sensors

Article
Joint High-Order Synchrosqueezing Transform and
Multi-Taper Empirical Wavelet Transform for Fault
Diagnosis of Wind Turbine Planetary Gearbox under
Nonstationary Conditions
Yue Hu ID
, Xiaotong Tu ID
, Fucai Li * and Guang Meng
State Key Laboratory of Mechanical System and Vibration, Shanghai Jiao Tong University, Shanghai 200240,
China; huyue_sjtu@sjtu.edu.cn (Y.H.); tormii@sjtu.edu.cn (X.T.); gmeng@sjtu.edu.cn (G.M.)
* Correspondence: fcli@sjtu.edu.cn; Tel.: +86-21-3420-6531

Received: 13 November 2017; Accepted: 4 January 2018; Published: 7 January 2018

Abstract: Wind turbines usually operate under nonstationary conditions, such as wide-range
speed fluctuation and time-varying load. Its critical component, the planetary gearbox, is prone
to malfunction or failure, which leads to downtime and repair costs. Therefore, fault diagnosis
and condition monitoring for the planetary gearbox in wind turbines is a vital research topic.
Meanwhile, the signals measured by the vibration sensors mounted in the gearbox exhibit
time-varying and nonstationary features. In this study, a novel time-frequency method based
on high-order synchrosqueezing transform (SST) and multi-taper empirical wavelet transform
(MTEWT) is proposed for the wind turbine planetary gearbox under nonstationary conditions.
The high-order SST uses accurate instantaneous frequency approximations to obtain a sharper
time-frequency representation (TFR). As the acquired signal consists of many components, like
the meshing and rotating components of the gear and bearing, the fault component may be
masked by other unrelated components. The MTEWT is used to separate the fault feature from the
masking components. A variety of experimental signals of the wind turbine planetary gearbox under
nonstationary conditions have been analyzed to demonstrate the effectiveness and robustness of the
proposed method. Results show that the proposed method is effective in diagnosing both gear and
bearing faults.

Keywords: wind turbine gearbox; fault diagnosis; synchrosqueezing transform; multi-taper;

empirical wavelet transform; time-frequency analysis

1. Introduction
Wind energy is the fastest growing renewable energy source due to its cleanness and
reproducibility [1]. However, wind turbines are subjected to dynamic excitations such as varying
wind speeds, electricity grid events, or sea waves [2,3]. They are more prone to failure than in-house
generation units. As wind turbines are usually located at remote land/water locations that may
be difficult to access, their maintenance is difficult, expensive, and challenging. Therefore, fault
diagnosis and condition monitoring of wind turbines are important for reducing downtime and
economic loss [4–7]. Furthermore, fault diagnosis and detection in wind turbines can be accomplished
by sensor-based methods [8–10] (also called signal-based methods) or analytical model-based
methods [11–14].
A planetary gearbox is widely used in wind turbines to meet the requirements of smaller size
or higher torque output. In a wind turbine, the planetary gearbox transfers the mechanical energy
from the rotor hub with a low rotational speed to the generator with a high speed. The planetary

Sensors 2018, 18, 150; doi:10.3390/s18010150 www.mdpi.com/journal/sensors

Sensors 2018, 18, 150 2 of 18

gearbox malfunctions easily due to severe working conditions, such as wide-range speed fluctuation,
time-varying load, extreme operational temperature, and humidity changes [10]. The downtime and
repair costs significantly increase the operation cost of wind turbines. Hence, the fault diagnosis
for planetary gearboxes in wind turbines is a vital research topic. Many researchers have made
considerable contributions to planetary gearbox fault diagnosis recently. Some scholars [15–19]
studied the effects of gear faults, manufacturing errors, and loading on the vibration responses
through dynamics modeling and analysis. Lei [20] and Bartelmus [21,22] respectively proposed
statistical indices for the condition monitoring of a planetary gearbox under constant and nonstationary
operations. Feng [23–25] summarized the spectral characteristics of planetary gearbox vibration signals.
The industrial trend of wind energy leans towards multi-megawatt and large wind turbines,
whose size has increased from 20 kW to 5 MW during the last two decades [26,27]. To capture
wind energy efficiently, most large wind turbines operate at varying speeds [10]. Consequently,
the signals collected from vibration sensors mounted in the wind turbine planetary gearbox are
characterized as nonstationary features. Extracting the time-variant fault information from these
signals is challenging for the fault diagnosis of a wind turbine planetary gearbox under nonstationary
conditions. Time-frequency analysis methods can effectively reveal the time-variant features and have
been commonly used to extract the fault information from nonstationary signals [28,29]. Tang [30–32]
presented different methods based on wavelet transform for extracting transient features for diagnosing
wind turbine gearbox faults. Liu [27] used local mean decomposition (LMD) to analyze the gearbox
vibration signal and to identify the fault characteristic frequency of the gear. Feng [33] employed
the demodulation analysis methods based on ensemble empirical mode decomposition (EEMD) and
energy separation to diagnose the chipped tooth fault of the sun gear. However, the short time Fourier
transform (STFT) and wavelet transform are limited by the Heisenberg uncertainty principle [34].
The best time localization and finest frequency resolution cannot be accomplished simultaneously.
The empirical mode decomposition method (EMD) has four drawbacks, i.e., lack of mathematical
theory, boundary effect, mode mixing, and stop criteria [35,36]. The LMD method is ineffective in
dealing with the signals containing components whose sidebands are closely spaced.
Recently, Daubechies [37] proposed synchrosqueezing transform (SST) to reveal the complicated
time-frequency features of nonstationary signals. SST has been successfully used in a variety of fields,
such as medicine [38], geophysics [39–41], and rotating machinery [42,43]. SST is a special reassignment
algorithm [44,45] and corresponds to a nonlinear operator that improves the time-frequency (TF)
resolution. In SST, the time-frequency coefficients are reallocated only in the frequency axis to
preserve the causality [46]. However, SST and most of its variations [47–50] could only sharpen
the time-frequency representation (TFR) for signals with “slowly varying” instantaneous frequency
(IF). A novel SST method, named high-order SST, was proposed by Pham et al. [51]. This method
defines new synchrosqueezing operators based on high order amplitude and phase approximations
and produces a highly concentrated TFR for both signals with fast varying IF and slowly varying IF.
In this study, a time-frequency method based on high-order synchrosqueezing transform and
multi-taper empirical wavelet transform is proposed for the fault detection of wind turbine planetary
gearboxes under nonstationary conditions. The high-order SST produces a highly concentrated TFR.
The signals of a wind turbine exhibit typically time-varying features due to the complex operating
conditions, such as speed fluctuation and load variation. The original signal is resampled with the
shaft speed to avoid the speed fluctuation and spectrum smearing. Generally, a tachometer is required
to obtain the instantaneous speed to sample signals equi-angularly. This requirement is barely satisfied
in many applications due to limited space and cost [52]. The instantaneous rotating speed is estimated
via the dynamic path optimization ridge detection (DPORD) algorithm [53] based on the TFR. Also,
the estimation precision is guaranteed by high-quality time-frequency distribution calculated by the
high-order SST. Planetary gearbox signals consist of many complex components. Unwanted gear (sun,
planet, and ring) and/or bearing components may exist and easily mislead the fault diagnosis [54].
Multi-taper empirical wavelet transform (MTEWT) is employed to separate the fault features of
Sensors 2018, 18, 150 3 of 18

the planetary gearbox from other irrelevant components. The multi-tapering approach can reduce
the variance and improve the accuracy of the power spectrum estimation. A signal is decomposed
into a set of principal modes by an adaptive filter bank constructed by the MTEWT. Both simulated
and experimental vibration signals have verified the performance of the proposed method on fault
diagnosis for planetary gearbox of wind turbines under nonstationary conditions.
The rest of this paper is organized as follows. Section 2 describes the proposed method in detail,
especially the high-order SST algorithm and the MTEWT algorithm. Section 3 gives the procedure of
the proposed method in detecting faults of the wind turbine planetary gearbox under nonstationary
conditions. In Section 4, the effectiveness and robustness of the proposed method is verified by two
experiments. Finally, conclusions are given in Section 5.

