ISA Transactions xxx (xxxx) xxx–xxx

Contents lists available at ScienceDirect

ISA Transactions
journal homepage: www.elsevier.com/locate/isatrans

Practice article

Output-only modal analysis of wind turbine tower based on vibration

response under emergency stop
Aijun Hu, Jun Zhao, Ling Xiang∗
Department of Mechanical Engineering, North China Electric Power University, Baoding 071003, Hebei Province, China

A R T I C LE I N FO A B S T R A C T

Keywords: The identification technique of output-only modal parameters is proposed for the large wind turbine tower under
Output-only modal analysis emergency stop. Compared with the response of regular operating conditions, the immediate tower structural
Vibration response response under emergency stop much more resembles a state of free vibration, which is more appropriate for the
Wind turbine tower modal identification of the wind turbine tower. The vibration response is measured in the nacelle, which is easy
Emergency stop
to perform in the field modal test. The variational mode decomposition (VMD) is applied to decompose the
VMD
vibration response into several band-limited intrinsic mode functions. The free responses of decomposed func-
tions are extracted by applying the random decrement technique (RDT). Finally, the modal damping ratio and
natural frequency are identified from each free modal response by using the Hilbert transform method.
Simulations and a 1.5 MW wind turbine field modal test results verify the effectiveness of the proposed iden-
tification method. The main modal parameters of wind turbine, including weak modes, are effectively extracted
by using output-only vibration responses under emergency stop. The modal parameter identification method is
provided for the large wind turbine structure under the engineering condition.

1. Introduction
Over the decades, wind energy has recently gained much attention
from researchers [1]. With the sizes of modern wind turbines in-
creasing, their dynamic behaviors get more complicated. Analyzing and
experimentally measuring the modal parameters, natural frequency and
modal damping, for instance, of wind turbine structures became more
important to successful design and operation. The traditional Experi-
mental Modal Analysis (EMA) techniques have been widely developed
and employed in wind turbine structures, which extract modal para-
meters from input and output data. The different methods such as im-
pact hammer modal testing, single-input multiple-output (SIMO) and
multiple-input multiple-output (MIMO) measurement were used to
extract modal parameters of the wind turbine [2]. A 600 kW wind
turbine with 40 m hub-height is excited by shakers connected with
cables in EMA [3]. Structural response characteristics and modal
parameters are reported using shake table test on a full-scale 65-kW
wind turbine with 22.6-m hub height [4]. Step-relaxation excitation
method is used by Carne [5] in the field modal test of a 110 m Darrieus
bladed vertical axis wind turbine (VAWT). The step-relaxation forces
are applied to the wind turbine with a high strength cable and a diesel
winch, which is located on the ground 100 m from the base of the
turbine. As the sizes of modern wind turbines increase, it is usually a
hard work to perform excitation in field testing of large wind turbines.
It is a substantial challenge to carrying out modal tests on large
wind turbine structures. The operational modal analysis (OMA) is re-
garded as a key technique for the large and flexible structures. The
approach was formally documented for modal testing in 1986 [5]. The
OMA of structures subject to environment and natural excitations under
operational states has attracted a lot of attention in engineering since
1990s [6]. The OMA method extracts modal parameters from me-
chanical system responses only, so it is also called output-only modal
analysis [7]. Several ambient excitations are considered in OMA test of
the wind turbine. Carne et al. have a series of study on modal parameter
extraction from operating turbines with wind excitation, for example,
the cross spectra analysis for a 110 m Darrieus bladed vertical axis wind
turbine (VAWT) [5], the natural excitation technique (NExT) for a 34 m
Darrieus VAWT [8] and a two-bladed upwind horizontal axis wind
turbine (HAWT) [9]. Wind excitation is also adopted in the OMA of a
600 kW wind turbine with 40 m tall hub [3], a two blade 11 kW down-
wind turbine with 18 m tubular steel tower [10], a 40 kW wind turbine
with 21 m steel tower height [11]. Human power is also utilized to pull
the structure and excite the free vibration response of a 21 m tall steel
tower [11]. It is reported that sometimes a few modes cannot be well
excited by the wind [3]. Ambient excitation (wave and wind) and
overspeed stop test are performed in the modal analysis for a 3 MW

∗
Corresponding author.
E-mail address: ncepuxl@163.com (L. Xiang).

https://doi.org/10.1016/j.isatra.2018.07.035
Received 30 November 2017; Received in revised form 24 July 2018; Accepted 26 July 2018
0019-0578/ © 2018 ISA. Published by Elsevier Ltd. All rights reserved.

Please cite this article as: Hu, A., ISA Transactions (2018), https://doi.org/10.1016/j.isatra.2018.07.035
A. Hu et al. ISA Transactions xxx (xxxx) xxx–xxx

