Structural Monitoring and Maintenance, Vol. 5, No.

4 (2018) 507-519
DOI: https:// doi.org/10.12989/smm.2018.5.4.507 507

Vibration-based damage detection in wind turbine towers using

artificial neural networks

Cong-Uy Nguyena, Thanh-Canh Huynhb and Jeong-Tae Kim
Department of Ocean Engineering, Pukyong National University,
599-1 Daeyon-3dong, Nam-gu, Busan 608-737, Republic of Korea

(Received November 11, 2018, Revised November 25, 2018, Accepted November 28, 2018)

Abstract. In this paper, damage assessment in wind-turbine towers using vibration-based artificial neural
networks (ANNs) is numerically investigated. At first, a vibration-based ANNs algorithm is designed for
damage detection in a wind turbine tower. The ANNs architecture consists of an input, an output, and hidden
layers. Modal parameters of the wind turbine tower such as mode shapes and frequencies are utilized as the
input and the output layer composes of element stiffness indices. Next, the finite element model of a real
wind-turbine tower is established as the test structure. The natural frequencies and mode shapes of the test
structure are computed under various damage cases of single and multiple damages to generate training
patterns. Finally, the ANNs are trained using the generated training patterns and employed to detect damaged
elements and severities in the test structure.
Keywords: wind turbine; vibration; frequency; mode shape; ANN; damage detection; finite element
model

1. Introduction

Wind energy is a promising alternative source in the world’s future because it produces no
greenhouse gas, which is the main cause of global warming. Following the global trend, Korea has
concentrated to develop small- and medium-sized wind farms. Especially Jeju Island has an
ambitious plan to be a Carbon Free Island by 2030 (Park 2015). In order to fulfill the clean energy
demand, more wind turbine towers will be installed in the coming years. The strong investment
into wind energy harvest leads to the consideration of safety and durability of the wind turbine
tower (WTT). During the lifetime, the slender vertical wind tower exposes to extreme wind
frequently, so it experiences large deflections and repeated stress cycles, which would cause
damage in the WTT (Benedetti et al. 2011, Park et al. 2015).
Structural health monitoring based on modal properties can be a solution to ensure the safety as
well as the serviceability of the WTT (Li et al. 2014, Kim et al. 2014, Nguyen et al. 2015,
Martinez-Luengo et al. 2016, Nguyen et al. 2017). It is well-proven that structural damage causes

Corresponding author, Professor, E-mail: [email protected]

a
Graduate Student, E-mail: [email protected]
b
Postdoctoral Researcher, E-mail: [email protected]

Copyright © 2018 Techno-Press, Ltd.

http://www.techno-press.org/?journal=smm&subpage=7 ISSN: 2288-6605 (Print), 2288-6613 (Online)
508 Cong-Uy Nguyen, Thanh-Canh Huynh and Jeong-Tae Kim

the change of mechanical parameters such as mass and stiffness. Consequently, it alters the
dynamic characteristics of the system including modal parameters (Pandey et al. 1994). In the past,
there were considerable efforts in vibration-based structural health monitoring which utilized
dynamic measurement to extract damage sensitive features. Many researchers employed the
change in natural frequencies and mode shapes for damage detection and structural identification
(Vandiver 1977, Farrar et al. 1994, Kim et al. 1995, Kim et al. 2003, Huynh et al. 2016, Li et al.
2016). Nguyen et al. (2015) investigated the potential of the frequency-based and mode
shape-based methods for health monitoring of a numerical WTT model. They later used the
frequency-based method to detect damaged joints in a lab-scaled WTT (Nguyen et al. 2017). From
their findings, it is shown that the high stiffness of the segmental joints in the WTT could cause
discontinuities in mode shapes and modal curvatures that may result in inaccurate damage
localization. Thus, an alternative method should be sought.
Along with various damage detection methods, the artificial neural network (ANN) has an
excellent performance in damage pattern recognition (Yun et al. 2004, Park et al. 2009), which has
been widely applied as the damage identification aid. Auto-associative neural networks were
developed to monitor the cable-stayed Kap Shui Mun bridge (Hong Kong) via measured modal
data from an online system (Ni et al. 2002). Li et al. (2007) identified damage in a beam via ANN
based on statistical properties of structural dynamic response. Transverse crack in a beam was
investigated via developing a neural network in which the input parameters were first three natural
frequencies and the output parameter were relative crack depth and relative crack location in
dimensionless forms (Sutar et al. 2015). Shu et al. (2012) employed changes of variances and
covariance of structural dynamic responses for training ANN to localize damage in railway bridge
under train-induced vibrations.
Despite those research efforts, the implementation of the ANNs to damage identification in the
WTT has not studied so far. A well-established and sufficiently trained ANNs model could be an
important component in the real-time health monitoring system of the WTT. This current study
focuses on the possibility of employing the vibration-based ANNs for the identification of damage
location and severity in the WTT. The remaining is arranged as follows. Firstly, a vibration-based
ANNs algorithm is established for single and multiple structural damages in the WTT. The modal
parameters including normalized natural frequencies and mode shapes are selected as the input,
while the output layer is composed of element stiffness indices. Next, the finite element model of a
real WTT is established as the test structure. Natural frequencies and mode shapes of the test
structure are extracted from the modal analysis under a number of damage cases. Finally, the
ANNs are trained using the training patterns generated from those damage cases. The trained
ANNs are then employed to detect damaged elements and their severities for the test structure.

