Applied Sciences

applied
sciences
Article
Diagnostics 101: A Tutorial for Fault Diagnostics of
Rolling Element Bearing Using Envelope Analysis
in MATLAB
Seokgoo Kim 1,2 , Dawn An 3, * and Joo-Ho Choi 4, *
1 Department of Mechanical and Aerospace Engineering, University of Florida, Gainesville, FL 32611, USA;
seokgoo.kim@ufl.edu
2 Department of Aerospace and Mechanical Engineering, Korea Aerospace University, Goyang,
Gyeonggi-do 10540, Korea
3 Daegyeong Division/Safety System R&D Group, Korea Institute of Industrial Technology, Yeongcheon-si,
Gyeongbuk-do 38822, Korea
4 School of Aerospace and Mechanical Engineering Korea Aerospace University, Goyang,
Gyeonggi-do 10540, Korea
* Correspondence: dawnan@kitech.re.kr (D.A.); jhchoi@kau.ac.kr (J.-H.C.)

Received: 7 September 2020; Accepted: 13 October 2020; Published: 19 October 2020
Abstract: This paper presents a MATLAB-based tutorial to conduct fault diagnosis of a rolling
element bearing. While there have been so many new developments in this field, no studies have
addressed the tutorial aspects in this field to help the engineers learn the concept and implement by
their own effort. The three most common techniques—the autoregressive model, spectral kurtosis,
and envelope analysis—are selected to demonstrate the bearing diagnosis process. Simulation signal
is introduced to help understand the characteristics of fault signal and carry out the process toward
the fault identification. The techniques are then applied to the two real datasets to demonstrate the
practical applications, one made by the authors and the other by the Case Western Reserve University,
which is known as a standard reference in testing the diagnostic algorithms.
Keywords: fault diagnosis; bearings; MATLAB code; envelope analysis; spectral kurtosis;
autoregressive model; diagnosis tutorial
1. Introduction
Prognostics and Health Management (PHM), which enables life extension while preventing
unexpected failures in various engineering systems, have received great attention over the last decades
for more efficient operation and maintenance. Generally, there are three main steps in the PHM
development: (1) signal processing, (2) diagnosis, and (3) prognosis. Signal processing is the step
in which the valuable information called features are extracted from noisy sensor signals such as
vibrations or acoustic emissions. Diagnosis focuses on detecting the faults, isolating their locations,
and identifying their severity based on the extracted features. Lastly, the prognosis is to predict the
Remaining Useful Life (RUL) based on the accumulated diagnostic information, which indicates how
long it will take until the system’s health reaches the failure threshold. Though the most important
step is the prognosis, its performance is of no doubt dependent on the quality of prior steps: the signal
processing and the diagnosis.
To date, the PHM has mainly focused on the vibration signal-based techniques for rotating
mechanical components such as gears [1,2] and bearings [3–7], since they usually lead to the critical
failures [1,8]. Among them, the rolling element bearings, which support the shaft during the rotation,
have been of greatest importance in view of its widespread use and often cause of another failures.
Appl. Sci. 2020, 10, 7302; doi:10.3390/app10207302 www.mdpi.com/journal/applsci

Appl. Sci. 2020, 10, 7302 2 of 23
For example, the bearings are known to cause around 70% of gearboxes and 21–70% of generator failures
in the wind turbine industries [9]. To prevent such failures, early fault detection and diagnosis of the
bearings is important. A great deal of research has been conducted on this topic. Many review papers
are available, addressing the state-of-the-art algorithms from various perspectives. Only a few are given
here; among others, Rai et al. reviewed several signal processing techniques for the rolling element
bearing fault diagnosis [10]. Cerrada et al. have reviewed fault severity assessment methods [11].
Hamadache et al. made review on artificial intelligence-based approaches including shallow and deep
learning [12]. Reviews are available, too, on the bearings for specific applications; Caesarendra et al.
presented feature extraction methods for the low-speed slew bearing [13]. de Azevedo et al. have
reviewed condition monitoring methods for bearings in the wind turbine [9].
As such, a great number of new algorithms have been introduced and reviewed over the last
decades. Despite the importance, however, no studies have dealt with the tutorial aspects except the
one by Randall and Antoni, which have introduced a number of techniques covering the essential
as well as advanced algorithms in the context of diagnostic procedure [14]. The paper was highly
informative and enlightening in several aspects, but it was in some sense close to a collection of their
recent achievements in a single article, and hence was less instructive in view of the self-explanatory
aspects of the tutorial. In real practice, there are many instances that engineers without enough
knowledge in this field fail to solve their problems due primarily to the lack of detailed understanding
of the algorithms and of how to implement them. Motivated by this, a tutorial is presented in this
paper to conduct the fault diagnosis of rolling element bearing, with the purpose to help the beginners
learn by themselves the basic concepts along with implementing the algorithms with minimum resort
to the external resources.
Among the techniques currently available for the bearing diagnosis, envelope analysis is the most
common and relatively well established, which enables to detect the fault with location by matching the
frequency spectrum of the envelope to the fault frequencies. By the way, the bearing signal is usually
masked by the resonant noise and those coming from other components (e.g., gears and engines),
which should be filtered out as a first step. Spectral kurtosis and autoregressive model are the most
commonly applied techniques for this purpose [14]. These are addressed in the context of diagnosis
process using the MATLAB scripts.
Simulation datasets are prepared to explain the characteristics of the raw signal and the process of
how to retrieve the fault information out of it. The techniques are then applied to the real datasets
made by the authors to demonstrate the practical applications. The datasets by the Case Western
Reserve University, which have been a standard reference in testing the diagnostic algorithms, are also
used to demonstrate the general nature of the techniques.
The paper is organized as follows. In Section 2, the basic theory and procedures in the bearing
diagnostics are introduced. Section 3 addresses how to implement the bearing diagnosis using the
simulation data. Practical diagnoses using the MATLAB codes are also performed with the real datasets
in Sections 4 and 5. Lastly, conclusions are presented in Section 6.
2. Bearing Diagnostics
In this section, fundamental characteristics of fault signal is addressed first in view of amplitude
modulation. Then, a virtual signal is generated to illustrate what the real signal is like in terms of its
components, which will be used in implementing the diagnosis procedure in the next section.
2.1. Fault Signal Given by Amplitude Modulation

Whenever the rolling element in the bearing strikes a local fault in the inner or outer race or
vice versa during operation, an impact signal is produced at the period of one of the bearing fault
Appl. Sci. 2020, 10, 7302 3 of 23
frequencies, such as Ball Pass Frequency Outer race (BPFO), Ball Pass Frequency Inner race (BPFI),
Fundamental Train
Appl. Sci. 2020, 10, x FORFrequency (FTF), and Ball Spin Frequency (BSF), which are defined as follows3 [15].
PEER REVIEW of 23
( )
n fs d
BPFO = 𝑛𝑓 1 + 𝑑 cosα (1)
𝐵𝑃𝐹𝑂 = 2 1 + D 𝑐𝑜𝑠𝛼 (1)
2 𝐷
( )
n fs d
BPFI = 𝑛𝑓 1 − 𝑑 cosα (2)
𝐵𝑃𝐹𝐼 = 2 1 − D 𝑐𝑜𝑠𝛼 (2)
2 𝐷
f𝑓s d𝑑
( )
𝐹𝑇𝐹 ==
FTF 11 −
− cosα
𝑐𝑜𝑠𝛼 (3)
(3)
22 𝐷
D
 !2 
D f
BSF = 𝐷𝑓  𝑑cosα 
s

 d 
1 − (4)

