Composites Science and Technology 88 (2013) 62–68

Contents lists available at ScienceDirect

Composites Science and Technology

journal homepage: www.elsevier.com/locate/compscitech

Neural network approach for damaged area location prediction

of a composite plate using electromechanical impedance technique
S. Na, H.K. Lee ⇑
Department of Civil and Environmental Engineering, Korea Advanced Institute of Science and Technology (KAIST), 373-1 Guseong-dong, Yuseong-gu, Daejeon 305-701, South Korea

a r t i c l e i n f o a b s t r a c t

Article history: Nowadays, breakthrough composite technologies are intensifying the complexity of structural compo-
Received 12 April 2013 nents every day and assuring the structural integrity is becoming more essential, thus creating challenges
Received in revised form 22 July 2013 for developing a cost effective and reliable non-destructive evaluation (NDE) technique. As conventional
Accepted 17 August 2013
NDE techniques usually require expensive equipments, trained experts and out-of-service period, such
Available online 30 August 2013
techniques may be inadequate for autonomous online health monitoring of structures. In this study, a
relatively new technique known as electromechanical impedance (EMI) technique is combined with a
Keywords:
neural network technique to predict the damaged areas on a composite plate. Regardless of the advanta-
A. Smart materials
A. Glass fibers
ges such as low cost, robustness, simplicity and online possibilities, this technique still has various prob-
C. Probabilistic methods lems to be solved. For one, locating a damaged area can be extremely difficult as this non-model based
D. Non-destructive testing technique heavily relies on the variations in the impedance signatures. The results show that the
D. Ultrasonics non-homogenous property is an advantage for the study, successfully identifying the damage location
for the prepared test specimen with an acceptable performance.
Ó 2013 Elsevier Ltd. All rights reserved.

1. Introduction the manual inspection methods as this can increase the overall
maintenance cost and the down time of the structure in service.
With the advancement in composite technologies, the complex- However, the EMI technique still has various unsolved problems
ity of structural components is increasing and independently including temperature effect, damage type identification, and dam-
assuring the structural integrity is becoming absolutely vital. Such age location prediction, limiting the technique to be used for
advancements create challenges for developing a cost effective and practical applications [1]. Especially, predicting the damaged area
reliable non-destructive evaluation (NDE) technique as traditional using the EMI technique is known to be extremely difficult as this
methods such as acoustic emission, X-ray, optical, laser-optical, technique relies on the changes in the impedance signatures which
radiography and other various techniques require expensive can be unpredictable at most times.
equipments and experts, often making it difficult for practical The 1-D model for the EMI technique shown in Eq. (1) was first
applications [1–3]. proposed by Liang et al. [10] where it shows the coupled relation-
On the other hand, a relatively new non-destructive evaluation ship between the electrical and mechanical impedance of the lead
technique known as the electromechanical impedance (EMI) tech- zirconate titanate (PZT) element and the structure, respectively.
nique is well known for its robustness and acceptable performance  
[4–9]. This non-model based technique uses a single piezoelectric Z s ðxÞ 2
YðxÞ ¼ ixa eT33 ð1  idÞ  d YE ð1Þ
material to act as an actuator and a sensor simultaneously, making Z s ðxÞ þ Z a ðxÞ 3x xx
it suitable for complex structures where damage identification is
Here, the equation proves that the electrical admittance (in-
achieved by monitoring the variations in the impedance signa-
verse of impedance), Y(x) of the PZT element is related to the
tures. In addition, the commercialized AD5933 evaluation board
mechanical impedance of the structure, Zs and PZT element, Za
from Analog Devices offers the ability to measure impedance with 2
[10]. The remaining variables x, a, eT33 , d, d3x and Y Exx represent
the current low cost of US$59, minimizing the equipment cost for
the excitation frequency geometric constant, dielectric constant,
conducting the EMI technique. Furthermore, the online monitoring
dielectric loss tangent, coupling constant and the complex Young’s
possibility of this technique makes it a powerful tool to be used for
modulus of the PZT element, respectively [10]. This equation
structural health monitoring of structures, eliminating the need for
proves that any changes in the structure can be identified by mon-
itoring the changes in the impedance of the PZT element [10].
⇑ Corresponding author. Tel.: +82 42 350 3623; fax: +82 42 350 3610. To date, damage detection of structures using Artificial Neural
E-mail address: leeh@kaist.ac.kr (H.K. Lee). Networks (ANN) has been vigorously researched as they have

