2010 Structures Congress © 2010 ASCE 3143

Local Damage Detection in Beam-Column Connections Using a Dense Sensor

Network

Elizabeth L. Labuz1, Minwoo Chang2, and Shamim N. Pakzad3

1
Graduate Student, Department of Civil and Environmental Engineering, Lehigh
University, 117 ATLSS Drive, Bethlehem, PA 18015; PH (610) 758-4543; FAX
(610) 758-5553; email: [email protected]
Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

2
Graduate Student, Department of Civil and Environmental Engineering, Lehigh
University, 117 ATLSS Drive, Bethlehem, PA 18015; PH (610) 758-6273; FAX
(610) 758-5553; email: [email protected]
3
P.C. Rossin Assistant Professor, Department of Civil and Environmental
Engineering, Lehigh University, 117 ATLSS Drive, Bethlehem, PA 18015; PH (610)
758-6978; FAX (610) 758-5553; email: [email protected]

ABSTRACT

Damage prognosis for structural health monitoring is a challenging and complex

research topic in civil engineering. Critical components of damage detection are
identifying the location and severity of damage in a structure, as well as its global
effect on the structure. Local damage can increase over time and have additional
adverse effects on the entire structure. Traditional damage detection methods using
sensor data are effective in recognizing the change in global properties of a structure.
However, these methods are neither effective nor sensitive in identifying local
damage. The use of dense clustered sensor networks provides promising applications
in analysis of structural components and identifying local damage. In this study, a
prototype beam-column connection was constructed and instrumented by a dense
sensor network. The column ends of the test specimen have fixed connections, and
the beam cantilevers from the centerline of the column. The beam was excited with
an actuator at its free end, and accelerometer sensors measured the response of the
members to dynamic excitations at several locations along the specimen. The
response at each sensor location was compared to that of other locations and pair-
wise influence coefficients were estimated. Damage is introduced to the system by
replacing a portion of the beam element with a smaller section, and thus reducing its
stiffness. New influence coefficients were calculated and compared to the undamaged
values. By statistically comparing the change in influence coefficients, the damage is
accurately and effectively identified.

INTRODUCTION

Structural health monitoring (SHM) plays an integral role in maintaining the integrity
of important civil, mechanical, and aerospace engineering systems. Structures
experience a number of dynamic influences on a daily basis ranging from typical
ambient vibrations to more extreme wind and earthquake loadings. Whether the
damaging effects of these load cases are visible immediately or appear more
gradually in time, it is important to be able to detect the damage before it becomes too

Structures Congress 2010

 2010 Structures Congress © 2010 ASCE 3144

detrimental to the entire structure and its surroundings. Not only will money be saved
in being able to repair more manageable damages at earlier detection, but also fatal
catastrophes can be prevented. As a number of structures are beginning to reach the
limit of their service lives, the need for effective, efficient, and affordable damage
detection methods is becoming more and more apparent.
Traditional non-destructive evaluation (NDE) techniques include but are not
limited to visual inspection, liquid penetrant, magnetic particle, radiography, eddy
currents, ultrasonic waves, acoustic emission, and infrared thermography (Trimm
Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

2007). While these methods can be useful in certain circumstances, they require a
priori knowledge of the location of damage and are also subject to the trained eye of
the inspector. Also, in order to implement these techniques one must have direct
access to the location of damage, which may be difficult to reach. Furthermore, NDE
techniques are be costly, difficult to use with complex equipment, and provide only a
temporary means of SHM.
Both advancements in sensor technologies as well as improved understanding
of the modal properties of structures have given rise to vibration-based damage
identification methods. These methods find a basis in the vibration behavior of a
structure, namely the connection between the modal parameters of a structural
system—natural frequencies, mode shapes, and modal damping—and its physical
properties—mass, stiffness, and damping (Doebling et al 1998; Alvandi and Cremona
2006). Parametric changes signify changes in the physical properties, i.e. structural
damage in the form of mass or stiffness loss. Also, the advancements of wireless
networks make it possible to implement these vibration-based methods on a semi-
permanent basis for continual monitoring of structures.
However, current SHM practices involving global-based damage detection
require knowledge of specific structural properties, including mass, stiffness, or
damping ratio, for which it is often difficult to determine correct values (Koh et al
1995; Morassi and Rovere 1997; Sohn and Law 1997; and Ratcliffe 1997).
Additionally, global detection techniques, which are based on global properties, are
not sensitive to local damage and, therefore, cannot identify damage or determine its
locations. Other proposed local damage detection methods, for example the damage
locating vector (DLV) method (Bernal 2002), also require the knowledge of structural
properties, or require homogeneity of the structural properties as in the two-
dimensional gapped smoothing method (Yoon et al 2005).
This paper presents an effective damage detection method that uses vibration
responses collected via densely clustered sensors to achieve localized damage
detection without the need for exact knowledge of structural properties. Influence
coefficients, obtained from linear regression between every two node responses, are
used as the index for determining the existence of damage. The accuracy of these
coefficients is confirmed by considering both estimation accuracy and normalized
estimation error. The change point of time-variant influence coefficients can also be
determined using a Bayesian statistical framework.
The effectiveness of influence coefficient as a damage indicator is
demonstrated through simulated examples as well as laboratory experiments. An
OpenSees model was implemented to demonstrate the use of this algorithm as applied
to a simple beam-column connection, simulating a local joint, subjected to damage.