2. The Proposed Method

2.1. High-Order SST

As a post-processing method, SST has been used for other TFRs [42,49,55]. In this paper,
the derivation process of the high-order SST is based on the STFT-based SST.
The signal acquired by the vibration sensor mounted in the wind turbine planetary gearbox can be
modeled as a typical amplitude-modulated and frequency-modulated (AM-FM) signal. The high-order
SST achieves a highly energy concentrated TFR for an AM-FM signal by using accurate approximations
of the amplitude and phase.
An AM-FM signal is modeled as:

s(τ ) = A(τ )ei2πφ(τ ) (1)

where A(τ) and φ(τ) are the amplitude and phase functions; both are real functions.
Then the STFT of this signal at time t and frequency f is defined as:
Z
g
Vs (t, f ) = s(τ ) g∗ (τ − t)e−i2π f (τ −t) dτ (2)

where g is the window function and g* is the complex conjugate of g.

The traditional SST is defined as follows:
1
Z
g,ε g
Ts (t, ω ) = Vs (t, f )δ(ω − ω̂s (t, f ))d f (3)
g ∗ (0) g
{|Vs (t, f )|>ε}

where ε is the threshold and δ is the Dirac distribution. ω̂s (t, f ) denotes the IF and is estimated as:

g
( )
∂t Vs (t, f )
ω̂s (t, f ) = R g (4)
i2πVs (t, f )

where R{Z} represents the real part of Z and ∂t is the partial derivative with respect to t.
g
In SST, the coefficients Vs (t, f ) are reassigned in the time-frequency plane. However, the SST
method uses a zero-order estimation to extract the true IF. It only works well in processing a signal
with a slowly time-varying IF. In the high-order SST method, the IF estimation is based on high order
Taylor expansions of the amplitude and phase of a signal to improve the energy concentration of
the TFR.
The Taylor expansion of the signal in Equation (1) for τ close to t is given as:

N
[log( A)](k) (t) + i2πφ(k) (t)
s(τ ) = exp( ∑ (τ − t)k ) (5)
k =0
k!

where F(k) (t) denotes the kth derivative of F at t.

Sensors 2018, 18, 150 4 of 18

Then Equation (2) is written as:

g
Vs (t, f ) = s(τ + t) g∗ (τ )e−i2π f τ dτ
R
N [log( A)](k) (t)+i2πφ(k) (t)) (6)
= exp( ∑ )τ k g∗ (τ )e−i2π f τ dτ
R
k!
k =0

g g
The STFT Vs (t, f ) is differentiated with respect to t and divided by i2πVs (t, f ). The above
equation is given as:

g t k −1 g
∂t Vs (t, f ) [log( A)]0 (t) N
[log( A)](k) (t) + i2πφ(k) (t) Vs (t, f )
g = + φ0 (t) + ∑ g (7)
i2πVs (t, f ) i2π k =2
i2π (k − 1)! Vs (t, f )

t k −1 g
where Vs (t, f ) is the STFT of the signal s(t) computed with window tk −1 g(t).
[N]
The Nth-order complex IF ω e s (t, f ) is defined as:

[N] [log( A)]0 (t)

e s (t, f ) =
ω + φ0 (t) (8)
i2π
 
[N] [N] [N]
Its real part ω̂s (t, f ) = R ωe s (t, f ) is the desired IF φ0 (t). Moreover, ω
e s (t, f ) can be
calculated as:
g N t k −1 g
∂t Vs (t, f ) [log( A)](k) (t) + i2πφ(k) (t) Vs (t, f )
−∑
[N]
e s (t,
ω f) = g g (9)
i2πVs (t, f ) k=2 i2π ( k − 1 ) ! Vs (t, f )

The modulation operator qe[k,N ] is introduced and defined as:

[log( A)](k) (t) + i2πφ(k) (t)

qe[k,N ] = (10)
i2π (k − 1)!

[N]
e s (t, f ) is written as:
Then the Nth-order complex IF ω

g N t k −1 g
∂t Vs (t, f ) V (t, f )
− ∑ qe[k,N ] s g
[N]
e s (t,
ω f) = g (11)
i2πVs (t, f ) k=2 Vs (t, f )

Furthermore, qe[k,N ] can be calculated as:


 y N (t, f )
 i f k = N;
[k,N ] N
qe = [ j,N ] i f k < N; (12)
 yk (t, f ) − ∑ x j,k (t, f )qe

j = k +1

where y j (t, f ) and xk,j (t, f ) are defined as follows:

 g
∂t Vs (t, f )
 g i f j = 1;
i2πVs (t, f )
y j (t, f ) = ∂ f y j−1 (t, f ) (13)

∂ f x j,j−1 (t, f )
i f j > 1;

t k −1 g

Vs (t, f )
i f j = 1;

 g
xk,j (t, f ) = Vs (t, f ) (14)
∂ f xk,j−1 (t, f )
i f j > 1;


∂ f x j,j−1 (t, f )
Sensors 2018, 18, 150 5 of 18

[N]
Then the Nth-order SST (SSTN) is defined by replacing ω̂s (t, f ) by ω̂s (t, f ) in Equation (3):
Sensors 2018, 18, 150 5 of 18
1
Z
g,ε g [N]
TN,s (t, ω )
= ∗ g
Vs (t, f )δ(ω − ω̂s (t, f ))d f (15)
g ( 0 ) {| }

g , 1 Vs ( t, f )|> ε
TN , s (t ,  )  * Vs (t , f ) (  ˆ s (t , f ))df
g [N ]
(15)
g (0) {|V (t , f )| } s
g

A simulated signal s(t) = A(t) cos(2πφ(t)) is used to evaluate the performance of the high-order
A simulated
SST. The signal signal
is sampled 𝑠(𝑡) =
at 4096 𝐴(𝑡)cos⁡
Hz. (2𝜋𝜙(𝑡)) is
Its amplitude andusedphase to are
evaluate
given theas: performance of the
high-order SST. The signal is (
sampled at 4096 Hz. Its amplitude and phase are given as:
A(t) = 10 + 20t22
 A(t )  10  20t (16)
φ(t) = 1200t − 5e−1.5t sin(30πt) (16)
 (t )  1200t  5e sin(30 t )
1.5t

The signal and its IF are shown in Figure 1. The IF ranges from 700 Hz to 1600 Hz and varies in
The signal and its IF are shown in Figure 1. The IF ranges from 700 Hz to 1600 Hz and varies in
one second. The large
one second. andand
The large fastfast
frequency
frequency variations
variationsmake
make time-frequency analysis
time-frequency analysis difficult.
difficult.

(a) 3
2
Amplitude

1
0
-1
-2
-3
0.0 0.2 0.4 0.6 0.8 1.0
Time [s]
(b) 2000
Frequency (Hz)

1500

1000

500

0
0.0 0.2 0.4 0.6 0.8 1.0
Time (s)

Figure 1. The
Figure 1. simulated signal:
The simulated (a)(a)
signal: ItsItstime
timedomain
domain waveform and(b)
waveform and (b)itsits instantaneous
instantaneous frequency.
frequency.

The TFR provided by different time-frequency methods are compared in Figure 2. In the TFR
The TFR provided by different time-frequency methods are compared in Figure 2. In the
provided by the STFT and the traditional SST methods, ridge smearing appears. The low
TFR provided by the
time-frequency STFT and
resolution the traditional
will greatly SST
decrease the methods,
precision of theridge smearing
IF estimation. Onappears. The low
the other hand,
time-frequency resolution
the TFR provided will
by the greatly decrease
third-order SST (SST3)theand
precision of the IFSST
the forth-order estimation. On theare
(SST4) methods other
much hand,
the TFR provided
sharpener thanbythose
the third-order
of the STFT SSTand (SST3) and theSST
the traditional forth-order
methods. SSTThe IF(SST4) methods
feature can be are much
easily
sharpener than those
interpreted based onof the
thesharpened
STFT andTFR theprovided
traditional SST
by the methods.
SST3 and SST4 The IF feature
approaches. can bethe
Therefore, easily
high-order SST method can effectively improve the energy concentration
interpreted based on the sharpened TFR provided by the SST3 and SST4 approaches. Therefore, of the TFR for the signal
with largeSST
the high-order and method
fast frequency variations.improve the energy concentration of the TFR for the signal
can effectively
SST3 and SST4 seem to
with large and fast frequency variations. behave similarly in terms of the energy concentration of the TFR.
Furthermore, computational costs of four time-frequency methods are considered. The algorithms
SST3 and SST4 seem to behave similarly in terms of the energy concentration of the TFR.
are run on a computer with an Intel Core i5-4460 CPU and 8.0 GB RAM. Then results of the
Furthermore, computational costs of four time-frequency methods are considered. The algorithms are
computational costs are listed in Table 1. The computational cost of the SST3 method is shorter than
run onthat
a computer
of the SST4with an Intel
method. Core
For the i5-4460
sake CPU and 8.0
of computational GB RAM.
efficiency, theThen
SST3 results
methodofis the computational
adopted in this
costs are listed
study. in Table 1. The computational cost of the SST3 method is shorter than that of the SST4
method. For the sake of computational efficiency, the SST3 method is adopted in this study.
Table 1. Computational costs of four time-frequency methods for a simulated signal.
Table 1. Computational costs of four time-frequency methods for a simulated signal.
Method STFT SST SST3 SST4
Computational cost (s) 4.01 7.45 22.12 42.52
Method STFT SST SST3 SST4
Computational cost (s) 4.01 7.45 22.12 42.52
Sensors 2018, 18, 150 6 of 18
Sensors 2018, 18, 150 6 of 18

Figure 2. The time-frequency representation (TFR) provided by (a) STFT; (b) SST; (c) SST3; and
Figure 2. The time-frequency representation (TFR) provided by (a) STFT; (b) SST; (c) SST3; and (d) SST4.
(d) SST4.