offshore wind turbine with 72 m with hub height [12]. The first natural
frequency and modal damping are obtained in the condition measure-
ment [12]. The damping ratios at different tower level are identified
[13]. “Rotor-stop” tests are also used to obtain the natural frequency
and modal damping of offshore wind turbines at four offshore wind
parks [14]. Rotor-stop condition can cause obvious decay process of the
wind turbine.
Emergency stop is an ambient operating condition of wind turbine.
It is can be triggered by many events, a fault within the control system,
high grid loading or manually triggered by the emergency stop buttons
[15], for instance. The global motion and dynamic structural response
of the wind turbine can be activated during emergency stop. Compared
with the response of regular operating conditions, the emergency stop
can be treated as a step-relaxation. The large amount energy can be
input to the wind turbine. The immediate tower responses represent a
wider frequency range of high signal energy. So the response data can
be used to identify the wind turbine structures. In this paper, the
output-only responses under emergency stops are proposed for the
modal parameters identification of a 1.5 MW onshore wind turbine with
70 m tower height.
Frequency domain and time domain technique still are the critical
groups of output-only system identification methods. For example,
Brincker et al. [16] used frequency domain decomposition (FDD) to
practice modal identification from ambient responses. In next year,
Brincker et al. [17] enhanced frequency domain decomposition to
identify damping. However, frequency domain methods are based on
Fourier transform, which is adapted to the time-period vibration sig-
nals, so the local characteristics for identification of time-varying
system may be erased. Because of the limitation, the time domain
methods have further progressed. Some classical methods such as
Ibrahim time domain (ITD) method, its upgrade method and other
methods are illustrated and applied in literature [18–25]. The time
domain methods can directly use the measured vibration responses
from time-varying system, therefore, the disadvantages of the frequency
domain methods, such as errors caused by aliasing, leakage and other
reasons could be avoided. Yang et al. [26,27] performed the modal
identification of linear structures through adopting Hilbert-huang
transform (HHT) method. Subsequently, the identification method of
the instantaneous characteristics of the nonlinear multi-degree of
freedom (MDOF) structures based on the Hilbert transform (HT) and
the empirical mode decomposition (EMD) was presented by Huang and
Lou [28]. On the basis of EMD, Pines and Salvino [29] extracted phase
information from transient signals in structural health monitoring.
However, some deficiencies related to the EMD process is found in Ref.
[30]. Bao et al. [31] provided a novel improved Hilbert-Huang trans-
form (HHT) algorithm for identification of time-varying systems and
analysis of nonlinear structural response with closely spaced modes.
The RDT-AMD (analytical mode decomposition, AMD) is reported for
building system identification [32]. The effectiveness of variational
mode decomposition (VMD) is validated by Abdollah et al. for shear
frame and footbridge system identification [33].
Based on the researches, this paper presents a new system identifi-
cation routine that uses output-only vibration responses under emer-
gency stop to identify the parameters of the wind turbine tower. The
VMD [34] is used to decompose the measured responses in time do-
main, which consists of a series of modal responses. Then, the random
decrement technique (RDT) is used to acquire the free responses from
each modal response. Finally, the natural frequency and the modal
damping ratio can be recognized from each modal free response by
using the Hilbert transform (HT) method. The field modal test can ea-
sily be performed using accelerometer in the nacelle under emergency
stop condition. The large amount energy can be supplied by the
emergency stop, which makes it suitable for the field test of large wind
turbine structure with low modal frequencies. The modal parameters of
the dominant mode and weak modes of the wind turbine tower are
identified by VMD-RDT method. The remainder of this paper is
Fig. 1. Schematic diagram of the proposed method.
organized as follow: Section 2 introduces the VMD algorithm, random
decrement technique (RDT) and Hilbert transform. Section 3 presents
the numerical analysis results of two responses, which are damped free
vibration response without noise and with white Gaussian noise of a
MDOF system. Section 4 dedicates the implementation procedure of the
proposed identification routine of the wind turbine tower, and the
modal analysis results are showed in this section. In the end, in section
5, the conclusions are demonstrated and some further study direction
proposals are discussed.
2. Identification algorithms
Fig. 1 illustrates the schematic diagram for the modal parameter
identification using VMD-RDT method under emergency stop. The
method consists of four main steps: 1) The vibration signal under
emergency stop is acquired by accelerometers mounted in the nacelle.
2) VMD is used to decompose the measured responses in time domain.
3) The random decrement technique (RDT) is used to obtain the free
responses from each modal response. Finally, the natural frequency and
the modal damping ratio can be recognized from each modal free re-
sponse by using the Hilbert transform (HT) method. Each of these steps
is described in the following sub-sections.
2.1. Variational mode decomposition (VMD) method
EMD is widely used in non-stationary signal process. However, as a
recursive decomposition model, EMD is known for limitations like
mode mixing, end effect and sensitivity to sampling and noise.
Recently, Dragomiretskiy and Zosso proposed a novel and non-recur-
sively adaptive time-frequency analysis method, the variational mode
decomposition (VMD). VMD can decompose a complex and real valued
multi-component signal into a discrete number of sub-signals whose
bandwidths are estimated using the H1 Gaussian smoothness of the
demodulation signal.
The decomposition process of VMD is the process of solving varia-
tional problems, which makes the sum of estimation bandwidth of each
mode minimum. The resulting constrained variational problem for any
given non-stationary input signal x(t) in VMD is written as
2
⎧ min ⎧∑ ∂ δ (t ) +
⎪
{uk },{ωk }⎨
⎩
k(t⎡
⎣
j
πt ) ∗u (t ) ⎤⎦ e
k
−jωk t
2⎬
⎭
⎫
⎨
⎪ s.t. ∑k uk = f
⎩ (1)
Where ωk = {ω1, ω2, ⋯, ωK } is the center frequency of each intrinsic
mode components, uk = {u1, u2, ⋯, uK } is the corresponding all intrinsic

2
A. Hu et al. ISA Transactions xxx (xxxx) xxx–xxx

Fig. 2. The simulated signal (a) time waveform of the exponentially decayed components, (b) time waveform and amplitude spectrum of the simulated MDOF
response.

K
mode components, ∑k uk = ∑k = 1 uk represents the summation of whole
intrinsic mode components, and K is the number of the achieved in-
trinsic mode components. By using the Lagrangian multipliers λ (t ) and
quadratic penalty term α the variational model can be obtained. The
augmented Lagrange can be calculated as
2 2
j ⎤
L ({uk }, {ωk }, λ ) = α ∑ ∂t ⎡ ⎛δ (t ) + ⎞ ∗uk (t ) + f (t ) − ∑ uk (t )
k
⎢
⎣ ⎝ πt ⎥
⎠⎦ k
2 2

+ λ (t ), f (t ) − ∑ uk (t )
k (2)

Where α is the balancing parameter, f (t ) is the original signal and δ (t )

Fig. 3. Center frequency at different K of the simulated signal decomposition.
is the Dirac function. Because of the sensitivity of α for decomposition
results of VMD, the selection of value of α needs to consult the mor-
phology of the vibration signal. Selecting reasonably value of α is cru-
cial in the practical application. In addition, the alternate direction
method of multipliers (ADMM) is used to describe the saddle point. In
other words, the constraint variational model (see Eq. (1)) could be

Fig. 4. Decompositions of numerical signal x(t) by VMD and EMD: (a) VMD of signal x(t), (b) EMD of signal x(t).