𝐵𝑆𝐹 = d 1 − D 𝑐𝑜𝑠𝛼 (4)
𝑑 𝐷
where D, d, α, n and fs are the pitch diameter, ball diameter, contact angle between the ball and
where 𝐷, 𝑑, 𝛼, 𝑛 and 𝑓 are the pitch diameter, ball diameter, contact angle between the ball and the
the cage, number of rolling elements, and the rotating speed of bearing (Hz), respectively. Usually,
cage, number of rolling elements, and the rotating speed of bearing (Hz), respectively. Usually, the
the impact signals at these fault frequencies are modulated by the much higher resonant frequencies of
impact signals at these fault frequencies are modulated by the much higher resonant frequencies of
the bearings, which results in the signal like the one in Figure 1, the exponentially decaying ringing
the bearings, which results in the signal like the one in Figure 1, the exponentially decaying ringing
occurring periodically at the fault frequencies [16].
occurring periodically at the fault frequencies [16].
Figure 1.
Figure 1. Bearing
Bearing geometry
geometry and
and impact
impact signal.
signal.
This phenomenon
This phenomenon is is mathematically
mathematically the the samesame asas the
the amplitude
amplitude modulation
modulation (AM), (AM), ofof which
which the
the
concept is
concept is illustrated
illustrated byby aa simple
simple example
example in in Figure
Figure 2.2. The fault and the resonance resonance signal
signal of
of the
the
bearings are assumed
bearings assumed to to be
be Figure
Figure 2a,b
2a,b which
which are are the
the modulating
modulating and and carrier
carrier signal
signal in
in the
the signal
signal
processingtheory,
processing theory, respectively.
respectively. Then,
Then, the AMthe signal
AM signal
becomes becomes
the carrierthesignal
carrierwithsignal with
variable variable
amplitude
amplitude
given by thegiven by the modulating
modulating signalin
signal as shown asFigure
shown2c. in Figure 2c. After
After taking thetaking
FFT, the thefrequency
FFT, the frequency
domains
domains
of of modulating,
modulating, carrier, andcarrier, andsignals
the AM the AMare signals
givenarein given
Figurein2e–g,
Figure 2e–g, respectively.
respectively. Theshows
The result result
shows
that thethat
AMthe AM spectrum
spectrum has peakshasatpeaks at the resonance
the resonance (carrier)(carrier)
frequency frequency
( fr ) with(𝑓the) with the sidebands
sidebands spaced
spaced
apart byapart by the
the fault fault (modulating)
(modulating) frequencyfrequency
( f f ) [17]. (𝑓
The) [17]. The AM
AM signal cansignal can be expressed
be expressed by the
by the following
following[18]:
equation equation [18]: n o
x(t) = [1 + f (t)] × r(t) = 1 + Msin 2π f t Acos(2π fr t) (5)
𝑥(𝑡) = 1 + 𝑓(𝑡) × 𝑟(𝑡) = 1 + 𝑀𝑠𝑖𝑛 2𝜋𝑓f 𝑡 𝐴𝑐𝑜𝑠(2𝜋𝑓 𝑡) (5)
where f (t) is the fault signal (plus DC offset to ensure that the mean amplitude of AM signals is zero)
where f(t) is the fault signal (plus DC offset to ensure that the mean amplitude of AM signals is zero)
with the amplitude M and frequency f f , r(t) is the resonance signal with the amplitude A and frequency
with the amplitude M and frequency 𝑓 , r(t) is the resonance signal with the amplitude 𝐴 and
fr , and x(t) is the AM signal. Equation (5) can be transformed to
frequency 𝑓 , and 𝑥(𝑡) is the AM signal. Equation (5) can be transformed to
𝐴A·M
·𝑀 h i
x(t)==𝐴𝑐𝑜𝑠(2𝜋𝑓
𝑥(𝑡) Acos(2π f𝑡)
r t)++ 𝑠𝑖𝑛 2𝜋 𝑓
sin2π +𝑓
fr + f f t𝑡−−sin2π
𝑠𝑖𝑛 2𝜋fr 𝑓 −−f f 𝑓t 𝑡 (6)
(6)
22
which shows that the spectrum of the AM signals consists of three components 𝑓 , 𝑓 ± 𝑓 , with the
amplitudes 𝐴 and 𝐴·𝑀/2, respectively, as shown in Figure 2f. Therefore, it can be concluded that,
when a bearing has fault, it appears as the sideband around the resonance signal. Then, the remaining
task is to extract or separate the original fault signal Figure 2a or its FFT (e) from the AM signal Figure
2c or its FFT (g). This is realized by the envelope analysis, which is an amplitude demodulation
technique as will be addressed in the subsequent section.
Appl. Sci. 2020, 10, 7302 4 of 23
which shows that the spectrum of the AM signals consists of three components fr , fr ± f f , with the
amplitudes A and A·M/2, respectively, as shown in Figure 2f. Therefore, it can be concluded that,
when a bearing has fault, it appears as the sideband around the resonance signal. Then, the remaining
task is to extract or separate the original fault signal Figure 2a or its FFT (e) from the AM signal
Figure 2c or its FFT (g). This is realized by the envelope analysis, which is an amplitude demodulation
technique as will be addressed in the subsequent section.
Appl. Sci. 2020, 10, x FOR PEER REVIEW 4 of 23
(a) (e)
(b) (f)
(c) (g)
(d) (h)
Figure 2. Amplitude modulation and envelope signal: (a) modulating signal, (b) carrier signal, (c) AM
Figure 2. Amplitude modulation and envelope signal: (a) modulating signal, (b) carrier signal, (c) AM
signal, (d) envelope signal, in the time domain; (e) modulating signal, (f) carrier signal, (g) AM signal,
signal, (d) envelope signal, in the time domain; (e) modulating signal, (f) carrier signal, (g) AM signal,
(h) envelope signal, in the frequency domain.
(h) envelope signal, in the frequency domain.
2.2. Simulation of Fault Signal
With this understanding, more complex signal closer to the reality is considered. Toward this
end, a virtual signal is generated using the MATLAB code given in Appendix A. A fault signal is
Appl. Sci. 2020, 10, 7302 5 of 23
2.2. Simulation of Fault Signal