0266-3538/$ - see front matter Ó 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.compscitech.2013.08.019
 S. Na, H.K. Lee / Composites Science and Technology 88 (2013) 62–68 63

emerged as a promising tool for identifying damages due to their as it has been experimentally proven to perform better compared
pattern recognition and interpolation capabilities [11,12]. Among to the imaginary part for detecting structural changes [18].
them, various researches have investigated using ANN in conjunc- X h i2 .XN 1=2
N
tion with the EMI technique in order to identify and classify dam- RMSD ¼ k¼1
ReðZ k Þj  ReðZ k Þi k¼1
½ReðZ k Þi 2 ð2Þ
age. Min et al. [13] used ANN combined with the EMI technique to
identify the damage type between a notch and loose bolt on a bolt- Also shown in Fig. 1 is the repeatability performance of the
joined aluminum beam and a lab-scaled pipe structure. In addition, AD5933 evaluation board. After 10 consecutive measurements of
the severity of each damage type was also investigated. Lopes et al. a test specimen, the maximum RMSD value between two signa-
[14] performed EMI technique on a 1/4 scale model of a steel tures was 0.06%. Since the difference is very small, possible varia-
bridge section and a space truss structure with ANN to assess the tion associated with repetitive measurements is ignored for this
state of the structures subjected to damage by loosening bolts. study.
Moura et al. [15] used ANN to classify the type of damage between
a hole and a crack of helicopter blades with the impedance based
3. Application of the damage enhancement technique
technique resulting in a promising outcome.
In this study, an attempt to locate damaged areas is performed
3.1. Implantation of a resonance frequency range
on a composite plate using the EMI technique. Probabilistic Neural
Network (PNN) is used with the idea of the damage enhancement
One of the major problems when using the EMI technique on a
technique for the EMI technique in [4] to locate the damaged area
non-homogenous material, such as composite or concrete struc-
on a composite plate. The damage enhancement technique was
tures is the difficulty in damage identification due to the lack of
previously developed by the authors to significantly enhance the
change in the impedance signatures [4]. In the previous work [4],
damage detection ability of the EMI technique on composite struc-
the authors enhanced the damage detection sensitivity by implant-
tures where the EMI technique failed to identify damage due to the
ing a resonance frequency range for the EMI technique to success-
absence of resonance peaks [4].
fully achieve damage identification [4]. This was achieved by
attaching a 10 mm  10 mm  0.508 mm sized PZT material on
2. Impedance measuring system one side of a circular metal piece having a 25 mm diameter and
a 2 mm thickness. The other side of the metal piece was then
The experimental setup is shown in Fig. 1. A commercialized partially attached onto the host structure (see Fig. 1) [4]. The
AD5933 evaluation board manufactured from Analog Devices Co. damage enhancement technique proposed in [4] was tested
is connected to the laptop where the evaluation software provided against de-bonding and artificial cut cases to prove its ability to de-
by the manufacturer allows one to measure the impedance signa- tect damage where the conventional method of attaching the PZT
tures up to 100 kHz [16]. The evaluation board is fully powered by material failed to achieve [4]. In this study, this idea is used for
the data cable alone, thus resulting in a portable system for mea- locating damaged areas of a glass fiber–epoxy composite plate
suring the impedance of a structure. The PZT material (model manufactured from Hankuk Carbon Co. where the properties are
PSI-5A4E) manufactured by Piezo Systems Inc. of a 0.508 mm as follows: heat resistance temperature of 134 °C, moisture content
thickness is used for this study where the positive and negative of 0.06%, tensile strength of 454 MPa, flexural strength of 602 MPa
wires from the evaluation board are connected to the PZT material and compressive strength of 385 MPa [19].
that is attached to the test specimen [17].
Once the impedance signature is retrieved, one of the statistical 3.2. Impedance signature observation of a metallic structure
techniques known as the root mean square deviation (RMSD) is
used to quantify the intensity of the damage [18]. The RMSD equa- Since the damage enhancement technique is designed for a
tion is shown below where (Zk)i and (Zk)j represents the reference non-homogenous material which usually results in a peakless sig-
signature of the PZT impedance and the corresponding impedance nature, this technique is inapplicable for homogenous materials
for each measurement time at the kth measurement point, respec- such as a metal plate that results in multiple peaks [4]. However,
tively [18]. In general, the real part of the impedance is used here the idea of finding the damaged location using the EMI technique
on a metal plate can be difficult due to too many peaks. Fig. 2
shows the impedance signature changes subjected to damage on
a metal plate of a size 250 mm  70 mm  0.3 mm. Two areas
are damaged on the metal plate, one close and one far away from
the PZT patch. Looking at the impedance signatures, there are a