3
A. Hu et al. ISA Transactions xxx (xxxx) xxx–xxx

Fig. 5. RDT and HT of BLIMF1 from VMD: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.

Fig. 6. RDT and HT of IMF3 by EMD: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.

Table 1
The theoretical and identified values of natural frequency and damping ratio.
Madal Parameters Theoretical value Identified results by VMD Identified results by EMD

Frequency/Hz Damping ratio/% Frequency/Hz Damping ratio/% Frequency/Hz Damping ratio/%

First mode 0.5 3.00 0.514 2.894 0.273 8.420

Second mode 1.8 1.00 1.796 0.921 0.794 2.262
Third mode 3.5 0.68 3.516 0.663 1.795 0.917

∞
obtained by using ADMM. Hence, in the spectral domain, the all band- ∫0 ω uˆ k (ω) 2 dω
limited intrinsic mode functions (BLIMFs) gained from solutions can ωkn + 1 = ∞ .
∫0 uˆ k (ω) 2 dω (4)
calculated in Eq. (3), as shown below.

λˆ (ω) In literature [34] and [35], the VMD method is completely illu-
fˆ (ω) − ∑i ≠ k uˆ i (ω) + 2 strated. The VMD method is proved to have the amazing performance in
uˆkn + 1 (ω) = .
1 + 2α (ω − ωk )2 (3) regard to existing mode decomposition models. This method still can
accurately process vibration response data even from the strong back-
Eq. (4) is the description of center frequency ωk . The wiener filter is ground noise.
embedded in the VMD method because of owning some optimality for
decreasing the noise.

4
A. Hu et al. ISA Transactions xxx (xxxx) xxx–xxx

Fig. 7. RDT and HT of BLIMF2 from VMD: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.

Fig. 8. RDT and HT of BLIMF3 from VMD: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.

2.2. Random decrement technique (RDT)
Random decrement technique (RDT), is a kind of method used to
extract the free response signal of structure. RDT is based on the su-
perposition principle of linear system, using the statistical average
method to eliminate the random part, and filter out the deterministic
free attenuation signal. RDT is simple to achieve and can realize the
real-time data analysis in the actual condition, so it has become a very
mature technology of signal attenuation in structure extraction and has
been widely used in the analysis of the vibration mode and structure
damage detection.
Using the RDT, the free modal response from each BLIMF modal
acceleration response x¨j (t ) is gained. It should be noted that in order to
obtain L segments the threshold level should be definition, where L is
Where tk is the starting time for each segment, M is an integer with
MΔt = T being the duration of each segment Δt is the time interval of
the modal response time history and x¨ jf (t ) is the jth free modal re-
sponse. Once the x¨ jf (t ) is obtained, the modal damping ratio and natural
frequency can be adopted by using the HT for the single-degree-of-
freedom (SDOF) system.
2.3. Hilbert transform (HT) modal parameters identification
The response of the system under ambient excitation is decomposed
by VMD, and each band-limited intrinsic mode functions (BLIMF) after
RDT can be expressed as
n

the number of segments which have relation to the threshold level ẍs . BLIMFi (t ) = ∑ Ai0 e−ξi ωi t cos(ωdi t + θi0)
The jth free modal response x¨ jf (iΔt ) can be calculated from the en- i=1 (6)
semble average of all segments as
Where A0 and θi are constants which depend on the intensity of im-
L pulsive loading and the mass of the system. ωdi , ωi are the damped
1
x¨ jf (iΔt ) = ∑ x¨j (tk + iΔt ), i = 1,2, ⋯, M circular frequency and the undamped natural circular frequency, re-
L k=1 (5) spectively. ξi is the damping ratio. The analytical signal can be

5
A. Hu et al. ISA Transactions xxx (xxxx) xxx–xxx

Fig. 9. Center frequency at different K with (a) SNR = 1 (b) SNR = 4/3 (c) SNR = 2 (d) SNR = 4.

differential operator to instantaneous phase, the modal parameters can

be solved as Eq. (9). At the end, the linear least-square fit procedures
could be applied to estimate natural frequency and damping ratio.

⎧ ln Ai (t ) = −ξi ωi t + ln Ai0
dθ (t )
⎨ ω (t ) = ωdi = dti (9)
⎩

3. Numerical simulation

In this section, VMD-RDT based method is used to recognize the

Fig. 10. Decompositions of numerical signal x(t) with Gaussian noise by VMD.
modal parameters of a simulated MDOF system. For a no-load linear
MDOF system, each of the damped free vibration response of the system
should be considered. All the modal responses are set according to as-
sumed to the real wind turbine tower parameters in the simulation. The
free vibration response comprises three exponentially decayed compo-
nents in the system, and the equation is shown below
−0.0942t sin(3.1397t + 0.314)
⎧ x1 (t ) = e
x2 (t ) = 2e−0.1131t sin(11.3091t )
⎨ −0.1495t sin(21.9905t + 0.628)
⎩ x3 (t ) = e (10)

The simulated signal is

calculated as below by using HT to each BLIMF.
x (t ) = x1 (t ) + x2 (t ) + x3 (t ) (11)
Zi (t ) = BLIMFi (t ) + jH (BLIMFi (t )) = Ai (t ) e−jθi (t ) (7) Fig. 2 shows time waveform and amplitude spectrum of the simu-
When the damping is relatively small, the instantaneous amplitude lated signal. Considering the existing of noises in the engineering test,
and the instantaneous phase can be written as the simulation is divided into two parts, response with and without
noise.
−ξ ω t
⎧ Ai (t ) = Ai0 e i i
⎨ 3.1. Free vibration response without noise of a MDOF system
⎩ θi (t ) = ωdi t + θi0 (8)

Introducing logarithmic operator to instantaneous amplitude and VMD is applied for decomposing the simulated MDOF response

6
A. Hu et al. ISA Transactions xxx (xxxx) xxx–xxx

Fig. 11. RDT and HT of BLIMF1 from VMD: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.

