energies

Article
Investigation on Planetary Bearing Weak Fault
Diagnosis Based on a Fault Model and Improved
Wavelet Ridge
Hongkun Li *, Rui Yang ID
, Chaoge Wang and Changbo He
School of Mechanical Engineering, Dalian University of Technology, Dalian 116024, China;
yr90223@mail.dlut.edu.cn (R.Y.); dutwcg@163.com (C.W.); strange@mail.dlut.edu.cn (C.H.)
* Correspondence: lihk@dlut.edu.cn; Tel.: +86-0411-8470-6561




Received: 12 April 2018; Accepted: 9 May 2018; Published: 17 May 2018


Abstract: Rolling element bearings are of great importance in planetary gearboxes. Monitoring their
operation state is the key to keep the whole machine running normally. It is not enough to just apply
traditional fault diagnosis methods to detect the running condition of rotating machinery when
weak faults occur. It is because of the complexity of the planetary gearbox structure. In addition,
its running state is unstable due to the effects of the wind speed and external disturbances. In this
paper, a signal model is established to simulate the vibration data collected by sensors in the event
of a failure occurred in the planetary bearings, which is very useful for fault mechanism research.
Furthermore, an improved wavelet scalogram method is proposed to identify weak impact features of
planetary bearings. The proposed method is based on time-frequency distribution reassignment and
synchronous averaging. The synchronous averaging is performed for reassignment of the wavelet
scale spectrum to improve its time-frequency resolution. After that, wavelet ridge extraction is carried
out to reveal the relationship between this time-frequency distribution and characteristic information,
which is helpful to extract characteristic frequencies after the improved wavelet scalogram highlights
the impact features of rolling element bearing weak fault detection. The effectiveness of the proposed
method for weak fault recognition is validated by using simulation signals and test signals.

Keywords: planetary bearing; fault model; improved wavelet scalogram; initial impact feature;
wavelet ridge

1. Introduction
Condition monitoring, fault diagnosis and periodic maintenance are indispensable steps to
maintain the normal operation of rotating machinery [1,2]. Fault diagnosis and pattern recognition
are very important and difficult parts of this task. At present, signal processing methods are the most
popular fault diagnosis techniques due to the fact vibration signals are easily collected by sensors,
but feature extraction is an essential and complex question for fault diagnosis of rotating machinery
based on vibration signal analysis due to the fact vibration signals are usually disturbed by the external
circumstances in practical engineering applications. Meanwhile, early fault signal characteristics
are very weak and may even be buried under strong background noise, so with traditional signal
processing methods it is difficult to analyze this kind of signal. Many researchers have focused their
attentions on weak impact feature extraction.
Due to their compact structure, accurate transmission ratio, smooth transmission, and high
efficiency, planetary gearboxes are widely used in automobile, aerospace, petrochemical and other
machinery manufacturing industries. Rolling element bearings are the most extensively available
components in planetary gearboxes [2]. The planetary gearbox differs from the traditional parallel

Energies 2018, 11, 1286; doi:10.3390/en11051286 www.mdpi.com/journal/energies

Energies 2018, 11, 1286 2 of 23

gearbox in that it consists of one or more planet gears, a sun gear, and a planet carrier. Planet gears are
distributed around the sun gear and driven by the planet carrier to rotate around the sun gear and
also rotate around their own rotation axis. Due to the complexity of the planetary gear structure and
operation environment, vibration signal components are no longer as simple as in a parallel gearbox.
In the event of planetary gearbox failure, vibration mechanism and vibration signal processing methods
are two main research areas. The study of mechanical equipment failure mechanisms is the basis for
other research. This provides theoretical support for investigating the changes of running component
parameters and vibration signal models before and after a failure.
The research on the vibration mechanism of gearboxes is mainly focused on dynamic models and
vibration signal response analysis. McFadden and Smith [3] indicated that asymmetrical sidebands
are contained in the vibration spectrum of planetary gearboxes and explained that this phenomenon
is affected by the number of gear ring teeth and the number of planet gears. Mosher [4] established
the kinetic equations of planetary gearboxes and predicted that the frequency component with larger
amplitude in the frequency spectrum is located at the meshing frequency and its multiplied frequency,
and there is a sideband near it. Inalpolat and Kahraman [5] developed a mathematical model to
study the sideband mechanism in the spectrum of planetary vibration signals, thereby classifying
planetary gearboxes into five types. Subsequently, they added dynamic meshing stiffness and damping
analysis to their original model to perfect the mathematical model of planetary gearboxes [6]. However,
they considered that the planet gears transmit vibration to sensors only within 1/3 rotation cycle and
the vibration signal intensity at other times was considered to be zero. Feng et al. [7] established a
vibration signal model of a planetary gearbox in normal operation state and in local and distributed
fault states. However, in the establishment of the vibration signal model, only the motion of a single
planetary gear is taken into consideration. Neither the manufacturing error nor the meshing phase
difference is analyzed in the meshing process. Liang et al. [8] simulated the vibration source signal
by constructing kinetic equations. Then, an improved Hamming window function is used to show
the influence of signal transmission path on the vibration signal model. Taking comprehensive
consideration of each source signal vibration, the vibration signal simulation model of a planetary
gearbox in a healthy state and with a cracked sun gear are built. Lei et al. [9] analyzed the vibration
signal transmission path and meshing phase of planetary gearboxes according to the transmission
mechanism and meshing principle and established the vibration signal model for normal operation
and fault conditions. Ding et al. [10] adopted the ring gear modal experimental to verify the variation
of vibration amplitude with the transmission path.
To summarize, it can be seen that the research on planetary gearboxes based on vibration
mechanisms is mostly carried out under the conditions of healthy state or gear failure. The literature on
failure mechanism of planetary bearing is very scarce. Most researchers are accustomed to establishing
kinetic equations to study vibration mechanisms. However, the kinetic equations are based on the
assumption of multiple parameters. In the event of a fault, the structural properties will change and
the value of parametric variation cannot be accurately measured. Therefore, it is only applicable
to the analysis of healthy gearbox and the fault characteristics cannot be described accurately and
completely [2].
In addition, it will take a long time from the start of the fault to when significant impact features
are discovered, which may lead to huge economic losses and even casualties. It would make no sense
to detect planetary bearing faults by using traditional fault diagnosis methods when the machine has
already broken down. Therefore, weak impact feature identification in rolling element bearings is
significant to reduce downtime and safety hazards during the running period. Nowadays, a large
number of modern approaches have been proposed, such as kurtograms [11,12], improved kurtograms
and spectral kurtosis [13–15], cyclostationary analysis [16,17], empirical mode decomposition [18],
synchronous averaging technology [19,20], sparse models [21,22], stochastic resonance [23], wavelet
spectra [24], etc.
Energies 2018, 11, 1286 3 of 23

When the impact characteristics of a signal are relatively weak or submerged in fault-independent
interference components, they require further analysis [13]. In some circumstances, time domain
analysis and frequency domain analysis are available for simple fault signal processing. However,
most bearing fault signals are both non-stationary and non-linear. As a result, time-frequency
distribution will be a good choice [25]. Nowadays, the main time-frequency distribution methods,
including Short Time Fourier Transform and Wigner-Ville distribution, are very useful for dynamic
vibration signal analysis. However, they have a poor performance in extracting weak impact features
in the case where the signal is disturbed by strong background noise. To solve this problem, another
method named Hilbert spectrum is studied in signal processing [26,27], although it has a limitation
in time-frequency resolution. The background noise can be ignored as an insignificant ingredient
because it has nothing to do with fault diagnosis. Based on this property, several new time-frequency
analysis methods are extensively used in weak fault feature extraction. For example, wavelet analysis
has multi-resolution characteristics and can detect well the local characteristics of signals. It is very
effective in analyzing non-stationary signals. Recently, lots of improved algorithms have been put
forward, that are very effective for a more complicated signal analysis. He proposed a novel method
which takes the time-frequency manifold as a template to do a correlation matching to enhance the
periodic faults’ impulse components [28]. Time domain averaging can reduce the random disturbance
and retain the periodic components. Time synchronous averaging carried out for wavelet coefficients
in the range of all scales obtained by wavelet transform [29] can improve the time-frequency resolution.
This is a feasible approach for noisy fault signal feature extraction. Besides that, time-frequency
distribution based on wavelet scalograms can reveal signal feature information. In some cases, it fails
to detect the fault symptoms and lacks concentration. The improved scalograms named reassigned
wavelet scalograms have a better concentration in time-frequency plane [30].
In this paper, to extract fault features more accurately, the fault mechanism of planetary bearings
is studied. A fault simulation model is established considering the influence of various factors. Then,
a new method of improved wavelet scalograms is proposed. It combines time-frequency distribution
reassignment with synchronous averaging. A wavelet ridge is also determined to improve the veracity
of mechanical fault feature recognition. The organization of the rest paper is as follows: in Section 2,
the fault mechanism is introduced and the simulation signal model is established. The theory of the
proposed method is introduced in Section 3. Section 4 presents the validity of the proposed method by
simulated signal and weak rolling element bearing fault data. Finally, some major concluding remarks
are given in Section 5.

2. Fault Model

2.1. Meshing Frequency and Fault Characteristic Frequency

During the operation of a planetary gearbox, the planet carrier and sun gear rotate in the same
direction. The planet gear and the sun gear mesh with each other, whilst the planet gears orbit around
their own rotation axis and the planet carrier, as shown in Figure 1. It is assumed that the rotation
direction of planet carrier is positive and its rotational frequency is f c , the planetary gear rotational
(r )
frequency is f p , the absolute rotation frequency of planet gear is f p . The gear ring is connected with
the case and it rotates with respect to the planet carrier at the frequency f c . The meshing frequency of
planetary gearbox is:
f m = Nr f c = Ns ( f s − f c ) = Np ( f p + f c ) (1)

wherein f m is the meshing frequency of planetary gearbox; Nr is the number of gear ring tooth; Ns is
the number of sun gear tooth; Np is the number of planet gear tooth; f s is the rotational frequency of
sun gear.
 Energies 2018, 11, 1286 4 of 23
Energies 2018, 11, x 4 of 23

2
1
4
5
6

3
Figure 1. Planetary gearbox model diagram.
Figure 1. Planetary gearbox model diagram.

Taking Taking
a fixed point
a fixed ononthe
point thehousing
housing asas reference
reference point,point, the rotational
the rotational frequencyfrequency
of planetary of
gearplanetary
gear inner
innerring
ring isisequal
equal toplanetary
to the the planetary carrier’s,
carrier’s, that is f pi = that f pi = fc frequency
f c . Theisrotational . The rotational
of planetaryfrequency
gear of
outer ring relative to the reference point is the absolute rotational frequency of planetary gear, that is
planetary gear(r) outer ring relative to the reference point is the absolute rotational frequency of
f po = f p = f c Zr /Z p . (r )
planetary gear, po = f p
that is ffrequency
The rotational =offplanet
c Zr / Zgear
p . bearing cage relative to this reference point is:

The rotational frequency of planet gear bearing cage relative

f (1− d cos α) d
to this reference point is:
pi D f po (1+ D cos α)
f cg = 2 + 2
fd(1− Dd cos α) f c Zr (1− Dddcos α )
f pi=(1 − c cos 2 α
) + f po (1 +
2Z p cos α, ) (2)
fcg = fD( Z + Z )+ f c + Dα
( Zr − Z p ) Dd cos
= c2 r p 2Z p 2
d d
The characteristic frequency offcplanet cosα
(1 − gear ) fc Z
bearing r (1 −
outer is: α )
ringcos
= D + D (2)
= N ( f po − f cg ) 2= N ( fZc Zp r − 2Z p −
d cos α )
f c (1− D d cos α )
f c Zr (1+ D
f BPFO 2 2Z p )
d cos α
f c ( Zr + Z p )+ f c ( Zr − Z p ) D
= f c Zr
N ( Zp − d , (3)
fc (Zr +Zp )d + fc2Z (Zp r − Zp ) )cos α
=
f ( Z − Z )(1− cos α)
N (= c r p2Z p D ) D
2Zp
,
The characteristic frequency of planet gear bearing inner ring is:
The characteristic frequency of planet gear bearing outer
d
ring is: d cos α )
f c (1− D cos α) f c Zr (1+ D
f BPFI = N ( f pi − f cg ) = N ( f c − 2 + 2Z p )
f c ( Zr + Z p )+ f c ( Zr − Z p )dD
d cos α d
= N( fc − fc Zr 2Zcp f (1 - cos α)
) fc Zr (1 + cos,α) (4)
f BPFO = N(f po - fcgf c ()Zp=−N( d - α) D - D )
Zr )(1+ D cos
= N 2Z pZ p 2 2Zp

wherein, N is the number of rolling elements, d

f Z fc (Zr + Zdp is) +thefc (Z
roller cosDαis the bearing pitch diameter,
diameter,
r - Zp )
α is the contact angle. f= is
cg N( thec rotational
r
- frequency of bearing D
cage. f BPFO
) and f BPFI represent the (3)
Z
characteristic frequencies of planetp gear bearing outer ring 2Z p and planet gear bearing inner ring.