Fig. 1. General setup for measuring impedance of a test specimen. Fig. 2. Impedance signatures for intact and damaged cases for a metal plate.
64 S. Na, H.K. Lee / Composites Science and Technology 88 (2013) 62–68

large number of peaks within the selected frequency range of ððxx0 Þ2 þðyy0 Þ2 þðzz0 Þ2 Þ
23–33 kHz and such property is desired when using the EMI tech- Z ¼ f ðx; y; zÞ ¼ e 3r2 ð3Þ
nique. Although impedance signatures as such can be very sensi- Here, x0, y0 and z0 are the trained data which are the RMSD
tive to damages, it can be difficult when one conducts the EMI values from the three PZT sensors attached on the composite plate
technique to specify the location of the damage. Referring to the (shown later). x, y and z are the input parameters which are ac-
above mentioned figure, when the RMSD values are calculated quired from the three PZT sensors. r is the constant that controls
for the two impedance signatures, close and far distanced damage, the width of the function that requires an educated guess based
the RMSD values are calculated to be 12.34% and 13.54% respec- on the knowledge of the data when determining the value. In this
tively. This shows that it is difficult to tell which area is damaged study, r of 0.3 is used.
closest to the PZT patch. In addition, the damage closest to the After the training process, the method is ready for predicting
PZT patch showed a lower RMSD value. This can occur as the damaged areas. When a specific area is damaged on a test speci-
impedance signatures change significantly each time, thus too men, the 3 RMSD values are retrieved (from the three PZT patches,
many peaks in the impedance signature can create an unwanted Sensor X, Sensor Y and Sensor Z) and inputted into the pattern units
outcome, resulting in difficulties when specifying the location of for each category where they are summed to create a summed
the damage. On the other hand, when the damage enhancement category unit. The maximum value is selected to represent which
technique is applied, the resonance frequency is dominated by area has been damaged for the prediction.
the metallic layer which avoids the impedance signatures from
changing significantly, decreasing the possibility of an undesired
outcome. 5. Experimental setup