Fig. 12. RDT and HT of BLIMF2 from VMD: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.

Fig. 13. RDT and HT of BLIMF3 from VMD: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.

7
A. Hu et al. ISA Transactions xxx (xxxx) xxx–xxx

Table 2
The identified results with different SNR value.
SNR 4 (K = 4) 2 (K = 4) 4/3 (K = 4) 1 (K = 5)

Modal parameters Frequency/Hz Damping ratio/% Frequency/Hz Damping ratio/% Frequency/Hz Damping ratio% frequency/Hz Damping ratio/%

1st mode 0.475 2.907 0.475 2.736 0.475 2.646 0.475 2.608
2nd mode 1.796 0.946 1.796 0.922 1.796 0.918 1.795 0.913
3rd mode 3.515 0.371 3.515 0.307 3.514 0.283 3.515 0.290

Fig. 14. The sensor placement of the wind turbine (a) schematic diagram (b) picture of the tail of the bed plate in the nacelle.

Fig. 15. The signals during emergency stop condition (a) the operation parameters, (b) vibration in the tail and middle bed plate.

signal into an ensemble of band-limited intrinsic mode functions
(BLIMFs). For comparison, the EMD results are also supplied.
In application, the parameters in VMD have significant effect on the
decomposition results. Dragomiretskiy and Zosso [34] indicated that
the number of modes K should be predefined, and the K could be tested
by checking the spectral overlap between modes. The number K was
selected based on experience in literature, for example, rotor-to-stator
rubbing [36] and seismic signal [37] analysis. The low and high fre-
quency vibrations that were not due to the structural vibration were
removed in value K selection [38]. The detrended fluctuation analysis
(DFA) was adopted as an indicator for defining the number K in Ref.
[35]. Based on References [34,39], we analyze the center frequency to
select the number of modes K in this paper. As the quadratic penalty
term α, Reference [39] demonstrated that VMD could well retrieve all
the components when α was fixed to 2000. The value of α is varied from
1000 to 10000, and it can be calculated in the simulation. According to
the results, α = 2000 is selected in this research.
The simulated signal (Eq. (11)) is decomposed by VMD with