2.2. Fault Simulation Model d

fc (Zr - Zp )(1 - cosα)
= N( D )
During the operation of planetary gearbox, the meshing points are shown as points 1 to 6 in
2Zp
Figure 1. In order to analyze the fault mechanism, the vibration signal model is established, in the case of
sensors are installed on the housing. In the paper, only time-varying transmission paths are considered,
Thethat
characteristic frequency
is shown in Figure 2. Theof planet of
positions gear bearing inner
all planetary ring
gears are is:
constantly changing during the
operation. The transmission path of vibration signal generated by each meshing points to sensor
d d
fc (1 - cosα) fc Zr (1 + cosα)
f BPFI = N(f pi - fcg ) = N(fc - D + D )
2 2Zp
d
fc (Zr + Zp ) + fc (Zr - Zp ) cosα
D (4)
case of sensors are installed on the housing. In the paper, only time-varying transmission paths are
considered, that is shown in Figure 2. The positions of all planetary gears are constantly changing
during the operation. The transmission path of vibration signal generated by each meshing points to
sensor isEnergies
time-varying.
2018, 11, 1286
Vibration transmission is mainly through these two paths. The first
5 of 23
time-varying path starts from planet gear bearing to planet gear, and then to gear ring. Finally, it
passes through the case to sensors. The second path has one more transmission of sun gear to planet
is time-varying. Vibration transmission is mainly through these two paths. The first time-varying
gear than the first one, and its transmission length does not change with the rotation of planet
path starts from planet gear bearing to planet gear, and then to gear ring. Finally, it passes through
carrier. Therefore, the second
the case to sensors. path path
The second is anhas
equal-proportional attenuation
one more transmission process
of sun gear to relative
planet gear to the first
than the
first one, that
one. Assuming and its
thetransmission
amplitudelength does not change
modulation effect with the rotation
produced of planet
by the wi (t) , and then
carrier.isTherefore,
first path
the second path is an equal-proportional attenuation process relative to the first one. Assuming that the
the amplitude modulation
amplitude modulation effect produced
effect produced by first
by the thepath
second
is w (path
t), andcan
thenbe
theexpressed as κwi (t) , where
amplitude modulation
i
0 < κ < 1effect
. produced by the second path can be expressed as κwi (t), where 0 < κ < 1.

Figure 2.Figure
Vibration signalsignal
2. Vibration transmission in in
transmission planetary
planetary gearbox.
gearbox. 1.1. From
Fromplanet
planet gear
gear to gear
to gear ring, ring, to
housing,tothen
housing, then to sensors;
to sensors; 2. From 2. sun
Fromgear
sun gear to planet
to planet gear,to
gear, to gear
gear ring,
ring,toto
housing, thenthen
housing, to sensors.
to sensors.

2.2.1. Amplitude Modulation Model

2.1.1. Amplitude Modulation Model
As the planetary gear moves from the bottom to the top of gearbox housing, it comes closer to
As the planetary
the sensor gear
and the moves
impact from thestrength
characteristic bottomoftovibration
the topsignals
of gearbox housing,
collected it comes
by the sensor showcloser to
the sensor
an and the impact
increasing trend. characteristic
As the planetarystrength
gear moves of vibration
from the topsignals collected
to the bottom by the sensor
of gearbox housing,show an
increasing
thetrend. Asgears
planetary the are
planetary gearfrom
farther away moves fromand
the sensor, thethe
top to thesignals
vibration bottom of gearbox
collected housing, the
by the sensor
are weaker. The amplitude modulation effect can be expressed as the following equation:
planetary gears are farther away from the sensor, and the vibration signals collected by the sensor
are weaker. The amplitude modulation β(mod(wceffect can
t+ψi ,2π )− π )2 be expressed as the following equation:
wi ( t ) = e (α − (1 − α) cos(2π f c t + ψi )), (5)
β( mod ( wc t +ψi ,2π ) -π )2
wherein, wi (t) represents wi (tthe
) = amplitude
e ( α - (effect
modulation 1 - α)cos( 2πfc t + ψgearbox
of planetary i )) , during operation. (5)
α and β indicate the intensity of modulation effect. ψi is the phase difference between the i − th planet
wi (t)
wherein,gear and represents
the first planetthegear.
amplitude modulation
The different modulation effect
effectsofcaused
planetary gearbox
by different during
values operation.
of α and β α
are shown in Figure 3.
and β indicate the intensity of modulation effect. ψi is the phase difference between the i-th planet
In Figure 3a, in the case of α > 1, all of the amplitudes are greater than 1. If α < 0.5, the amplitudes
gear andare
theless
first planet
than gear. of
0. Neither The different
these modulation
two conditions effects
satisfies caused by different
the requirements. Therefore,values of α and β
it can be
are shown in Figure
determined that3.α is in the range of [0.5, 1]. In Figure 3b, when α is fixed and β > 0, the amplitudes are
greater than 1. The larger the value of β, the greater the effect of amplification. When 0.5 < α < 1 and
β < 0, all of the amplitude concentrated between [0, 1], that meets the requirements.
Energies 2018, 11, 1286 6 of 23
Energies 2018, 11, x 6 of 23

(a) (b)

(c)

Figure 3. The influence of different α and β values on exponential decay function; (a) The
Figure 3. The influence of different α and β values on exponential decay function; (a) The exponential
exponential
decay effectdecay effect of
of different different
values of α;values α ; (b) Thedecay
(b) Theofexponential exponential decay
effect of effectvalues
different of different values
of β; (c) The
of β ; (c) The exponential decay effect of different
exponential decay effect of different values of α and β.values of α and β .

2.2.2.InMeshing
Figure Force the case of α > 1 , all of the amplitudes are greater than 1. If α < 0.5 , the
3a, inModel
amplitudes are less than 0. Neither of these two conditions satisfies the requirements. Therefore, it
can be determined that α is in the range of J [0.5, 1]. In Figure 3b, when α is fixed and β > 0 , the
F ( t ) = A 1∑
amplitudes are greater than ir1. The larger thecos ( j2π fof
value mt +
βφ (t) greater
, irthe + θ j ) the effect of amplification.
j =1
When 0.5 < α < 1 and β < 0 , all of the amplitude
J concentrated between , (6)
[0, 1], that meets the
= A1 ∑ cos( j2πZr f c t + φir (t) + θ j )
requirements. j =1

J
2.2.2. Meshing Force Model F (t) = A2 ∑ cos( j2π f m t + φis (t) + θ j )
is
j =1
J J , (7)
j=1
Fir (t )== A1A2 ∑
j=1
cos
cos( ( j2πZ
j2πf f t + φis (j )t) + θ j )
m t +rφcir ( t ) + θ

J
, (6)
1  cos( j2πZr f c t + φir ( t ) + θ j )
ir ( t ) =
=φA 2πZr (i − 1)/3, (8)
j =1
φis (t) = 2πZs (i − 1)/3, (9)

with Fir (t) and Fis (t) meaning the meshingJ forces generated by the i − th planetary gear engaged with
Fis (t ) = A2  cos( j2πfm t + φis (t ) + θ j )
the ring gear and the sun gear, respectively. A and A2 indicate the size of meshing force. f m is meshing
j=1 1
frequency. J is the maximum harmonic order, J j = 1, 2 · · · J. θ j is the (7)
, phase of the j − th harmonic.
= A 2
φir (t) and φis (t) are the meshing phase differences cos( j2πZ f t
generated
r c + φ ( t ) +
is by the θ j i − th planet gear engaged with
)
the ring gear and the sun gear, respectively.j=1This is affected by whether the number of ring gear teeth
and the number of planet gears are integral multiples and whether the phase difference between the
adjacent planet gears is equal. The meshing φir (phase
t) = 2πZ , be calculated by Equations (8) and(8)(9).
r (i -1) / 3can
difference

2.2.3. Fault Step-Impact Model φis (t) = 2πZs (i -1) / 3 , (9)

with InFirthe
(t ) case
andwhere
Fis (t ) ameaning
bearing outer ring is cracked
the meshing forces or peeled off,
generated byitsthe
geometry will change.
i-th planetary gearAs shown
engaged
in Figure 4, when the rolling element is in contact with the fault position, the components of the
with the ring gear and the sun gear, respectively. A1 and A2 indicate the size of meshing force.
f m is meshing frequency. J is the maximum harmonic order, j = 1,2  J . θ j is the phase of the
phase difference between the adjacent planet gears is equal. The meshing phase difference can be
calculated by Equations (8) and (9).

2.2.3. Fault Step-Impact Model

Energies 2018, 11, 1286 7 of 23
In the case where a bearing outer ring is cracked or peeled off, its geometry will change. As
shown in Figure 4, when the rolling element is in contact with the fault position, the components of
vibration signal
the vibration willwill
signal change.
change. Reference
Reference [31] proposed
[31] proposedthatthatthe
thebearing
bearingfault
faultsignal
signalisis composed
composed of
two parts. A step signal will be generated when the rolling element enters the fault position and an
impact signal will be generated when the rolling element leaves the fault position. The fault size can
be quantified
quantifiedbased
basedon onthethetime
timedifference
differencebetween the the
between stepstep
signal and the
signal andimpact signal.signal.
the impact In thisInpaper,
this
the research mainly focuses on bearing faults occurred at an early stage, that is the
paper, the research mainly focuses on bearing faults occurred at an early stage, that is the case of a case of a small fault
size. The
small rolling
fault size. element
The rollingdoeselement
not come intonot
does contact
comewith
intothe lowest
contact endthe
with of the bearing
lowest endouter
of thering fault
bearing
when
outer it passes
ring faultthrough
when itthe fault through
passes area. It isthe
assumed that the
fault area. It isfrequency
assumed of bearing
that cage is f cgof
the frequency , the outer
bearing
ring fault
cage is fcgwidth is l0 and
, the outer ring D o is the
fault widthouter
is ring
l0 anddiameter.
Do is theTherefore, thediameter.
outer ring time difference between
Therefore, the
the time
rolling element entering and leaving the fault area is:
difference between the rolling element entering and leaving the fault area is:
ll00
∆t
Δt0 0== , (10)
(10)
2πDoo f cgcg

(a) (b)

Figure
Figure 4.
4. Schematic
Schematicdiagram
diagram of of
rolling element
rolling faultfault
element position. (a) Schematic
position. diagram
(a) Schematic of the rolling
diagram of the
element enteringentering
rolling element the faultthe
location; (b) Schematic
fault location; diagram diagram
(b) Schematic of the rolling
of theelement
rolling leaving
elementthe fault
leaving
location.
the fault location.

The step response model can be expressed as:

The step response model can be expressed as:
e -c( t-kT ) cos( 2πfn (t - kT )) + e -c( t-kT ) t - kT + Δt0 > 0
(t) =−c(t−kT )
xstep( , (11)
e  cos(2π f n0(t − kT )) + e−c(t−kT )t - kT Δt0+<∆t
t −+kT 00 > 0 ,
xstep (t) =  (11)
0 t − kT + ∆t0 < 0
The impact signal model can be expressed as:
The impact signal model can be expressed as:
e -c( t-kT ) cos( 2πfn / 2.5(t - kT )) t - kT > 0
t) = 
ximp (( , (12)
e−c(t−kT ) cos0(2π f n /2.5(t − kT ))t - kT t −<kT
0 >0
ximp (t) = , (12)
0 t − kT < 0
The bearing failure vibration signal model:
The bearing failure vibration signal
x(tmodel:
) = k1 xstep (t) + k2 ximp (t) , (13)

wherein, c is the attenuation coefficient,

x (t) = k1 xkstep (t) + k ximp (tthe
1 and k 22 are
), contributions of the above two
(13)
vibration signal models. The step signal, impact signal, and bearing fault simulation model for a
wherein, c is the attenuation coefficient, k1 and k2 are the contributions of the above two vibration
rotating cycle are shown in Figure 5a–c, respectively.
signal models. The step signal, impact signal, and bearing fault simulation model for a rotating cycle
are shown in Figure 5a–c, respectively.
Energies 2018, 11, 1286 8 of 23

Energies 2018, 11, x 8 of 23

Energies 2018, 11, x 8 of 23

(a) (b)

(a) (b)

(c)

Figure 5. Single-cycle bearing fault signal model. (a) The time domain waveform of step signal; (b)
Figure 5. Single-cycle bearing fault signal model. (a) The time domain waveform of step signal; (b) The
The time domain waveform of impact signal; (c) (c) The time domain simulation waveform for
timequantitative
domain waveform
analysis ofof impactfaults.
bearing signal; (c) The time domain simulation waveform for quantitative
Figure 5. Single-cycle bearing fault signal model. (a) The time domain waveform of step signal; (b)
analysis of bearing faults.
The time domain waveform of impact signal; (c) The time domain simulation waveform for
The simulated signal is determined as c = 100 , T = 0.1 , fn = 500 . The multi-cycle time domain
quantitative analysis of bearing faults.
waveform of bearing
The simulated signal attenuation impactassignal
is determined is shown
c = 100, in Figure
T = 0.1, f n = 6a.
500.InThe
Figure 6b, the frequency
multi-cycle time domain
spectrum
waveform The analysis
bearing isattenuation
ofsimulated continued,
signal and theassignal
is determined
impact carrier frequencies
c = 100
is ,shown , offFigure
T = 0.1 in 500 HzTheand
n = 500 .6a.
200 Hz are
multi-cycle
In Figure the
time
6b, center
domain
the frequency
frequencies
waveform ofwith the attenuation
bearing impact characteristic
impact frequency
signal is shown 10inHzFigure and6a.
itsInmultiplication
Figure 6b, the frequency
frequency
spectrum analysis is continued, and the carrier frequencies of 500 Hz and 200 Hz are the center
component
spectrum is the side
analysis frequency and
is continued, spread over thefrequencies
thefrequency
carrier entire frequency axis.
Hz Inandthe envelope
are spectrum,
frequencies with the impact characteristic 10 Hz andofits500 200
multiplication Hz
frequency the component
center
only the impact
frequencies characteristic
with the impactfrequency is demodulated.
characteristic frequency 10 Hz and its multiplication frequency
is the side frequency spread over the entire frequency axis. In the envelope spectrum, only the impact
component is the side frequency spread over the entire frequency axis. In the envelope spectrum,
characteristic frequency is demodulated.
only the impact characteristic frequency is demodulated.