2. Vibration-based damage detection method using artificial neural networks

2.1 Schematic of damage detection method

The vibration-based damage detection method using ANNs is schematized in Fig. 1. The
algorithm used feed-forward neural networks wherein connections between the units do not form a
cycle. The ANNs architecture consists of an input, an output, and hidden layers. The activation
functions are sigmoid in the hidden layers and linear function in the output layer. The scheme of
the method includes two parts: (a) Training modal properties-based ANNs and (b) Damage
 Vibration-based damage detection in wind turbine towers using artificial neural networks 509

detection for the WTT using the trained ANNs.

The training of modal properties-based ANNs is performed in the following five steps: (1) a
baseline finite element model of a WTT is established from structural analysis by ANSYS S/W.
This FE model is coded via ANSYS Parametric Design Langue (APDL) in order to auto-analyze a
large number of damage scenarios. (2) single and multiple damage scenarios are selected on the
basis of the potential damage in the WTT. Each damage scenario contains two important
parameters which are location and severity. The single damage can locate at any segments of the
FE model while the multiple damages are assumed only at 4 connection flanges to reduce the
computational cost. (3) the modal analysis of the WTT model with every damage scenario is
proceeded. (4) modal parameters (i.e., mode shapes and natural frequencies) are acquired using
modal identification methods such as frequency domain decomposition, or stochastic subspace
identification (Yi and Yun 2004, Qu et al. 2017) and saved into a database for single and multiple
damages detections, independently. (5) these data sets are later used for training modal
parameters-based ANNs. Four ANNs are built for single and multiple damage detection of the
WTT. They are frequency-based ANNs for single damage detection (FAS), frequency-based ANNs
for multiple damage detection (FAM), mode shape-based ANNs for single damage detection
(MAS), mode shape-based ANNs for multiple damage detection (MAM).
The damage detection for the WTT is performed as follows: first, acceleration signals from
sensors on the WTT are recorded during operation; second, the modal parameters are extracted and
classified as natural frequencies and mode shapes separately; and third, the WTT status can be
monitored via the FAS and the MAS for single damage detection. Simultaneously, the flange
connections are checked up via the FAM and the MAM for multiple damage detection.
The output of the FAS and MAS are the damage severity of each element in the WTT model as
defined in Eq. (2). Meanwhile, the FAM and MAM provide damage severities of all four flange
connections. The damage severity is introduced as the reduction of element stiffness k. The
element stiffness 𝑆𝑖 index is defined as
𝑆𝑖 = 𝑘𝑖,𝑑 /𝑘𝑖,𝑢 (1)

Fig. 1 Scheme of vibration-based damage detection method using ANNs

510 Cong-Uy Nguyen, Thanh-Canh Huynh and Jeong-Tae Kim

where i denotes element number; d and u denotes damaged and intact case. Besides, the segment
damage severity is defined as
𝛼 = 1 − 𝑆𝑖 (2)