5.1. Signature changes subjected to damages in different locations

4. Application of probabilistic neural networks
To understand how the impedance signatures change subjected
In this study, probabilistic neural networks (PNN) introduced by to damages at different locations, a glass fiber–epoxy composite
Donald F. Specht (1990) is used to determine the location of the plate of a size 250 mm  70 mm  3.5 mm is prepared with a
damage on a composite plate [20]. Compared to other ANN meth- PZT patch of 20 mm  20 mm  0.5 mm attached near the end of
ods, the training process is much faster, especially compared to the plate with the application of the damage enhancement tech-
backpropagation. In addition, it is guaranteed to converge to an nique using a metallic layer of a size 25 mm diameter and 3 mm
optimal classifier as the size of the representative training set thickness as shown in Fig. 4. The 3 areas (A, B and C) are marked
increases. Another advantage of PNN is that training samples can on the test specimen and 5 holes are randomly drilled at each area
be added or removed without extensive retraining, which suits where the impedance signatures are measured before and after
well with the EMI technique on civil structures as retraining could creating a hole. The holes are created using a hand drill with a
be required due to the change in the structural integrity with time diameter of 6 mm where a total number of 15 holes are drilled
and the surrounding environment, affecting the impedance to investigate the relationship between the changes in the imped-
signatures. ance signatures against damaged area locations. For measuring the
Fig. 3 depicts the classification algorithm for classifying three- impedance, the frequency range of 23–33 kHz was used as reso-
dimensional data with 6 categories. All the categories have 2 units nance was observed in this range.
where each of them represents an area in the composite plate.
Thus, meaning that only 2 training data for each area will be used
5.2. Damage location prediction using PNN algorithm
for this study. Each trained data point corresponds to the units
which is a Gaussian function with a peak centered on the parame-
In this part of the experiment, a single composite plate of a size
ters location where the equation used for this study is shown
200 mm  200 mm  3.5 mm is used and 3 PZT patches are
below [cf. 20].
attached onto the composite plate with the application of the
damage enhancement technique as shown in Fig. 5. The sizes of
the 3 PZT patches and the metallic layers are identical to the setup
mentioned in Section 5.1. As shown in the figure, the 3 PZT patches
are attached to the 3 corners of the composite plate labeled, Sensor
X, Sensor Y and Sensor Z. Each square area is sized
66.67 mm  66.67 mm (one third of 200 mm) where damage is
introduced from Area 1 to Area 6. The impedance signatures for
Sensor X, Sensor Y and Sensor Z are shown in Fig. 6. By observation,
all the signatures have few peaks and the resonance frequency
ranges are within the frequency range of 23–33 kHz. This reason
lies with the metallic layer used for the damage enhancement
technique as the resonance frequency range is dominated by the
metal piece [4].

5.2.1. Training
For each area, a 6 mm diameter hole is drilled with impedance
signatures being measured before and after the drilling to calculate
the RMSD value each time. After drilling a hole, another hole is
drilled on the next area where the impedance signature is
measured to obtain the RMSD value. After drilling a hole into Area
1–6, another set of 6 holes are created, resulting in two holes for
Fig. 3. Probabilistic neural networks algorithm for the study. each area. The RSMD values are shown in Table 1 where the two
 S. Na, H.K. Lee / Composites Science and Technology 88 (2013) 62–68 65

Fig. 4. Composite plate with a single PZT patch attached.

Fig. 5. Composite plate with three PZT patches attached for damage location prediction.

5.2.2. Damage location prediction

In this part of the experiment, a hole is drilled from Area 1–6
and this process is repeated 4 more times to obtain 5 RMSD values
for each of the PZT patches per area. This results in 15 RMSD values
per area, totaling 90 RMSD values for the experiment. Then using
the PNN algorithm, these results are inserted for evaluating the
performance of the PNN method to predict the damage locations
for the 6 areas. In general, damaged areas close to the PZT patches

Table 1
Training data for the three PZT patches.