(a)

(a)

(b)

(b)

(c)

Figure 6. The analytic results of bearing multi-cycle fault signal; (a) The time domain waveform; (b)
The frequency domain waveform; (c) The envelope (c)spectrum.

Figure 6. The analytic results of bearing multi-cycle fault signal; (a) The time domain waveform; (b)
Figure 6. The analytic results of bearing multi-cycle fault signal; (a) The time domain waveform;
The frequency domain waveform; (c) The envelope spectrum.
(b) The frequency domain waveform; (c) The envelope spectrum.
Energies 2018, 11, x 9 of 23
Energies 2018, 11, 1286 9 of 23
2.2.4. Bearing Load Zone Pass Effect
2.2.4.During
Bearingthe operation
Load of planetary
Zone Pass Effect gearbox, bearing capacity can be considered as the radial
force During
along the
theradial direction.
operation In the case
of planetary of thebearing
gearbox, rotationcapacity
frequency
canof
bebearing outer
considered asring is fo(s) force
the radial , the
(s) (s)
angle
along deviating
the radial from the vertical
direction. centerline
In the case is β = 2πf
of the rotation o t . The of
frequency function
bearingexpression
outer ring of
is fits
o ,load over
the angle
(s)
time is: from the vertical centerline is β = 2π f o t. The function expression of its load over time is:
deviating
(  11 (s)( s ) (s) ( s )
Qmax
Qmax
{1{1−- 2ε cos((2πf
cos 2πof ot )}t,)}, 2πf
2πo fto < tβ<max| β max |
, (14)
Q
Q(t) = ( t ) =  2ε , (14)
 0, 0, other other

wherein, Q
wherein, is the maximum load density; ε is the load distribution factor; β factor;
max is the maximum load density; ε is the load distribution max
Qmax is theβmax
maximum
is the
angle of load distribution area. In the process of bearing rotation, there is not always load
maximum angle of load distribution area. In the process of bearing rotation, there is not always load distribution
in one period.
distribution The period.
in one load will
Thebe load
zero will
when bethe rotation
zero angle
when the exceeds
rotation the exceeds
angle maximum theload distribution
maximum load
distribution area. The maximum angle of load distribution area can be controlled by changing of
area. The maximum angle of load distribution area can be controlled by changing the size theε.
In addition, the magnitude of modulated vibration signal by load distribution is changed
size of ε . In addition, the magnitude of modulated vibration signal by load distribution is changed by resizing
theresizing
by maximum the load
maximumdensity.
load density.
(s)
Assuming tt ==(0(0: n): /n6000
Assuming )/6000(n(nisisthe
the sampling length), f ofo(s) ==100
sampling length), 100, Qmax= =
, Q 1 . 1.InInthe
thecase
case of
of
max
ε = 0.6/0.5/0.4/0.3, the signal model is shown in Figure 7. As can be seen from figure, ε < 0.5 meets
ε = 0.6 / 0.5 / 0.4 / 0.3 , the signal model is shown in Figure 7. As can be seen from figure, ε < 0.5
the requirements.
meets the requirements.

Figure 7. The effect of load distribution factor on vibration waveform.

Figure 7. The effect of load distribution factor on vibration waveform.

2.2.5. Outer Ring Fault Vibration Signal Model

2.2.5. Outer Ring Fault Vibration Signal Model
The vibration signal model of planet gear bearing outer ring fault is established based on the
aboveThe vibration
analysis. signal
In the processmodelofofmodeling,
planet gear bearing
bearing outer
load zonering fault iseffect,
passing established based signal
step-impact on the
above analysis. In the process of modeling, bearing load zone passing
model, and the effect of transmission paths are all taken into account. Considering only one planet effect, step-impact signal model,
and bearing
gear the effect of transmission
outer ring failed, the paths are allsignal
vibration takenmodelinto account.
is shownConsidering
as follows: only one planet gear
bearing outer ring failed, the vibration signal model is shown as follows:
xo (t ) = wi (t )Q(t )[ k1 xstep (t ) + k2 ximp (t )] , (15)
xo (t) = wi (t) Q(t)[k1 xstep (t) + k2 ximp (t)], (15)
β(mod(wc t +ψi ,2 π)-π)2
wherein, wi (t ) = e ( α - (1 - α)cos(2πfc t + ψi ))
2
wherein, wi (t) = e β(mod(wc t+ψi ,2π )−π ) (α − (1 − α) cos(2π f c t + ψi ))
 1 (s)
(Q {1n- cos(2πfo t)} (s) o
Q(t) =  Qmax 2ε 1
max 1 − 2ε cos(2π f o t )
Q(t) =  0 q(t) < 0
0 q(t) < 0
(s)
f c = 4 , β = -3 , α = 0 .6 , θ 、 φ 、 ψ = 0 , ε = 0 . 2 , Qmax = 0.2 , fn = 500 , f o = 6 . The repeat period
(s)
of = 4, β =
f c impact −3, α = 0.6,
simulation = 0,
θ, φ,isψ100
signal ε =The
ms. = 0.2, f and
Qmaxdomain
0.2,time n = 500, f o = domain
frequency 6. The repeat period of
waveforms of
impact simulation signal is 100 ms.
simulated signal are shown in Figure 8. The time domain and frequency domain waveforms of simulated
signal are shown in Figure 8.
Energies 2018, 11, 1286 10 of 23
Energies 2018, 11, x 10 of 23

0.01

Amplitude(g)
0

-0.01

0 0.5 1 1.5 2
Time (s)
(a)
-3
x 10
5
f o( s ) − f c
f ob
4 fc
Amplitude(g)

3 fob + f o( s ) − fc
(s)
2 fc
f fob + f c
2 o

0
0 5 10 15 20 25 30 35 40 45 50
Frequency(Hz)
(b)

Figure 8. The analytic results of planet gear bearing outer ring fault simulation signal. (a) Time
Figure 8. The analytic results of planet gear bearing outer ring fault simulation signal. (a) Time domain
domain waveform of simulation signal; (b) The envelope spectrum of simulation signal.
waveform of simulation signal; (b) The envelope spectrum of simulation signal.

By analyzing the time domain waveform and frequency spectrum of the planet gear bearing
outerBy analyzing
fault simulationthesignal,
time domain
it can bewaveform
seen that and frequency
a series spectrum
of impact signalsofarethe planet gear
generated bearing
in the time
domain signal when the bearing fails. The load of the bearing is changed due to the rotation oftime
outer fault simulation signal, it can be seen that a series of impact signals are generated in the the
domain
planet signal
gear when
bearing the ring,
outer bearing fails.causing
thereby The load of the bearing
an intermittent is changed
distribution dueimpact
of the to the signal.
rotationTheof
the planet
natural gear bearing
frequency outer
of the ring, thereby
bearing and its causing an intermittent
side frequency distribution
components appear of the impact
in the signal.
envelope
The natural frequency of the bearing and its side frequency components appear
spectrum. In addition, there are side frequency components that are the characteristic frequency of in the envelope
spectrum.
bearing, In addition,
carrier frequencythere
andare side frequency
bearing components
rotation frequency. Inthat are the when
summary, characteristic
the outer frequency
ring of
of bearing, carrier frequency and bearing rotation frequency. In summary, when( s)the outer ring of
planetary bearing fails, the frequency peak mainly appears in fn ± lfob ± mfc ± nfo ((ls), m , n are
planetary bearing fails, the frequency peak mainly appears in f n ± l f ob ± m f c ± n f o (l, m, n are
positive integers).
positive integers).
3.
3.Theory
Theoryof
ofProposed
ProposedMethod
Method

3.1.
3.1.Improved
ImprovedWavelet
WaveletScalogram
Scalogram
As
As aa method
method of of time-frequency
time-frequency analysis,
analysis, wavelet
wavelet transform
transform is is adaptive.
adaptive. ItIt can
can eliminate
eliminate noise
noise
and identify weak fault information. The different scale value corresponds to different
and identify weak fault information. The different scale value corresponds to different frequency frequency
band
band[32].
[32]. Select
Selectthe
theappropriate
appropriatescale
scalevalue
valueaccording
accordingto to the
the frequency
frequency characteristics
characteristicsof of the
the signal
signal
being analyzed. Suppose ψ(t) is a wavelet basis function, then it must satisfy finite
being analyzed. Suppose ψ(t) is a wavelet basis function, then it must satisfy finite energy theorem, energy
i.e., ψ(t) ∈
theorem, L2 (ψ(t)
i.e., ∈ L2 (R)
R) and its Fourier
and itstransform ψ̂(ω ) should
Fourier transform ψ̂(ω)respect
shouldtherespect
following rule:
the following rule:
22
Z 
CψC=
ψ = ψ̂ˆ (ω ) //ω
ψ(ω)
 ω <<
|ωd|dω ∞ ,∞, (16)
(16)

wherein
wherein ψ (t) isisaawavelet
ψ(t) waveletbasic
basicfunction,
function,also
alsocalled
called mother
mother wavelet.
wavelet. InIn general,
general, the
the signal
signal analysis
analysis
results
resultsdepend
dependon onthe ψ(t).. Reference [33] discussed the effect of
the ψ(t) of different
different mother
mother wavelets
wavelets on
on the
the
analysisresult
analysis resultofofbearing
bearingimpact
impact failure
failure signal.
signal. ItItisisverified
verifiedthat
thatthe
theMorlet
Morletwavelet
waveletcan
canbetter
bettermatch
match
the characteristic
the characteristic ofof bearing
bearing impact
impact faults.
faults. Therefore,
Therefore,with withaaMorlet
Morletwavelet
waveletas as mother
mother wavelet
wavelet to
to
perform wavelet transform in the following section, a series of wavelet functions
perform wavelet transform in the following section, a series of wavelet functions can be obtained bycan be obtained by
translating and
translating and stretching
stretchingthe
themother
mother wavelet
waveletψ(t)ψ(t). Taking a as its
. Taking a scale
as itsfactor
scaleand b as and
factor the translation
b as the
translation factor, the deformation of a basic wavelet function ψ a,b (t) can be obtained by changing
the values of a and b according to the following Equation (17):
Energies 2018, 11, 1286 11 of 23

factor, the deformation of a basic wavelet function ψa,b (t) can be obtained by changing the values of a
and b according to the following Equation (17):

t−b
 
−1/2
ψa,b (t) = | a| ψ , (17)
a

ψa,b (t) is not unique, and its shape and position vary in relation to the values of a and b. The
continuous wavelet transform was carried out on x (t) ∈ L2 ( R), that can be expressed in the general
form of Equation (18):
t−b
Z  
Wx ( a, b; ψ) = a−1/2 x (t)ψ∗ dt, (18)
a
wherein, ψ(t) is an analytical wavelet, ψ∗ (t) is the complex conjugate of ψ(t). This definition is
analogous to the Fourier transform. What makes them different is the basis function: a Fourier
transform is performed on cosine functions and wavelet transform depends on the wavelet function.
The wavelet transform is reversible and the original signal can be restored by using the wavelet
transform coefficients:
1 x −2
x (t) = a Wx ( a, b; ψ)ψa,b (t)dadb, (19)
Cψ
wherein SG ( a, b; ψ) can be defined as the modulus of continuous wavelet transform coefficients,
which is given by Equation (20).