Structures Congress 2010

 2010 Structures Congress © 2010 ASCE 3145

Laboratory experiments were performed on a prototype of a similar beam-column

connection with unknown structural properties subjected to different loading profiles.
Both the simulated and experimental results, combined with the investigation of
estimation accuracy and normalized estimation error, support the efficacy and
reliability of this proposed method in detecting damage and its location on a local
scale.

METHODOLOGY
Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

Influence Coefficients and Normalized Estimation Error. The algorithm

employed in this paper, which is outlined in Figure 1, utilizes regression analysis to
determine the pair-wise relationship between node responses in order to predict the
location of structural damage. By calculating influence coefficients, αij, between two
nodes i and j, based on vibration-induced acceleration response data, one can
determine the correlation between these responses as follows:

The comparison of the resulting influence coefficients from the initial

undamaged state with that of the damaged state of the structure serves as a ―damage
indicator‖ when it yields a significant change in the value of the coefficients from
state to state. More specifically, the influence coefficients exhibit a much more
significant change when nodes i and j are located on opposing sides of the damaged
segment versus when they are on the same side. This characteristic allows for the
identification of the damage location by comparing which influence coefficients
exhibit significant changes.
Once the coefficients have been calculated the accuracy of the data must be
assessed and verified before damage detection can be performed. This is done
through consideration of both the accuracy of the pair-wise coefficients and the
estimation error. The product of influence coefficients αij and αji, yields the evaluation
accuracy, EAij, of these coefficients, indicating which node responses are linearly
related to one another with the least amount of error, εij, and thus are more accurate
predictors. An evaluation accuracy of 1.0 signifies a strong accuracy of estimation,
while a product of less than 1.0 corresponds to progressively higher values of noise
and nonlinear behavior of the physical structure.
The second parameter that is used for data verification is normalized
estimation error, which is calculated by

Normalized estimation error allows for a direct comparison of the amount of

error associated with the estimation of each influence coefficient as a damage
indicator. This parameter is used to determine which influence coefficients should be
used for damage detection. A low estimation error will correspond to a more accurate
predictor. Once the accuracy and error have been assessed for each coefficient, post-
processing of the data can be performed for damage identification and localization.

Structures Congress 2010

 2010 Structures Congress © 2010 ASCE 3146

Statistical Framework. The final portion of this damage detection algorithm applies
a statistical framework. A Bayesian statistic is used to determine the change point
(Chen and Gupta 2000), the point at which the data indicates damage, at 95%
confidence level. This statistical inference method tests the hypothesis,

against the one-sided alternative hypothesis,

Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

The variables r, representing the change point at which the parameter α may have
changed, mean μ, and standard deviation σ are all unknown. N represents the number
of tests. The Bayesian statistic that is used is
.
Because the standard deviation is unknown, it is estimated as the standard error, .
The statistic that is used to test the aforementioned hypothesis is

and has a t-distribution with N-2 degrees of freedom. The confidence level to which
this hypothesis is tested can be designated as desired. The examples presented in this
paper assume the commonly used 95% confidence level. The physical significance of
this hypothesis test is such that the alternative hypothesis, HA, indicates that the
structure has incurred damage, while the null hypothesis, H0, means that there is not
adequate evidence to establish that damage exists. These hypotheses are tested for
those node pairs that have been identified as significant damage indicators in the
assessment and verification stage of the method.

Figure 1. Flowchart of Methodology.