2.2. Multi-Taper Empirical

2.2. Multi-Taper Wavelet
Empirical Transform
Wavelet Transform
The multi-taper empirical
The multi-taper empirical wavelet
wavelettransform algorithmisisproposed
transform algorithm proposed to identify
to identify the fault
the fault features
features
from from a planetary
a planetary gearbox
gearbox vibration
vibration signal.
signal. As the As original
the original empirical
empirical wavelet
wavelet transform
transform (EWT)(EWT)
method
method yields unexpected segmentation due to false peaks and burrs induced
yields unexpected segmentation due to false peaks and burrs induced by noisy and non-stationary by noisy and
non-stationary
factors, factors, isthe
the multi-tapering usedmulti-tapering
to overcome these is used to overcome
shortcomings. Thethese shortcomings.
multi-tapering The can
approach
multi-tapering approach can reduce the variance and improve the accuracy of the power spectrum
reduce the variance and improve the accuracy of the power spectrum estimation. In this approach,
estimation. In this approach, several orthonormal functions are used as tapers to estimate the
several orthonormal functions are used as tapers to estimate the corresponding power spectrum.
corresponding power spectrum. The multi-taper spectral estimate is defined as:
The multi-taper spectral estimate is defined as:
1 L 1 mt
S mt ( f )  L S (f) (17)
1L k− 1 k
S ( f ) = ∑ Skmt ( f )
mt 0
(17)
L k =0
where

  hk (n) x(n)e
2
 i 2 fn
where Skmt ( f )  2 (18)
Skmt ( f ) = ∑ hk (n) x (n)e−i2π f n  (18)
 
And hk(n) is the kth taper orthogonal to other tapers:
And hk (n) is the kth taper orthogonal to other tapers:
 hk (n)hj (n)   kj (19)

In the spectrum segmentation part, ∑ hkboundaries

(n)h j (n) =ofδkjeach segment are determined by the (19)
spectrum peaks indicating dominant components. These peaks are picked out based on the
In the spectrum
spectrum segmentation
shape. Then part,isboundaries
the boundary described asof the
eachminimum
segmentinare determined
two byfrequency
consecutive the spectrum
peakspeaks.
indicating
Detailsdominant components.
of the spectrum segmentationThese
canpeaks
be found areinpicked out [56,57].
References based on the spectrum shape.
Then the Inspired
boundary by is
thedescribed as the
construction minimum in two
of Littlewood-Paley andconsecutive frequency
Meyer’s wavelets, peaks. Details
the empirical waveletsof the
are constructed as bandpass filters. The empirical
spectrum segmentation can be found in References [56,57]. scaling function and empirical wavelets are given
by the following expressions:
Inspired by the construction of Littlewood-Paley and Meyer’s wavelets, the empirical wavelets
are constructed as bandpass filters. The empirical scaling function and empirical wavelets are given by
the following expressions:
Sensors 2018, 18, 150 7 of 18


 1 h 
 i if |ω | ≤ ωn − τn
Φn (ω ) = cos π2 ν 2τ1n (|ω | − ωn + τn ) if ωn − τn ≤ |ω | ≤ ωn + τn (20)

0 otherwise




 1 h  i if ωn + τn ≤ |ω | ≤ ωn+1 − τn+1
1

 cos π v (| ω | − ωn+1 + τn+1 ) if ωn+1 − τn+1 ≤ |ω | ≤ ωn+1 + τn+1

2 2τ
Ψn (ω ) = h  n +1 i (21)
1


 sin π
2 v 2τn (| ω | − ω n + τn ) if ωn − τn ≤ |ω | ≤ ωn + τn

0 otherwise.


To obtain a wavelet tight frame, the parameter τ n satisfies the following equation:

ω n +1 − ω n
τn < ωn minn ( ) (22)
ω n +1 + ω n

According to the traditional wavelet theory, both approximated coefficients and detailed
coefficients are given by the inner products:
Z
W εf (0, t) = h f , φ1 i = f (τ )φ1 (τ − t) dτ (23)
Z
W εf (n, t) = h f , ψn i = f (τ )ψn (τ − t) dτ (24)

Then the empirical modes are defined as:

f 0 (t) = W εf (0, t) ∗ φ1 (t) (25)

f k (t) = W εf (k, t) ∗ ψk (t) (26)

An example is given to compare the performance of the EWT and the MTEWT method. The test
signal is defined as: (
cos(1.5t) · sin(100πt) 0 < t ≤ 0.5
signal = (27)
0.1 sin(400t) 0.5 < t ≤ 1

This signal and its constituents are shown in Figure 3. The low-frequency and high-frequency
components only exist in certain time periods. Then the EWT and MTEWT methods are used to
decompose the signal. The Fourier and the multi-taper spectrum along with boundaries detected by
both decomposition methods are depicted in Figure 4. The empirical modes decomposed by the EWT
and MTEWT are shown in Figure 5.
In Figure 4a, some false peaks occur in the Fourier spectrum due to the data truncation or other
nonstationary factors, which result in the improper boundary. In the EWT decomposition results shown
in Figure 5a,b, the low-frequency component and the high-frequency component are mixed together.
In Figure 4b, the curve of the multi-taper spectrum is smooth and the boundary detected by the
MTEWT is reasonable. In Figure 5c,d, both the low-frequency and high-frequency components are
clearly recovered by the MTEWT methods. Therefore, the MTEWT method performs better than the
EWT method in decomposing the nonstationary signal.
Sensors 2018, 18, 150 8 of 18
Sensors 2018,
Sensors 2018, 18,
18, 150
150 88 of
of 18
18

(a)1.0
(a) 1.0

Amplitude
Amplitude
0.5
0.5
0.0
0.0
-0.5
-0.5
-1.0
-1.0
0.0
0.0 0.2
0.2 0.4
0.4 0.6 0.8 1.0
(s) 0.6
Time (s)
Time
0.8 1.0
(b)1.0
(b) 1.0 (c)0.2
(c) 0.2

0.5
0.5 0.1
0.1
Amplitude

Amplitude
Amplitude

Amplitude
0.0
0.0 0.0
0.0
-0.5
-0.5 -0.1
-0.1
-1.0
-1.0 -0.2
-0.2
0.0 0.2
0.0 0.2 0.4
0.4 0.6
0.6 0.8
0.8 1.0
1.0 0.0 0.2
0.0 0.2 0.4
0.4 0.6
0.6 0.8
0.8 1.0
1.0
Time (s)
Time (s) Time (s)
Time (s)

Figure
Figure 3.(a)
Figure3.
3. (a)The
(a) Thetest
The testsignal;
test signal;(b)
signal; (b)its
(b) itslow-frequency
its low-frequencycomponent
low-frequency componentand
component and(c)
and (c)its
(c) itshigh-frequency
its high-frequencycomponent.
high-frequency component.
component.