SG ( a, b; ψ) = |Wx ( a, b; ψ)|2 , (20)

As mentioned earlier, a wavelet scalogram lacks concentricity just like other traditional
time-frequency analysis method. Taking a simple chirp signal as an example, its time domain waveform
is shown in Figure 9a, its time-frequency spectra obtained by a wavelet scalogram is shown in Figure 9b.
It can be seen that the time-frequency ridge spreads over a wide area. Applying the reassigned wavelet
scalogram to calculate the time-frequency ridge and the result is shown in Figure 9c. The picture
displays that the time-frequency feature concentrated in a fine straight line.
Given the problem that the time-frequency resolution of wavelet scalogram is low, a reassigned
method was put forward. The theoretical definition of improved wavelet scalogram is expressed as:
x
RSG ( â, b̂; ψ) = ( â/a)2 SG ( a, b; ψ) × δ(b̂ − b0( a, b))δ( â − a0( a, b))dadb, (21)

wherein, b0 ( a, b), ψ0 (t), ψ̂(t) and ω0 /a0 ( a, b) can be calculated by the following equations:
( )
0 Wx ( a, b; ψ0 )Wx∗ ( a, b; ψ)
b ( a, b) = b − Re a , (22)
|Wx ( a, b; ψ)|2
( )
ω0 ω0 Wx ( a, b; ψ̂)Wx∗ ( a, b; ψ)
= + Im , (23)
a0 ( a, b) a 2πa|Wx ( a, b; ψ)|2
dψ
ψ0 (t) = tψ(t), ψ̂(t) = ( t ), (24)
dt
wherein, ω0 is the central angular frequency. SG ( ai , b j ; ψ) represents the average of the energy density
of a region which takes (ω0 /ai , ai t + b j ) as geometric center. Due to the fact the energy of this
region is not geometrically symmetric and it is unreasonable to concentrate the average energy at the
geometric center, at this time, some people advise placing the average energy at the local energy center
(ω0 / âi , âi t + b̂ j ). Then, a reassigned wavelet scalogram is put forward based on this principle.
Energies 2018, 11, 1286 12 of 23
Energies 2018, 11, x 12 of 23

Amplitude
0
-1
0 100 200 300 400 500 600 700 800 900 1000
Time(s)
(a)
0.5

0.4

Frequency(Hz)
0.3

0.2

0.1

0
0 200 400 600 800 1000
Time(s)
(b)
0.5

0.4
Frequency(Hz)

0.3

0.2

0.1

0
0 200 400 600 800 1000
Time(s)
(c)

Figure 9. Simulation signal analysis. (a) Time waveform for a line modulation signal; (b) Wavelet
Figure 9. Simulation signal analysis. (a) Time waveform for a line modulation signal; (b) Wavelet
scalogram for the line modulation signal; (c) Reassigned wavelet scalogram.
scalogram for the line modulation signal; (c) Reassigned wavelet scalogram.

For a detailed description about reassigned wavelet scalograms readers can consult [30]. In
For9caisdetailed
Figure easier todescription
identify how about reassigned
the frequency wavelet
changes with scalograms
time compared readers canFigure
with consult9b. [30].
The
In Figure 9c is easier to identify how the frequency changes with time
reassigned wavelet scalogram is a spectrogram that can improve readability after the average value compared with Figure 9b.
The reassigned wavelet
is concentrated scalogram
at the local energyis center.
a spectrogram that can improve
The mechanical vibrationreadability after theexhibits
signal generally averagea
value is concentrated at the local energy center. The mechanical vibration signal
periodic characteristic, and its cycle is variable with the change of device operating state. Moreover, generally exhibits a
periodic characteristic, and its cycle is variable with the change of device
the operation of equipment is vulnerable to external interferences. Therefore, vibration signals operating state. Moreover,
the operation
collected from ofrotating
equipment is vulnerable
machinery to external
are often mixedinterferences.
with externalTherefore,
interference vibration
that issignals collected
independent of
from rotating machinery are often mixed with external interference
fault information. The fault feature is weak when the external interference is strong [16,17]. that is independent of fault
The
information.
traditional time Thedomain
fault feature is weak when
and frequency the external
domain analysisinterference
method is unable is strong to[16,17].
diagnoseTheweak
traditional
faults.
time domain and frequency domain analysis method is unable to diagnose
Then, a series of new methods are exploited to improve the accuracy of the fault diagnosis, such weak faults. Then, a series
as
of new methods are exploited to improve
the synchronous averaging technique [19,34]. the accuracy of the fault diagnosis, such as the synchronous
averaging
Let ustechnique
suppose [19,34].
x(n) is a periodic vibration signal that was collected by a transducer mounted
Let us suppose x (n) is a periodic vibration signal that was collected by a transducer mounted on a
on a rotary machine, and its sampling point is N , where n = 1, 2, N . If the vibration signal x is
rotary machine, and its sampling point is N, where n = 1, 2, · · · N. If the vibration signal x is formed by
formed by M stationary components with the period is P , then N = P × M . We choose the Morlet
M stationary components with the period is P, then N = P × M. We choose the Morlet wavelet as the
wavelet as the basic wavelet to perform wavelet transform on x(n) . After that, we can implement
basic wavelet to perform wavelet transform on x (n). After that, we can implement the reassignment
the reassignment
process to obtain the process to obtain
reassigned wavelet thescalogram
reassigned wavelet
RTFR ( f , p) scalogram
according toRTFR(f,
Equation according
p) (21), where to f
Equation (21),
represents where f and
the frequency represents the frequency
p represents the time. and represents
If thep signal the time.
contains If the signal
F frequencies, thencontains F
the size of
frequencies,
RTFR ( f , p) isthen P. size of RTFR(f, p) is F × P .
F × the
Based on the reassigned wavelet scalogram, we can use time synchronous averaging for further
analysis. Each row corresponds to a frequency value, which is made up of repeating M
components with a period P . An average for each cycle and time synchronous RTFR(f, p) can be
obtained.
Energies 2018, 11, 1286 13 of 23

Based on the reassigned wavelet scalogram, we can use time synchronous averaging for further
analysis. Each row corresponds to a frequency value, which is made up of repeating M components
with a period P. An average for each cycle and time synchronous RTFR( f , p) can be obtained.
Energies 2018, 11, x 13 of 23
M −1
1
1∑
RTFR( f , p) = M -1 RTFR ( f , p + mP ), (25)
M

RTFR(f, p) = m=0 RTFR(f, p + mP) ,
M m =0
(25)

The process diagram of the improved wavelet scalogram is displayed in Figure 10.
The process diagram of the improved wavelet scalogram is displayed in Figure 10.

Figure 10. Multi-scales synchronous averaging of reassigned wavelet scalogram.

Figure 10. Multi-scales synchronous averaging of reassigned wavelet scalogram.
The value of P directly affects the synchronous averaging results. The characteristic frequency
The
is related thePrunning
valuetoof directlyspeed
affects the rolling
of the synchronous
element averaging results. in
bearings. However, The characteristic
some frequency
cases, they are not
integer multiple relationships, so a one-step rounding process is usually added.
is related to the running speed of the rolling element bearings. However, in some cases, they are not
integer multiple relationships, so a one-step rounding  T process
-T is usually added.
NUM = round  K 1 × fS  , (26)
 K - 1  
TK − T1
NU M = round
We can intercept the K impact series to calculate S ×f , (26)
K − 1 the average points per cycle. In normal
circumstances, we can select a proper signal length according to the signal characteristics for analysis
We can intercept
in order to improvethetheKcomputational
impact seriesefficiency,
to calculate thenotaverage
whilst affectingpoints per cycle.
the analysis InThe
results. normal
computational
circumstances, weefficiency
can selectof Equation
a proper (26)signal
will belength
low in the case where
according K issignal
to the large. In addition, the for
characteristics
calculation
analysis in order accuracy will decrease
to improve in the case where
the computational K is small.
efficiency, Therefore,
whilst when defining
not affecting the value
the analysis results.
of K , the calculation accuracy and computational complexity should both be taken into
The computational efficiency of Equation (26) will be low in the case where K is large. In addition, account.
the calculation accuracy will decrease in the case where K is small. Therefore, when defining the value
3.2. Wavelet Ridge
of K, the calculation accuracy and computational complexity should both be taken into account.
The wavelet ridge [35] is a sequence points extracted from the maximum of modulus values of
wavelet coefficients of time-frequency spectra at each moment. By analyzing the wavelet ridge, we
3.2. Wavelet Ridge
can explore the composition of signal components, like signal envelope, instantaneous amplitude,
The
and wavelet ridge phase.
instantaneous [35] is Suppose
a sequence x(t)points
is a extracted fromsignal.
simple cosine the maximum
It can beofexpressed
modulus as
values
of wavelet coefficients
x(t) = A(t)cos(φ(t)) andofitstime-frequency
analytic signal canspectra at each
be calculated moment.
by the following By analyzing the wavelet
expression.
ridge, we can explore the composition of signal components, like signal envelope, instantaneous
amplitude, and instantaneous phase.  Suppose
x(t) = x(t) + jxxH(t(t)
) is= aAsimple
x (t)e
jφx (t)
cosine
, (27)
signal. It can be expressed as
x (t) = A(t) cos(φ(t)) and its analytic signal can be calculated by the following expression.
wherein, xH (t) is the Hilbert transform of x(t) . A(t) > 0 is the instantaneous amplitude and φ(t) is
the instantaneous phase angle. The complex form of the mother wavelet
jφx (t) is denoted as Equation (28).
xe(t) = x (t) + jx H (t) = A x (t)e , (27)
 = A (t)e
ψ(t)
jφψ (t)
, (28)
wherein, x H (t) is the Hilbert transform of x (t). A(ψt) > 0 is the instantaneous amplitude and φ(t) is
the instantaneous  phase
Select ψ(t) angle.
as the Thebasis
wavelet complex form transform
for wavelet of the mother wavelet
to analyze is denoted
signal  .
x(t) as Equation (28).

e(t) = Aψ (t)e jφψ (t) ,

ψ (28)
Energies 2018, 11, 1286 14 of 23

e(t) as the wavelet basis for wavelet transform to analyze signal xe(t).
Select ψ

t−b
 
1 1
Z Z
WTx ( a, b; ψ) = √ e∗
xe(t)ψ dt = √ A a,b (t)e jφa,b (t) dt, (29)
2 a a 2 a
R R
(
A a,b (t) = A x (t) Aψ ∗ ( t−a b )
, (30)
φa,b (t) = φx (t) − φψ ( t−a b )

wherein, ψe∗ (t) is the conjugate of ψ

e(t).
The Morlet wavelet is selected as the basis function to match the vibration signal. That is
determined by its downhill feature in a short time. This is very similarity to the bearing fault impulse
signal. The expression of a Morlet complex wavelet is:

1 t2
ψ(t) = p exp(− )exp(i2π f c t), (31)
π fb fb

wherein, f b is the frequency band range and f c is wavelet center frequency. There is a one-to-one
correspondence between wavelet transform scale and frequency. Suppose sampling frequency is f s
and the corresponding frequency of scale a is:

fs fc
f = , (32)
a
After obtaining the wavelet ridge line, the wavelet coefficients corresponding to wavelet ridge
line position can be deduced by inverse transform formula [28,36].
p
ar ( b ) _∗
WTx ( ar (b), b) = × A x (b) × exp(iφx (b)) × ( φ ψ (ωc ) + ξ ( a, b)), (33)
2
_
ξ ( a, b) is the modification item, which is usually small and often can be ignored. φ ψ (ω ) is the Fourier
_
transform of φψ (t), φ (ω ) will reach maximum at the point of center frequency ωc . With a slight
modification for Equation (33), the amplitude of continuous wavelet transform is approximately
equal to: p
ar ( b ) _∗
|WTx ( ar (b), b)| ≈ × A x ( b ) × φ ψ ( ω c ), (34)
2
In wavelet function expression, the parameter b represents time shift. Then, b can be taken as t.
Accordingly, WTx ( ar (b), b) can be represented as WTx ( ar (t), t). In this circumstance, the instantaneous
amplitudes of signal envelope are estimated by:

2|WTx ( ar (t), t)|

A x (t) ≈ p , (35)
ar ( t )

In the same way, Equation (20) can be transformed into Equation (36).

ar ( t )
SGx ( ar (t), t) ≈ ( A x (t))2 , (36)
4

Afterwards, instantaneous amplitudes of signals x (t) can be approximated by Equation (37).

p
2|WTx ( ar (t), t)| 2 SGx ( ar (t), t)
A x (t) ≈ p = p , (37)
ar ( t ) ar ( t )

It can be seen from Equation (37) that the wavelet ridge extraction process is actually a
demodulation process. In the same way, the instantaneous amplitudes of signal can be estimated on
Energies 2018, 11, 1286 15 of 23