With this understanding, more complex signal closer to the reality is considered. Toward this end,
a virtual signal is generated using the MATLAB code given in Appendix A. A fault signal is created as
0 ) to represent the transient impact
shown in Figure
created 3a byin
as shown anFigure
exponential
3a by an function, x1 (t)function,
exponential = A · exp (𝑡) =
𝑥(−αt 𝐴 ∙ 𝑒𝑥𝑝(−𝛼𝑡 ) to represent the
signaltransient
with theimpact
frequency signal
f f with
= 10 the where t =𝑓mod
Hz, frequency 0 = 10𝐻𝑧
t, 1/, fwhere
f and A 𝑡 and
= 𝑚𝑜𝑑(𝑡, 1/𝑓 ) and
α represent the A and α of
magnitude
impulse represent
signalthe andmagnitude
the damping of impulse signal of
parameter andthethe damping
bearing, parameter ofThey
respectively. the bearing,
are givenrespectively.
by 5 and 75.4
in thisThey are given
example. Thebymod 5 and 75.4 infor
stands this example.
the modulus Theafter
moddivision,
stands forwhich
the modulus afterrepeated
is to make division, signal
which over
is to make repeated signal over time [19]. The FFT of this signal becomes Figure 3d, where the
time [19]. The FFT of this signal becomes Figure 3d, where the fundamental frequency is 10 Hz with
fundamental frequency is 10 Hz with amplitude 1.0 with the subsequent ones being its harmonics.
amplitude 1.0 with the subsequent ones being its harmonics. This is modulated by a carrier signal given
This is modulated by a carrier signal given by the sinusoidal function 𝑥 (𝑡) = 𝑠𝑖𝑛(2𝜋𝑓 𝑡) with
by the sinusoidal function x2 (t) =
frequency 𝑓 = 600𝐻𝑧 which
sin(2π fr t) with
is much higher
frequency f = 600 Hz which is much higher than
than f , as shownr in Figure 3b–c. This corresponds to
f f , asthe
shown in Figure 3b–c. This corresponds
resonance of the bearing. Then, the modulated signal to the resonance
becomes of 𝑥 the
(𝑡) =bearing.
𝑥 (𝑡) ∗ 𝑥Then,
(𝑡), andthe
is modulated
shown
Figure 3cx3in(twhich
signalinbecomes ) = x1the
(t) ∗carrier
x2 (t),contains
and is shown
the faultin in
Figure 3c in which
its amplitude [20].the
Its carrier contains
FFT becomes the 3f,
Figure fault in
its amplitude
where the[20]. Its FFT
sidebands arebecomes Figure600
present around 3f, Hz.
where the sidebands are present around 600 Hz.
(a) (d)
(b) (e)
(c) (f)
Figure 3. Modulation of simulation fault signal: (a) fault signal, (b) resonance signal, (c) AM signal,
Figure 3. Modulation of simulation fault signal: (a) fault signal, (b) resonance signal, (c) AM signal,
in the time domain; (d) fault signal, (e) resonance signal, (f) AM signal, in the frequency domain.
in the time domain; (d) fault signal, (e) resonance signal, (f) AM signal, in the frequency domain.
Added to this, the signal may include a certain discrete signals from other components such as
Added to this, the signal may include a certain discrete signals from other components such
gears and shafts as well as the background noise. In this example, they are represented by another
as gears and shafts as well as the background noise. In this example, they are represented by
low-frequency sinusoidal function 𝑥 (𝑡) = 𝑠𝑖𝑛(2𝜋𝑓 𝑡) with amplitude 2 and frequency 𝑓 =
another
13 Hz and Gaussiansinusoidal
low-frequency white noisefunction x4 (𝜎t)) =
𝑥 (𝑡)~𝑁(0, withsin (2π fd t) 𝜎with
variance amplitude
= 0.5 2 and
, which are frequency
given by
Appl. Sci. 2020, 10, 7302 6 of 23

fd = 13 Hz and Gaussian white noise x5 (t) ∼ N 0, σ2 with variance σ2 = 0.52 , which are given
by Figure 4a,b, and their FFTs (d) and (e), respectively. Putting all these together, the resulting raw
(t) =4a,b,
signal xFigure x3 (tand
) +their
x4 (tFFTs
) + x(d)
5 (tand
) becomes Figure Putting
(e), respectively. 4c withallits FFTtogether,
these being (f). Note that
the resulting rawthe governing
signal
sinusoidal = 𝑥 (𝑡)
𝑥(𝑡) wave + 𝑥 (𝑡)in
shown 𝑥 (𝑡) becomes
+ Figure 4c is the discrete
Figure signal
4c with x4 . The
its FFT beingamplitudes
(f). Note that atthe
thegoverning
high-frequency
sinusoidal
region in Figure wave shown inthe
4f represent Figure 4c is the
carrier discrete
at 600 Hz signal 𝑥 . The and
(resonance) amplitudes at the high-frequency
surrounding sidebands (faults),
region in Figure 4f represent the carrier at 600 Hz (resonance) and surrounding sidebands (faults),
which is a more complex version of the AM mentioned in Figure 2. The frequency amplitudes of
which is a more complex version of the AM mentioned in Figure 2. The frequency amplitudes of the
the discrete and AM signal are 2 and 0.6, respectively, which means that the fault is buried by the
discrete and AM signal are 2 and 0.6, respectively, which means that the fault is buried by the carrier
carrier and anotherdiscrete
and another discrete signals.
signals. This This
agreesagrees with
with the the that
finding finding that the
the signals fromsignals from
the gears andthe gears and
shafts
shafts often
oftenshow
show higher amplitudes
higher amplitudes than
than those
those of faulty
of faulty bearing;
bearing; typically,
typically, the amplitudes
the amplitudes of the shaft
of the shaft
rotation,rotation, gear meshing
gear meshing and and
the the bearing
bearing faultsignals
fault signals are
areononthe
theorder
orderof 0.1, 10 and
of 0.1, 10 0.001 g [21,22].
and 0.001 g [21,22].
(a) (d)
(b) (e)
Raw signal
10
-5
-10
0 0.5 1 1.5 2
Time[sec]
(c) (f)
Figure 4. Simulation of bearing fault signal: (a) discrete signal, (b) white noise, (c) raw signal, in the
Figure 4. Simulation of bearing fault signal: (a) discrete signal, (b) white noise, (c) raw signal, in the
time domain; (d) discrete signal, (e) white noise, (f) raw signal, in the frequency domain.
time domain; (d) discrete signal, (e) white noise, (f) raw signal, in the frequency domain.
3. Diagnosis Using Simulation Data
3. Diagnosis Using Simulation Data
The goal of the diagnosis is to isolate the fault information from the raw signal mixed with other
Thenoisy
goalsources.
of the The
diagnosis
steps areistaken
to isolate the fault
in reverse order information
of the way thefrom
virtualthe
rawraw signal
signal mixed
is made and with
are other
noisy sources.
summarizedThe in
steps are5.taken
Figure in reverse
It consists of threeorder
steps: of theisway
First the virtual
the discrete signalraw signal to
separation is remove
made and are
summarized
signals in Figure
from 5. Itcomponents.
the other consists ofSecond
three is
steps: First is the band
the demodulation discrete signal
selection separation
to find to remove
out the band
signals from the other components. Second is the demodulation band selection to find out the band
containing resonance frequencies, using which the band pass filtering is carried out to obtain only
the signal in the band while removing those elsewhere. Last is the envelope analysis to restore the
original fault signal out of the modulated signal around the resonance frequencies. Usually, the first
Appl. Sci. 2020, 10,