2.2 Frequency-based ANNs

The frequency-based ANNs algorithms (i.e., FAS and FAM) for damage detection are designed
with feed-forward neural networks. The input layer is composed of natural frequencies of the WTT.
For better training results, the normalized natural frequencies from first three modes along the
wind direction (X direction) and cross-wind direction (Y direction) are selected as the input of
frequency-based ANN, which is described in Eq. (3).
𝑓 = [𝑓𝑥1 𝑓𝑥2 𝑓𝑥3 𝑓𝑦1 𝑓𝑦2 𝑓𝑦3 ] (3)
The normalization of natural frequencies is defined as in Eq. (4).
‖𝑓0 ‖2 = 𝑓0𝑇 𝑓0 = 1; ‖𝑓𝑑 ‖2 = 𝑓𝑑𝑇 𝑓𝑑 = 1 (4)
where subscripts 0 and d denote the intact and damaged cases, respectively. In single damage
detection for the WTT, only one segment is randomly damaged with the severity ranging from 5%
to 50% with a step of 5%. Totally, there are 71 cylinder segments can be damaged in the tower.
Therefore, the FAS for single damage detection is trained with 710 patterns. Meanwhile, the
multiple damage detection only considers flange connections as potential locations. It is noted that
they are the most vulnerable parts of the WTT during operation. Because the multiple damages
with random severity ranging from 0% to 45% with a step of 5% can appear at any of four flange
connections, there are 10000 training patterns for the FAM. The number of hidden layers and
neurons in each layer for the FAS and FAM are summarized as in Table 1.
As shown in Fig. 2, the FAS has 6 neurons for normalized natural frequencies in the input layer
and 71 neurons for the segmental damage severity in the output layer. There are 2 hidden layers in
the FAS. The first hidden layer is composed of 35 neurons, while the second one is composed of
71 neurons. Fig. 3 shows that the FAM has 6 neurons for normalized natural frequencies in the
input layer and 4 neurons for the flange damage severity in the output layer. The hidden layer of
FAM is composed of 50 neurons. After a sufficient training, the frequency-based ANNs including
FAS and FAM are employed to identify damage with severity staying out the training values.

Fig. 2 The FAS architecture

 Vibration-based damage detection in wind turbine towers using artificial neural networks 511

Fig. 3 The FAM architecture

Table 1 Properties of frequency-based ANNs for damage detection

ANN Properties FAS FAM
No. Training Patterns 710 10000
No. Input Neurons 6 6
No. Hidden 1 Neurons 35 50
No. Hidden 2 Neurons 71 -
No. Output Neurons 71 4

2.3 Mode shape-based ANNs

The mode shape-based ANNs algorithms (i.e., MAS and MAM) for damage detection are
designed with feed-forward neural networks. The input layer is composed of mode shapes of the
WTT. For better training results, the normalized mode shapes from the first three modes in the
along-wind direction are selected as the input of mode shape-based ANN, which is described in Eq.
(5).
𝜙 = [𝜙𝑥1 𝜙𝑥2 𝜙𝑥3 ] (5)
where the 𝜙𝑥𝑖 is the eigenvector of the ith mode in the along-wind direction. Near the support of
WTT, the displacement around zero are excluded from the input vector.
The normalization of mode shapes are defined as in Eq. (6).
‖𝜙0 ‖2 = 𝜙0𝑇 𝜙0 = 1; ‖𝜙𝑑 ‖2 = 𝜙𝑑𝑇 𝜙𝑑 = 1 (6)
where subscripts 0 and d denote the intact and damaged cases, respectively. In single damage
detection for the WTT, only one segment is randomly damaged with the severity ranging from 5%
to 50% with a step of 5%. Similar to the FAS, the MAS for single damage detection is trained with
710 patterns. Meanwhile, the multiple damage detection only considers flange connections as
potential inflicted locations. Because the damage with random severity ranging from 0% to 45%
with a step of 5% can appear at four flange connections, the MAM is trained with 10000 patterns
as in the FAM. The number of hidden layers and neurons in each layer for the MAS and MAM are
summarized as in Table 2.
512 Cong-Uy Nguyen, Thanh-Canh Huynh and Jeong-Tae Kim

Table 2 Properties of mode shape-based ANNs for damage detection

ANN Properties MAS MAM
No. Training Patterns 710 10000
No. Input Neurons 203 203
No. Hidden 1 Neurons 69 50
No. Hidden 2 Neurons 71 4

Fig. 4 The MAS architecture Fig. 5 The MAM architecture

As presented in Fig. 4, the MAS has 203 neurons for normalized mode shapes in the input layer
and 71 neurons for the segmental damage severity in the output layer. Meanwhile, the hidden layer
is composed of 69 neurons. Besides, Fig. 5 shows that the MAM has 203 neurons for normalized
mode shapes in the input layer and 4 neurons for the flange damage severity in the output layer.
The hidden layer of FAM is composed of 50 neurons. After sufficient training, the frequency-based
ANNs including MAS and MAM are employed to identify damage with severity staying out the
training values.