Energies 2018,wavelet
the improved 11, x scalogram according to Equation (37). There are two different ways 15 to
of extract
23

wavelet ridge. The one way is based on wavelet coefficient phase values and the other one is based on
It can be seen from Equation (37) that the wavelet ridge extraction process is actually a
modulus values. By
demodulation comparing
process. In the both
sameanalysis
way, the results, it can amplitudes
instantaneous be concluded that the
of signal canmethod on the
be estimated on basis
of wavelet coefficient modulus is more accurate. Because the local maximum
the improved wavelet scalogram according to Equation (37). There are two different ways to extract value of the wavelet
scalewavelet
spectralridge.
coefficients
The one and
way isthe continuous
based on wavelet wavelet transform
coefficient coefficients
phase values and theis the one
other same point that
is based
is theon
wavelet
modulus ridge point
values. ( ar (b), b) [28].
By comparing both In this paper,
analysis results,the wavelet
it can ridge isthat
be concluded extracted
the methodby calculating
on the
basis of coefficient
the wavelet wavelet coefficient
modulus modulus
maximum is more
and accurate.
a new time Because
seriesthe
waslocal maximum value
reconstructed. of the [37]
Reference
wavelet scale spectral coefficients and the continuous wavelet transform
proposed a method to extract the wavelet ridge which minimizes the cost function. This method coefficients is the same only
point that is the wavelet ridge point ( a (b), b ) [28]. In this paper, the wavelet
needs the model information to extract ther ridge line. There are n − 1 cost functions along the wavelet ridge is extracted by
ridgecalculating
line described the wavelet coefficient equation.
in the following modulus maximum and a new time series was reconstructed.
Reference [37] proposed a method to extract the wavelet ridge which minimizes the cost function.
This method only needs
CFk =the−|model
W ( a(kinformation
), k)|2 +| a(kto) −
extract
ar (k the
− 1ridge
)|2 , k line.
= 2, There are n -1 cost functions (38)
3, · · · n,
along the wavelet ridge line described in the following equation.
The ridge point can be determined CFk = -|W(a(k),according
k)|2 +|a(k) to
- arthe Equation
(k - 1)|2
n , If the (k − 1)th ridge
(38).
, k = 2, 3, (38) point
( ar (k − 1), k − 1) was derived, then the k − th ridge point {( ar (k), k)} can be determined by minimizing
The ridge point can be determined according to the Equation (38). If the ( k − 1) th ridge point
the CFk . During run time, each choice is evaluated in order. When all of the items (from k = 2 to k = n)
) was derived,
( ar ( k - 1) , k - 1calculated,
are completely thethen the k -th ridge
instantaneous point {( arof
amplitudes k )} signal
(k),the can becandetermined by minimizing
be estimated according to
the CF
Equation k . During
(37). Due torun thetime, each fault
bearing choicesignal
is evaluated
presents in periodic
order. When all ofcharacteristics,
impact the items (from which
k = 2 to
can be
k = n ) are completely calculated, the instantaneous amplitudes of
identified from the instantaneous amplitude. In general, a small scale band is chosen for analysis, the signal can be estimated
therebyaccording
improving to Equation (37). Due
the accuracy of to the bearing
wavelet ridgefault signal presents
extraction results. periodic impact characteristics,
which can be identified from the instantaneous
It is indicated that 3 dB bandwidth is the most suitable choiceamplitude. In general, a small scale
according band
to the is chosen of
characteristic
for analysis, thereby improving the accuracy of wavelet ridge extraction results.
signal being analyzed. The region where the wavelet coefficients modulus is large usually corresponds
It is indicated that 3 dB bandwidth is the most suitable choice according to the characteristic of
to the range of the wavelet ridge. Besides, selecting a suitably small scale to analyze can reduce the
signal being analyzed. The region where the wavelet coefficients modulus is large usually
calculation
corresponds time. to the range of the wavelet ridge. Besides, selecting a suitably small scale to analyze can
reduce the calculation time.
3.3. Incipient Feature Extraction Method
3.3. Incipient
The processFeature Extraction
flow diagram ofMethod
rolling element bearing fault diagnosis based on improving wavelet
scalogramThe andprocess
waveletflow
ridge is shown
diagram of in Figure
rolling 11. This
element method
bearing is readily
fault available
diagnosis based for
on rolling
improvingelement
wavelet
bearing faultscalogram and wavelet
feature extraction. Toridge
beginiswith,
shown theinimproved
Figure 11. wavelet
This method is readily
scalogram available
plays for
an important
rolling
role in element
extracting bearing
and fault feature
enhancing extraction.
the impact To begin with,
characteristics. the improved
In addition to that,wavelet scalogram
the wavelet ridge can
plays
identify an important
fault frequencyrole in extracting
more and method
clearly. This enhancing the impact
does characteristics.
not require any priorInknowledge
addition to of
that,
rolling
the wavelet ridge can identify fault frequency more clearly. This method does not require any prior
element bearing fault information, and can be applied to determine the weak fault characteristics for
knowledge of rolling element bearing fault information, and can be applied to determine the weak
rolling element bearings.
fault characteristics for rolling element bearings.

Data acquisition for

vibration and speed signals

Continuous wavelet for

Rolling element decomposition and
bearing type ressigned wavelet scalogram

Multi-cycle sychronous
Rolling bearing averaging for reassigned
operational condition wavelet scalogram

Wavelet ridge determination

Rolling bearing
based on improved wavelet
characteristic frequency
scalogram

Early fault classification for rolling bearing

based on characteristic frequency analysis

Figure 11. Flowchart for rolling element bearing fault diagnosis.

Figure 11. Flowchart for rolling element bearing fault diagnosis.
Energies 2018, 11, 1286 16 of 23

Energies 2018, 11, x 16 of 23

4. Algorithm Verification
4. Algorithm Verification
4.1. Simulation
Energies 2018,Analysis
11, x 16 of 23
4.1.order
In Simulation Analysis
to prove the validity of the wavelet ridge line method based on improved wavelet
4. Algorithm Verification
scalogram, Inthe simulation
order to prove signal that isofestablished
the validity the waveletbyridge
researching the characteristics
line method of rolling
based on improved element
wavelet
scalogram,
bearing
4.1.taken asthe
Simulation simulation
analytical signal
anAnalysis object.that
Weisadd established
Gaussian bywhite
researching the characteristics
noise with of rolling
a signal-to-noise ratio size
element
of −15 dBInintobearing taken as
thetoplanet an analytical object.ring
We fault
add Gaussian white noisemodel
with a which
signal-to-noise
order prove gear bearingofouter
the validity the wavelet ridgesignal simulation
line method based on improved is shown in
wavelet
ratio
Figure size of −15 dB into the planet gear bearing outer ring fault signal simulation model which is
scalogram, the simulation signal that is established by researching the characteristics of rolling can
8a. The time-domain impact characteristic signal is almost invisible in Figure 12, which
shown in Figure 8a. The time-domain impact characteristic signal is almost invisible in Figure 12,
represent
elementas abearing
weak impact
taken assignal.
an analytical object. We add Gaussian white noise with a signal-to-noise
which can represent as a weak impact signal.
ratio size of −15 dB into the planet gear bearing outer ring fault signal simulation model which is
shown in Figure 8a. The time-domain
0.04 impact characteristic signal is almost invisible in Figure 12,
which can represent as a weak
0.02 impact signal.
AmplitudeAmplitude

0
0.04
-0.02
0.02
-0.04
00 0.25 0.5 0.75 1 1.25 1.5
Time (s)
-0.02
(a)
-0.04 -4
x0 10 0.25 0.5 0.75 1 1.25 1.5
3 Time (s)
AmplitudeAmplitude

2 (a)
-4
x 10
3
1
2
0
0 50 100 150 200 250 300
1 Frequency(Hz)
(b)
0
0 50 100 150 200 250 300
Frequency(Hz)
Figure 12. The analytic results of simulation signal with Gaussian white noise. (a) The time domain
Figure 12. The analytic results of simulation signal with Gaussian white noise. (a) The time domain
waveform for noisy simulation signal; (b) The envelope
(b) spectrum for noisy simulation signal.
waveform for noisy simulation signal; (b) The envelope spectrum for noisy simulation signal.
Figure 12. The analytic
Subsequently, results spectrum
the envelope of simulation signal with
is applied on Gaussian white noise.
the noise-added (a) The time
simulation domain
signal. It can be
waveform for
seen that the envelope noisy simulation signal;
analysisspectrum
is infeasible(b) The envelope
for weakon spectrum
impact for noisy simulation
feature extraction. signal.
Then, asignal.
waveletIt can
Subsequently, the envelope is applied the noise-added simulation
transform
be seen that is further
the envelope performed on it. Its reassigned wavelet scalogram was calculated as shown in
Subsequently, the analysis
envelope is infeasible
spectrum for weak
is applied on theimpact featuresimulation
noise-added extraction. Then,
signal. a wavelet
It can be
Figure
transform 13. Many
is further spectral
performed lines are
on it. scattereded
Its reassigned in the wavelet
wavelet scalogram
scalogram spectrum. It
was calculated is difficult to
seen that the envelope analysis is infeasible for weak impact feature extraction. Then, a as shown in
wavelet
recognize the impact information by adopting the reassigned wavelet scalogram method under the
Figure 13. Many
transform spectral
is further lines areon
performed scattereded
it. Its reassigned in thewavelet
wavelet scalogram
scalogram wasspectrum.
calculated as It shown
is difficult
in to
situation of strong noise. There are many interfering components in the spectrum because of outside
recognize
Figurethe13.impact information
Many spectral lines areby adopting
scatterededthe in reassigned wavelet scalogram
the wavelet scalogram spectrum. It method under
is difficult to the
noise. In order to eliminate more irrelevant interference information, the scalogram is further
recognize
situation of the impact
strong noise. information
There are by adopting
many interfering the reassigned
components wavelet
in scalogram
the spectrum method
becauseunder of the
outside
processed by synchronous averaging where the synchronous cycle is 400 ms. The improved wavelet
situation
In orderofto
noise.scalogram strong noise.more There are manyinterference
interfering components in the spectrum because of outside
of eliminate
the noisy simulation irrelevant
signal is shown in Figure information,
14. Thethe scalogram
impact is further
characteristics from processed
the
noise.
by synchronous In order to
averaging eliminate more irrelevant interference information, the scalogram is further
time-frequency spectrumwhere can bethe synchronous
easily distinguished. cycleAllisof400
thems. Thecomponents
impact improved wavelet scalogram of
in the simulation
processed by synchronous averaging where the synchronous cycle is 400 ms. The improved wavelet
the noisy
signalsimulation signalinis the
are concentrated shown in Figure
frequency axis 14.of 500TheHz.
impact characteristics
The time from the
interval between time-frequency
the two impact
scalogram of the noisy simulation signal is shown in Figure 14. The impact characteristics from the
positions is 100 ms. That is to say that the impact period
spectrum can be easily distinguished. All of the impact components in the simulation signal are is 100 ms and the impact characteristic
time-frequency spectrum can be easily distinguished. All of the impact components in the simulation
frequencyin
concentrated is 10
theHz. This is consistent
frequency axis of with
500 the Hz. theoretical value of thebetween
simulation signal.
signal are concentrated in the frequency axisThe time
of 500 Hz.interval
The time interval the two impact
between the twopositions
impact is
100 ms. That is to say that the impact
positions is 100 ms. That is to say that period is 100 ms and the impact characteristic
the impact period is 100 ms and the impact characteristic frequency is 10 Hz.
55
This isfrequency
consistent is 10with the theoretical
Hz. This is consistent valuewith the oftheoretical
the simulationvalue ofsignal.
the simulation signal.
44
Frequency(Hz)
Frequency(kHz)

3355

2244
Frequency(Hz)
Frequency(kHz)

1133

002
20 50 150 200
100 150
50 100 200 250
250 300 350 400
300 350 400
Time(ms)
Time (ms)
11
Figure 13. Time frequency 0reassignment
0
distribution for the simulation signal.
0 50 150 200
100 150
50 100 200 250
250 300 350 400
300 350 400
Time(ms)
Time (ms)

Figure 13. Time frequency reassignment distribution for the simulation signal.
Figure 13. Time frequency reassignment distribution for the simulation signal.
Energies 2018, 11, 1286 17 of 23
Energies 2018, 11, x 17 of 23