containing 7302
resonance 7 of 23
frequencies, using which the band pass filtering is carried out to obtain only
the signal in the band while removing those elsewhere. Last is the envelope analysis to restore the
original fault signal out of the modulated signal around the resonance frequencies. Usually, the first
two steps are denoted as the signal pre-processing. Note that the raw signal includes the mixture of
two steps are denoted as the signal pre-processing. Note that the raw signal includes the mixture of
four signals: bearing fault signal, resonance signal, and discrete signal with frequencies f f , fr and fd
four signals: bearing fault signal, resonance signal, and discrete signal with frequencies 𝑓 , 𝑓 and 𝑓
and the background noise.
and the background noise.
Figure5.5.Procedure
Figure Procedure for
for bearing
bearing fault
faultdiagnosis.
diagnosis.
Regardingthe
Regarding thealgorithms
algorithmsin ineach
each step,
step, autoregressive
autoregressive (AR)
(AR)models
modelsare
areapplied
applied[14,23] forfor
[14,23] thethe
discrete signal separation at first. Spectral kurtosis (SK) and Short Time Fourier Transform
discrete signal separation at first. Spectral kurtosis (SK) and Short Time Fourier Transform (STFT) (STFT) are
areemployed
employedinin the second
the secondstep to to
step select frequency
select frequencyband where
band the bearing
where resonance
the bearing is dominant.
resonance In
is dominant.
the third step, Hilbert transform is applied for the envelope analysis to obtain the analytic
In the third step, Hilbert transform is applied for the envelope analysis to obtain the analytic signal. signal.
Then, its magnitude becomes the envelope signal. By taking the FFT of the envelope signal, the fault
Then, its magnitude becomes the envelope signal. By taking the FFT of the envelope signal, the fault
information is retrieved from the amplitudes at the fault frequencies. More details are explained in
information is retrieved from the amplitudes at the fault frequencies. More details are explained in
the following sub-sections. MATLAB code with 63-lines is provided in the Appendix B to implement
the following sub-sections. MATLAB code with 63-lines is provided in the Appendix B to implement
the process, which consists of four parts: (1) problem definition, (2) discrete signal separation, (3)
the process, which consists of four parts: (1) problem definition, (2) discrete signal separation,
demodulation band selection, and (4) envelope analysis.
(3) demodulation band selection, and (4) envelope analysis.
3.1. Problem Definition (Lines 1–7)
3.1. Problem Definition (Lines 1–7)
In the code, six parameters are required in the problem definition. The first four are related to
In the code, six parameters are required in the problem definition. The first four are related to
the data acquisition and bearing dimensions, and thus should be fixed for the same applications. Raw
thedatadataareacquisition
stored in the and bearing
variable dimensions,
rawData. and thus
Simulation datashould
generated be byfixed for the Asame
Appendix applications.
are loaded into
Raw data are stored in the variable rawData. Simulation data generated
this variable. The parameters sampRate, rpm and bearFreq denote the sampling rate in Hz, by Appendix A are shaft
loaded
into this variable.
rotation speed in The
rpm parameters
and bearing sampRate, rpm Note
fault frequency. and bearFreq
that there aredenote
four thefaultsampling
frequenciesrate
inin
theHz,
shaftbearing, which are calculated by the Equations (1)–(4) based on the shaft frequency in rpm and in
rotation speed in rpm and bearing fault frequency. Note that there are four fault frequencies
thebearing
bearing, which are calculated
specifications. by the however,
In the simulation, Equationsa single
(1)–(4)fault
based on the shaft
frequency frequency
is directly in rpm
assigned, whichand
is 𝑓 specifications.
bearing = 10Hz as shown In the simulation,
in Figure however,the
3a. Therefore, a single
rpm isfault frequency
just set is directly
at 60 here. The last assigned, which is
two parameters
= 10and
f f maxP Hz windLeng
as shown in areFigure 3a. the
to control Therefore, the rpm
performance is just
of the set at 60
algorithm: here.is The
maxP last two parameters
the maximum order of
maxP theand
AR windLeng
model, which areistogiven
controlby the
300 performance
from the ratioofofthethealgorithm:
sampRatemaxP is the
= 3000 maximum
and bearFreq order of the
= 10.
ARThis is based
model, on the
which criteria
is given bythat
300the model
from the order should
ratio of not be greater
the sampRate = 3000thanand
the number
bearFreq of data within
= 10. This is
baseda fault
on interval [23].that
the criteria Thethe
parameter shouldisnot
windLeng
model order a row vector ofthan
be greater the the
lengths
numberin theofwindow function
data within a fault
of STFT. In this example, four different window lengths [2^4 2^5 2^6 2^7]
interval [23]. The parameter windLeng is a row vector of the lengths in the window function of STFT. are used. More details
In about how to set
this example, optimum
four differentwindow
window length can[2ˆ4
lengths be referred
2ˆ5 2ˆ6to in2ˆ7]
Ref. are
[24].used. More details about how
to set optimum window length can be referred to in Ref. [24].
3.2. Discrete Signal Separation (Lines 8–22)
3.2. Discrete Signal
The first Separation
step (Linessignal
is the discrete 8–22) separation that removes the signals from the other sources
such
Theasfirst
gears andisshafts,
step which we
the discrete call separation
signal the discrete that
(deterministic)
removes thesignal. The from
signals discrete
thesignal
othercan be
sources
obtained by the autoregressive (AR) model, which is to estimate the current value of time series
such as gears and shafts, which we call the discrete (deterministic) signal. The discrete signal can be data
obtained by the autoregressive (AR) model, which is to estimate the current value of time series data
via weighted sum of the past values. Then, the residual signal is obtained by subtracting the AR model,
i.e., the discrete part, from the raw signal, as shown in Figure 6.
via Sci.
Appl. weighted sum
of the past values. Then, the residual signal is obtained by subtracting the
2020, 10, 7302 8 of AR
23
model, i.e., the discrete part, from the raw signal, as shown in Figure 6.
Figure6.6.Flow
Figure Flowchart
charton
onhow
howthe
theresidual
residualsignals
signalsare
areobtained.
obtained.
Let
Letususconsider
considera atime seriesx(n)
timeseries withnn==1,1,N,
x(n)with N,which
whichisisthe
theraw
rawvibration
vibrationsignal
signalacquired
acquiredwith
with
sampRate
sampRate perper
seconds.
seconds.TheThe
ARARmodel
modelis then constructed
is then to predict
constructed the the
to predict nextnext
value in a in
value time series
a time by
series
using a series of previous observations, which is described by the following equation [22] (line
by using a series of previous observations, which is described by the following equation [22] (line 15): 15):
𝑥 (𝑛) = − ∑X p 𝑎(𝑘)𝑥(𝑛 − 𝑘) or in a matrix form 𝐱 = −𝐗𝐚 (7)

xp (n) = − a(k)x(n − k) or in a matrix form xp = −Xa (7)
where 𝑥 (𝑛) is the predicted value k = 1 at nth time step as a weighted sum of its p previous values in the
sequence, p is the order of the model, a(k) is the weighting coefficients, and x(n-k) is n-kth data. The
where (n)isisanthe𝑁 predicted
matrixxp𝐗 × 𝑝 matrixvalue at nth of
consisting timethestep
pastasseries
a weighted values of sum𝑥 of p previous
its 13-14).
(line The values in
coefficient
the sequence, p is the order of the model, a(k) is the
𝑎(𝑘) can be obtained by solving the Yule–Walker equation [14,22] (line 12): weighting coefficients, and x(n-k) is n-kth data.
The matrix X is an N × p matrix consisting of the past series values of x (line 13–14). The coefficient a(k)
𝑟 (0) 𝑟 (−1) ⋯ 𝑟 (−𝑝 + 1) 𝑎(1) 𝑟 (1)
can be obtained by solving the Yule–Walker equation [14,22] (line 12):
𝑟 (1) 𝑟 (0) ⋯ 𝑟 (−𝑝 + 2) 𝑎(2) 𝑟 (2)
= − (8)
⋮ ⋮ ⋱ ⋮  ⋮ ⋮ 
 r𝑟xx ((𝑝 0) rxx (−1) · · · rxx (−p + 1)  a(1)  ( )
 
r
 𝑟 (𝑝)  1
− 1) 𝑟 (𝑝 − 2) ⋯ 𝑟 (0)  𝑎(𝑝)  xx 
 rxx (1) rxx (0) · · · rxx (−p + 2)  a(2)   r (2) 

 xx 
where 𝑟 (𝑘) is the autocorrelation .. function
. . of x(n), .. which ..  = by
 is defined − (8)

..    .. 
 . . . .  . 
  