Structures Congress 2010

 2010 Structures Congress © 2010 ASCE 3147

SIMULATED EXAMPLE

The algorithm was first validated using a simulated model of a beam-column

connection, representing a local joint. The simulation was created using OpenSees
software (Mazzonih et al. 2000). The model has two members (representing the
portion of the beam and column close to the connection) of uniform, hollow, square
cross-section with 3.05 mm thickness, 272.5 mm2 area, and a moment of inertia of
23,114 mm4. The steel is assumed to have an elastic modulus of 200 GPa. The
Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

structure is assumed to be massless, as one assumption of the method is that the mass
has negligible effect on the performance of the algorithm. The boundary conditions of
the column are fixed on either end, while the beam cantilevers from the centerline of
the column. Two simulations were performed including an undamaged condition, the
properties of which have been described, and a damaged condition. Damage was
simulated by reducing the beam stiffness by 40%. The damaged properties of the
beam include a wall thickness of 1.52 mm, a cross-sectional area of 145.5 mm2 to a
moment of inertia of 13,885 mm4. For each of these models, displacement data was
simulated for a white noise excitation applied at the free end of the beam, and
responses were collected for each of the nine node locations designated along the
beam and column as shown in Figure 2. Measurement noise was also accounted for
by adding a Gaussian noise with a standard deviation equal to 5% of the root mean
square of each response signal.

fixed

yk

actuator

fixed

: location of accelerometer
: location of damage
Figure 2. Simulated model of beam-column connection with nine node locations.

Structures Congress 2010

 2010 Structures Congress © 2010 ASCE 3148

The algorithm was then applied to simulated data and the parameters were
extracted. The relative change in the influence coefficients between the undamaged
and damaged states is shown for each pair-wise node relationship in Table 1. The
influence coefficients all experience very small (less than 1%)
changes between the undamaged and damaged states. This implies that the physical
properties between these nodes have not changed. However, the coefficients of nodes
1 through 6 paired with nodes 7, 8, and 9 show relative changes of between 30-40%.
Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

When nodes are on opposite sides of the damage, i.e. nodes 1 through 6 are located
on the undamaged column, while nodes 7, 8 and 9 are located on the damaged beam,
the physical properties between the paired nodes changes. This physical change is
reflected in a more significant relative change in the value of influence coefficients.
Furthermore, the influence coefficients also experience a noticeable
change in coefficients (about 3-10%). This signifies that the physical properties of the
structure between α78, α79, and α89 have changed. Therefore, damage must exist
between these nodes. This is consistent with the simulated damage: a 40% stiffness
reduction of the beam.

Table 1. Relative change in influence coefficients, αij, for simulated structure.

1 2 3 4 5 6 7 8 9
1 0.24% 0.12% 0.03% 0.12% 0.09% 30.6% 39.5% 43.3%
2 -0.23% -0.10% -0.19% -0.12% -0.14% 30.3% 39.2% 43.0%
3 -0.12% 0.13% -0.06% -0.01% -0.04% 30.5% 39.4% 43.1%
4 -0.06% 0.19% 0.08% 0.07% 0.03% 30.6% 39.5% 43.2%
5 -0.11% 0.12% -0.01% -0.07% -0.03% 30.5% 39.4% 43.1%
6 -0.07% 0.16% 0.04% -0.04% 0.03% 30.5% 39.4% 43.2%
7 -23.4% -23.3% -23.4% -23.4% -23.4% -23.4% 6.81% 9.70%
8 -28.3% -28.2% -28.3% -28.3% -28.2% -28.3% -6.38% 2.71%
9 -30.2% -30.1% -30.1% -30.2% -30.1% -30.2% -8.85% -2.63%

EXPERIMENTAL SETUP

The algorithm was further verified through laboratory experiments on a specimen

similar to that of the simulated model. The connection was constructed such that the
two ends of the column have fixed supports, while the beam cantilevers from the
centerline of the column. Again, the prototype represents a portion of the beam and
column members as they come to a local joint. There were nine accelerometer sensor
locations on the specimen, as shown in Figure 3. The free end of the cantilever was
attached to an actuator and excited by harmonic force at various dictated
frequencies—5 Hz, 10 Hz, 15 Hz, and 20 Hz—and the acceleration responses were
collected at each of the nine sensor nodes. The influence coefficients were then
computed. This portion of the experiment served to establish a base response for the
undamaged structure.
For the second portion of the experiment, the beam member was replaced by a
member of 40% reduced wall thickness in order to simulate damage to the structure.
The excitations were repeated, and the damaged state influence coefficients were

Structures Congress 2010

 2010 Structures Congress © 2010 ASCE 3149

computed. The resulting data was then compared between the damaged and
undamaged states to verify the detection of damage.
Data was collected at a 200 Hz sampling rate and 0.005 sec time step, with
each test lasting 10 seconds total. Both the undamaged and damaged structures were
tested 50 times, for a total of 100 tests, each containing 2000 data points per sensor
location. These 1.8 million data points were then processed through the algorithm to
detect the occurrence of damage.
Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

Figure 3. Experimental beam-column prototype.