55
Energies 2018, 11, x 17 of 23
Energies 2018, 11, x 44 17 of 23

Frequency(Hz)
Frequency(kHz)
33 55
55
44

Frequency(Hz)
22 44

Frequency(Hz)
Frequency(kHz)
100 ms

Frequency(kHz)
33
11 33

00 222
2

0 50 100
50 100 ms 150
100 150 200 250
200 250 300
300 350
350 400
400
100 ms Time (ms)
11
11
Time(ms)
00
Figure 000 50 100 150 200 250
200 250 300
300 350
350 400
Figure14.
14.The
Thescalogram under synchronous averaged.
50 100 150 400
scalogram
0 50
50 100
100 under
150 (ms) synchronous
Time200
150
Time(ms)
200 250
250 300
300 350 400 averaged.
350 400
Time (ms)
Time(ms)
Figure
Although it can be seen that 14. The scalogram
the14.wavelet under
scale spectrum synchronous averaged.distribution characteristic in
has a periodic
Although it can be seenFigure
that the The scalogram
wavelet scaleunder
spectrumsynchronous
has aaveraged.
periodic distribution characteristic
Figure 14, the impact information is distributed in an area of the block rather than a point or a line.
in Figure Although
14, the impact
it can beinformation is distributed
seen that the wavelet in an has
scale spectrum areaa periodic
of the block rathercharacteristic
distribution than a point in or a
The fault Although
period cannotit can bebeseen extracted from Figure
that the wavelet 14 correctly
scale spectrum using distribution
has a periodic the synchronous averaging
characteristic in
Figure
line. The 14, period
fault the impact information
cannot be is distributed
extracted from in an area
Figure 14 of the block
correctly rather
using the than a point or aaveraging
synchronous line.
Figure
scalogram. 14,
There the impact
is a cannot information
lack ofbeintuition is distributed
and veracity.in an area
In this of the
paper, block rather
further than a point or
analysis is performed a line. by
The fault
scalogram. period
There is acannot extractedand
lack ofbeintuition from Figure 14
veracity. In correctly
this paper, using the synchronous
further averaging by
analysis is performed
The fault
usingscalogram.
the wavelet period
ridge extraction extracted
method. from
The Figure
analysis14 correctly
scale areausing
is the synchronous
limited to between averaging
6 and 29 in
usingscalogram.
the wavelet There is a extraction
ridge lack of intuition and veracity.
method. The analysisIn thisscale
paper,area
further analysistois between
is limited performed6 by and 29
termsusing
of thethe There
characteristic is a lack of intuition and veracity. In this paper, further analysis is performed by
wavelet ridgefrequency
extraction range
method. obtained by the
The analysis computational
scale area is limitedequation
to between of 6wavelet
and 29 in ridge.
in terms of the characteristic frequency range obtained by the computational
using the wavelet ridge extraction method. The analysis scale area is limited to between 6 and 29 in equation of wavelet
The wavelet
terms of theridge extractedfrequency
characteristic from Figure 14 is shown
range obtained in Figure 15.
by the computational Then, ofthe
equation instantaneous
wavelet ridge.
ridge.terms
The of the characteristic
wavelet ridge extractedfrequency fromrange obtained
Figure 14 isbyshown
the computational
in Figure 15. equation
Then,ofthe wavelet ridge.
instantaneous
amplitude and theridge
The wavelet corresponding
extracted from frequency
Figure can
14 isbeshown
calculated according
in Figure to Equation
15. Then, (37) and the
the instantaneous
The wavelet
amplitude and theridge extracted from
corresponding Figure can
frequency 14 isbeshown in Figure
calculated 15. Then,
according to the instantaneous
Equation (37) and the
Fourieramplitude
transform, and the
as corresponding
shown in Figure frequency
16a,b. canThebe calculated
impact according tofrequency
characteristic Equation (37) 10 and
Hz the
and its
Fourieramplitude
transform, and the corresponding
as shown in inFigurefrequency
16a,b. can
The be calculated according
impactcharacteristic
characteristic to Equation
frequency (37) and
10 and
Hz and the
Fourier
multiplied transform,
frequency as shown
components Figure
are 16a,b.
clearly The impact
displayed in Figure 16b. frequency 10 Hz its its
Fourier transform, as shown in Figure 16a,b. The impact characteristic frequency 10 Hz and its
multiplied frequency
multiplied frequencycomponents
components areareclearly
clearlydisplayed
displayed in in Figure
Figure16b.16b.
multiplied frequency components are clearly displayed in Figure 16b.
1.4
1.4
1.4
1.4
1.2
1.21.4
1.4
1.2
1.2
Frequency(Hz)

111.2
Frequency(Hz)

1.2
11
Frequency(kHz)
Frequency(Hz)
Frequency(kHz)

11
0.8
0.8
Frequency(kHz)

0.8
0.8
0.8
0.8
0.6
0.60.6
0.6
0.6
0.6
0.4
0.40.4
0.4
0.4
0.4
0.20.2
0.2 0.2
0.2
0.2
000 000000 100 200
100 200 300 200 300 400
400
0 00 100
100 200
100 Time (ms) 300
300 400
400
100 Time(s)
Time (ms) 300
200
200
Time(ms)
Time(s)
300 400
400
Time(ms)
Time (ms)
Time(s)
Time(ms)
Figure 15. Wavelet ridge for improved wavelet scalogram.
Figure
Figure
Figure Wavelet
15.15.
15. Wavelet ridge
ridge
Wavelet for
for
ridge improved
forimproved waveletscalogram.
improved wavelet
wavelet scalogram.
scalogram.
(a) 0.02 -3
(a) 0.02(a) 0.02 b) 1.5 x 10 -3-3
100 ms 100 ms b) 1.5x10x1010
Hz
100 ms 100 ms b) 1.5 1020
HzHz
50 Hz
0. 015100 ms 100 ms 10 Hz20 Hz40 Hz5060
Hz
Hz
Amplitude(g)

0. 015 30 Hz 50 Hz
20 Hz 40 Hz 60 Hz
Amplitude(g)

0. 015
Amplitude(g)

1 3040
Hz Hz 7 0 Hz
60 Hz80 Hz
Amplitude(g)

Amplitude(g)

1 30 Hz 7 0 Hz
Amplitude(g)

0.01 1 80 Hz
7 0 Hz
0.01 80 Hz
0.01 0.5
0. 005 0.5
0. 005 0.5
0. 005 0
0 00 50 100 150 200 250
00 100 200 300 400 0 50 Frequency
100 150 (Hz) 200 250
0 100 Time200
(ms) 300 400
Time (ms) 0 Frequency (Hz)
0 150 200 250 0 50 100
0 100 (a)
200 (b) 300 400
Time (a)
(ms) Frequency (Hz)
(b)
Figure 16. The analytic (a)results of simulation signal’s instantaneous amplitude. (a) Time-domain
Figure 16. The analytic results of simulation signal’s instantaneous(b) amplitude. (a) Time-domain
waveform ofanalytic
simulation signal’s instantaneoussignal’s
amplitude; (b) The frequency domain (a)waveform of
Figurewaveform of simulation signal’s simulation
16. The results of instantaneous
instantaneous amplitude; amplitude.
(b) The frequency Time-domain
domain waveform of
Figure 16. The
simulation
waveform analytic
signal’s
of simulation results of
instantaneous simulation
amplitude.
signal’s instantaneous signal’s instantaneous amplitude. (a) Time-domain
simulation signal’s instantaneous amplitude. amplitude; (b) The frequency domain waveform of
waveform
simulationof simulation
signal’s signal’s instantaneous
instantaneous amplitude. amplitude; (b) The frequency domain waveform of
When
simulation the signal is disturbed by strong background noise, fault information cannot be separated
Whensignal’s instantaneous
the signal is disturbedamplitude.
by strong background noise, fault information cannot be separated
from the time domain waveform directly. The weak impact characteristics of the noisy simulation
from
When the time
the signal domain
is and waveform
disturbed directly. The weak impact characteristics of the cannot
noisy simulation
signal the
When were clearly
signal exactly by
is disturbed
strong by
extracted background
adopting thenoise, fault information
improved wavelet scalogram be separated
cannotalgorithm
signal were clearly and exactlybyextracted
strong background
by adopting noise, fault information
the improved wavelet scalogram be separated
algorithm
from which
the time is domain
proposedwaveform directly.
in this paper. The The weak impact
reassigned waveletcharacteristics
scalogram wasofused the noisy simulation
to enhance impactsignal
from the
whichtimeis domain
proposedwaveform
in this paper.directly. The weakwavelet
The reassigned impactscalogram
characteristics
was used of to
theenhance
noisy simulation
impact
were features
clearly firstly.
and Then,extracted
exactly the outside
by interference
adopting thecomponents
improved were decreased
wavelet scalogram after synchronous
algorithm which is
signal features
were clearly
firstly.and
Then, exactly extracted
the outside by adopting
interference the improved
components wavelet scalogram
were decreased algorithm
after synchronous
averaging is performed
proposed in thisispaper. The onon the reassigned
reassigned wavelet
wavelet scalogram.
scalogram Finally,
was Finally, the visibility
used to enhance and the
impactthe accuracy
features firstly.
whichaveraging
is proposed performed
in this paper. the reassigned wavelet
The reassigned scalogram.
wavelet scalogramthe visibility
was used and accuracy
to enhance impact
features firstly. Then, the outside interference components were decreased after synchronous
averaging is performed on the reassigned wavelet scalogram. Finally, the visibility and the accuracy
Energies 2018, 11, 1286 18 of 23

Then, the outside interference components were decreased after synchronous averaging is performed
on the reassigned wavelet scalogram. Finally, the visibility and the accuracy of time-frequency
spectrum were improved by extracting the wavelet ridge. The wavelet coefficients characterize the
similarity between the signal being analyzed and the basic wavelet. The larger the wavelet coefficients,
Energies 2018, 11, x 18 of 23
the stronger the impact characteristic of the signal and the more obvious the fault feature. The modulus
Energies 2018, 11, x 18 of 23
of wavelet coefficients of
of time-frequency every frequency
spectrum were improved points at each moment
by extracting the waveletcanridge.
be highlighted by extracting the
The wavelet coefficients
waveletcharacterize
ridge on improved
the similarity
of time-frequency wavelet
spectrum between scalogram.
the signal
were improved By further
by being analyzed
extracting analysis
and the
the wavelet ofThe
the
basic
ridge. wavelet
wavelet.
wavelet The ridge, it is easily
larger the
coefficients
characterize
wavelet the similarity
coefficients, the between
stronger the
the signal
impact being analyzed
characteristic and
of thethe basic
signal
to extract the time domain and frequency domain information of fault impulses. By analyzing the wavelet.
and the The
more larger
obviousthe the
wavelet
fault coefficients,
feature. The modulusthe stronger the impact
of wavelet characteristic
coefficients of everyof the signalpoints
frequency and the more moment
at each obvious the
can be
simulation signals, it can be concluded that the proposed method is suitable for detecting the weak
fault feature.
highlighted by The modulus
extracting theofwavelet
wavelet ridge
coefficients of every frequency
on improved points at each
wavelet scalogram. momentanalysis
By further can be of
impact features.
thehighlighted
wavelet ridge,by extracting thetowavelet
it is easily extractridge on improved
the time domain andwavelet scalogram.
frequency By further
domain analysis
information ofoffault
the wavelet
impulses. By ridge, it is easily
analyzing to extract the
the simulation time domain
signals, it can and frequency domain
be concluded that theinformation
proposed of fault is
method
4.2. Experimental
impulses.
suitable
Study
By analyzing
for detecting the simulation
the weak signals, it can be concluded that the proposed method is
impact features.
suitable for detecting the weak impact features.
In order to further demonstrate the validity of the method presented in this research, signals
4.2. Experimental Study
were collected from a planet
4.2. Experimental Study gear rolling element bearing experiment rig for analysis. The schematic
diagram of the In order
bearingto further demonstrateexperiment
test-to-failure the validity of the method presented in this 17.
research, signals
In order to further demonstrate the validity ofsystem is shown
the method presented in in
Figure
this research,Onsignals
the laboratory
were collected
were collected from a planet gear
from a planet drives rolling
gear rolling element bearing experiment rig for analysis. The schematic
bench, a frequency transformer oneelement
motorbearing experiment
connected to therig for analysis.housing,
bearing The schematicthe planetary
diagram
diagramofofthe thebearing
bearingtest-to-failure
test-to-failure experiment systemisisshown
shownininFigure
Figure 17.17.
OnOn thethe laboratory
gearboxbench,
and the magnetic
a afrequency powder brakeexperiment
through system
motorthe coupling. The planet gear laboratory
bearing outer ring is
bench, frequencytransformer
transformer drives one motor
drives one connected
connected totothethe bearing
bearing housing,
housing, thethe planetary
planetary
machined gearbox and the magnetic powder brake through the coupling. The planet gear bearing outer ring is is
with
gearbox a
and 1 mm
the wide
magnetic fault and
powder the
brake bearing
through surface
the structure
coupling. The after
planet machining
gear bearing failure
outer ring is shown
in Figure 18.
machinedPlanetary
with a 1gearbox
mm wide
machined with a 1 mm wide fault and parameters
fault and the are shown
bearing in
surface Table
structure 1. After
after calculation,
machining the
failure
bearing surface structure after machining failure is shown isparameters
shown of
in Figure
in Figure 18.
18.Planetary
Planetary gearbox
gearbox parameters
parameters
each component during operation are shown in Table 2. are
are shown
shown in
in Table
Table 1.1. After
After calculation,
calculation, thethe parameters
parameters of of
each component
each componentduring
duringoperation
operation are
are shown inTable
shown in Table2.2.

(a)
(a)

(a)
(a)
(b)
(b)
Electric
motor
Electric
motor

Coupling
Coupling Coupling
Coupling Magnetic
Coupling Planetary Coupling
Frequency Plummer block gearbox Powder Brake
transformer housing Planetary Magnetic
Frequency Plummer block gearbox Powder Brake
transformer housing(b)

(b)
Figure 17. A bearing test-to-failure experiment system. (a) System structure; (b) Test system.
Figure 17. A bearing test-to-failure experiment system. (a) System structure; (b) Test system.
Figure 17. A bearing test-to-failure experiment system. (a) System structure; (b) Test system.

Figure 18. Planetary bearing outer ring fault surface map.

Figure 18. Planetary bearing outer ring fault surface map.

Figure 18. Planetary bearing outer ring fault surface map.
Energies 2018, 11, x 19 of 23

Table 1. Parameters for planetary gearbox and planetary bearing.