 . 

   
rxx (p − 1) rxx (p − 2)1 · · · rxx (0) a(p) rxx (p)

𝑟 (𝑘) = 𝑥(𝑛)𝑥(𝑛 − 𝑘) , 𝑘 = 0, … , 𝑝 − 1 (9)
𝑁
where rxx (k) is the autocorrelation function of x(n), which is defined by
In order to illustrate the process, an example is given when p = 2. Then, Equation (8) becomes
N
𝑟 (0) 1 X 𝑟 x(−1)
rxx (k) = (n)x(n𝑎(1)− k), k= =− 0,𝑟 . .(1) .,p−1 (9)
(10)
𝑟 (1) N
n=1 𝑟 (0) 𝑎(2) 𝑟 (2)
where
In order to illustrate the process, an example is given when p = 2. Then, Equation (8) becomes
𝑟 (0)
" = 𝑥(1)𝑥(1) +#"𝑥(2)𝑥(2) # + ⋯ "+ 𝑥(𝑁)𝑥(𝑁)# /𝑁
rxx (0) rxx (−1) a(1) rxx (1)
𝑟 (1) = 𝑥(1)𝑥(0) = − … + 𝑥(𝑁)𝑥(𝑁 − 1) /𝑁 (10)
rxx (1) +rxx𝑥(2)𝑥(1)
(0) a+
(2𝑥(3)𝑥(2)
) rxx (2)
𝑟 (−1) = 𝑥(1)𝑥(2) + 𝑥(2)𝑥(3) + ⋯ + 𝑥(𝑁 − 1)𝑥(𝑁) + 𝑥(𝑁)𝑥(𝑁 + 1) /𝑁
where
𝑟 (2)rxx
=(0)𝑥(1)𝑥(−1)
= [x(1)x+
(1)𝑥(2)𝑥(0)
+ x(2)x(+2)𝑥(3)𝑥(1)
+ · · · + x…
(N+)x𝑥(𝑁)𝑥(𝑁
(N )]/N − 2) /𝑁
In the equations,
rxx (1)the [x(1)x𝑥(0)exceeding
= items + x(2)x(1the) + range
x(3)x([1,
2) .N]
..+dox(not
N )xexist,
(N − 1hence,
)]/N are assigned zero.
Then, the AR model at n step becomes by Equation (7)
th
rxx (−1) = [x(1)x(2) + x(2)x(3) + · · · + x(N − 1)x(N ) + x(N )x(N + 1)]/N

𝑥 (𝑛) = −𝑎(1)𝑥(𝑛 − 1) − 𝑎(2)𝑥(𝑛 − 2) (11)
rxx (2) = [x(1)x(−1) + x(2)x(0) + x(3)x(1) . . . + x(N )x(N − 2)]/N
Once the deterministic signal 𝑥 is obtained from the AR model, the residual signal is obtained
In the equations,
by subtracting the items
the prediction x exceeding
from 16).range [1, N] do not exist, hence, are assigned zero.
this (line the
Then, Since
the ARthe
model
order at pnth step becomes
affects by Equation
the prediction (7)
performance, it should be determined to make the
residual signal represent the fault as best as it can. Depending on its uses, there are several different
xp (n) = −a(1)x(n − 1) − a(2)x(n − 2) (11)
Appl. Sci. 2020, 10, 7302 9 of 23
Once the deterministic signal xp is obtained from the AR model, the residual signal is obtained by
subtracting the prediction from this (line 16).
Since the order p affects the prediction performance, it should be determined to make the residual
signal represent the fault as best as it can. Depending on its uses, there are several different criteria
to determine
Appl. Sci. 2020, x FOR p,
the10,order such
PEER as Akaike information criteria (AIC) [25–28], Bayesian information
REVIEW 9 of 23
criterion (BIC) [29], Minimum description length (MDL) [30], and Partial Autocorrelation Function
criteria to determine the order p, such as Akaike information criteria (AIC) [25–28], Bayesian
(PACF) [31]. In the bearing fault diagnosis, the order p is usually determined such that the kurtosis of
information criterion (BIC) [29], Minimum description length (MDL) [30], and Partial Autocorrelation
the residual
Function signal
(PACF) is maximized, as arefault
[31]. In the bearing addressed
diagnosis, in the
manyorderliteratures
p is usually (e.g., [23]). The
determined suchlatter
that is
theused
here because it is simpler to calculate and the kurtosis is the impulsiveness indicator
kurtosis of the residual signal is maximized, as are addressed in many literatures (e.g., [23]). The latter of fault signal.
For this objective, the kurtosis is calculated by increasing the order p from 1 to maxP
is used here because it is simpler to calculate and the kurtosis is the impulsiveness indicator of fault = 300 (lines 10–18).
The result
signal.isFor
in Figure 7a, where
this objective, the maximum
the kurtosis is calculated = 82, hence,
is atbyp increasing is used
the order for optP
p from (line =19).
1 to maxP 300Then,
the AR(lines
model 10–18). The result isagain
is constructed in Figure 7a,this,
using where the maximum
which is to repeatis atlines
p = 82, hence,Inis this
12–15. usedcase,
for optP (line it is
however,
19). Then,by
implemented thetheARMATLAB
model is constructed againwhich
function filter, using does
this, which
the sameis tobut
repeat lines 12–15.
in more simple Inway
this case,
(line 21).
however, it is implemented by the MATLAB function filter, which does
Residual signal thus obtained can be plotted by entering plot(rawData.t (optP+1:end),e) in the the same but in more simple
way (line 21). Residual signal thus obtained can be plotted by entering plot(rawData.t
MATLAB command window, and is shown in Figure 7b along with its kurtosis value. Additionally,
(optP+1:end),e) in the MATLAB command window, and is shown in Figure 7b along with its
the raw signal and its kurtosis are given in Figure 7c for comparison. Kurtosis is higher by about four
kurtosis value. Additionally, the raw signal and its kurtosis are given in Figure 7c for comparison.
times Kurtosis
after removing
is higherthe by discrete
about four signals, reflecting
times after removingthe the
fault presence
discrete more
signals, clearly.the
reflecting Forfault
more advanced
presence
results, however, another criteria as mentioned above may be applied
more clearly. For more advanced results, however, another criteria as mentioned above may be to determine the model order
p and applied
choose the one with better
to determine the modelperformances
order p and after comparison.
choose the one with This may
betterbeperformances
beyond the after scope of
this paper.
comparison. This may be beyond the scope of this paper.
(a)
(b)
Kurtosis=2.9715
5
-5
0 0.5 1 1.5 2
Time[sec]
(c)
FigureFigure 7. Kurtosis
7. Kurtosis variation
variation in in terms
terms ofofmodel
modelorder
order and
and corresponding
corresponding signals: (a) (a)
signals: model order
model order
selection based on maximum kurtosis, (b) residual signal for p = 82, and (c) raw
selection based on maximum kurtosis, (b) residual signal for p = 82, and (c) raw signal. signal.
3.3. Demodulation Band Selection (Lines 23–49)

As the second step of the signal pre-processing, a frequency band enclosing the resonance
frequencies is sought from the residual signal. It is determined based on the spectral kurtosis (SK),
which is the kurtosis at each frequency line in a time- frequency diagram. When the fault occurs, the
Appl. Sci. 2020, 10, 7302 10 of 23
3.3. Demodulation Band Selection (Lines 23–49)

As the second step of the signal pre-processing, a frequency band enclosing the resonance
frequencies is sought from the residual signal. It is determined based on the spectral kurtosis (SK),
which is the kurtosis at each frequency line in a time- frequency diagram. When the fault occurs, the
Appl. Sci. 2020, 10, x FOR PEER REVIEW
SK
10 of 23
becomes higher around the region of resonance frequencies onto which the fault signal is modulated.
The
SK SK is defined
becomes by the
higher functions
around of shortof
the region time Fourier transform
resonance (STFT)
frequencies onto [14,32]
which(line
the 39):
fault signal is
modulated. The SK is defined by the functions
of short4 time Fourier transform (STFT) [14,32] (line
39): S(t, f )
K( f ) = 2 〉 − 2 (12)