3. Finite element modeling of wnd turbine tower

The finite element model of a real WTT structure is established as the test model. The real
WTT is located in the Hankyung II Wind Park, Jeju Island, Korea. The type of the WTT is V90-3.0
MW with nominal rating 3000 kW. The cut-in wind speed is 4 m/s while the cut-out and 25 m/s.
There are 3 blades up-wind direction. Nacelle and Rotor weigh 68 and 39.8 ton, respectively. The
real structure is 80 m high including the hub. There are 4 main segments in the tower. Each main
segment has a flange at 2 tips so each one can combine together through bolt connections. The first
segment is embedded in the foundation and the only 0.55 m is above the foundation surface. The
second segment is 19.21m long while two remaining segments are around 29 m. Each main
segment is formed by several sections with the thickness changing along elevation. Table 3 lists
the variation of the section’s thickness with respect to the increasing height. The tubular tower
material is S355 J2G31. The top-level diameter is 2.316 m and the bottom level is 4.150 m.
 Vibration-based damage detection in wind turbine towers using artificial neural networks 513

As shown in Fig. 6, the finite element WTT model is simulated in ANSYS. The rotor and
nacelle are simulated as lump masses on the top of the tower in the right dimension. These masses
are linked rigidly to the top flange of the WTT. In order to generate a database of modal
parameters in each damage scenario, the finite element model is divided into 71 cylinder segments.
Each cylinder segment is composed of 36 shell elements along the perimeter. The height of
each element depends on the diameter of each segments ranging from 1.8 m in the bottom to 1 m
in the top. The modal analysis of the intact WTT shows the first three mode shapes along-wind
direction and cross-wind direction, as in Fig. 7. The rotor and nacelle axis is always perpendicular
with flange 4 in every mode. This confirms the assumption of rigid links between those elements.

Fig. 6 Finite element modeling of the WTT

Table 3 Cross-sectional thickness of the WTT model

0 5.4 21.9 30.6 36.4 42.2 50.9 53.8 56.7  59.6 
Height (m)
5.4 21.9 30.6 36.4 42.2 50.9 53.8 56.7 59.6 77.3
Thickness
40 26 24 23 22 21 19 18 17 16
(mm)
514 Cong-Uy Nguyen, Thanh-Canh Huynh and Jeong-Tae Kim

(a) Along-wind modal shape

(b) Cross-wind modal shape

Fig. 7 Modal analysis results for the WTT model

4. Damage detection in wind turbine tower using vibration-based ANNs

4.1 Detection of single damage

Damage detection via FAS

Six test cases are selected to demonstrate the performance of single damage detection via the
trained FAS, as listed in Table 4. The infliction is located randomly from the bottom to the top of
the WTT model with severity ranging from 12% to 18%. The first three cases (Cases 1-3) are
trained patterns, while the 3 remaining cases (Cases 4-6) are untrained patterns. The damage
location and severity predicted via the FAS are presented in Fig. 8. As shown in the figure, the
damage locations are recognized correctly via the FAS. However, the estimated damage severities
are below the inflicted ones.

Damage detection via MAS

Six test cases are selected to demonstrate the performance of single damage detection via the
MAS, as listed in Table 5. The infliction is located randomly from the bottom to the top of the
WTT model with severity ranging from 10% to 25%. The first three cases (Cases 1-3) are trained
patterns, while the 3 remaining cases (Cases 4-6) are untrained patterns. The damage location and
severity predicted via the MAS are shown in Fig. 9. It is observed that the damage locations are
recognized correctly via the MAS. However, the estimated damage severities are below the
inflicted ones.
Vibration-based damage detection in wind turbine towers using artificial neural networks 515

Fig. 8 Damage detection in the WTT model via the FAS

Fig. 9 Detection results for single damage in the WTT via the MAS
516 Cong-Uy Nguyen, Thanh-Canh Huynh and Jeong-Tae Kim

Table 4 Test cases for the FAS

Trained Patterns Untrained Patterns
Damaged Element
Case 1 Case 2 Case 3 Case 4 Case 5 Case 6
ID 8 32 71 8 32 71
Severity 15% 15% 15% 18% 12% 16%

Table 5 Test cases for the MAS

Trained Patterns Untrained Patterns
Damaged Element
Case 1 Case 2 Case 3 Case 4 Case 5 Case 6
ID 7 38 68 7 38 68
Severity 15% 10% 25% 18% 24% 18%

4.2 Detection of multiple damage

Damage detection via FAM

Six test cases are selected to demonstrate the performance of multiple damage detection via the
FAM, as shown in Table 6. The infliction is located randomly in any of the four flange connections
with severity ranging from 5% to 45%. The first three cases (Cases 1-3) are trained patterns, while
the 3 remaining cases (Cases 4-6) are untrained patterns. The damage location and severity
estimated by the FAM are presented in Fig. 10. As shown in the figure, the damage locations and
severity are recognized correctly via the FAM.