Energies 2018, 11, 1286 Configuration Parameter Parameter Values 19 of 23

The number of sun gear tooth 17
The number of planet gear tooth 35
Table 1. Parameters for planetary
The number of annular gear tooth gearbox and planetary
88 bearing.
The number of planet gear 3
Configuration Parameter Parameter Values
Transmission ratio 6.176
The number of sun gear tooth 17
Type of planetary bearing
The number of planet gear tooth
204 NJ
35
The number of planetary
The number bearing
of annular roller
gear tooth 88 11
The number
The bearing of planet
pitch gear
diameter 3 mm
34
Transmission ratio 6.176
The planetary rolling element diameter
Type of planetary bearing
7.493 mm
204 NJ
The planetary
The number bearing contact
of planetary bearingangle
roller 11 0°
The bearing pitch diameter 34 mm
The planetary rolling element diameter 7.493 mm
Table 2. Parameters
The planetary bearingfor failure
contact angleexperiment 0system.
◦

Characteristic Frequency Frequency Values

Table 2. Parameters for failure experiment system.
Sampling Frequency 12,800 Hz
The rotation speed
Characteristic Frequency
1800 rpm
Frequency Values
The rotation frequency of sun gear 30 Hz
Sampling Frequency 12,800 Hz
The rotation frequency
The rotationofspeed
planet carrier 18004.86
rpm Hz
The relativeThe
rotation
rotationfrequency
frequency ofof
sunplanet
gear gear 3012.21
Hz Hz
The rotation frequency of planet carrier 4.86 Hz
The absolute rotation frequency of planet gear 7.36 Hz
The relative rotation frequency of planet gear 12.21 Hz
BearingThe
outer ring rotation
absolute fault characteristic frequency
frequency of planet gear 31.53
7.36 Hz Hz
BearingBearing
inner outer
ring ring
faultfault characteristic frequency
characteristic frequency 31.53 Hz Hz
49.38
Bearing inner ring fault characteristic frequency 49.38 Hz

In the event of a failure in the outer ring of planet gear, the fault characteristics of the original
In the
acceleration eventmeasured
signal of a failureby
in sensors
the outerare
ringweak
of planet gear, of
because thethe
faultlong
characteristics
and complex of the original
transmission
acceleration signal measured by sensors are weak because of the long and complex transmission
path between the vibration source and measuring points. There is no impact characteristic in its time
path between the vibration source and measuring points. There is no impact characteristic in its
domain waveform as shown in Figure 19a. Only the rotation frequency component of the shaft and
time domain waveform as shown in Figure 19a. Only the rotation frequency component of the shaft
its differential frequency
and its differential components
frequency withwith
components the the
planetary
planetarycarrier
carrier exist inFigure
exist in Figure 19b.
19b. TheThe
faultfault
characteristic frequency
characteristic cannot
frequency bebe
cannot determined
determinedfromfrom the frequencydomain
the frequency domain waveform.
waveform.

(a) 20
Amplitude(g)

10

-10

-20
0 1 2 3 4 5
Time (s )

(a)
(b) 0.5
f1=fs-fc
f4 f 5 f2=fr
0.4
f3 f3=2fr
f4=3fr
Amplitude (g)

f2 f7
0.3 f5=4fr
f6=5fr
0.2 f f7=6fr
1 f6 f8
f8=7fr -4f c
f9
f9=7fr
0.1

0
0 100 200 300 400 500
Frequency (Hz)

(b)

Figure 19. Cont.

 Energies 2018, 11, 1286 20 of 23

Energies 2018, 11, x 20 of 23

(c) 1.25
Energies 2018, 11, x 0.97 s 20 of 23
1

Amplitude(g)
(c) 1.25
0.75 0.97 s
1
0.5

Amplitude(g)
0.75
0.25
0.5
0
0.25
-0.125
0 0.2 0.4 0.6 0.8 1
0 Time (s)
-0.125
0 0.2 0.4 0.6 0.8 1
(c)Time (s)
(c)
Figure 19. The analytic results of planetary bearing outer ring fault. (a) The time domain waveform
Figure 19. The analytic results of planetary bearing outer ring fault. (a) The time domain waveform of
of planetary
planetarybearing
Figure 19.bearingouter
outerring
The analytic ringfault;
results (b)
(b) The
of planetary
fault; envelope spectrum
bearing outer
The envelope ring of
spectrum ofplanetary
planetary
fault. bearing
(a) Thebearing
time outer
domain
outer ring
waveform
ring fault;
fault;
of
(c) The planetary
key phase bearing
signal outer
of sunring fault;
gear (b)
input The envelope
shaft.
(c) The key phase signal of sun gear input shaft. spectrum of planetary bearing outer ring fault;
(c) The key phase signal of sun gear input shaft.
It canIt be
canconcluded
be concluded that thethe
that traditional
traditionalspectral
spectral analysis methodisisinapplicable
analysis method inapplicableforfor planetary
planetary
It can be concluded that the traditional spectral analysis method is inapplicable for planetary
bearing fault fault
bearing diagnosis. According
diagnosis. to thetokey
According thephase signalsignal
key phase shown in Figure
shown 19c, the
in Figure 19c,sun
thegear
sun rotation
gear
bearing fault diagnosis. According to the key phase signal shown in Figure 19c, the sun gear rotation
frequency
rotation offrequency
the planetary
of the gearbox
planetaryisgearbox
approximately 29.7 Hz.
is approximately Defining
29.7 a value
Hz. Defining of 127
a value ms ms
of 127 as the
frequency of the planetary gearbox is approximately 29.7 Hz. Defining a value of 127 ms as the
synchronization cycle,
as the synchronization afterward, a continuous
cycle, afterward, wavelet
a continuous transform
wavelet was
transform carried out
was carried
synchronization cycle, afterward, a continuous wavelet transform was carried out on the planet
on
out the planet
on the
planet
bearing outerbearing outer ringsignals.
ring ring
vibration vibration signals. Twenty synchronization
Twenty period signals were for
selected
bearing outer vibration signals. Twentysynchronization periodsignals
synchronization period signals were
were selected
selected analysis.
for analysis.
for analysis.
The wavelet The
scalogram wavelet
andand scalogram
improved and improved wavelet scalogram results are displayed in
The wavelet scalogram improvedwavelet
wavelet scalogram results are
scalogram results aredisplayed
displayedin in Figure
Figure 20a,b,
20a,b,
Figure
respectively. 20a,b, respectively.
respectively.

（a） （a）
66 66 （b）6
（b）6
55 55
555
Frequency(Hz)

5
Frequency(Hz)

Frequency(Hz)
Frequency(kHz)
Frequency(kHz)

Frequency(Hz)
Frequency(kHz)
Frequency(kHz)

44 4
44 44
33 33
33 3
22 22
22 22
11 1
11 1
00 00
0 20
20 40
40 60
60 80
80 100
100 120
120
00 20
20 40
40 60
60 80
80 100
100 120
120
00 Time (ms)
60Time(ms) 00
40Time(ms)
0 20
20 40
40 60 80
80 100 120
100 120 Time (ms)
Time (ms) 00 20
20 40 60
60 80
80 100
100 120
120
Time(ms)
(a) Time(ms)
Time (ms)
(b)
(a) (b)
Figure 20. Improved wavelet scalogram analysis. (a) Reassigned scalogram; (b) Scalogram under
Figure 20. Improved wavelet scalogram analysis. (a) Reassigned scalogram; (b) Scalogram under
Figure 20. Improved
synchronous
synchronous wavelet scalogram analysis. (a) Reassigned scalogram; (b) Scalogram under
averaged.
averaged.
synchronous averaged.
The wavelet ridges extracted from the improved wavelet scalogram are shown in Figure 21. The
The wavelet
instantaneous ridges extracted
amplitude and the from the improved
corresponding wavelet
frequency were scalogram
calculated are shown intoFigure
according Equation 21.
The
The wavelet
instantaneousridges extracted from the improved wavelet scalogram are shown in Figure 21.
to The
(37) and the Fourieramplitude
transform and the corresponding
formula, as shown in Figure frequency were
22. It can be calculated
observed that according
the planet
instantaneous
Equation amplitude
(37) and and
thefault
Fourierthe corresponding
transform formula, frequency
as shown were calculated according to Equation
bearing outer ring characteristic frequency is 31.53in HzFigureand22.itsIt multiplied
can be observed that
frequency
(37) and
the the Fourier
planet bearing transform
outer ring formula,
fault as shown
characteristic in Figure
frequency is 22. ItHz
31.53 can
andbeitsobserved
multiplied that the planet
frequency
components are clearly displayed in the partial enlarged view in the upper right corner of Figure
bearing outer ring
components are gearfaultdisplayed
clearly characteristic frequency is view
31.53 Hzupperand right
its corner
multiplied frequency
22b. For planet bearing faultin the partial
signals with enlarged
weak impact in the
characteristics, of Figure
the fault features 22b.
cannot
components
be directly extracted by envelope spectrum analysis. Fortunately, the interference informationFigure
For planet are clearly
gear bearingdisplayed
fault in
signals thewithpartial
weak enlarged
impact view in
characteristics,the upper
the fault right corner
features cannot of beis
directly
22b. For planetextracted
gear by envelope
bearing spectrum
fault signals analysis.
with Fortunately,
weak impact the interference
characteristics,
weakened by using the improved wavelet scalogram analysis. This method is very effective for information
the fault is weakened
features cannot
by using
be directly
enhancing thetheimproved
extracted wavelet scalogram
by characteristics.
impact envelope spectrum analysis.
analysis.
Subsequently, This method
ridgeisextraction
Fortunately,
wavelet very
theeffective
is alsofor
interference enhancing asthe
information
considered an is
impact
weakened characteristics. Subsequently, wavelet ridge extraction is also considered
improved envelope extraction method and the faint fault information under external interference is for
by using the improved wavelet scalogram analysis. This method isas an
very improved
effective
envelope
extracted extraction method and theThe faint faultdescription
informationindicates
under external
that theinterference is extracted
enhancing the by a series
impact of transforms.
characteristics. above
Subsequently, wavelet ridge extraction proposed method
is also considered has asa an
by a
betterseries of
performancetransforms.
in The
planet above
gear description
bearing weak indicates
fault that
diagnosis. the proposed
Based on method
the above has a better
simulation
improved envelope extraction method and the faint fault information under external interference is
performance
signalbyand in planetsignal
vibration gear bearing
analysis, weak fault
itabove
can diagnosis.
concluded Based
bedescription that onproposed
the the abovemethod
simulation signal and
is effective for
extracted
vibration a series of transforms.
signal analysis, The indicates that the
it can be concluded that the proposed method is effective for rolling proposed method
element has a
rolling element bearing analysis.
betterbearing
performance
analysis. in planet gear bearing weak fault diagnosis. Based on the above simulation
signal and vibration signal analysis, it can be concluded that the proposed method is effective for
rolling element bearing analysis.
Energies 2018, 11, 1286 21 of 23

Energies 2018, 11, x 21 of 23

Energies 2018, 11, x 21 of 23
4.5
4.5
4.5
4.5

Frequency(Hz)
Frequency(kHz)
44

Frequency(Hz)
Frequency(kHz)
44

3.5
3.5
3.5
3.5

330
3300 20
20 40
40 60
60 80
80 100
100 120
120
0 20
20 40 Time(ms)
40 60
60 (ms) 80
Time 80 100
100 120
120
Time(ms)
Time (ms)

Figure 21. Wavelet ridge of bearing signals for improved wavelet scalogram.
Figure 21.
Figure Wavelet ridge
21. Wavelet ridge of
of bearing
bearing signals
signals for
for improved
improved wavelet
wavelet scalogram.
scalogram.
3.5
(a)4.2
3.5
(a)4.2
3.93
3.93
Amplitude(g)
Amplitude(g)
Amplitude(g)

3.6
2.5
Amplitude(g)

3.6
2.5
3.32
3.32
3
1.5
3
1.5
2.71
2.710 20
20 40
40 60
60 80
80 100
100 120
120
0 20
20 40 Time(ms)
40 60
60 (ms) 80
Time 80 100
100 120
120
Time(ms)
Time (ms)
(a)
(b)0.7
0.7 (a)
(b)0.7
0.7 (b) 1
0.60.6 (b) 1 f1=31.53=fob
Amplitude(g)

0.6
0.6 0.8
f1 f3
ff2=63.05=2f
1=31.53=fob
ob

0.5
0.5 ff3=94.58=3f
Amplitude(g)

Amplitude(g)

0.8 f2 f 2=63.05=2fob
ob
0.6 f1 f3 4 ff4=126.1=4f
Amplitude(g)

0.5
0.5 f5 3=94.58=3fob
ob
Amplitude(g)

f2 f4 ff5=189.2=6f
0.4
0.4 4=126.1=4fob
Amplitude(g)