8
A. Hu et al.
Fig. 16. The vibration response of wind turbine tower from X1 sensor.
Fig. 17. Decompositions of the vibration signal from X1 sensor using VMD.
Fig. 18. RDT and HT of BLIMF1 from X1: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.
9
ISA Transactions xxx (xxxx) xxx–xxx
different K value. The relationship between the center frequency and K
is shown in Fig. 3. It shows that the center frequencies of the first three
modes are the same value when K≥ 3. Since the center frequency is
estimated as the center-of-gravity of the mode's power spectrum [34],
the number K = 3 is selected in the VMD decompositions of the simu-
lated signal. Fig. 4 (a) shows the extracted BLIMFs by VMD. The red
dash lines depicts the original simulated signal in Eq. (10) and blue line
indicates the decomposed components. The BLIMF1, BLIMF2 and
BLIMF3 in Fig. 4 (a), which are decomposed by VMD are identical with
the original components x1 (t ) , x2 (t ) and x3 (t ) . Fig. 4 (b) shows the three
of six IMFs resulting from EMD. It can be seen that IMF1 and IMF2
cover more than one frequency components, and other IMFs are also
generated. Only IMF3 recovers the original component x1 (t ) approxi-
mately. Moreover, the endpoint effect of EMD method is apparently
obvious. It is verified that VMD is better than EMD in such signal de-
composing.
Further, the RDT and HT are applied to identify the modal para-
meters of the signal. What should be paid attention to is that low order
IMFs represent high frequency modes, and high order IMFs represent
low frequency modes in EMD. While in VMD, which is exactly the
opposite. Fig. 5 shows the recognition results of the first mode from
VMD, and Fig. 6 presents the results of IMF3 from EMD. The wave of
BLIMF1 after RDT is shown in Fig. 5 (a). Fig. 5 (b) shows that the in-
stantaneous frequency, amplitude of the mode and the corresponding
linear least-squares fit. It is found that the modal information is suc-
cessfully identified by applying the VMD. The first frequency is com-
puted to 0.514 Hz and the damping ratio is 2.894%. The results are
close to theoretical values which is shown in Table 1. But in Fig. 6 (b) it
is shown that the extracted mode from IMF3 decomposed by EMD come
into distortion phenomenon after 10 s, leading to the recognition errors.
Using this method, the first frequency is computed to 0.273 Hz and the
damping ratio is 8.420%, which have differences with theoretical va-
lues.
The second and third order modal parameters are further identified
using the VMD. Fig. 7 shows the recognition results of the second mode
from VMD. The wave of BLIMF2 after RDT is shown in Fig. 7 (a). Fig. 7
(b) shows the instantaneous frequency of the mode and the corre-
sponding linear least-squares fit. It can be also seen that the modal
information is successfully identified. The second frequency is com-
puted to 1.796 Hz and the damping ratio is 0.921%. The recognition
results of the third mode from VMD are shown in Fig. 8. The wave of
BLIMF3 after RDT is shown in Fig. 8 (a). The amplitude of the mode,
instantaneous frequency and the corresponding linear least-squares fit
A. Hu et al.
Fig. 19. RDT and HT of BLIMF2 from X1: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.
Fig. 20. RDT and HT of BLIMF3 from X1: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.
are represented as shown in Fig. 8 (b). The third frequency is computed
to 3.516 Hz and the damping ratio is 0.663%. Table 1 shows the
identified modal parameters of the signal x(t), so it is demonstrated that
the VMD can effectively identify a MDOF system without noise.
3.2. Damped vibration response with white Gaussian noise of a MDOF
system
The decomposition performance of VMD was better than that of
EMD in signal denoising [34]. In order to study the robustness of the
proposed method, the white Gaussian noise with different SNR value is
added to the simulated signal as shown in Eq. (11).
The simulated signal with different SNR value is decomposed by
VMD. The relationship between the center frequency and K is shown in
Fig. 9. It shows that the center frequencies of the first three modes are
the same value when K≥ 5 (Fig. 9 (a)). Therefore, the number K = 5 is
selected under SNR = 1. For other SNR signals, the number K = 4 is
selected from the Fig. 9 (b)-(d).
Fig. 10 depicts the VMD decompositions of the signal x(t) with
10
ISA Transactions xxx (xxxx) xxx–xxx
SNR = 2 Gaussian noise. When the number K is set to 4, the BLIMF1,
BLIMF2 and BLIMF3 recover the original components x1 (t ) , x2 (t ) and
x3 (t ) . Further, the RDT and HT are applied to identify the modal
parameters of the signal. Fig. 11 shows the recognition results of the
first mode from VMD. The wave of BLIMF1 after RDT is shown in
Fig. 11 (a). The instantaneous frequency, amplitude of the mode and
the corresponding linear least-squares fit are shown in Fig. 11 (b). It can
be seen the modal information is successfully identified. The first fre-
quency is computed to 0.475 Hz and the damping ratio is 2.736%.
Fig. 12 shows the recognition results of the second mode from VMD.
The second frequency is computed to 1.796 Hz and the damping ratio is
0.922%. Fig. 13 shows the recognition results of the third mode from
VMD. The third frequency is computed to 3.515 Hz and the damping
ratio is 0.307%.
The identification results in different SNR are shown in Table 2. It
can be seen the SNR affects the results of system identification. The
natural frequencies are well identified for different SNR value. But in
the third modal information, the damping ratios have differences with
theoretical values due to the influence of noise.
A. Hu et al.
Fig. 21. The vibration response of wind turbine tower from X2 sensor.
Fig. 22. Decompositions of the vibration signal from X2 sensor using VMD.
Fig. 23. RDT and HT of BLIMF1 from X2: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.
11
ISA Transactions xxx (xxxx) xxx–xxx
4. Modal identification of wind turbine tower
In working process of wind turbine, when the driving frequency
from engine or wind turbine blade is closed to the natural frequency of
wind turbine tower, the turbine tower will produce large vibration,
even resonance. The resonance will cause unimaginable harm, so it is
necessary to recognize the modal parameters of the wind turbine tower.
The operating field tests of a 1.5 MW onshore wind turbine with 70 m
hub height are explored by using accelerometer, and the vibration re-
sponses under emergency stop condition are analyzed by applying VMD
and RDT method to identify the modal parameters.
In the test, four ENDEVCO 7290A variable capacitance accel-
erometers are mounted in the middle and end of the bed plate in the
wind turbine nacelle. Fig. 14 shows the schematic diagram and the
picture of sensors placement. X is the wind direction (fore-aft (FA) di-
rection), Y is perpendicular to the nose direction (side-to-side (SS) di-
rection). The emergency stop is artificially triggered through the con-
trol system. The operation parameters and the vibration signals are
recorded with a sampling ratio of 1280 Hz. Fig. 15 shows the recorded
signals during the complete emergency stop process. The recorded
operation parameters (pitch angle, generator shaft rotating speed,
torque and yaw angle Fig. 15 (a)) are collected from the analog term-
inals of the control system in the test. During this test, the emergency
stop is triggered at about 26 s and the hydraulic brake is applied to stop
the generator shaft immediately. The pitch angle is changed from −1.2
to 88.8°and the rotating speed is decreased from about 1500 to 0 rpm in
a couple of seconds. Because of this sudden collective pitch variation,
the thrust dramatic change excites the tower mainly in the wind (FA)
direction. The vibration response of the SS direction accelerometer is
fainter than that of the FA direction (Fig. 15 (b)). Therefore the vi-
bration signals of the two FA direction accelerometers are used to
further analysis. At 39 s, the brake process is completed and the gen-
erator shaft rotating speed is decreased to 0 rpm. The rotor-gearbox-
generator system is at rest, so only the response of tower arousing by
emergency shutdown process and noise are included in the measured
vibration signal. The recorded data after 39 s of the two FA direction
accelerometers are selected to identify the modal information of the
tower.
The FA direction sensor in the middle of wind turbine's bed plate is
named X1, and the FA direction sensor in the tail is named X2. Fig. 16
shows the time waveform and amplitude spectrum of the vibration
response of the wind turbine tower from X1 sensor. The first three peaks
are marked in the amplitude spectrum. It can be seen the dominant
frequency of the dynamic response is clear, which is 0.3875 Hz, in
A. Hu et al. ISA Transactions xxx (xxxx) xxx–xxx

Fig. 24. RDT and HT of BLIMF2 from X2: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.

Fig. 25. RDT and HT of BLIMF3 from X2: (a) mode wave by RDT, (b) the corresponding linear least-squares fit.