DISCUSSION OF RESULTS

The experiments were conducted with the undamaged and damaged states at four
different forcing frequencies: 5 Hz, 10 Hz, 15 Hz, and 20 Hz harmonic vibrations.
While the algorithm is applicable for data resulting from all four frequencies, the 15
Hz data exhibited the least amount of actuator noise error, and thus is presented
hereafter.

Accuracy Assessment and Verification. Once the 72 influence coefficients have

been calculated from the acceleration data, the data must be assessed to identify the
most significant indicators, which will be used for damage detection. This involves
two parameters: evaluation accuracy, EA, and normalized estimation error, γ. By
inspection of these parameters, eight different trends can be identified in the
undamaged and damaged parameters, with lower estimation errors coinciding with
greater evaluation accuracies and vice versa. These trends have been designated as
eight different regions, whose mean values are presented in Table 2. Region 1 in the
table corresponds to the least estimation error and highest accuracy, and region 8
corresponds to the greatest estimation error and lowest accuracy. Parameters in region
1, consisting of , are the most accurate and have the least error. This
is a reasonable outcome as nodes 7, 8, and 9 are all located in a row along the same

Structures Congress 2010

 2010 Structures Congress © 2010 ASCE 3150

structural element (the beam). Therefore, the accelerations at node 7, for example,
will be more highly correlated to those at node 8 or 9.
On the contrary, region 8, which consists of parameter α16, exhibits the
poorest accuracy and the greatest estimation error. This can be accounted for by the
fact that each of these nodes is located at either end of the column near the fixed
connections. These boundary conditions create large additional noise, which
contributes to the fact that α12 or α56 exhibit lower accuracy than α78 or α89 despite the
similar configuration of the nodes with respect to one another. Figures 4(i) and (ii)
Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

show an example of α, EA, and γ results from regions 1 and 8, respectively. Figure
4(i) shows that the EA is almost equal to unity and the γ is almost equal to 0 for
region 1, while Figure 4(ii) shows a much lower EA and a noticeably higher γ for
region 8. Based on similar data for all 8 regions, it can be concluded that regions 1
through 3 contain useful damage indicators. On average, these influence coefficients
exhibit accuracy greater than 95% and estimation error less than .0044.

Table 2. Average Estimation Error (γ) and Evaluation Accuracy (EA).

Region Influence Coefficients γij Average EAij Average

1 α78, α79, and α89 0.0002 0.9999
2 α23 and α45 0.0015 0.9947
α27, α28, α29, α37, α38, α39, α47, α48, α49, α57,
3 0.0044 0.9554
α58, and α59
4 α12, α13, α46, and α56 0.0057 0.9377
5 α17, α18, α19, α67, α68, and α69 0.0074 0.8997
6 α24, α25, α34, and α35 0.0093 0.8379
7 α14, α15, α26, and α36 0.0114 0.7903
8 α16 0.0161 0.6626

15Hz Sine wave, 78

15Hz Sine wave,
3.2 -0.6
16
16
78

3 -0.8

2.8 -1
1 1
0.9 0.9
EA16
EA78

0.8 0.8
0.7
0.6
EA78 ≈ 1.0 0.7
0.6
0.5 0.5
γ78 ≈ 0 0.03 0.03
78

16

0.02 0.02
0.01 0.01
0 0
0 20 40 60 80 100 0 20 40 60 80 100
Run Number Run Number
(i) (ii)
Figure 4. Comparison of α, EA, and γ Results: (i) Region 1 and (ii) Region 8.