Fig. 10 Damage detection in the WTT via the FAM

 Vibration-based damage detection in wind turbine towers using artificial neural networks 517

Table 6 Test cases for the FAM

Damaged Trained Patterns Untrained Patterns
Flanges Case 1 Case 2 Case 3 Case 4 Case 5 Case 6
1 5% 30% 5% 8% 27% 9%
2 - - 15% - - 17%
3 25% 10% 45% 22% 12% 42%
4 - 15% 40% - 17% 36%

Damage detection via MAM

Six test cases are selected to demonstrate the performance of multiple damage detection via the
MAM, as shown in Table 7. The infliction is located randomly in any of the four flange
connections with severity ranging from 5% to 42%. The first three cases (Cases 1-3) are trained
patterns, while the 3 remaining cases (Cases 4-6) are untrained patterns. The damage location and
severity estimated by the MAM are presented in Fig. 11. As shown in the figure, the damage
locations are recognized correctly, meanwhile, the damage severities are estimated around true
values via the MAM.

Fig. 11 Damage detection in the WTT via the MAM

Table 7 Test cases for the MAM

Damaged Trained Patterns Untrained Patterns
Flanges Case 1 Case 2 Case 3 Case 4 Case 5 Case 6
1 15% 20% 5% 8% 27% 9%
2 - - 15% - - 17%
3 25% 10% 25% 22% 12% 42%
4 - 15% 35% - 17% 36%
518 Cong-Uy Nguyen, Thanh-Canh Huynh and Jeong-Tae Kim

5. Conclusıons

In this study, damage detection of a wind turbine tower (WTT) model using vibration-based
artificial neural networks (ANNs) was presented. At first, a vibration-based ANNs algorithm was
designed for damage detection in the WTT model. The ANNs architecture consisted of an input, an
output, and hidden layers. Modal parameters of the wind turbine tower such as mode shapes and
frequencies were utilized as the input and the output layer composes of element stiffness indices.
Next, the finite element model of a real wind-turbine tower was established as the test structure.
The natural frequencies and mode shapes of the test structure were computed under various
damage cases to generate training patterns. Finally, the ANNs were trained using the generated
training patterns and employed to detect damaged elements and severities in the test model.
The first three natural frequencies in the along-wind and cross-wind directions were employed
for the frequency-based ANNs. Meanwhile, only the first three mode shapes in the along-wind
were employed for the mode shape-based ANNs. The structural damages were introduced as the
reduction of element stiffness in the WTT. In case of single damage, the infliction was simulated
randomly at any segment in total 71 locations; whereas multiple damages only located at 4 flange
connections. The designed vibration-based ANNs were successful in indicating single-damage as
well as multi-damage locations in the WTT structure. However, the structural damage severity was
only identified correctly via the frequency-based method while the remaining ANNs estimated
infliction severity around the true values for all cases. It is noted that the frequency-based ANNs
were quite better than mode shape-based ANNs in damage location and severity estimation. Future
studies remain to evaluate the accuracy of the FE model by comparing with the experimental result,
and to detect damage below 5% severity in the WTT.

Acknowledgements

This work was supported by a Research Grant of Pukyong National University (2017 year).