Sensor X Sensor Y Sensor Z

1st hole 2nd hole 1st hole 2nd hole 1st hole 2nd hole
Area 1 0.405 0.604 0.27 0.326 0.56 0.519
Area 2 0.815 0.9 0.422 0.509 0.391 0.374
Fig. 6. Baseline impedance signatures for the three sensors.
Area 3 0.582 0.455 0.611 0.704 0.728 0.688
Area 4 0.48 0.272 0.305 0.224 0.413 0.246
Area 5 1.019 0.616 0.602 0.784 0.376 0.352
RMSD values for each PZT patch are used as the training data,
Area 6 0.401 0.417 0.734 1.043 1.08 0.737
which are inserted into x0, y0, z0 of the pattern layer.
66 S. Na, H.K. Lee / Composites Science and Technology 88 (2013) 62–68

will result in a higher RMSD values compared to the damaged areas area, the RMSD values for Sensor X varies from 0.409 to 1.214,
further away from the PZT patches. Sensor Y from 0.264 to 0.926 and Sensor Z from 0.304 to 1.138
depending on the damaged area.
Investigating the results for each PZT patch, for Sensor X, the
6. Results and discussions closest Area 2 and Area 5 showed the highest RMSD values of
1.214 and 0.884, respectively. Rest of the areas 1, 3, 4 and 6 showed
6.1. Impedance signature changes of three damaged areas the RMSD values of 0.409, 0.498, 0.459 and 0.528, respectively. For
Sensor Y, the closest Area 5 and Area 6 showed the highest RMSD
For Area A, the 5 RMSD values were 1.217, 1.475, 1.012, 1.07 values of 0.75 and 0.93 respectively. The furthest away area, Area 1
and 1.047 with an average of 1.16. For Area B, the 5 RMSD values showed the lowest RMSD value of 0.26 as expected and the middle
were 2.7, 1.324, 1.595, 2.12 and 2.219 with an average of 1.99. area, Area 4 resulted in 0.34. Area 3 showed a relatively high RMSD
For Area C, the 5 RMSD values were 2.72, 1.74, 3.276, 3.07 and value of 0.72 and Area 2 showed a RMSD value of 0.47. Although
2.008 with an average of 2.56. This shows that the RMSD values that Area 3 and Area 2 have approximately the same distance from
are generally high for the damages done close to the PZT patch Sensor Y, Area 3 has a RMSD value 1.5 times higher than Area 2. To
where the area furthest away from the PZT patch shows lowest understand why such large difference has occurred between the
RMSD values. For each area, the differences in the RMSD values two areas, 2 PZT patches of a size 10 mm  10 mm  0.5 mm were
are clearly observed. The lowest RMSD values for Area A, Area B attached to Area 2 and Area 3 and a function generator (Agilent
and Area C are 1.012, 1.595 and 1.74 and the highest RMSD values 33250A) was connected to Sensor Y. Using the function generator,
are 1.475, 2.70 and 3.276 respectively. Here, the differences be- a standing wave with a peak–peak voltage of 5 V was applied from
tween the lowest and the highest RMSD values are large for Area 23 kHz to 33 kHz and the maximum voltage received by the PZT
A and small for Area C. Such differences in the values are expected patches from Area 2 to Area 3 was retrieved using an oscilloscope
as the holes are drilled randomly at each area. Furthermore, a large (Agilent MSO6034A) in 200 Hz steps. The result is shown in Fig. 8
difference between the RMSD values, such as the RMSD values of where the maximum voltage gain from Area 3 is considerably
1.74 and 3.276 in Area C can possibly be introduced by damaging larger than Area 2. Since piezoelectric materials produce voltage
either a node or an anti-node. When conducting the EMI technique upon mechanical stress, it is safe to assume that the vibration level
on a test specimen, a standing wave is created at each frequency at Area 3 is much larger than Area 2. Larger vibration means that
range, thus if one were to damage an anti-node, the standing wave the amplitudes of the sanding wave are larger, possibly resulting
can dramatically change resulting in a large change in the imped- in a higher RMSD values subjected to damage. For Sensor Z, Area
ance signature, however, if one were to damage a node, the stand- 3 and Area 6 have the highest RMSD values out of the 6 areas. This
ing wave may experience a small change, possibly resulting in a is expected as they are closest to the PZT patch. The middle area,
vague change in the impedance signature. Nevertheless, the aver- Area 4 resulted in the lowest RMSD value of 0.304, and furthest
age of the 5 RMSD values shows clear difference between different away Area 2 and Area 5 showed RMSD values of 0.375 and 0.491
areas for damage location identification. respectively. Area 1 showed the averaged RMSD value of 0.614,
To observe how the impedance signatures change subjected to which is relatively high compared to the rest of the areas.
damage, Fig. 7 is drawn for Area A for the first 5 cumulative dam- When these data are inserted into the PNN algorithm, the
ages on the plate. By observation, the changes in the impedance algorithm selects the area with the highest value as the damaged
signatures maintained its general shape as damage occurred. This area. Table 3 shows the performance of the algorithm where 24
shows that although that the resonance frequency is dominated predictions were successful (out of 30), which is 80% accuracy.
by the metallic layer, the reflected waves from the holes have a de- Except for Area 1 and 4, all the areas have one or two results that
cent impact on the impedance signatures, which agrees well with
the authors previous work [4].