Post-Processing and Damage Detection. The results of the relative changes in the
influence coefficients, αij, from the 15 Hz harmonic excitation of the undamaged and
damaged tests are shown in Table 3. Observing these coefficient changes is the first
step in post-processing the data. The changes associated with α78, α79, and α89, which
are 8%, 12%, and 4% respectively, are indicators of a property change between nodes

Structures Congress 2010

 2010 Structures Congress © 2010 ASCE 3151

7 and 9. This is consistent with the damage, or stiffness reduction, that was created
along the full length of the beam portion of the test structure.
The coefficients of region 2, α23 and α45, experienced 3% and 8% changes.
This is less consistent with what would be expected in comparison to the very low
(less than 1%) changes that were seen in the simulated results. One major cause of
this is noise, which is difficult to control in an experimental setting, but even more so
in a real structure.
Influence coefficients from region 3 were more in line with the prediction of
Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

the simulated results, with the exception of α49 (2%) and α57 (3%), which were low
compared to the 5-40% range of the others. Again, these two discrepancies are likely
accountable to noise. However, the majority of the coefficients in this region showed
noticeable fluctuations from the undamaged state to the damaged state. These
changes, along with the model results, show that nodes on opposite sides of the
damage location experience the largest changes. This is because as damage occurs,
the relationship of two points in the actual physical structure deviates slightly from
linear behavior. Again, the changes in coefficients point to the location of damage in
the structure (i.e. damage between nodes 2 and 7, 3 and 8, and so forth).
While the results for the first three regions were mostly consistent with the
expectations set out by the simulated structure, the remaining regions were not as
consistent. A prime example is region 6, consisting of parameters α24, α25, α34, and
α35. According to the model, these coefficients, whose nodes are all located on the
same side of the damage, should experience very little fluctuation from the
undamaged to damaged states. However, their experimental changes range from 12-
20%. Recall that in the previous verification stage region 6 showed the third lowest
accuracy and the third highest error.

Table 3. Relative change in influence coefficients, αij, for experimental structure.

1 2 3 4 5 6 7 8 9
1 2.3% 5.6% 17.9% 12.1% 11.2% 19.7% 29.5% 34.6%
2 0.5% 2.9% 19.8% 14.6% 16.4% 22.4% 32.5% 37.7%
3 1.3% 2.2% 16.4% 12.4% 14.5% 22.7% 32.9% 38.1%
4 36.8% 36.9% 41.8% 8.4% 7.0% 12.7% 5.5% 1.8%
5 28.7% 28.5% 33.4% 8.6% 1.8% 3.0% 5.0% 9.0%
6 32.3% 30.4% 35.0% 5.6% 2.5% 5.4% 2.4% 6.4%
7 22.7% 22.3% 26.0% 5.2% 1.5% 0.4% 8.3% 12.5%
8 28.6% 28.3% 31.7% 2.8% 9.0% 7.3% 7.6% 3.9%
9 31.2% 30.9% 34.2% 6.5% 12.4% 10.8% 11.1% 3.8%

A further comparison of the simulated and experimental results for select

coefficients is show in Figure 6. From this schematic it is evident that the results of
the experiments were not as symmetric as the simulation suggested. Specifically, α25
and α58 were inconsistent between the simulated and experimental cases. For
example, the experimental α25, which belongs to region 6, suggests that damage exists
between the nodes 2 and 5, which is not the case. Thus, this example, in conjunction
with the results in Table 3, shows the importance of using parameters EA and γ to
identify accurate damage indicators.

Structures Congress 2010

 2010 Structures Congress © 2010 ASCE 3152
Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

Figure 3. Comparison of relative change of coefficients between simulated and

experimental structures.
As is evident, the difference between undamaged and damaged structures can
be indicative of the existence and location of the damage. However, in an actual
structure, one will not necessarily know when damage has occurred, or if in fact
damage has occurred at all. Therefore, another element must be added for complete
damage detection: the statistical framework.
The hypothesis testing graphically shows the change point of damage, the
point at which damage is identified at a certain confidence level, by plotting the test
statistic against the test run number. A graph in which the data crosses the confidence
bounds, either positive or negative, corresponds to a positive hypothesis, previously
defined as HA, indicating the detection of damage. If the accuracy and estimation
error associated with the nodes being considered are high and low respectively, the
prediction of the hypothesis test will be more exact and will cross the confidence
bounds closer to the occurrence of damage. In order to demonstrate this behavior, the
test statistic from the 50 damaged state tests were plotted against their run number.
Because damage exists for all of the plotted data, the most accurate damage indicators
will yield a plot in which the confidence bounds are crossed closest to the occurrence
of damage. This can be seen in comparing Figures 6(i) and (ii), the hypothesis test
results for a region 1 coefficient, an accurate indicator, versus that of a region 8
coefficient, a poor indicator. The plot in Figure 6(i) crosses the 95% confidence
bound much sooner than that of Figure 6(ii). Furthermore, these plots also
demonstrate that damage is detected by hypothesis testing, although less precisely, for
even the poorest damage indicating coefficient. Therefore, this method is a reliable
means of damage detection.