References

Benedetti, M., Fontanari, V. and Zonta, D. (2011), “Structural health monitoring of wind towers: remote
damage detection using strain sensors”, Smart Materi. Struct., 20, 1-13.
Farrar, C.R. (1997), “System identification from ambient vibration measurements on a bridge”, J. Sound Vib.,
205(1), 1-18
Huynh, T.C., Park, J.H. and Kim, J.T. (2016), “Structural identification of cable-stayed bridge under
back-to-back typhoons by wireless vibration monitoring”, Measurement, 88, 385-401.
Kim, J.T. and Stubbs, N. (1995), “Model uncertainty and damage detection accuracy in plate-girder bridges”,
J. Struct. Eng., 121(10), 1409-1417
Kim, J.T., Huynh, T.C. and Lee, S.Y. (2014), “Wireless structural health monitoring of stay cables under two
consecutive typhoons”, Struct. Monit. Maint., 1(1), 47-67.
Kim, J.T., Ryu, Y.S., Cho H.M. and Stubbs, N. (2003), “Damage identification in beam-type structures:
Frequency-based method vs mode-shape-based method”, Eng. Struct., 25, 57-67.
Lee, J.J., Lee, J.W., Yia, J.H, Yun, C.B. and Jung, H.Y. (2004), “Neural networks-based damage detection for
bridges considering errors in baseline finite element models”, J. Sound Vib., 280, 555-578
Li, H.N., Li, D.S., Ren, L., Yi, T.H., Jia, Z.G. and LI, K.P. (2016), “Structural health monitoring of
innovative civil engineering structures in Mainland China”, Struct. Monit. Maint., 3(1), 1-32.
 Vibration-based damage detection in wind turbine towers using artificial neural networks 519

Li, H.N., Yi, T.H., Ren, L., Li, D.S. and Huo, L.S. (2014), “Review on innovations and applications in
structural health monitoring for infrastructures”, Struct. Monit. Maint., 1(1), 1-45.
Li. Z.X. and Yang. X.M. (2008), “Damage identification for beams using ANN based on statistical property
of structural response”, Comput. Struct., 86(1), 64-71.
Martinez-Luengo, M., Lolios, A. and Wang, L. (2016), “Structural health monitoring of offshore wind
turbines: A review through the statistical pattern recognition paradigm”, Renew. Sust. Energ. Rev., 64,
91-105.
Nguyen, C.U., Huynh, T.C., Dang, N.L. and Kim, J.T. (2017), “Vibration-based damage alarming criteria for
wind turbine towers”, Struct. Monit. Maint., 4(3), 221-236.
Nguyen, T.C., Huynh, T.C. and Kim, J.T. (2015), “Numerical evaluation for vibration-based damage
detection in wind turbine tower structure”, Wind Struct., 21(6), 657-675.
Nguyen, T.C., Huynh, T.C., Yi, J.H. and Kim, J.T. (2017), “Hybrid bolt-loosening detection in wind turbine
tower structures by vibration and impedance responses”, Wind Struct., 24(4), 385-403.
Ni, Y.Q., Zhou, X.T., Ko, J.M. and Wang, B.S. (2002), “Vibration-based damage localization in Ting Kau
Bridge using probabilistic neural network”, Adv. Struct. Dynamics. 2, 1069-1076.
Pandey, A.K. and Biswas, M. (1994), “Damage detection in structures using changes in flexibility”, J. Sound
Vib., 169(1), 3-17.
Park, J. (2015), Annual Report on Wind Energy Industry of Korea 2015, Korea Wind Energy Industry
Association
Park, J.H., Huynh, T.C., Choi, S.H. and Kim, J.T. (2015), “Vision-based technique for bolt-loosening
detection in wind turbine tower”, Wind Struct., 21(6), 709-726.
Park, J.H., Kim, J.T., Hong, D.S., Ho, D.D. and Yi, J.H. (2009), “Sequential damage detection approaches
for beams using time-modal features and artificial neural networks”, J. Sound Vib., 323(1-2), 451-474.
Qu, C.X., Yi, T.H., Yang, X.M. and Li, H.N. (2017), “Spurious mode distinguish by eigensystem realization
algorithm with improved stabilization diagram”, Struct. Eng. Mech., 63(6), 743-750.
Shu, J., Zhang, Z., Gonzalez, I. and Karoumi, R. (2012), “The application of a damage detection method
using Artificial Neural Network and train-induced vibrations on a simplified railway bridge model”, Eng.
Struct., 52, 408-421.
Sutar, M.K., Sarojrani. P. and Jayadev, R. (2015), “Neural based controller for smart detection of crack in
cracked cantilever beam”, Proceeding of Materials Today, 2, 2648-2653.
Vandiver, J.K. (1977), “Detection of structural failure on fixed platforms by measurement of dynamic
response”, J. Petrol. Technol., 29(3), 305-310.
Yi, J.H. and Yun, C.B. (2004), “Comparative study on modal identification methods using output-only
information”, Struct. Eng. Mech., 17(3-4), 445-466.

TY