(a) 1.0
(a) 1.0
Fourier spectrum
Fourier spectrum
EWT boundary
EWT boundary
Amplitude
Amplitude

0.5
0.5

0.0
0.0
00 100
100 200
200 300
300 400
400 500
500
Frequency (Hz)
Frequency (Hz)
(b) 1.0
(b) 1.0
multi-taper spectrum
multi-taper spectrum
0.8
0.8 MTEWT boundary
Amplitude

MTEWT boundary
Amplitude

0.6 0.02
0.02
0.6
0.4
0.4 0.01
0.01

0.2
0.2 0.00
0.00
190
190 200
200 210
210
0.0
0.0
00 100
100 200
200 300
300 400
400 500
500
Frequency (Hz)
Frequency (Hz)
Figure4.4.
Figure
Figure 4.TheThe
TheEWTEWT
EWT and
andand MTEWT
MTEWT
MTEWT detected
detected
detected boundaries:
boundaries:
boundaries: (a)
(a)
(a) The The Fourier
The
FourierFourier spectrum
spectrum
spectrum and the
and
and the EWT the EWT
EWT
detected
detected boundary;
detected boundary;
boundary; (b) the multi-taper
(b) the multi-taper
(b) the multi-taper spectrum
spectrum spectrum and the MTEWT
and thedetected
and the MTEWT detected
MTEWTboundary. boundary.
detected boundary.
Sensors 2018, 18, 150 9 of 18
Sensors 2018, 18, 150 9 of 18

(a) 1.0 EWT (b) 1.0 EWT

s1 s2
0.5 0.5
Amplitude

Amplitude
0.0 0.0

-0.5 -0.5

-1.0 -1.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Time (s) Time (s)

(c) 1.0 MTEWT (d) 1.0 MTEWT

s1 s2
0.5 0.5
Amplitude

Amplitude
0.0 0.0

-0.5 -0.5

-1.0 -1.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
Time (s) Time (s)

Figure5. 5.
Figure TheThe
EWT EWT and MTEWT
and MTEWT decomposition
decomposition results:
results: (a) (a)EWT
The first Themode
first and
EWT themode and the
low-frequency
low-frequency
component; component;
(b) the (b) mode
second EWT the second EWT
and the mode and the
high-frequency high-frequency
component; component;
(c) the first (c) the
MTEWT mode
firstthe
and MTEWT mode and
low-frequency the low-frequency
component; component;
(d) the second MTEWT mode (d) the
andsecond MTEWT mode
the high-frequency and the
component.
high-frequency component.
3. Procedure of the Proposed Method
3. Procedure of the Proposed Method
To detect the faults of the wind turbine planetary gearbox under nonstationary conditions, a new
To detect the
time-frequency faultsbased
method of theonwind turbineSST
high-order planetary gearbox
and MTEWT is under
proposed nonstationary conditions,
in this study. The faulta
new time-frequency
features are separated method based components
from masking on high-order or SST and MTEWT
speed-related is proposed
components. Theinprocedure
this study. The
of the
fault features are separated from
proposed method is summarized as follows: masking components or speed-related components. The procedure
of the proposed method is summarized as follows:
(1) Obtain the time-frequency representation via the high-order SST
(1) Obtain the time-frequency representation via the high-order SST
A sharpened TFR is obtained via the high-order SST. Compared with other time-frequency
A sharpened
methods, TFR is
the high-order SSTobtained
provides viaa highly
the high-order SST. time-frequency
concentrated Compared withrepresentation,
other time-frequency
which
methods, the
guarantees thehigh-order
precision ofSST provides a highly concentrated time-frequency representation, which
IF estimation.
guarantees the precision of IF estimation.
(2) Estimate the IF and resample the signal synchronous to the shaft instantaneous speed
(2) Estimate the IF and resample the signal synchronous to the shaft instantaneous speed
The IF is a crucial feature to recognize the time-frequency pattern of nonstationary signals in
The turbine.
the wind IF is a crucial
In thefeature
presentto study,
recognize
the the
IF istime-frequency pattern
estimated directly fromof the
nonstationary signalsasinthe
vibration signal, the
wind turbine. In the present study, the IF is estimated directly from the vibration
provision of a tachometer with sufficient resolution is not always possible due to technological or cost signal, as the
provision of a tachometer with sufficient resolution is not always possible
reasons. The IF is estimated based on the DPORD algorithm. Details about the DPORD algorithm can due to technological or
cost
be reasons.
found The IF [53,58].
in References is estimated based on the DPORD algorithm. Details about the DPORD
algorithm
Under nonstationaryin
can be found References
conditions, [53,58].
the fault features of the vibration signal vary with time. Therefore,
the direct application of spectrum-basedthe
Under nonstationary conditions, fault features
methods leads to of the vibration
spectrum smearing signal
and vary
false with time.
diagnosis.
Therefore,
The originalthe direct
signal applicationatof
is resampled spectrum-based
constant methods leads
angular increments of thetoestimated
spectrumshaft
smearing and false
instantaneous
diagnosis. The original signal is resampled at constant angular increments
speed. In this way, the spectrum smearing phenomenon is avoided and speed-related components of the estimated shaft
instantaneous
are separated. speed. In this way, the spectrum smearing phenomenon is avoided and speed-related
components are separated.
(3) Apply the MTEWT algorithm to extract the fault feature
(3) Apply the MTEWT algorithm to extract the fault feature
The wind turbines exhibit complex dynamic responses due to extreme loading and widely
varying operating conditions. The fault features are masked by unrelated components originating
Sensors 2018, 18, 150 10 of 18

The wind turbines exhibit complex dynamic responses due to extreme loading and widely varying
operating conditions. The fault features are masked by unrelated components originating from other
machine elements like shafts or gears. The resampled signal is decomposed by the MTEWT method
and the fault features are separated from masking components.
Sensors 2018, 18, 150 10 of 18
Sensors 2018, 18, 150 10 of 18
4. Experiments
from other machine elements like shafts or gears. The resampled signal is decomposed by the
from
The other machine
effectiveness elements like shafts
of the or gears. The resampled signal is decomposed by the
MTEWT method and the proposed method
fault features is validated
are separated by vibration
from masking components.signals collected from two
MTEWT method and the fault features are separated from masking components.
experiments. The first experiment is carried out on a drivetrain dynamics simulator. The second
4. Experiments
experiment is performed on a wind turbine planetary gearbox of an actual wind turbine. Meanwhile,
4. Experiments
the proposed method
The is also
effectiveness of compared
the proposed with the fast
method kurtogram
is validated method
by vibration to show
signals its effectiveness.
collected from two
The effectiveness of the proposed method is validated by vibration signals collected from two
experiments. The first experiment is carried out on a drivetrain dynamics simulator. The second
experiments. The first experiment is carried out on a drivetrain dynamics simulator. The second
4.1. Experiment
experiment1: is
A performed
MachineryonFault Simulator
a wind turbine planetary gearbox of an actual wind turbine. Meanwhile,
experiment is performed on a wind turbine planetary gearbox of an actual wind turbine. Meanwhile,
the proposed method is also compared with the fast kurtogram method to show its effectiveness.
Thethevibration
proposeddata areiscollected
method fromwith
also compared a Spectra
the fast Quest’s drivetrain
kurtogram method to dynamics simulator, which is
show its effectiveness.
used as 4.1.
a fault simulation
Experiment of the wind
1: A Machinery turbine planetary gearbox. As shown in Figure 6, the simulator
Fault Simulator
4.1. Experiment 1: A Machinery Fault Simulator
contains an Thealternating current (AC) motor, a planetary gearbox, a fixed shaft gearbox, and a magnetic
vibration data are collected from a Spectra Quest’s drivetrain dynamics simulator, which is
brake. used The
A tachometervibration data are collected
and an accelerometer from a Spectra Quest’s
are planetary
used todrivetrain
collect the dynamics
shaftsimulator,
speed which isand the
signal
as a fault simulation of the wind turbine gearbox. As shown in Figure 6, the
used as a fault simulation of the wind turbine planetary gearbox. As shown in Figure 6, the
gear vibration
simulatorsignal,
containsrespectively. The shaft
an alternating current (AC) rotating speed isgearbox,
motor, a planetary manipulated by gearbox,
a fixed shaft a speedand controller.
simulator contains an alternating current (AC) motor, a planetary gearbox, a fixed shaft gearbox, and
a magnetic
The sampling rate brake.
is A tachometer
12,800 Hz, and and
the an accelerometer
time length for are
the used
record to collect
is 20 the
s. Theshaft speed signal
schematic viewandand gear
a magnetic brake. A tachometer and an accelerometer are used to collect the shaft speed signal and
the gear
parameters vibration
of the signal,gearbox
planetary respectively.
are The
given shaft
in rotating
Figure speed
7. is manipulated by a speed controller.
the gear vibration signal, respectively. The shaft rotating speed is manipulated by a speed controller.
The sampling rate is 12,800 Hz, and the time length for the record is 20 s. The schematic view and
TheThe
effectiveness
sampling rateofis the proposed
12,800 Hz, and the method is verified
time length for theby detecting
record is 20 s. the
The tooth crack
schematic in and
view a sun gear
gear parameters of the planetary gearbox are given in Figure 7.
gear
which is given parameters of the planetary gearbox are given in Figure 7.
The in Figure 8. of
effectiveness Thethesun gear fault
proposed method order with reference
is verified by detecting tothethetooth
rotating
crack sun geargear
in a sun is 2.5.
The The effectiveness
collected vibration ofsignal
the proposed
and itsmethod
spectrumis verified by detecting
are reference
depicted the tooth9.crack in a sun gear
which is given in Figure 8. The sun gear fault order with tointheFigure
rotating sunMany
gear peaks
is 2.5. appear in
which is given in Figure 8. The sun gear fault order with reference to the rotating sun gear is 2.5.
the Fourier The
spectrum
collectedand it is difficult
vibration signal and to associate
its spectrum these peaks with
are depicted fault9.characteristic
in Figure Many peaks appear frequencies
in or
The collected vibration signal and its spectrum are depicted in Figure 9. Many peaks appear in
the
orders due Fourier spectrum
to the speed and it
variation. is difficult to associate these peaks with fault characteristic frequencies
the Fourier spectrum and it is difficult to associate these peaks with fault characteristic frequencies
To or
or
orders
reveal
orders
due
the to thefault,
crack
due
speed variation.
to the speedthe proposed method is employed to analyze the vibration signal. Firstly,
variation.
To reveal the crack fault, the proposed method is employed to analyze the vibration signal.
the TFR of the To signal is calculated
reveal the crack fault, viathe the high-order
proposed methodSST and shown
is employed in Figure
to analyze the 10.
vibration signal.
Firstly, the TFR of the signal is calculated via the high-order SST and shown in Figure 10.
Firstly, the TFR of the signal is calculated via the high-order SST and shown in Figure 10.