Table 3
The identified results of the vibration responses from X1 and X2.
Sensor X1 X2

Modal Frequency/Hz Damping Frequency/Hz Damping

parameters ratio/% ratio/%

1st mode 0.40 0.69 0.40 0.68

2nd mode 1.70 0.49 1.70 0.24
3rd mode 3.13 1.41 3.49 1.15

agreement with the first mode natural frequency. However, the other
order modal frequencies are faint and difficult to be recognized in ex-
periments. The proposed modal identification method above is applied
to obtain the modal parameters from the vibration responses.
Fig. 26. Decomposed result of EMD.
Fig. 17 shows the decompositions of vibration signal from X1 sensor
using VMD. The K value is set to 4 according to the selecting method in
simulation part. The original vibration signal from X1 is decomposed parameters of the signal. According to the principle of RDT, the selec-
into four modes. Fig. 18 shows the recognition results of the first mode tion of sample segments should contain the entire stationary modal
from VMD. Further, the RDT and HT are applied to identify the modal response for obtaining good estimation accuracy [40,41]. The duration
of the free decay response is chosen corresponding to frequency of the

12
A. Hu et al. ISA Transactions xxx (xxxx) xxx–xxx

Fig. 27. EMD-RDT identification: (a) the first mode wave by RDT (b) the linear least-squares fit of the first mode (c)the second mode wave by RDT (d) the linear least-
squares fit of the second mode (c)the third mode wave by RDT (d) the linear least-squares fit of the third mode.

modal. The wave of BLIMF1 after RDT is shown in Fig. 18 (a). The
amplitude of the mode, instantaneous frequency and the corresponding
linear least-squares fit are exhibited in Fig. 18 (b). It shows that the first
modal information is identified. The first frequency is computed to
0.40 Hz and the damping ratio is 0.69%. Fig. 19 shows the recognition
results of the second mode from VMD. The wave of BLIMF2 after RDT is
shown in Fig. 19 (a). The amplitude of the mode, instantaneous fre-
quency and the corresponding linear least-squares fit are exhibited in
Fig. 19 (b). The second frequency is computed to 1.70 Hz and the
damping ratio is 0.49%. Fig. 17 shows the recognition results of the
third mode. The wave of BLIMF3 after RDT is shown in Fig. 20 (a). The
amplitude of the mode, instantaneous frequency and the corresponding