〈|𝑆(𝑡, 𝑓)| 2
𝐾(𝑓) =

S ( t, f ) −2 (12)
〈|𝑆(𝑡, 𝑓)| 〉
whereh·i〈∙〉is is
where thethe
time-averaging
time-averaging operator, andand
operator, S(t,f) is the
S(t,f) is STFT
the STFTof the
ofresidual signal.
the residual The STFT
signal. is a local
The STFT is a
Fourier transform
local Fourier for a finite
transform for a time
finitewindow
time windowat timeatt,time
which from tfrom
variesvaries
t, which = 0 𝑡to=tend0 to 𝑡 27–37).
(lines (lines
The process
27–37). The isprocess
illustrated in Figure in
is illustrated 8, in which8,the
Figure in top
which figure
the istop
thefigure
raw signal,
is the and
rawmiddle
signal, figure is the
and middle
window function to restrict the signal to the range at each time
figure is the window function to restrict the signal to the range at each0 time point t = t ~ t w point t = t0 ~ tw, local
, from which the from
signals are local
which the madesignals
as shown in theas
are made figures
shownbelow.
in the Then,
figuresone obtains
below. Then,theonebottom figure
obtains the by taking
bottom the
figure
Fourier transform for each signal. Among various window functions,
by taking the Fourier transform for each signal. Among various window functions, this paper this paper employs the Hann
window
employsas thefollows
Hann with
window length L [24] (line
as follows with27):
length L [24] (line 27):
l𝑙
!!
𝑤(𝑙)
w(l) == 0.5
0.5 11 −
− 𝑐𝑜𝑠 2𝜋
cos 2π , 0, ≤
0 l ≤𝑙 L −𝐿1− 1 (13)
(13)
L𝐿−−1 1
Figure 8. Concept
Figure8. Concept of
of STFT.
STFT.
The time interval of Fourier transform is given by the half of the window length. Remembering
that the sampling rate of the signal is 3000 Hz, which means 3000 data are in a second, when the
window length is 27 = 128; for example, the first signal at 𝑡 = 0 in the middle figure resides in the
range of 0 𝑡 0.0427 (1 𝑛 128), and the second at 𝑡 = 0.0213 (𝑛 = 65), in the range of
Appl. Sci. 2020, 10, 7302 11 of 23
The time interval of Fourier transform is given by the half of the window length. Remembering
that the sampling rate of the signal is 3000 Hz, which means 3000 data are in a second, when the
window length is 27 = 128; for example, the first signal at t0 = 0 in the middle figure resides in
the
11 of 23
range of 0≤ t ≤ 0.0427 (1 ≤ n ≤ 128), and the second at t1 = 0.0213 (n = 65), in the range of
0.0213The≤ t results
≤ 0.0640 of(FFT
65 ≤in n Figure
≤ 192), 8etc.
can also be plotted in the 2D form called spectrogram, as shown
The results of FFT in Figure
in Figure 9a, which shows the frequency 8 can also be plotted
spectrum in the
where the2D formycalled
x and spectrogram,
axes are asfrequency.
the time and shown in
Figure 9a, which shows the frequency spectrum where the x and y axes are the
Finally, the SK can be computed using Equations (1) and (2) to obtain as a function of frequency, as time and frequency.
Finally,
shown in theFigure
SK can 9b be computed
(lines using on
38–40). Based Equations
the figure,(1)the
and (2) to obtain
frequency rangeasfor
a high
function of frequency,
SK can be selected
as shown in Figure 9b (lines 38–40). Based on the figure, the frequency range for
visually, which is around 400-800 Hz in this case (lines 43–47). The frequency band thus obtained high SK can be selectedis
visually, which is around 400-800 Hz in this case (lines 43–47). The frequency band
used to band-pass filter the residual signal, which leaves only the spectrum in the band (lines 48–49). thus obtained is
used to band-pass filter the residual signal, which leaves only the spectrum in the band (lines 48–49).
(a) (b)
Figure9.9.Results
Figure ResultsofofSTFT
STFTand
andSK:
SK:(a)
(a)STFT
STFTS(t,
S(t,f )f)and
and(b)
(b)SK
SKK(f
K(f).
).
By
Bythe
theway,
way,thethe
results of STFT
results may may
of STFT vary by different
vary window
by different lengths,lengths,
window which leads whichto the
leadsdifferent
to the
SK. For better
different results,
SK. For betteritresults,
is necessary to set a proper
it is necessary to set a window length in
proper window the STFT.
length in the In caseInthe
STFT. fault
case the
interval can be can
fault interval identified visibly as
be identified is found
visibly 0.1 s (nas =
as isasfound s (𝑛in=Figure
0.1300) 300) 7b, it is advised
in Figure 7b, it that the length
is advised that
be
thesmaller
length tobe some
smaller degree than
to some the number
degree than the ofnumber
data withinof datathewithin
interval,thewhich
interval, is the same
which as maxP.
is the same
In
aspractice,
maxP. Inhowever,
practice,since the signal
however, since is the
not so clear,isseveral
signal not so window lengths
clear, several are usually
window attempted
lengths [14],
are usually
which are in[14],
attempted thiswhich
exampleare given
in thisarbitrarily
example given by [2ˆ4 2ˆ5 2ˆ6
arbitrarily [2^4(line
by 2ˆ7] 2^57).2^6 Upon obtaining
2^7] (line 7).the SK
Upon
with each window
obtaining the SK withlengths
each(lines 26–42),
window they(lines
lengths are plotted
26–42),in superposition
they are plotted (line 45), from which
in superposition (line the
45),
frequency
from which band
the for higher SK
frequency bandis for
visually
higherdetermined
SK is visually (lines 47). The result
determined (lines is shown
47). in Figure
The result 10a,
is shown
where demodulation
in Figure bands are set between
10a, where demodulation bands are[500 700] for[500
set between the range
700]of forhigher
the rangekurtosis in common.
of higher kurtosis
This should be
in common. input
This in the
should becommand
input in the window
command as shown
window in Figure
as shown 10b.inThen,
Figurethe residual
10b. signal
Then, the goes
residual
through band-pass
signal goes through filtering to result
band-pass in the to
filtering signal in Figure
result in the 10c of which
signal the kurtosis
in Figure increases
10c of which theby twice
kurtosis
to 23.1865 by
increases from 10.5885.
twice to 23.1865 from 10.5885.
Appl. Sci. 2020, 10, 7302 12 of 23
15
window size = 16
window size = 32
window size = 64
window size = 128
10
0
0 500 1000 1500
Frequency[Hz]
(a)
(b)
(c)
Figure
Figure 10. Demodulationband
10. Demodulation band selection:
selection: (a) (a) spectral
spectral kurtosis
kurtosis for different
for different windowwindow lengths,
lengths, (b) (b)
frequency
frequency bandinput
band selection selection input in MATLAB,
in MATLAB, and (c) signal
and (c) vibration vibration
aftersignal after band-pass
band-pass filtering. filtering.
3.4.Envelope
3.4. EnvelopeAnalysis
Analysis(Lines
(Lines50–64)
50–64)
Aftertwo
After twopre-processing
pre-processingtechniques—discrete
techniques—discretesignal signalseparation
separationand andband-pass
band-passfiltering—one
filtering—one
obtainsthe
obtains theAM
AMsignal
signalthat
thatincludes
includesonlyonlycomponents
componentsof ofresonance
resonance(carrier)
(carrier)and
andfault
fault(modulation)
(modulation)
frequency. Then, the envelope analysis is used as the last step for the diagnosis
frequency. Then, the envelope analysis is used as the last step for the diagnosis of bearing faults, of bearing faults,
whichconsists
which consistsof ofthree
threesub-steps:
sub-steps:First,
First,Hilbert
Hilberttransform
transformisisapplied
appliedtotothetheAMAMsignal
signaltotoobtain
obtainthe
the
analytic signal [14,15,17] (line 51). Next, its magnitude is calculated to obtain
analytic signal [14,15,17] (line 51). Next, its magnitude is calculated to obtain the envelope signal, the envelope signal,
whichisisthe
which theabsolute
absolutevalue
valueof ofthe
theanalytic
analyticsignal
signal(line
(line52).
52).ItItisisfurther
furthertranslated
translatedto tohave
havezero
zeromean
mean
byremoving
by removingthe theDCDCoffset
offset (line
(line 53).
53). Last,
Last, FFTFFTisis taken
taken toto the
the envelope
envelopesignal
signalto
toretrieve
retrievethe
thefault
fault
information from the amplitudes at the fault frequencies
information from the amplitudes at the fault frequencies (lines 54–55). (lines 54–55).
TheHilbert
The Hilberttransform referredtotoasasa a90°
transformisisreferred 90◦ phase
phase shifter
shifter and
and results
resultsininthe
theanalytic
analyticsignals,
signals,
expressedas
expressed asfollows
follows[18]:
[18]:
𝑎 (𝑡) = 𝑥(𝑡) + 𝑗𝑥 (𝑡) = 𝑥 (𝑡) + 𝑥 (𝑡)𝑒 ( ) = 𝐴 (𝑡)𝑒 ( ) (14)