6.2. Performance of the damaged area prediction Table 2

The RMSD results for the 5 sets of damage cases.
The RMSD value results for this part of the experiment are 1st 2nd 3rd 4th 5th Av. Std.
shown in Table 2 with the average and standard deviation values
Sensor X
calculated for analysis. Looking at the averaged values for each Area 1 0.358 0.566 0.348 0.406 0.367 0.409 0.09
Area 2 0.929 0.736 1.608 1.109 1.689 1.214 0.419
Area 3 0.398 0.247 0.805 0.562 0.48 0.498 0.207
Area 4 0.575 0.419 0.453 0.5 0.347 0.459 0.086
Area 5 1.127 0.859 0.858 0.849 0.725 0.884 0.147
Area 6 0.865 0.451 0.508 0.4 0.414 0.528 0.193

Sensor Y
Area 1 0.259 0.33 0.179 0.255 0.296 0.264 0.056
Area 2 0.207 0.554 0.666 0.48 0.424 0.466 0.171
Area 3 0.982 0.756 0.725 0.652 0.495 0.722 0.177
Area 4 0.343 0.355 0.255 0.345 0.409 0.341 0.055
Area 5 0.416 0.654 0.794 0.786 1.175 0.765 0.275
Area 6 1.108 0.673 0.801 1.204 0.845 0.926 0.222

Sensor Z
Area 1 0.644 0.744 0.594 0.587 0.501 0.614 0.089
Area 2 0.366 0.408 0.371 0.408 0.323 0.375 0.035
Area 3 0.839 0.998 0.622 0.616 0.906 0.796 0.171
Area 4 0.293 0.332 0.186 0.372 0.337 0.304 0.072
Area 5 0.779 0.336 0.286 0.638 0.414 0.491 0.21
Area 6 0.829 0.904 1.077 1.04 1.841 1.138 0.406
Fig. 7. Impedance signature changes for the first 5 damages on Area A.
 S. Na, H.K. Lee / Composites Science and Technology 88 (2013) 62–68 67