13
A. Hu et al.
Fig. 28. VMD identification: (a) the wave of BLIMF1 (b) the linear least-squares fit of BLIMF1 (c) the wave of BLIMF2 (d) the linear least-squares fit of BLIMF2 (c) the
wave of BLIMF3 (d) the linear least-squares fit of the wave of BLIMF3.
linear least-squares fit are exhibited in Fig. 20 (b). The third frequency
is computed to 3.13 Hz and the damping ratio is 1.41%.
X2 FA direction sensor is mounted in the tail of wind turbine's bed
plate. Fig. 21 shows the time waveform and amplitude spectrum of the
vibration response from X2 sensor. It indicates that the dominant nat-
ural frequency of the dynamic response is clear, and the value is
14
ISA Transactions xxx (xxxx) xxx–xxx
0.3875 Hz which is same as Fig. 16. Likewise, the other order modal
frequencies are faint and cannot be recognized, so the proposed modal
identification method is applied to obtain the modal parameters from
the vibration responses. Fig. 22 shows the decompositions of vibration
signal from X2 sensor using VMD, where the original signal from X2 is
decomposed into four modes. Fig. 23 shows the recognition results of
A. Hu et al.
Table 4
Identification results of EMD-RDT-HT and VMD-HT.
method EMD-RDT-HT VMD-HT
Modal Frequency/Hz Damping Frequency/Hz Damping
parameters ratio/% ratio/%
1st mode 0.40 0.61 0.40 0.53
2nd mode 1.64 0.32 1.69 0.21
3rd mode failed failed 3.45 0.14
the first mode from VMD. The wave of BLIMF1 after RDT is shown in
Fig. 23 (a). The instantaneous frequency and amplitude of the mode
and the corresponding linear least-squares fit are shown in Fig. 23 (b). It
is found that the first modal information is identified applying the VMD.
The first frequency is computed to 0.40 Hz and the damping ratio is
0.68%. Fig. 24 shows the recognition results of the second mode from
VMD. The wave of BLIMF2 after RDT is shown in Fig. 24 (a). The
amplitude of the mode, instantaneous frequency and the corresponding
linear least-squares fit are exhibited in Fig. 24 (b). The second fre-
quency is computed to 1.70 Hz and the damping ratio is 0.24%. Fig. 25
shows the recognition results of the third mode from VMD. The wave of
BLIMF3 after RDT is shown in Fig. 25 (a). The amplitude of the mode,
instantaneous frequency and the corresponding linear least-squares fit
are exhibited in Fig. 25 (b). The third frequency is computed to 3.49 Hz
and the damping ratio is 1.15%.
The identification results of X1 and X2 sensors are shown in Table 3.
It can be seen that the first two modal frequencies and the first mode
damping ratio identified from two sensors in same order are consistent
exactly, and the third mode frequency and the last two modal damping
ratios have a little difference between the two sensors. Combining the
spectrum of the two response signals, it can be seen that the first mode
is dominant in the spectrum. The amplitude of the first mode frequency
is much higher than that of the second and third mode. Due to the
environment noise, small difference occurs in the modal parameters
identification of the low amplitude mode. It can be concluded that the
first three mode parameters, including two weak modes (second and
third) are identified by the proposed VMD-RDT method through two
accelerometers located on the different position and the results are
coincident approximately. The sensors are commonly mounted on the
turbine tower at different level in modal parameter test [12,13]. Con-
sidering the large wind turbine with 70 m tower height, it is a hard
work in sensor mounting and long signal cable locating. With the
proposed method, the operational mode parameters of the wind turbine
tower can be identified using only one accelerometer in the nacelle
under emergency stop condition. It can be easily performed in the field
modal test.
The EMD-RDT-HT and VMD-HT are used to identify the structure
parameters. The measured vibration responses of X2 are processed in
the two methods for comparison. Fig. 26 shows the three components of
eight IMFs decomposition resulting from EMD. The identified results of
EMD-RDT-HT are presented in Fig. 27. The first frequency is computed
to 0.40 Hz and the damping ratio is 0.61%. It can be seen the first modal
frequency is in accordance with the result in Table 3. And the damping
ratio is close to 0.68% which is identified by the proposed method. The
first modal parameter is identified by the EMD-RDT-HT. The second
frequency and damping ratio is computed to 1.64 Hz and 0.32% by the
EMD-RDT-HT (Fig. 26 (c) and (d)), respectively. The results are ap-
proximate to the values in Table 3. However, the third modal para-
meters are failed to be identified by EMD-RDT-HT method.
According to the VMD-HT method [33], the BLIMFs decomposed by
VMD (Fig. 22) were directly used to parameter identification. The
amplitude of the mode, instantaneous frequency and the corresponding
linear least-squares fit are exhibited in Fig. 28. The identification results
of EMD-RDT-HT and VMD-HT are listed in Table 4. It can be seen the
identified modal parameters by VMD-HT are very close to those of the
15
ISA Transactions xxx (xxxx) xxx–xxx
VMD-RDT-HT method except the third modal damping ratio. Thus, the
validity of the proposed method is proved by the comparison result. At
the same time, the large fluctuation is presented in the amplitude of the
first modal response of VMD-HT (left in Fig. 28 (b)) and the attenuation
trend of the second and third modal response is not significant in
Fig. 28 (c) and (e).
The spectrum of the vibration responses (Figs. 16 and 21) indicates
that the first mode frequency is dominant. The amplitude of the second
and third mode is very small. However, the second modal response is
extracted after EMD (Fig. 27(c)) and the second and third modal re-
sponses are also extracted by VMD (Fig. 28 (c) and (e)). The results
indicate the existence of the weak modes. The identification results
show that the third modal parameters are failed to be identified by
EMD-RDT-HT method and the attenuation trend of the second and third
modal response is not significant in the VMD-HT result. In fact, the
purpose of RDT is to eliminate the random part and extract the de-
terministic free attenuation signal from the response. Therefore, the
identification accuracy of the decomposed responses by VMD can be
improved after RDT process.
5. Conclusions
Emergency stop is utilized to excite the response of large wind
turbine in field modal testing. The global motion and dynamic struc-
tural response of the wind turbine are activated during emergency stop.
The output-only response under emergency stop is adopted for the
operational modal parameter identification of large wind turbine tower.
The VMD-RDT modal parameter identification method is proposed in
this paper. The output-only vibration responses under emergency stop
are analyzed. In simulations, it is shown that the VMD scheme clearly
outperforms EMD with regards to sampling and noise robustness. The
field test shows that the first mode of the wind turbine tower is domi-
nant and other modes are weak. The main modal parameters, including
two weak modes of a 1.5 MW onshore wind turbine with 70 m hub
height are obtained through the proposed identification procedure. It
indicates that field modal test can easily be performed using accel-
erometer in the nacelle under emergency stop condition. The effec-
tiveness is verified through comparison analysis. The proposed identi-