10

Structures Congress 2010

 2010 Structures Congress © 2010 ASCE 3153

Bayesian Test for Bayesian Test for

78 16
10 5
t Statistic

t Statistic
5
0
0
Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

-5 -5
0 10 20 30 40 50 0 10 20 30 40 50
Run Number After Damage Run Number After Damage
(i) (ii)
Figure 6. Comparison of Hypothesis Test Results: (i) Region 1 and (ii) Region 8.

CONCLUSION

In this paper, a densely clustered sensor network was successfully implemented for
local damage detection of both a simulated model as well as an experimental
prototype of a steel beam-column connection. Linear regression was used to estimate
influence coefficients from vibration-induced acceleration responses of the structure.
Two parameters, evaluation accuracy and normalized estimation error, demonstrated
that certain influence coefficients are more reliable damage indicators than others.
The influence parameters change when the structural properties change due to
damage, such as a reduction in stiffness or mass. Therefore, this change point was
used to detect and locate this damage. This method was verified through its
application to a simulated model and an experimental prototype, both of which
represented the local joint of a steel beam-column connection with nine sensor nodes.
Damage equivalent to a 40% reduction in stiffness was created in a portion of the
beam close to the joint for comparison to the initial state of the structures. The
simulated model was excited by white noise, while the experimental model was
excited by harmonic force at various forcing frequencies. The experimental
acceleration data was then compared to the simulated response data. Although the
experimental structure exhibited more noise error than the simulated results, both the
simulation and the experiments identified the same influence coefficients (parameters
in regions 1, 2, and 3) as the most reliable damage indicators. By statistically
comparing these influence coefficients, damage was accurately and effectively
diagnosed to a 95% confidence bound.

ACKNOWLEDGMENTS

The research described in this paper is supported by the National Science Foundation
through Grant No. CMMI-0926898 by Sensors and Sensing Systems program. The
authors thank Dr. Shih-Chi Liu for his support and encouragement.

REFERENCES

Alvandi, A., and Cremona, C. ―Assessment of vibration-based damage.‖ Journal of

Sound and Vibration 292 (2006) 179–202.

11

Structures Congress 2010

 2010 Structures Congress © 2010 ASCE 3154

Bernal, D., M.ASCE. ―Load Vectors for Damage Localization.‖ Journal of

Engineering Mechanics (Jan 2002): 7-14.
Chen, J., and Gupta, A.K. (2000). Parametric change point analysis. Birkhäuser,
Boston.
Doebling, Scott W., Farrar, Charles R., and Prime, Michael B. ―A Summary Review
of Vibration-Based Damage Identification Methods.‖ The Shock and Vibration
Digest 30 (Mar 1998): 91-105.
Koh, C.G., See, L.M., and Balendra, T. ―Damage Detection of Buildings: Numerical
Downloaded from ascelibrary.org by Syracuse University Library on 05/20/13. Copyright ASCE. For personal use only; all rights reserved.

and Experimental Studies.‖ J. Struct. Engrg. 121(8), (Aug 1995): 1155-1160.

Mazzonih, S., McKenna, F., Scott, M.H., and Fenves, G.L. (May 2009). OpenSees
Command Language Manual. <http://opensees.berkeley.edu/OpenSees/
manuals/usermanual/index.html>.
Morassi, A., and Rovere, N. ―Localizing a Notch in a Steel Frame from Frequency
Measurements.‖ J. Engrg. Mech. 123(5) (May 1997): 422-432.
Ratcliffe, C.P. ―Damage Detection Using a Modified Laplacian Operator on Mode
Shape Data.‖ Journal of Sound and Vibration 204(3) (1997): 505-517.
Sohn, H., and Law, K.H. ―A Bayesian Probabilistic Approach for Structure Damage
Detection.‖ Earthquake Engineering and Structural Dynamics 26 (1997): 1259-
1281.
Trimm, Marvin. ―An overview of nondestructive evaluation methods.‖ Journal of
Failure Analysis and Prevention 3(3) (2003): 17-31.
Yoon, M.K., Heider, D., Gillespie Jr., J.W., et al. ―Local damage detection using the
two-dimensional gapped smoothing method.‖ Journal of Sound and Vibration
279 (2005): 119-139.

12

Structures Congress 2010