Figure 6. The test rig of the wind turbine planetary gearbox.

Figure
Figure 6. The
6. The testtest
rigrigofofthe
thewind
wind turbine
turbineplanetary gearbox.
planetary gearbox.

Planet gear
Planet gear
Zp= 40
Zp= 40

Sun gear
Sun gear
Zs= 20
Zs= 20

Ring gear
Ring gear
Zr= 100
Zr= 100

Figure
Figure . Schematic of
7. 7Schematic of the
theplanetary
planetarygearbox.
gearbox.
Figure 7. Schematic of the planetary gearbox.
Sensors 2018, 18, 150 11 of 18
Sensors 2018, 18, 150 11 of 18
Sensors 2018, 18, 150 11 of 18
Sensors 2018, 18, 150 11 of 18

Figure 8. The faulty sun gear with tooth crack: (a) The top view and (b) the side view.
8. The8.faulty
FigureFigure sunsun
The faulty gear with
gear tooth
with toothcrack: (a)The
crack: (a) Thetop
top view
view andand (b)side
(b) the theview.
side view.
Figure 8. The faulty sun gear with tooth crack: (a) The top view and (b) the side view.
(a) 1.0
(a) 1.0
(a) 1.0
(g)(g) (g)

0.5
Amplitude

0.5
0.5
Amplitude

0.0
Amplitude

0.0
0.0
-0.5
-0.5
-0.5
-1.0
-1.00 5 10
Time (s)
15 20
-1.00
(b) 0.03 5 10 15 20
0 5 Time10(s) 15 20
(b) 0.03 Time (s)
(b) 0.03
(g)(g) (g)

0.02
Amplitude

0.02
Amplitude

0.02
Amplitude

0.01
0.01
0.01
0.00
0.000 50 100 150 200 250
0.00 0 50 Frequency (Hz)
100 150 200 250
0 50 100 150
Frequency (Hz) 200 250
Frequency
Figure 9. The sun gear fault signal: (a) Its (Hz)waveform and (b) its spectrum.
time domain
Figure Figure
9. The9.
Figure
Thegear
9.sun
sun gear
fault
The sun gear
fault signal:(a)
signal:
fault
(a) Itstime
signal: (a)Its
time domain waveform
Its time domain waveform
domain waveform
andand
(b) its
(b)spectrum.
and (b) its its spectrum.
spectrum.

Figure 10. The TFR of the sun gear fault signal.

Figure 10. The TFR of the sun gear fault signal.
Figure 10. The TFR of the sun gear fault signal.
Figure 10. The TFR of the sun gear fault signal.

Secondly, the IF is estimated via the DPORD method based on the calculated TFR. Thirdly,
the original signal is resampled synchronous to the estimated IF. The resampled signal and its
order spectrum are shown in Figure 11a,b. However, the order spectrum fails to provide sufficient
information about the sun gear crack fault.
 Sensors 2018, 18, 150 12 of 18

Sensors 2018,Secondly,
18, 150 the IF is estimated via the DPORD method based on the calculated TFR. Thirdly, the
12 of 18
original signal is resampled synchronous to the estimated IF. The resampled signal and its order
spectrum are shown in Figure 11a,b. However, the order spectrum fails to provide sufficient
information
Finally, about the sun
the resampled gear crack
signal fault.
is decomposed into 13 empirical modes via the MTEWT method.
Finally, the resampled signal is decomposed
As shown in Figure 11c, the sun gear fault order can into
be13 empiricalin
identified modes via the
the order MTEWT of
spectrum method.
empirical
As shown in Figure 11c, the sun gear fault order can be identified in the order spectrum of empirical
mode 6. The identified fault order, marked with a red arrow, is 2.52, which is very close to the
mode 6. The identified fault order, marked with a red arrow, is 2.52, which is very close to the
theoretical value of 2.50. Therefore, the sun gear fault can be effectively diagnosed.
theoretical value of 2.50. Therefore, the sun gear fault can be effectively diagnosed.

(a) 1.0

0.5
Amplitude (g)

0.0

-0.5

-1.0
0 100 200 300 400 500 600 700
Angle (rad)
(b) 0.10
Amplitude (g)

0.05

0.00
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
Order
(c)0.015

Sun Gear Fault Order

Amplitude (g)

0.010

0.005

0.000
0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
Order

FigureFigure 11.The
11. (a) (a) The resampled
resampled sunsun gear
gear signal;(b)
signal; (b)the
the order
order spectrum
spectrumofof
the resampled
the signal;
resampled and (c)
signal; and (c)
the order spectrum of empirical
the order spectrum of empirical mode 6. mode 6.

4.2. Experiment 2: A Wind Turbine Planetary Gearbox

4.2. Experiment 2: A Wind Turbine Planetary Gearbox
The second set of experimental data is collected from a wind turbine planetary gearbox and is
The second
provided by set
the of experimental
diagnosis data isthe
contest during collected from aConference
International wind turbine planetary
on Condition gearbox of
Monitoring and is
provided by theindiagnosis
Machinery contest
Non-Stationary during the
Operations International
(CMMNO) in 2014.Conference on Condition Monitoring of
MachineryThe in vibration
Non-Stationary
signal is Operations
measured on(CMMNO)
the gearbox in 2014.of a wind turbine in the radial direction.
casing
The vibration
The kinematicssignal
of the wind turbine planetary
is measured gearboxcasing
on the gearbox is described in Figure
of a wind 12. The
turbine vibration
in the radialsignal
direction.
is sampled at 5 kHz with a time length of approximately 550 s. The amplitude
The kinematics of the wind turbine planetary gearbox is described in Figure 12. The vibration signal of this signal is is
normalized to unitary maximum amplitude. A tachometer is located on the high-speed
sampled at 5 kHz with a time length of approximately 550 s. The amplitude of this signal is normalized shaft
(carrying gear #7 in Figure 12) to give a reference speed signal. A fault is set on the inner race of the
to unitary maximum amplitude. A tachometer is located on the high-speed shaft (carrying gear #7
in Figure 12) to give a reference speed signal. A fault is set on the inner race of the planet bearing
and generates an angular periodic impulsive phenomenon at 0.191 high-speed shaft orders in the
signal [59]. And parameters of the planetary gearbox are given in Table 2.
Sensors 2018, 18, 150 13 of 18

planet
Sensors
bearing
Sensors2018,
2018, 18,150
18,
and
generates an angular periodic impulsive phenomenon at 0.191 high-speed13
150 13shaft
of18
of 18
orders in the signal [59]. And parameters of the planetary gearbox are given in Table 2.
planet bearing and generates an angular periodic impulsive phenomenon at 0.191 high-speed shaft
High Speed
orders in the signal [59]. And parameters of the planetary Shaft
gearbox are given in Table 2.

High Speed Shaft

Figure 12.
Figure 12. The
The planetary
planetary gearbox
gearbox kinematics
kinematics [59].
[59].

