Nonlinear Equations Classics in Applied MathematicsDiunggah olehmohammad_albooyeh
- EconometricsDiunggah olehMichel André Breau
- Testing of Survey Network AdjustmentsDiunggah olehGiora Rozmarin
- lec10Diunggah olehclubyz
- Valladolid Research Presentation.Diunggah olehlucdah
- Reading Stories to Enhance English Grammar Intake-Correlational AnalysisDiunggah olehIJ-ELTS
- lavaanDiunggah olehJordano Bruno
- hsc-commerce-2015-march-maths2.pdfDiunggah olehShradha Rohan Bayas
- 1506.06099 finite elementDiunggah olehGrantHerman
- FullReport IBBSDiunggah olehtengkek2002
- 10.Estimation - IntroductionDiunggah olehsparsh
- ESTIMATION OF REGRESSION COEFFICIENTS USING GEOMETRIC MEAN OF SQUARED ERROR FOR SINGLE INDEX LINEAR REGRESSION MODDiunggah olehAdam Hansen
- CDB 3093 Data Handling, Statistic and ErrorsDiunggah olehJc Jackson
- Cost and Decision MakingDiunggah olehAnonymous qAegy6G
- Pearsons.pdfDiunggah olehMei Astrid Ardiani
- Samper 2006 Physics and Chemistry of the Earth, Parts ABCDiunggah olehMisael Ramírez
- Simple RegressionDiunggah olehDlo Perera
- On the Least-squares Residual Anomaly Determination - Abdelrahman1985Diunggah olehHadyan Syahputra
- Vaz, BrendonDiunggah olehClaudia
- hw1Diunggah olehMorokot Angela
- cHapter 2Diunggah olehSey Sey
- Final Business ReportDiunggah olehFahad Ahmed
- pas_2007Diunggah olehandhracolleges
- CorrelationDiunggah olehSyed Raheel Hassan
- Aberdeen Growth StudyDiunggah olehCristian Issac
- Q.LRDiunggah olehRozen Tareque Hasan
- training assessment ERIC_ED221710Diunggah olehneeraj00715925
- Could&Physical&Activity&Improve&Healthy&Lifestyles&and&Diunggah olehec16043
- WorksheetDiunggah olehfieqa

- Wang 2013Diunggah olehpaulkohan
- Sezen 2012Diunggah olehpaulkohan
- 1-s2.0-S0141029610000209-mainDiunggah olehpaulkohan
- Digital Image Correlation to Reinforced Concrete FractureDiunggah olehjs kalyana rama
- 1-s2.0-S0141029612004609-mainDiunggah olehpaulkohan
- 99-s80Diunggah olehpaulkohan
- widianto2010Diunggah olehpaulkohan
- 106-s32Diunggah olehpaulkohan
- 1-s2.0-S0263822315009976-mainDiunggah olehpaulkohan
- 98s11.pdfDiunggah olehpaulkohan
- Wind Loads On Non-Building Structures Using ASCE 7-10 .pdfDiunggah olehgulilero_yo
- 10.1002@stc.2273Diunggah olehpaulkohan
- 2014 - Damage identification-Modal Curvature.pdfDiunggah olehpaulkohan
- 2013 - TurquiaDiunggah olehpaulkohan
- 2013 - Chile.pdfDiunggah olehpaulkohan
- 1-s2.0-S2214399815000089-main.pdfDiunggah olehpaulkohan
- 1-s2.0-S026382230500022X-mainDiunggah olehpaulkohan
- 101103366-Concrete-Bridges-Design-and-Construction.pdfDiunggah olehpaulkohan
- Merta 2010Diunggah olehpaulkohan
- 1-s2.0-S0141029610000209-mainDiunggah olehpaulkohan
- 81-46Diunggah olehpaulkohan
- 96-S13Diunggah olehpaulkohan
- 96-S13Diunggah olehpaulkohan
- Bhattacharya 2014Diunggah olehpaulkohan

- TL1 Dr KasikciDiunggah oleheaaziz
- FEA 2016 v11 Release NoteDiunggah olehKlLee
- Document 1Diunggah olehJojen
- Module08 New.docDiunggah olehMiltiadis Zabelas
- Sokolnikoff Theory of ElasticityDiunggah olehAdam Taylor
- chapter15-1Diunggah olehchangaboy24
- FIZ101E_1vDiunggah olehcenkj
- June 2015 QP - M1 EdexcelDiunggah olehRishita Singh
- SENSOR BANNER CONTRASTE.pdfDiunggah olehdieim
- Chemical EquilibriaDiunggah olehicegg
- A Finite Element Calculation of Stress Intensity Factors of Cruciform and Butt Welded Joints for Some Geometrical ParametersDiunggah olehAhmed M. Al-Mukhtar
- remotesensing-09-00043Diunggah olehLeandro Gregorio
- Electricity Homework SolutionsDiunggah olehDascaliuc Daniel
- Bis(Pinacolato)DiboronDiunggah olehYogesh Chandrasekaran
- 151 0548 FS2017 K3 Introduction to Polymer MaterialsDiunggah olehLu Hkarr
- Akışkanlar Mekaniği Kitap ÇözümleriDiunggah olehsefacancakmak
- Kinemtaic-2DDiunggah olehRezaCfc
- 1 Sonometer and StringsDiunggah olehrkvenkatgmail
- ME211.F2014.PbSet2 (1)_2Diunggah olehKhan AB
- The Densitiy of Cement Phases WorksheetsDiunggah olehHazem Diab
- Cold bending of laminated glass panels.pdfDiunggah olehJanardhan Reddy
- RADAR FLASHLIGHT for Throught the Wall Detection of HumansDiunggah olehbingwazzup
- 176.Ibm - Thinkpad x30, x31, x32Diunggah olehSoporte Tecnico Buenos Aires
- F648.1889433-1Diunggah olehKandido Aca
- Fly Ash Usage at Marine Structures to Resist Chloride and Sulfate AttacksDiunggah olehSuresh Rao
- Dyeing Efficiency of Polyester containing BlendsDiunggah olehAditya Shrivastava
- CE130_Oct 11Diunggah olehDevin Brar
- Development of Spin Coating System Based on AC Universal Motor for Deposition of Polymer FilmsDiunggah olehArie Amri Marta
- Program Explain at Ion EngDiunggah olehapi-3822608
- Alex Schiffer - Joe Cell - Experimenters Guide to the Joe CellDiunggah olehAnonymous UwXe23x