q
ax (t) = x(t) + jx̂(t) = x2 (t) + x̂2 (t)e jφx (t) = Ax (t)e jφx (t) (14)
where 𝑎 indicates the analytic signal, in which the real part 𝑥 (t) is the original AM signal,
imaginary part 𝑥 (𝑡)
where ax indicates the analytic
is its Hilbert in which the𝜙real
signal, transform, (𝑡)part
is x(t)
the isinstantaneous
the original AM signal,and
phase, 𝐴 is part
imaginary the
x̂(t) is its Hilbert
instantaneous transform,
amplitude, i.e.,φthe
x (t)envelope
is the instantaneous
signal. The phase,
resulting Ax is thesignal
andenvelope instantaneous
with the amplitude,
DC offset
isi.e.,
shownthe envelope
in Figure signal.
11a. TheThe resulting
envelope envelope
spectrum signal
after takingwith
FFTthe DC offset
is given is shown
in Figure 11b, in
in Figure
which the11a.
The envelope
fault spectrum
is identified after taking
by examining FFT is given
the amplitude in Figure
at the given 11b,
faultinfrequencies
which the (lines
fault is56–64).
identified by
In the
examining
figure, the amplitude
the peaks near the faultat the given fault
frequency frequencies
10 Hz (lines 56–64).
and its harmonics In the figure,
are apparent, whichthe peaks that
indicates near
thefault
the faulthas
frequency
occurred 10indeed.
Hz andCompare
its harmonics are apparent,
the Figure 11a,b withwhichthe indicates that in
original ones theFigure
fault has
3a,doccurred
to find
indeed.
out Compare
that result of thetheenvelope
Figure 11a,b with the
spectrum original which
is different, ones inwe Figure 3a,d to find
call smearing. The outreason
that result of the
is because
the sampling rate is low and the length of residual signal is changed from the original one after the
Appl. Sci. 2020, 10, 7302 13 of 23

envelope spectrum is different, which we call smearing. The reason is because the sampling rate is low
and
ARthe lengthThough
filtering. of residual signal is changed
not addressed here, it from
is leftthe original
to the readerone
to after the ARand
implement filtering.
figure Though
out whatnotthe
addressed
result willhere, it is left
be when to theisreader
the fault to nonexistent
small or implement and figure outsmaller
by assigning what the
valueresult
to A.will
Forbe when the
comparison,
fault is small
envelop or nonexistent
analysis results of by assigning
healthy smaller
bearing are also to A.in
valuegiven For comparison,
Figure envelop
11c,d, which cananalysis
be easilyresults
made
ofbyhealthy bearing are also given in Figure 11c,d, which can be easily made by
assigning A = 0 in the MATLAB code (Bearing fault simulation). Comparing Figure 11a,b with assigning A = 0 in the
the
MATLAB
fault, the code (Bearing
normal bearing fault simulation).
does not showComparing Figurerepresenting
periodic impulse 11a,b with the thefault,
faultthe
in normal
the timebearing
domain
does
and not
the show periodic
peak at the faultimpulse representing
frequency the faultdomain
in the frequency in the time domain
either. and the the
Furthermore, peak at the fault
kurtosis value
frequency
is about 3ininthe
thefrequency domain which
normal bearing, either. is
Furthermore,
much smaller thethan
kurtosis
that value
of faultis about
and the3 intypical
the normal
of the
bearing,
Gaussian which is much smaller than that of fault and the typical of the Gaussian distribution.
distribution.
(a) (b)
(c) (d)
Figure11.
Figure 11.Result
Resultofofenvelope
envelopeanalysis:
analysis:envelope
envelopesignal
signalinin(a)
(a)time
timedomain
domain(b)
(b)frequency
frequencydomain
domainofof
faulty bearing, in (c) time domain and (d) frequency domain of normal
faulty bearing, in (c) time domain and (d) frequency domain of normal bearing. bearing.
4.4.Diagnosis
Diagnosisof
ofKorea
KoreaAerospace
AerospaceUniversity
University(KAU)
(KAU)Bearing
BearingFaults
Faults
InInthis
thissection,
section,diagnosis
diagnosis is performed usingusing thethereal
realvibration
vibrationdata
datamade
madebybythethe authors’
authors’ test
test rig
rig in the Korea Aerospace University (KAU). It is shown
in the Korea Aerospace University (KAU). It is shown in Figure 12 in Figure 12 which consists of DC motor,
of DC motor,
support
supportbearings,
bearings,andandtwo
tworoller
rollerbearings
bearingsfor
forthe
thetest.
test.Vertical
Verticalload
loadisisapplied
appliedto
tothe
thesupport
supportbearing.
bearing.
The
Thetest
testbearing
bearingisisgiven
givenbybyNSK
NSKNJ NJ2306,
2306,with
withitsitsgeometry
geometryand anddimensions
dimensionsininTable
Table1.1.Fault
Faultisisseeded
seeded
atatthe inner race with an artificially induced notch (0.2 mm depth, 0.3 mm width). The
the inner race with an artificially induced notch (0.2 mm depth, 0.3 mm width). The accelerometer accelerometer is
installed on on
is installed toptop
of the testtest
of the bearing’s
bearing’shousing
housing using
usingbolt toto
bolt acquire
acquirethe vibration
the vibrationsignals.
signals.The
Thetest
testisis
performed under four different conditions as listed
performed under four different conditions as listed in Table 2.in Table 2.
Appl. Sci. 2020, 10, 7302 14 of 23
Figure 12. Korea Aerospace University (KAU) bearing test rig.

Figure 12. Korea Aerospace University (KAU) bearing test rig.
Table 1. Bearing dimensions.
Table 1. Bearing dimensions.
Parameter Name Value Parameter Name Value
Parameter Name Value Parameter Name Value
BearingBearing
type type NJ 2306
NJ 2306 Width
Width(w)(w) 27mm
27 mm
Inner diameter (id) 30 mm Dynamic load rating 51,500 N
Inner diameter (id) 30 mm Dynamic load rating 51500 N
Outer diameter (od) 72 mm Static load rating 51,000 N
Outer diameter
Number of rollers (n) (od) 11 72 mm Static load rating 51000 N
Number of rollers (n) 11
Table 2. Operating conditions.
Table 2. Operating conditions.
Data Name RPM LOAD
Data Name RPM LOAD
bearing 1 1
bearing 1200 rpm
1200 rpm 0 N0 N
bearing 2 1200 rpm 200 N
bearing 3 1000 rpm 0N
bearing 3 1000 rpm 0N
In order to diagnose
In order the the
to diagnose bearing fault,
bearing the
fault, thefault
faultfrequencies
frequencies asasstated
statedininEquations
Equations (1)–(4)
(1)–(4) should
should
be available. In practice,
be available. however,
In practice, however, most
mostofofthe
thebearing
bearing manufacturers
manufacturers do donot
not provide
provide thethe values
values of of
ballpitch
ball and and pitch diameter.
diameter. To solve
To solve thisthis problem,
problem, a commercialbearing
a commercial bearingfault
fault frequency
frequency calculator
calculator[33]
[33] is
adoptedis adopted
in this in this study
study and theandresults
the results
are are listed
listed in in Table3.3.The
Table The bearing
bearing data
dataare acquired
are acquiredthrough a a
through
sampling frequency
sampling frequency of 51.2 kHz.of 51.2 kHz.
Table 3. Bearing fault frequencies.