Table
Table 2.2.12.
Figure Gear
Gear parameters
parameters
The of
planetaryof theplanetary
the planetary
gearbox gearbox.
gearbox.
kinematics [59].
Gear 1 2 3 4 5 6 7 8 9
GearTable 2.1Gear parameters
2 3 of4 the planetary
5 6 gearbox.
7 8 9
Gear teeth 123 50(3) 21 93 22 120 29 63 23
Gear teeth 123 50(3) 21 93 22 120 29 63 23
Gear 1 2 3 4 5 6 7 8 9
The first goal is to estimate the instantaneous angular speed of the high-speed shaft directly
Gear teeth 123 50(3) 21 93 22 120 29 63 23
from the vibrational signal in order to address the case where no provision of a tachometer with
The first goal is to estimate the instantaneous angular speed of the high-speed shaft directly from
sufficient resolution
The first goal isto made
is in available
estimate for technological
the instantaneous angular or speed
cost reasons.
ofofthe The estimated
high-speed shaft result is
directly
the vibrational signal order to address the case where no provision a tachometer with sufficient
compared with the reference
from the vibrational signal forinspeed
order signal. The second
to address goal
the reasons.
case is tono
where diagnose
provision theof bearing faults inwiththe
resolution is made available technological or cost The estimated resultaistachometer
compared with
wind turbine planetary gearbox under nonstationary working conditions inThethe estimated
industrial result
field. In
the reference speed signal. The second goal is to diagnose the bearing faults in the wind turbineis
sufficient resolution is made available for technological or cost reasons.
this realistic
compared case,
with fault
theunder
referencefeatures
speed dosignal.
not correspond
The second togoal
priori
is expectations, leading to potential
planetary gearbox nonstationary working conditions in tothediagnose
industrial thefield.
bearing faults
In this in the
realistic
misinterpretation
wind turbine or falsegearbox
planetary diagnosis.
under nonstationary working conditions in the industrial field. In
case, fault features do not correspond to priori expectations, leading to potential misinterpretation or
The vibration
this diagnosis. signal and its spectrum are given in Figure 13. The
realistic case, fault features do not correspond to priori expectations, leading to potentialhigh-order SST algorithm is
false
employed to give a sharpened
misinterpretation TFR, which is shown in Figure 14a.
The vibrationor falseand
signal diagnosis.
its spectrum are given in Figure 13. The high-order SST algorithm is
The vibration signal
employed to give a sharpened TFR, and its spectrum
which isareshowngiven in in Figure
Figure 14a.13. The high-order SST algorithm is
(a) 1.0
employed to give a sharpened TFR, which is shown in Figure 14a.
0.5
Amplitude

(a) 1.0
0.0
0.5
Amplitude

-0.5
0.0
-1.0
-0.5 0 100 200 300 400 500
Time (s)
(b) 0.003
-1.0
0 100 200 300 400 500
Time (s)
Amplitude

(b)0.002
0.003
Amplitude

0.002
0.001

0.001
0.000
0 500 1000 1500 2000 2500
Frequency (Hz)
0.000
0 bearing fault
Figure 13. (a) The planet 500 signal provided
1000 by 1500 2000on Condition
the Conference 2500 Monitoring of
Frequency (Hz)
Machinery in Non-Stationary Operations (CMMNO) and (b) its envelope spectrum.
Figure13.
Figure 13.(a)
(a) The
The planet
planet bearing
bearing fault
fault signal
signal provided
providedbybythe
theConference
ConferenceononCondition
ConditionMonitoring
Monitoringof
of
Machinery in Non-Stationary Operations (CMMNO) and (b) its envelope
Machinery in Non-Stationary Operations (CMMNO) and (b) its envelope spectrum. spectrum.
Sensors 2018, 18, 150 14 of 18
Sensors
Sensors2018,
2018,18,
18,150
150 1414ofof1818

Figure 14. The

Figure TFR ofofthe CMMNO vibration signal.
Figure 14.14.The
TheTFR
TFRof theCMMNO
the CMMNOvibration
vibration signal.
signal.

Then the
Thenthe the DPORD
DPORD algorithm
algorithm isis employed
employed toto estimate
estimate the
the IF.IF. Furthermore,
Furthermore, the signal
signal isis
Then DPORD algorithm is employed to estimate the IF. Furthermore, the signal istheresampled
resampled in an angular domain to avoid the influences of speed fluctuation. The resampled signal
inresampled
an angularindomain
an angular domain
to avoid the to avoid theofinfluences
influences of speed fluctuation.
speed fluctuation. The resampled The signal
resampled signal
is shown
isis shown
shown along
along with its
with order envelope
its envelope order spectrum in Figure 15a,b. The resampled signal is
along with its envelope spectrumorder spectrum
in Figure 15a,b. inTheFigure 15a,b.signal
resampled The isresampled
decomposed signal
intois
decomposed
decomposed into
intoseven
seven modes.
modes.15c,As show
As show in Figure 15c,
in Figurefault the
15c,orderinner-race
the inner-race fault order
fault order can be identified
canenvelope
be identified in
seven modes. As show in Figure the inner-race can be identified in the orderin
the
the envelope
envelope order
order spectrum
spectrum ofofempirical
empirical mode
mode 5.5.The
The detected
detected inner-race
inner-race fault
faultorder
order isis0.1903,
0.1903,
spectrum of empirical mode 5. The detected inner-race fault order is 0.1903, which is very close to the
which is very
which is value close
very close to the
to the theoretical
theoretical value of 0.191.
value of 0.191. Therefore,
Therefore, the
the proposed
proposed method
method reveals
reveals the
the
theoretical of 0.191. Therefore, the proposed method reveals the presence of the inner-race fault
presence
presence of
ofthe
the inner-race
inner-race fault
fault ininone
one ofofthe
the planet
planet bearings.
bearings.
in one of the planet bearings.

Figure
Figure
Figure 15.
15.
15.(a)
(a)
(a)The
The
Theresampled
resampled
resampled signal;
signal;
signal;(b)
(b)(b)the
the envelope
envelope
the order
order
envelope spectrum
spectrum
order of of
spectrum ofthe
the theresampled
resampled signal;
signal;
resampled andand
signal; (c)
and
(c)
the the envelope
(c)envelope order
orderorder
the envelope spectrum
spectrum of empirical
of empirical
spectrum mode
modemode
of empirical 5.
5. 5.
Sensors 2018, 18, 150 15 of 18
Sensors
Sensors2018,
2018,18,
18,150
150 15
15of
of18
18

For
For comparison, thethe fast kurtogram method is also used to analyze the resampled signal to extractto
For comparison,
comparison, the fast fast kurtogram
kurtogram method
method isis also
also used
used toto analyze
analyze the the resampled
resampled signal
signal to
the fault
extract feature. The kurtogram is shown in Figure 16 and an optimal filter with the center frequency
extract the
the fault
fault feature.
feature. The
The kurtogram
kurtogram isis shown
shown in in Figure
Figure 1616 and
and an
an optimal
optimal filter
filter with
with the
the center
center
offrequency
3.21 and the3.21 bandwidth of 0.09 is determined. Then the resampled signal is filtered based on
frequency of of 3.21 and
and the
the bandwidth
bandwidth of of 0.09
0.09 isis determined.
determined. Then
Then the
the resampled
resampled signalsignal isis filtered
filtered
the above parameters
based and the envelope order spectrum of the filtered signal is shown inisFigure 17.
based onon the
the above
above parameters
parameters and and the
the envelope
envelope order
order spectrum
spectrum of of the
the filtered
filtered signal
signal is shown
shown in in
However,
Figure the inner-race fault order is indistinguishable in the envelope order spectrum of the filtered
Figure17.17.However,
However,the theinner-race
inner-racefault
faultorder
orderisisindistinguishable
indistinguishablein inthe
theenvelope
envelopeorder orderspectrum
spectrumof of
signal.
the The proposed The method is moremethod efficient than the efficient
fast kurtogramthe method in extracting the faultin
the filtered
filtered signal.
signal. The proposed
proposed method isis more more efficient thanthan the fast fast kurtogram
kurtogram method
method in
feature in this experiment.
extracting
extractingthe thefault
faultfeature
featurein inthis
thisexperiment.
experiment.

Figure
Figure
Figure 16.
16.
16. The
The
The kurtogramofof
kurtogram
kurtogram of the
thetheresampled
resampled
resampled signal.
signal.
signal.

Figure
Figure17. The envelope order spectrum of the filtered signal.
Figure 17.17.
TheThe envelope
envelope order
order spectrum
spectrum ofof
thethe filtered
filtered signal.
signal.