0.6 ob
0.4
f5f6 ff6=220.7=7f
=189.2=6fob
0.4
0.4
0.4 f6
5
f6=220.7=7fob
ob

0.3
0.3 0.2
0.3
0.3 0.2
0.2
0.2 0
0 100 200 300 400 500 600
0.2
0.2 0
0 100
Frequency
200 300
(Hz)
400 500 600
0.1
0.1 Frequency (Hz)
0.1
0.1
0000 1000
1000 2000
2000 3000
3000 4000
4000 5000
5000 6000
6000
0000 1000
1000 2000Frequency(Hz)
2000 3000
Frequency 4000
3000 (Hz)
4000 5000
5000 6000
6000
Frequency(Hz)
Frequency (Hz)
(b)
(b)
Figure 22. Time domain waveforms and spectrum of instantaneous amplitude. (a) Waveform of
Figure 22. Time domain waveforms and spectrum of instantaneous amplitude. (a) Waveform of
Figure 22. Time
instantaneous domain (b)
amplitude; waveforms
Spectrumand spectrum of instantaneous
for instantaneous amplitude. amplitude. (a) Waveform of
instantaneous amplitude; (b) Spectrum for instantaneous amplitude.
instantaneous amplitude; (b) Spectrum for instantaneous amplitude.
5. Conclusions
5.
5. Conclusions
Conclusions
A vibration signal model is established by studying the vibration mechanism of planetary
A vibration signal
A vibration signal model
model is established by
is established studying the
by studying the vibration
vibration mechanism
mechanism of planetary
of planetary
gearboxes, which is the theoretical basis of diagnosing the running state of planetary bearings based
gearboxes,
gearboxes, which isisthe
whichsignal thetheoretical
theoreticalbasis of diagnosing
basis of the running state state
of planetary bearings based
on the vibration analysis method. In diagnosing
addition, bearing the running
fault signal of planetary
weak impactbearings
feature
on the on
based vibration
the is signal signal
vibration analysis method.
analysis In addition, bearing bearing
fault signal
faultweak
signalimpact feature
determination very difficult because of method.
the excessive In addition,
interference. This paper applied weak impact
synchronous
determination is
feature determinationvery difficult because of the excessive interference. This paper applied synchronous
averaging to reduce the is very difficult
disturbed because
components of aofreassigned
the excessive waveletinterference. This paper
scalogram which applied
is obtained by
averaging
synchronous to reduce
averagingthe disturbed components
to reduce the disturbed ofcomponents
a reassignedofwavelet scalogram which is obtained by
multi-scale analysis of a continuous wavelet transformation, thusaimproving
reassigned wavelet
the wavelet scalogram which
time-frequency
multi-scale analysis
is obtained As of
by multi-scale a continuous
analysis wavelet transformation,
of aiscontinuous thus improving the wavelet time-frequency
resolution. well, a wavelet ridge extracted wavelet
to improve transformation,
the visibilitythusandimproving
the accuracythe wavelet
of the
resolution. As well,
time-frequency resolution. a wavelet ridge
As well, is extracted
a wavelet to
ridge improve the
is extracted visibility
to improve and the accuracy of the
the
time-frequency spectrum. The results indicated that the newly proposedthe visibility
method hasandgood
time-frequency
accuracy of the spectrum. Thespectrum.results indicated that the newly
that theproposed method has good
performance in time-frequency
detecting of planet gear The results
bearing indicated
weak fault signals newly
that areproposed
usually method
masked hasby
performance
good in
performance detecting
in of
detecting planet
of gear
planet bearing
gear bearing weakweak fault signals
fault signals that
thatare
are usually
usually masked
masked byby
external disturbances. This method can contribute to weak impact characteristic determination in
external disturbances.
externalfault
disturbances. This method can contribute to weak impact characteristic determination in
bearing analysis. This method can contribute to weak impact characteristic determination in
bearing fault analysis.
bearing fault analysis.
Author Contributions: R.Y. built the simulation model, analyzed the data and wrote the paper; H.L. conceived
Author Contributions:
Contributions: R.Y.
R.Y. built
built the
the simulation
simulation model,
model, analyzed thethe data and
and wrote the
the paper; H.L.
H.L. conceived
conceived
Author
the signal processing method and guided the writing ofanalyzed
paper; C.W. data
and C.H. wrote
designed paper;
the test stand and the
the signal
signal processing
processing method
methodandandguided
guidedthethewriting
writingofofpaper;
paper;C.W. and
C.W. C.H.
and designed
C.H. thethe
designed testtest
stand andand
stand the
experiments.
the experiments.
experiments.
Acknowledgments: This
This research
research was funded
funded by Chinese
Chinese National Science
Science Foundation (No.
(No. 51575075) and
Acknowledgments:
Acknowledgments: This research was
was funded by by Chinese National
National Science Foundation
Foundation (No. 51575075)
51575075) and
and
Chinese
Chinese National
National Science
Science Foundation
Foundation (No.(No. 51175057),
51175057), and
and Collaborative
Collaborative Innovation
Innovation Center
Center of
of Major
Major Machine
Machine
Chinese National Science Foundation (No. 51175057), and Collaborative Innovation Center of Major Machine
Manufacturing in Liaoning for this research is gratefully acknowledged.
Manufacturing
Manufacturing in Liaoning for this research is gratefully acknowledged.
Conflicts of
Conflicts of Interest: The authors
Interest: The authors declare
declare no
no conflict
conflict of
of interest.
interest.
Conflicts of Interest: The authors declare no conflict of interest.
Energies 2018, 11, 1286 22 of 23

References
1. Lihovd, E.; Johannssen, T.I.; Steinebach, C.; Rasmussen, M. Intelligent diagnosis and maintenance
management. J. Intell. Manuf. 1998, 9, 523–537. [CrossRef]
2. Lei, Y.G.; Lin, J.; Zuo, M.J.; He, Z. Condition monitoring and fault diagnosis of planetary gearboxes: A review.
Measurement 2013, 48, 292–305. [CrossRef]
3. Mcfadden, P.D.; Smith, J.D. An Explanation for the Asymmetry of the Modulation Sidebands about the Tooth
Meshing Frequency in Epicyclic Gear Vibration. Proc. Inst. Mech. Eng. Part C 1985, 199, 65–70. [CrossRef]
4. Mosher, M. Understanding Vibration Spectra of Planetary Gear Systems for Fault Detection. In Proceedings
of the ASME 2003 International Design Engineering Technical Conferences and Computers and Information
in Engineering Conference, Chicago, IL, USA, 2–6 September 2003; pp. 645–652.
5. Inalpolat, M.; Kahraman, A. A theoretical and experimental investigation of modulation sidebands of
planetary gear sets. J. Sound Vib. 2009, 323, 677–696. [CrossRef]
6. 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]
7. Feng, Z.; Zuo, M.J. Vibration signal models for fault diagnosis of planetary gearboxes. J. Sound Vib. 2012, 331,
4919–4939. [CrossRef]
8. Liang, X.; Zuo, M.J.; Hoseini, M.R. Vibration signal modeling of a planetary gear set for tooth crack detection.
Eng. Fail. Anal. 2015, 48, 185–200. [CrossRef]
9. Lei, Y.G.; Tang, W.; Kong, D.T. Vibration Signal Simulation and Fault Diagnosis of Planetary Gearboxes
Based on Transmission Mechanism Analysis. J. Mech. Eng. 2014, 50, 61–68. [CrossRef]
10. Huang, Y.H.; Ding, K.; He, G.L. Mathematical Model of Planetary Gear Sets’ Vibration Signal and
Characteristic Frequency Analysis. J. Mech. Eng. 2016, 52, 46–53. [CrossRef]
11. Antoni, J. The spectral kurtosis: A useful tool for characterising non-stationary signals. Mech. Syst.
Signal Process. 2006, 20, 282–307. [CrossRef]
12. Antoni, J. Fast computation of the kurtogram for the detection of transient faults. Mech. Syst. Signal Process.
2007, 21, 108–124. [CrossRef]
13. Lei, Y.G.; Lin, J.; He, Z.J.; Zi, Y.Y. Application of an improved kurtogram method for fault diagnosis of rolling
element bearings. Mech. Syst. Signal Process. 2011, 25, 1738–1749. [CrossRef]
14. Wang, Y.X.; Liang, M. An adaptive SK technique and its application for fault detection of rolling element
bearings. Mech. Syst. Signal Process. 2011, 25, 1750–1764. [CrossRef]
15. Chen, X.; Feng, F.; Zhang, B. Weak Fault Feature Extraction of Rolling Bearings Based on an Improved
Kurtogram. Sensors 2016, 16, 1482. [CrossRef] [PubMed]
16. Antoni, J.; Daniere, J.; Guillet, F. Effective vibration analysis of IC engines using cyclostationarity. Part I—A
methodology for condition monitoring. J. Sound Vib. 2002, 257, 815–837. [CrossRef]
17. Antoni, J.; Daniere, J.; Guillet, F.; Randall, R.B. Effective vibration analysis of IC engines using
cyclostationarity. Part II—New results on the reconstruction of the cylinder pressures. J. Sound Vib. 2002,
257, 839–856. [CrossRef]
18. Do, V.T.; Le, C.N. Adaptive Empirical Mode Decomposition for Bearing Fault Detectionp. J. Mech. Eng. 2016,
62, 281–290. [CrossRef]
19. McFadden, P.D.; Toozhy, M.M. Application of synchronous averaging to vibration monitoring of rolling
element bearings. Mech. Syst. Signal Process. 2000, 14, 891–906. [CrossRef]
20. Zhu, L.M.; Ding, H.; Zhu, X.Y. Synchronous averaging of time-frequency distribution with application to
machine condition monitoring. J. Vib. Acoust. 2007, 129, 441–447. [CrossRef]
21. Zhang, H.; Chen, X.; Du, Z.; Yan, R. Kurtosis based weighted sparse model with convex optimization
technique for bearing fault diagnosis. Mech. Syst. Signal Process. 2016, 80, 349–376. [CrossRef]
22. Ding, X.; He, Q. Time-frequency manifold sparse reconstruction: A novel method for bearing fault feature
extraction. Mech. Syst. Signal Process. 2016, 80, 392–413. [CrossRef]
23. Qian, Y.; Yan, R.; Hu, S. Bearing Degradation Evaluation Using Recurrence Quantification Analysis and
Kalman Filter. IEEE Trans. Instrum. Meas. 2014, 63, 2599–2610. [CrossRef]
24. Liu, J.; Wang, W.; Golnaraghi, F. An Extended Wavelet Spectrum for Bearing Fault Diagnostics. IEEE Trans.
Instrum. Meas. 2008, 57, 2801–2812.
Energies 2018, 11, 1286 23 of 23

25. Qin, Y.; Mao, Y.F.; Tang, B.P. Vibration signal component separation by iteratively using basis pursuit and its
application in mechanical fault detection. J. Sound Vib. 2013, 332, 5217–5235. [CrossRef]
26. Huang, N.E.; Shen, Z.; Long, S.R. The empirical mode decomposition and the Hilbert spectrum for nonlinear
and non-stationary time series analysis. Proc. R. Soc. Lond. A 1998, 254, 903–995. [CrossRef]
27. Soualhi, A.; Medjaher, K.; Zerhouni, N. Bearing Health Monitoring Based on Hilbert-Huang Transform,
Support Vector Machine, and Regression. IEEE Trans. Instrum. Meas. 2015, 64, 52–62. [CrossRef]
28. He, Q.B.; Wang, X.X. Time-frequency manifold correlation matching for periodic fault identification in
rotating machines. J. Sound Vib. 2013, 332, 2611–2626. [CrossRef]
29. Halim, E.B.; Choudhury, M.A.A.S.; Shah, S.L.; Zuo, M.J. Time domain averaging across all scales: A novel
method for detection of gearbox faults. Mech. Syst. Signal Process. 2008, 22, 261–278. [CrossRef]
30. Peng, Z.K.; Chu, F.L.; Tse, P.W. Detection of the rubbing-caused impacts for rotor-stator fault diagnosis using
reassigned scalogram. Mech. Syst. Signal Process. 2005, 19, 391–409. [CrossRef]
31. Sawalhi, N.; Randall, R.B. Vibration response of spalled rolling element bearings: Observations, simulations
and signal processing techniques to track the spall size. Mech. Syst. Signal Process. 2011, 25, 846–870.
[CrossRef]
32. Peng, Z.K.; Chu, F.L. Application of the wavelet transform in machine condition monitoring and fault
diagnostics: A review with bibliography. Mech. Syst. Signal Process. 2004, 18, 199–221. [CrossRef]
33. Yan, R.; Gao, R.X. Impact of wavelet basis on vibration analysis for rolling bearing defect diagnosis.
In Proceedings of the 2011 IEEE International Instrumentation and Measurement Technology Conference,
Beijing, China, 10–12 May 2011; pp. 1–4.
34. Siegel, D.; Al-Atat, H.; Shauche, V.; Liao, L.X.; Snyder, J.; Lee, J. Novel method for rolling element bearing
health assessment—A tachometer-less synchronously averaged envelope feature extraction technique.
Mech. Syst. Signal Process. 2012, 29, 362–376. [CrossRef]
35. Qin, Y.; Tang, B.; Mao, Y. Adaptive signal decomposition based on wavelet ridge and its application.
Signal Process. 2015, 120, 480–494. [CrossRef]
36. Mallat, S. A Wavelet Tour of Signal Processing; China Machine Press: Beijing, China, 2003; pp. 67–81.
37. Liu, H.; Cartwright, A.N.; Basaran, C. Moire interferogram phase extraction: A ridge detection algorithm for
continuous wavelet transforms. Appl. Opt. 2004, 43, 850–857. [CrossRef] [PubMed]

© 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/).