sults showed that the area close to the PZT patch resulted in higher
RMSD values than areas further away from the PZT patch as ex-
pected. The important observation made here is that when damage
was made in the same area, the differences in the RMSD values
were high, especially for Area C. Such can happen as damaging
an anti-node can have a decent impact on the impedance signa-
tures where damaging a node can have a vague impact on the
impedance signatures, resulting in a smaller RMSD value. This is
an important factor to know as the EMI technique is a non-model
based technique that heavily relies on the variations in the imped-
ance signatures.
The second part of the experiment involved using 3 PZT patches
attached near the ends of each corners of the test specimen for
damage location prediction. Using the first 2 holes from each area
as the training data, 30 holes were drilled into the composite plate
Fig. 8. Voltage gains from the PZT patches at Area 2 and Area 3. and Probabilistic Neural Networks (PNN) algorithm was used to
predict the damage location. The results showed over 80% accuracy
with all the damage locations predicted correctly for Area 1 and
Table 3 Area 4, which had the lowest standard deviation values compared
Results for the damage location prediction using PNN.
to the rest of the areas. From this study, one can assume that the
Set 1 Set 2 Set 3 Set 4 Set 5 accuracy of correctly identifying the damaged location is highly
Area 1 Area 1 Area 1 Area 1 Area 1 Area 1 influenced by the acquired training data. In addition, the manner
Area 2 Area 2 Area 2 Area 5 Area 2 Area 4 in how the areas are divided can also have a decent impact on
Area 3 Area 3 Area 6 Area 3 Area 3 Area 3 the accuracy performance as having larger and less divided areas
Area 4 Area 4 Area 4 Area 4 Area 4 Area 4
for the test specimen would have increased the accuracy perfor-
Area 5 Area 4 Area 5 Area 5 Area 5 Area 5
Area 6 Area 3 Area 6 Area 6 Area 6 Area 3 mance for this study.

Acknowledgement
led to a wrong prediction of the damaged areas. Area 2 has two,
The authors would like to thank the National Research Founda-
Area 3 has one, Area 5 has one and Area 6 has 2 two. However,
tion (2012-008855) and the Ministry of Land, Transport and
the predictions for Area 1 and Area 4 alone was 100% with 5 RMSD
Maritime Affairs (Development of the intelligent green bridge
values and such outcome can be explained with the aid of the stan-
technology) of the Korean government for the support.
dard deviation values shown in Table 2. For all 3 PZT patches, the
standard deviations are 0.09, 0.056, 0.089 and 0.086, 0.055, 0.072
for Area 1 and Area 4 respectively. With such low standard devia- References
tion, the PNN algorithm successfully predicted the location of the
[1] Park G, Sohn H, Farrar CR, Inman DJ. Overview of piezoelectric impedance-
damage. Thus, if the standard deviations were low for all the other based health monitoring and path forward. Shock Vib Digest
areas, the successful chance of correct damage location prediction 2003;35(6):451–63.
may increase. The accuracy of correctly identifying the damaged [2] McCann DM, Forde MC. Review of NDT methods in the assessment of concrete
and masonry structures. NDT & E Int 2001;34(2):71–84.
location can dramatically decrease if an unreliable set of training [3] Chae SR, Moon JH, Yoon SY, Bae SC, Levitz P, Winarski R, et al. Advanced
data is obtained since the accuracy of the results are dependent nanoscale characterization of cement based materials using X-ray synchrotron
on the training data. radiation: a review. Int J Concr Struct Mater 2013;7(2):95–110. http://
dx.doi.org/10.1007/s40069-013-0036-1.
Thus increasing the number of the training data can increase [4] Na S, Lee HK. Resonant frequency range utilized electro-mechanical
the successful prediction rate as only two training data were used impedance method for damage detection performance enhancement on
for this study. Another important factor which affects the accuracy composite structures. Compos Struct 2012;94(8):2383–9. http://dx.doi.org/
10.1016/j.compstruct.2012.02.022.
is the manner in which the areas were divided. If one were to
[5] Park G, Cudney H, Inman DJ. Feasibility of using impedance based damage
divide the test specimen into 3 large areas instead of 6 areas used assessment for pipeline systems. Earthquake Eng Struct Dyn
in this study, the accuracy would likely to increase. But, if the 2001;30(10):1463–74. http://dx.doi.org/10.1002/eqe.72.
[6] Na S, Lee HK. Steel wire electromechanical impedance method using a
number of areas are increased (e.g. more than 6), the accuracy of
piezoelectric material for composite structures with complex surfaces. Compos
successfully locating the damaged area would decrease due to Struct 2013;98:79–84. http://dx.doi.org/10.1016/j.compstruct.2012.10.046.
the smaller size in area. [7] Na S, Lee HK. A technique for improving the damage detection ability of
electro-mechanical impedance method on concrete structures. Smart Mater
Struct 2012;21(8):085024. http://dx.doi.org/10.1088/0964-1726/21/8/
7. Conclusions 085024.
[8] Panigrahi R, Bhalla S, Gupta A. A low-cost variant of electro-mechanical
impedance (EMI) technique for structural health monitoring. Exp Tech
In this study, an approach to use the Electromechanical Imped- 2009;34(2):25–9. http://dx.doi.org/10.1111/j.1747-1567.2009.00524.x.
ance (EMI) technique to predict the damage location of a compos- [9] Na S, Tawie R, Lee HK. Electro-mechanical impedance method of fiber–
reinforced plastic adhesive joints in corrosive environment using a reusable
ite plate using Probabilistic Neural Networks is proposed. Since piezoelectric device. J Intell Mater Syst Struct 2012;23(7):737–47.
EMI technique experiences difficulties against composite materials [10] Liang C, Sun FP, Rogers CA. Coupled electro-mechanical analysis of adaptive
when detecting damage due to vague changes in the impedance material system-determination of the actuator power consumption and
system energy transfer. J Intell Mater Syst Struct 1994;5(1):12–20. http://
signatures, the damage enhancement technique introduced by
dx.doi.org/10.1177/1045389X9400500102.
the authors’ previous work is applied when attached the lead [11] Lopes Jr V, Pereira JA, Weber HI. Using a model updating technique to train
zirconate titanate (PZT) materials onto the test specimens. neural networking for fault detection. In: 16th Biennial conference on
The first part of the experiment involved using a single PZT mechanics vib. and noise-symposium on system health monitoring, in CD-
ROM 1997. ASME conference, September.
patch to investigate how the damages at different locations had [12] Shanker R, Bhalla S, Gupta A. Integration of electro-mechanical impedance and
effect on the root mean square deviation (RMSD) values. The re- global dynamic technique for improved structural health monitoring. J Intell
68 S. Na, H.K. Lee / Composites Science and Technology 88 (2013) 62–68