5.
5.Conclusions
Conclusions
5. Conclusions
In
Inthis
thisstudy,
study,aanovel
noveltime-frequency
time-frequencymethod
methodbased
basedon
onhigh-order
high-orderSST SSTand
andMTEWT
MTEWTisisproposed
proposed
In this study, a novel time-frequency method based on high-order SST and MTEWT is
for
for the
the fault
fault diagnosis
diagnosis of
of aa wind
wind turbine
turbine planetary
planetary gearbox
gearbox under
under nonstationary
nonstationary conditions.
conditions. The
The
proposed for the fault diagnosis of a wind turbine planetary gearbox under nonstationary conditions.
high-order SST defines new synchrosqueezing operators based on high order amplitude
high-order SST defines new synchrosqueezing operators based on high order amplitude and phase and phase
The high-order SST defines new synchrosqueezing operators based on high order amplitude and phase
approximations.
approximations. Compared
Compared with with other
other time-frequency
time-frequency analysis
analysis methods,
methods, the
the high-order
high-order SST
SST
approximations. Compared with other time-frequency analysis methods, the high-order SST produces
produces
produces aa highly
highly concentrated
concentrated TFRTFR for
for both
both signals
signals with
with fast
fast varying
varying IFIF and
and slowly
slowly varying
varying IF.
IF.
Meanwhile, the MTEWT method is proposed to separate the fault feature from masking
Meanwhile, the MTEWT method is proposed to separate the fault feature from masking components components
originating
originatingfromfromother
othermachine
machineelements
elementssuch
suchasasshafts
shaftsor
orgears.
gears.
Sensors 2018, 18, 150 16 of 18

a highly concentrated TFR for both signals with fast varying IF and slowly varying IF. Meanwhile,
the MTEWT method is proposed to separate the fault feature from masking components originating
from other machine elements such as shafts or gears.
The effectiveness of the proposed method is validated by a variety of time-varying experimental
signals measured by the vibration sensors mounted in the wind turbine planetary gearbox under
nonstationary conditions. The calculated TFR is highly energy-concentrated and the estimated IF
matches very well with the actual IF. The proposed method can successfully detect the sun gear crack
fault, the planet gear crack fault, multi gear faults, and the planet bearing fault. Results indicate that
the proposed method is a promising tool for the fault diagnosis and condition monitoring of a wind
turbine planetary gearbox.

Acknowledgments: This work is supported by the National Natural Science Foundation of China under
No. 11372179, the National Key Technologies R & D Program of China (No. 2015BAF07B03) and Innovation
Project of Shanghai (No. 15JC1402600). The authors would like to thank the CMMNO contest committee for
sharing the vibration data of a wind turbine planetary gearbox.
Author Contributions: Yue Hu proposed this method, analyzed all the data, and wrote and edited the manuscript.
Xiaotong Tu carried out the first experiment and acquired the corresponding data. Fucai Li, the corresponding
author, guaranteed the integrity of the entire study and revised the manuscript. Guang Meng provided some
advice on the background, introduction, and conclusion.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Teng, W.; Ding, X.; Zhang, X.; Liu, Y.; Ma, Z. Multi-fault detection and failure analysis of wind turbine
gearbox using complex wavelet transform. Renew. Energy 2016, 93, 591–598. [CrossRef]
2. Peeters, C.; Guillaume, P.; Helsen, J. Vibration-based bearing fault detection for operations and maintenance
cost reduction in wind energy. Renew. Energy 2018, 116, 74–87. [CrossRef]
3. Helsen, J.; Guo, Y.; Keller, J.; Guillaume, P. Experimental investigation of bearing slip in a wind turbine
gearbox during a transient grid loss event. Wind Energy 2016, 19, 2255–2269. [CrossRef]
4. Bououden, S.; Chadli, M.; Filali, S.; El Hajjaji, A. Fuzzy model based multivariable predictive control of a
variable speed wind turbine: LMI approach. Renew. Energy 2012, 37, 434–439. [CrossRef]
5. Chibani, A.; Chadli, M.; Shi, P.; Braiek, N.B. Fuzzy Fault Detection Filter Design for T-S Fuzzy Systems in
Finite Frequency Domain. IEEE Trans. Fuzzy Syst. 2017, 25, 1051–1061. [CrossRef]
6. Li, F.; Shi, P.; Lim, C.-C.; Wu, L. Fault Detection Filtering for Nonhomogeneous Markovian Jump Systems
via Fuzzy Approach. IEEE Trans. Fuzzy Syst. 2017. [CrossRef]
7. Bououden, S.; Filali, S.; Chadli, M. Fuzzy predictive control of a variable speed wind turbine. Energy Procedia
2013, 42, 357–366. [CrossRef]
8. Brevdo, E. Efficient Representations of Signals in Nonlinear Signal Processing with Applications to Inverse
Problems. Ph.D. Thesis, Princeton University, Princeton, NJ, USA, 2011.
9. Sandryhaila, A.; Moura, J.M.F. Discrete Signal Processing on Graphs. IEEE Trans. Signal Process. 2013, 61,
1644–1656. [CrossRef]
10. Yang, W.; Tavner, P.J.; Tian, W. Wind Turbine Condition Monitoring Based on an Improved Spline-Kernelled
Chirplet Transform. IEEE Trans. Ind. Electron. 2015, 62, 6565–6574. [CrossRef]
11. Badihi, H.; Zhang, Y.; Hong, H. Fuzzy gain-scheduled active fault-tolerant control of a wind turbine.
J. Frankl. Inst. 2014, 351, 3677–3706. [CrossRef]
12. Tabatabaeipour, S.M.; Odgaard, P.F.; Bak, T.; Stoustrup, J. Fault detection of wind turbines with uncertain
parameters: A set-membership approach. Energies 2012, 5, 2424–2448. [CrossRef]
13. Odgaard, P.F.; Stoustrup, J. A benchmark evaluation of fault tolerant wind turbine control concepts.
IEEE Trans. Control Syst. Technol. 2015, 23, 1221–1228. [CrossRef]
14. Casau, P.; Rosa, P.; Tabatabaeipour, S.M.; Silvestre, C.; Stoustrup, J. A set-valued approach to FDI and FTC of
wind turbines. IEEE Trans. Control Syst. Technol. 2015, 23, 245–263. [CrossRef]
15. Chaari, F.; Fakhfakh, T.; Haddar, M. Dynamic analysis of a planetary gear failure caused by tooth pitting and
cracking. J. Fail. Anal. Prev. 2006, 6, 73–78. [CrossRef]
Sensors 2018, 18, 150 17 of 18