fication technique can be applied to the large and flexible structures.
The more veracious mode parameter identification method of the very
low energy mode is necessary to be further investigated.
Acknowledgements
This work was supported by the National Natural Science
Foundation of China (Grant No. 51475164 and No. 51675178).
References
[1] Liu WY, Tang BP, Jiang YH. Status and problems of wind turbine structural health
monitoring techniques in China. Renew Energy 2010;35:1414–8.
[2] Baqersad J, Poozesh P, Niezrecki C, et al. Comparison of modal parameters ex-
tracted using MIMO, SIMO, and impact hammer tests on a three-bladed wind tur-
bine. Conference proceedings of the society for experimental mechanics series, to-
pics in modal analysis II. vol. 8. 2014. p. 185–97. [Orlando, Florida USA].
[3] Osgood R, Bir G, Mutha H, et al. Full-scale modal wind turbine tests: comparing
shaker excitation with wind excitation. Structural Dynamics and Renewable Energy
2011;1:113–24. Springer New York.
[4] Prowell I, Uang CM, Elgamal A, et al. Shake table testing of a utility-scale wind
turbine. J Eng Mech 2012;138(7):900–9.
[5] Carne TG, Lauffer JP, Gomez AJ. Modal testing of a very flexible 110m wind turbine
structure. Proceedings of the sixth international modal analysis conference. 1988. p.
848–55. Kissimmee, Florida, USA.
[6] Zhang L, Wang T, Tamura Y. A frequency–spatial domain decomposition (FSDD)
method for operational modal analysis. Mech Syst Signal Process
2010;24(5):1227–39.
[7] Richardson M, Schwarz B. Modal parameter estimation from operating data. Sound
Vib 2003:28–36.
[8] James GH, Carne TG, Lauffer JP. The natural excitation technique (NExT) for modal
parameter extraction from operating structures. Int J Anal Exp Modal Anal
1995;10(4):260–77.
A. Hu et al.
[9] Carne TG, James GH. The inception of OMA in the development of modal testing
technology for wind turbines. Mech Syst Signal Process 2010;24:1213–26.
[10] Molinari M, Pozzi M, Zonta D, et al. In-field testing of a steel wind turbine tower.
Structural Dynamics and Renewable Energy 2011;1:103–12.
[11] Fujiyama C, Yonetsu K, Maeshima T, et al. Identifiable stress state of wind turbine
tower-foundation system based on field measurement and FE analysis. Procedia
Engineering 2014;95:279–89.
[12] Shirzadeh R, Devriendt C, Bidakhvidi MA, et al. Experimental and computational
damping estimation of an offshore wind turbine on a monopile foundation. J Wind
Eng Ind Aerod 2013;120(120):96–106.
[13] Devriendt C, Jordaens PJ, Sitter GD, et al. Damping estimation of an offshore wind
turbine on a monopile foundation. Renewable Power Generation 2013;7(4):401–12.
[14] Damgaard M, Ibsen LB, Andersen LV, et al. Cross-wind modal properties of offshore
wind turbines identified by full scale testing. J Wind Eng Ind Aerod
2013;116(5):94–108.
[15] Bas J, Smith J, Carriveau R, et al. Structural response of a commercial wind turbine
to various stopping events. Wind Eengineering 2012;36(5):553–69.
[16] Brincker R, Zhang L, Andersen P. Modal identification from ambient responses
using frequency domain decomposition. Proceedings of the 18th IMAC. 2000. San
Antonio, Texas, USA.
[17] Brincker R, Ventura C, Andersen P. Damping estimation by frequency domain
Decomposition. Proceedings of IMAC XIX. 2001. Kissimmee, USA.
[18] Ibrahim SR, Mikulcik EC. A method for the direct identification of vibration para-
meters from the free response. Shock Vib Bull 1977;47(4):35–42.
[19] Fukuzono K. Investigation of multiple-reference Ibrahim time domain modal
parameter estimation technique MS. thesis University of Cincinnati; 1986.
[20] Ku CJ, Cermak JE, Chou LS. Random decrement based method for modal parameter
identification of a dynamic system using acceleration responses. J Wind Eng Ind
Aerod 2007;95(6):389–410.
[21] Juang JN, Pappa RS. An eigensystem realization algorithm for modal parameters
identification and model reduction. J Guid Contr Dynam 1985;8(5):620–7.
[22] Brown DL, Allemang RJ, Zimmerman R, Mergeay M. Parameters estimation tech-
niques for modal analysis. Society of Automotive Engineers Technique Paper
790221 1979.
[23] Vold H, Kundrat J, Rocklin GT, Russel RA. Multi-input modal estimation algorithm
for mini-computers SAE Technical Paper Series, No. 820194 1982.
[24] Gautier PE, Gontier C, Smail M. Robustness of an ARMA identification method for
modal analysis of mechanical systems in the presence of noise. J Sound Vib
1995;179(179):227–42.
[25] Peeters B, De Roeck G. The Identification of Mechanical Systems:stochastic system
identification for operational modal analysis: a review. J Dyn Syst Meas Contr
16
ISA Transactions xxx (xxxx) xxx–xxx
2001;123(4):659–67.
[26] Yang JN, Lei Y, Pan SW, Huang N. System identification of linear structures based
on Hilbert–Huang spectral analysis. Part I: normal modes. Earthq Eng Struct Dynam
2003;32(9):1443–67.
[27] Yang JN, Lei Y, Pan SW, Huang N. System identification of linear structures based
on Hilbert–Huang spectral analysis. Part II: complex modes. Earthq Eng Struct
Dynam 2003;32(10):1533–54.
[28] Huang TL, Lou ML. System identification of nonlinear structures based on HHT.
Earthq Eng Eng Vib 2006;26(3):80–3.
[29] Pines DJ, Salvino LW. Structural health monitoring using empirical mode decom-
position and the Hilbert phase. J Sound Vib 2006;294(1–2):97–124.
[30] Peng ZK, Peter WT, Chu FL. An improved Hilbert–Huang transform and its appli-
cation in vibration signal analysis. J Sound Vib 2004;286:187–205.
[31] Bao CX, Hao H, Li ZX, et al. Time-varying system identification using a newly im-
proved HHT algorithm. Comput Struct 2009;87(23–24):1611–23.
[32] Wang Z, Chen G. Analytical mode decomposition with Hilbert transform for modal
parameter identification of buildings under ambient vibration. Eng Struct
2014;59(2):173–84.
[33] Bagheri Abdollah, Ozbulut Osman E, Harris Devin K. Structural system identifica-
tion based on variational mode decomposition. J Sound Vib 2018;417:182–97.
[34] Dragomiretskiy K, Zosso D. Variational mode decomposition. IEEE Trans Signal
Process 2014;62(3):531–44.
[35] Liu YY, Yang GL, Li M, et al. Variational mode decomposition denoising combined
the detrended fluctuation analysis. Signal Process 2016;125:349–64.
[36] Wang YX, Markert R, Xiang JW, et al. Research on variational mode decomposition
and its application in detecting rub-impact fault of the rotor system. Mech Syst
Signal Process 2015;60–61:243–51.
[37] Xue YJ, Cao JX, Wang DX, et al. Application of the variational-mode decomposition
for seismic time-frequency analysis. IEEE Journal of Selected Topics in Applied
Earth Observations and Remote Sensing 2016;9(8):3821–31.
[38] Zhang YC, Liu KP, Qin L, et al. An Deterministic and probabilistic interval predic-
tion for short-term wind power generation based on variational mode decomposi-
tion and machine learning methods. Energy Convers Manag 2016;112:208–19.
[39] Wang Y, Markert R. Filter bank property of variational mode decomposition and its
applications. Signal Process 2015;120:509–21.
[40] Yang JN, Lei Y, Lin S, et al. Identification of natural frequencies and dampings of in
situ tall buildings using ambient wind vibration data. J Eng Mech
2004;130(5):570–7.
[41] Perez-Ramirez CA, Amezquita-Sanchez JP, Adeli H, et al. New methodology for
modal parameters identification of smart civil structures using ambient vibrations
and synchrosqueezed wavelet transform. Eng Appl Artif Intell 2016;48:1–12.