Mater Syst Struct 2010;21(2):285–95. http://dx.doi.org/10.1177/ [16] Analog devices. AD5933 evaluation board. [accessed on 18.07.13]. <http://
1045389X09356609. www.analog.com/en/evaluation/EVAL-AD5933/eb.html>.
[13] Min J, Park S, Yun CB, Lee CG, Lee CG. Impedance-based structural health [17] Piezo systems, Inc. PSI-5A4E piezoceramic sheets. [accessed on 18.07.13].
monitoring incorporating neural network technique for identification of <http://www.piezo.com/prodsheet1sq5A.html>.
damage type and severity. Eng Struct 2012;39:210–20. http://dx.doi.org/ [18] Sun FP, Chaudhry Z, Liang C, Rogers CA. Truss structure integrity identification
10.1016/j.engstruct.2012.01.012. using PZT sensor–actuator. J Intell Mater Syst Struct 1995;6(1):134–9. http://
[14] Lopes V, Park G, Cudney HH, Inman DJ. Impedance-based structural health dx.doi.org/10.1177/1045389X9500600117.
monitoring with artificial neural networks. J Intell Mater Syst Struct [19] Hankuk carbon Co. Ltd. [accessed on 18.07.13]. <http://www.hcarbon.com/
2000;11(3). http://dx.doi.org/10.1106/H0EV-7PWM-QYHW-E7V. product/develop.asp>.
[15] Moura Jr JRV, Steffen V, Inman DJ. A damage classification technique for [20] Specht DF. Probabilistic neural networks. Neural Networks 1990;3(1):109–18.
impedance based health monitoring of helicopter blades. Proc SPIE 2008;6932. http://dx.doi.org/10.1016/0893-6080(90)90049-Q.
69323S-1.