16. Inalpolat, M.; Kahraman, A. A theoretical and experimental investigation of modulation sidebands of
planetary gear sets. J. Sound Vib. 2009, 323, 677–696. [CrossRef]
17. Mark, W.D. Stationary transducer response to planetary-gear vibration excitation II: Effects of torque
modulations. Mech. Syst. Signal Process. 2009, 23, 2253–2259. [CrossRef]
18. Inalpolat, M.; Kahraman, A. A dynamic model to predict modulation sidebands of a planetary gear set
having manufacturing errors. J. Sound Vib. 2010, 329, 371–393. [CrossRef]
19. Chaari, F.; Fakhfakh, T.; Hbaieb, R.; Louati, J.; Haddar, M. Influence of manufacturing errors on the dynamic
behavior of planetary gears. Int. J. Adv. Manuf. Technol. 2006, 27, 738–746. [CrossRef]
20. Lei, Y.; Kong, D.; Lin, J.; Zuo, M.J. Fault detection of planetary gearboxes using new diagnostic parameters.
Meas. Sci. Technol. 2012, 23, 55605. [CrossRef]
21. Bartelmus, W.; Zimroz, R. Vibration condition monitoring of planetary gearbox under varying external load.
Mech. Syst. Signal Process. 2009, 23, 246–257. [CrossRef]
22. Bartelmus, W.; Zimroz, R. A new feature for monitoring the condition of gearboxes in non-stationary
operating conditions. Mech. Syst. Signal Process. 2009, 23, 1528–1534. [CrossRef]
23. Feng, Z.; Zuo, M.J. Vibration signal models for fault diagnosis of planetary gearboxes. J. Sound Vib. 2012, 331,
4919–4939. [CrossRef]
24. Feng, Z.; Liang, M. Complex signal analysis for wind turbine planetary gearbox fault diagnosis via iterative
atomic decomposition thresholding. J. Sound Vib. 2014, 333, 5196–5211. [CrossRef]
25. Feng, Z.; Chen, X.; Liang, M. Iterative generalized synchrosqueezing transform for fault diagnosis of wind
turbine planetary gearbox under nonstationary conditions. Mech. Syst. Signal Process. 2015, 52–53, 360–375.
[CrossRef]
26. Nejad, A.R.; Guo, Y.; Gao, Z.; Moan, T. Development of a 5 MW reference gearbox for offshore wind turbines.
Wind Energy 2016, 19, 1089–1106. [CrossRef]
27. Liu, W.Y.; Zhang, W.H.; Han, J.G.; Wang, G.F. A new wind turbine fault diagnosis method based on the local
mean decomposition. Renew. Energy 2012, 48, 411–415. [CrossRef]
28. Feng, Z.; Qin, S.; Liang, M. Time-frequency analysis based on Vold-Kalman filter and higher order
energy separation for fault diagnosis of wind turbine planetary gearbox under nonstationary conditions.
Renew. Energy 2016, 85, 45–56. [CrossRef]
29. Feng, Z.; Chen, X.; Liang, M. Joint envelope and frequency order spectrum analysis based on iterative
generalized demodulation for planetary gearbox fault diagnosis under nonstationary conditions. Mech. Syst.
Signal Process. 2015, 76–77, 242–264. [CrossRef]
30. Jiang, Y.; Tang, B.; Qin, Y.; Liu, W. Feature extraction method of wind turbine based on adaptive Morlet
wavelet and SVD. Renew. Energy 2011, 36, 2146–2153. [CrossRef]
31. Tang, B.; Liu, W.; Song, T. Wind turbine fault diagnosis based on Morlet wavelet transformation and
Wigner-Ville distribution. Renew. Energy 2010, 35, 2862–2866. [CrossRef]
32. Tang, B.; Song, T.; Li, F.; Deng, L. Fault diagnosis for a wind turbine transmission system based on manifold
learning and Shannon wavelet support vector machine. Renew. Energy 2014, 62, 1–9. [CrossRef]
33. Feng, Z.; Liang, M.; Zhang, Y.; Hou, S. Fault diagnosis for wind turbine planetary gearboxes via demodulation
analysis based on ensemble empirical mode decomposition and energy separation. Renew. Energy 2012, 47,
112–126. [CrossRef]
34. Wang, S.; Chen, X.; Cai, G.; Chen, B.; Li, X.; He, Z. Matching demodulation transform and synchrosqueezing
in time-frequency analysis. IEEE Trans. Signal Process. 2014, 62, 69–84. [CrossRef]
35. Wang, Y.; Markert, R. Filter bank property of variational mode decomposition and its applications.
Signal Process. 2016, 120, 509–521. [CrossRef]
36. Lei, Y.; Lin, J.; He, Z.; Zuo, M.J. A review on empirical mode decomposition in fault diagnosis of rotating
machinery. Mech. Syst. Signal Process. 2013, 35, 108–126. [CrossRef]
37. Daubechies, I.; Lu, J.; Wu, H.T. Synchrosqueezed wavelet transforms: An empirical mode decomposition-like
tool. Appl. Comput. Harmon. Anal. 2011, 30, 243–261. [CrossRef]
38. Wu, H.T.; Chan, Y.H.; Lin, Y.T.; Yeh, Y.H. Using synchrosqueezing transform to discover breathing dynamics
from ECG signals. Appl. Comput. Harmon. Anal. 2014, 36, 354–359. [CrossRef]
39. Herrera, R.H.; Tary, J.B.; Van Der Baan, M.; Eaton, D.W. Body wave separation in the time-frequency domain.
IEEE Geosci. Remote Sens. Lett. 2015, 12, 364–368. [CrossRef]
Sensors 2018, 18, 150 18 of 18

40. Mousavi, S.M.; Langston, C.A.; Horton, S.P. Automatic microseismic denoising and onset detection using
the synchrosqueezed continuous wavelet transform. Geophysics 2016, 81, V341–V355. [CrossRef]
41. Mousavi, S.M.; Langston, C.A. Automatic noise-removal/signal-removal based on general cross-validation
thresholding in synchrosqueezed domain and its application on earthquake data. Geophysics 2017, 82,
V211–V227. [CrossRef]
42. Wang, S.; Chen, X.; Tong, C.; Zhao, Z. Matching Synchrosqueezing Wavelet Transform and Application to
Aeroengine Vibration Monitoring. IEEE Trans. Instrum. Meas. 2016, 66, 1–13. [CrossRef]
43. Wang, S.; Chen, X.; Selesnick, I.W.; Guo, Y.; Tong, C.; Zhang, X. Matching synchrosqueezing transform:
A useful tool for characterizing signals with fast varying instantaneous frequency and application to machine
fault diagnosis. Mech. Syst. Signal Process. 2018, 100, 242–288. [CrossRef]
44. Auger, F.; Flandrin, P. Improving the Readability of Time-Frequency and Time-Scale Representations by the
Reassignment Method. IEEE Trans. Signal Process. 1995, 43, 1068–1089. [CrossRef]
45. Yang, H. Synchrosqueezed wave packet transforms and diffeomorphism based spectral analysis for 1D
general mode decompositions. Appl. Comput. Harmon. Anal. 2015, 39, 33–66. [CrossRef]
46. Kowalski, M.; Meynard, A.; Wu, H. tieng Convex Optimization approach to signals with fast varying
instantaneous frequency. Appl. Comput. Harmon. Anal. 2015, 1, 1–34.
47. Huang, Z.L.; Zhang, J.; Zhao, T.H.; Sun, Y. Synchrosqueezing S-Transform and Its Application in Seismic
Spectral Decomposition. IEEE Trans. Geosci. Remote Sens. 2016, 54, 817–825. [CrossRef]
48. Ahrabian, A.; Mandic, D.P. A Class of Multivariate Denoising Algorithms Based on Synchrosqueezing.
IEEE Trans. Signal Process. 2015, 63, 2196–2208. [CrossRef]
49. Daubechies, I.; Wang, Y.; Wu, H. ConceFT: Concentration of Frequency and Time via a multitapered
synchrosqueezed transform. Philos. Trans. R. Soc. A 2015, 374. [CrossRef] [PubMed]
50. Oberlin, T.; Meignen, S.; Perrier, V. Second-order synchrosqueezing transform or invertible reassignment?
Towards ideal time-frequency representations. IEEE Trans. Signal Process. 2015, 63, 1335–1344. [CrossRef]
51. Pham, D.-H.; Meignen, S. High-Order Synchrosqueezing Transform for Multicomponent Signals
Analysis—With an Application to Gravitational-Wave Signal. IEEE Trans. Signal Process. 2017, 65, 3168–3178.
[CrossRef]
52. Zhao, M.; Lin, J.; Xu, X.; Lei, Y. Tacholess envelope order analysis and its application to fault detection of
rolling element bearings with varying speeds. Sensors 2013, 13, 10856–10875. [CrossRef] [PubMed]
53. Iatsenko, D.; McClintock, P.V.E.; Stefanovska, A. Extraction of instantaneous frequencies from ridges in
time-frequency representations of signals. Signal Process. 2016, 125, 290–303. [CrossRef]
54. He, G.; Ding, K.; Li, W.; Jiao, X. A novel order tracking method for wind turbine planetary gearbox vibration
analysis based on discrete spectrum correction technique. Renew. Energy 2016, 87, 364–375. [CrossRef]
55. Fourer, D.; Auger, F.; Czarnecki, K.; Meignen, S.; Flandrin, P. Chirp Rate and Instantaneous Frequency
Estimation: Application to Recursive Vertical Synchrosqueezing. IEEE Signal Process. Lett. 2017, 24,
1724–1728. [CrossRef]
56. Gilles, J. Empirical wavelet transform. IEEE Trans. Signal Process. 2013, 61, 3999–4010. [CrossRef]
57. Hu, Y.; Li, F.; Li, H.; Liu, C. An enhanced empirical wavelet transform for noisy and non-stationary signal
processing. Digit. Signal Process. 2017, 60, 220–229. [CrossRef]
58. Hu, Y.; Tu, X.; Li, F.; Li, H.; Meng, G. An adaptive and tacholess order analysis method based on enhanced
empirical wavelet transform for fault detection of bearings with varying speeds. J. Sound Vib. 2017, 409,
241–255. [CrossRef]
59. Leclère, Q.; André, H.; Antoni, J. A multi-order probabilistic approach for Instantaneous Angular Speed
tracking debriefing of the CMMNO’14 diagnosis contest. Mech. Syst. Signal Process. 2016, 81, 375–386.
[CrossRef]

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).