Table 3. Bearing fault frequencies.
Frequency Name Value Frequency Name Value
Frequency NameBPFO Value
4.4423 × 𝑓 Frequency
FTF Name 0.4038 × Value
f
BPFO BPFI 6.5577
4.4423 × fr × 𝑓 BSF
FTF 5.0079 0.4038
×f × fr
BPFI 6.5577 × fr rotating speed [Hz]
𝑓 : shaft BSF 5.0079 × fr
fr : shaft rotating speed [Hz].
In order to illustrate the process, the dataset for “bearing 1” is taken, which is operated with a
Inrotating
order speed
to illustrate
of 1200 the
rpmprocess,
and a 0 Nthe dataset
vertical load for “bearing
(note 1” isbytaken,
that the load which
the weight is operated
of bearing with a
housing
rotating stillspeed
exists).
ofFigure 13 shows
1200 rpm and athe
0Nraw signal,load
vertical of which
(notethe kurtosis
that is 4.9923.
the load by theThe four datasets
weight listed
of bearing in
housing
Table 2 can be downloaded from https://www.kau-sdol.com/ and https://sites.google.com/view/sg-
still exists). Figure 13 shows the raw signal, of which the kurtosis is 4.9923. The four datasets listed in
Table 2kim.can be downloaded from https://www.kau-sdol.com/ and https://sites.google.com/view/sg-kim.
Appl. Sci. 2020, 10, 7302 15 of 23
Figure13.
Figure 13.Raw
Rawvibration
vibrationsignals
signals(bearing
(bearing1).1).
Figure 13. Raw vibration signals (bearing 1).
Theraw
The rawdatadataforforbearing
bearing11are areloaded
loadedby bythe
thenamenamebearing1.
bearing1. Referring to Table 2, sampRate sampRate
andrpm
and The
rpm raw
areare data
set at for
set51.2e3 bearing
at 51.2e3 and 1200, 1 are loaded
and respectively. by the
1200, respectively. name
Bearing fault bearing1.
Bearing Referring
fault
frequencies to Table
frequencies
are assigned arein2,assigned
sampRate
bearFreq, in
and rpm
bearFreq,
which arewhich
is [4.4423 set 6.5577
atis51.2e3
[4.4423
0.4038 and5.0079]
1200, respectively.
6.5577 for0.4038
BPFO, BPFI, Bearing
5.0079]FTF, fault
and frequencies
for BPFO,
BSF, BPFI, areFTF,
respectively, assigned onin
and BSF,
based
bearFreq,
respectively,
Table which
based on
3. The maximum is Table
[4.4423
order 3.
ofThethe6.5577
AR model0.4038
maximum order
maxP is ofset 5.0079]
the atAR
390 model formaxP
based BPFO,
on BPFI,
is set
the at of
ratio FTF,
390the andonBSF,
based
sampling the
respectively,
ratio of the
frequency versusbased
theon
sampling Table
highest 3. The
frequency
fault maximum
versus
frequency the order
highest of the ARfrequency
fault
(51200/(6.5577*1200/60)) modelas maxP is set atbefore.
390 based
(51200/(6.5577*1200/60))
mentioned on of
Vector the
as
ratio of
mentioned the sampling
before. Vector frequency
of the versus
window the
lengths highest
windLeng fault
the window lengths windLeng for STFT are given by [2ˆ4 2ˆ5 2ˆ6 2ˆ7 2ˆ8] under the constraint offrequency
for STFT are(51200/(6.5577*1200/60))
given by [2^4 2^5 as
2^6
mentioned
2^7 2^8]
maximum before.
under
length Vector
the
being of the
constraint
smaller thanwindow
of maximum lengths
maxP (390). AswindLeng
length being the
a result, for
smaller STFTthanare
commands given
maxP
in by [2^4
the(390).
Problem 2^5 2^6
As adefinition
result, the
2^7
(lines 2^8]
commands
1–7) areunder the constraint
inmodified
the Problem of maximum
definition
as follows: (lines 1–7) length being smaller
are modified than maxP (390). As a result, the
as follows:
commands in the Problem definition (lines 1–7) are modified as follows:
- - LineLine2:2:rawData
rawData= = load('bearing1');
load('bearing1');
- Line 2: 3: rawData
sampRate ==load('bearing1');
51.2e3;
- Line 3: sampRate = 51.2e3;
- Line 3: sampRate
4: rpm = 1200; = 51.2e3;
- Line 4: rpm = 1200;
- Line 4: 5: rpm = 1200;
bearFreq = [4.4423 6.5577 0.4038 5.0079]*rpm/60;
- Line 5: bearFreq = [4.4423 6.5577 0.4038 5.0079]*rpm/60;
- Line 5: 6: bearFreq
maxP = 390; = [4.4423 6.5577 0.4038 5.0079]*rpm/60;
- - LineLine6:6: maxP = =390;
7: maxP
windLeng 390; = [2^4 2^5 2^6 2^7 2^8].
- - Line 7:
Line 7: windLeng= [2ˆ4
windLeng = [2^4 2ˆ5 2^5 2ˆ6 2^6
2ˆ7 2ˆ8].
2^7 2^8].
After going through the discrete signal separation (lines 8–22), one obtains the optimum order
optPAfter
After going
being going
318through
for thetheAR
through discrete
the discrete
model signal
by separation
signal separation
maximizing (lines 8–22),
the(lines oneofobtains
8–22),
kurtosis one the optimum
the obtains
residual thesignal. order
optimum Then,optP
order
the
being
optP 318
resulting for
being the318
residual ARfor
model by
ARmaximizing
the becomes
signal model by the
Figure 14kurtosis
and its of
maximizing the residual
the
kurtosiskurtosis signal.
of the
is 44.0245, Then, the
residual
roughly resulting residual
signal. Then,
a nine-fold the
increase.
signal becomes
resulting
Demodulation residualFigure
band 14
signal and its (lines
becomes
selection kurtosis
Figure is 14
23–49) 44.0245,
and
is thenitsroughly
kurtosis
carried aout
nine-fold
is 44.0245,
using theincrease.
roughly Demodulation
four setsa nine-fold
of SK made band
increase.
with
selection (lines
Demodulation
five window lengths, 23–49) is then
band selection
which are carried
(lines
givenout using
23–49) the
is then15a.
in Figure four sets
carried
The outof SK
proper made
using with five
the four band
frequency window lengths,
sets offorSKhigher
made SK with
is
which
five are given
window
selected by visualin Figure
lengths, 15a. The
which
examination areproper
this frequency
ofgiven in Figure
figure, band
which 15a.isfor
The higher
given proper SKfrequency
is selected
arbitrarily by
band
as [1.8e4 visual
for examination
higherby
2.3e4] SKthe
is
ofselected
this figure,
authors. by which
Thevisual
residual is signal
given arbitrarily
examination after ofapplying as [1.8e4
this figure, which
band-pass 2.3e4]
isfilterbyisthe
given authors.
arbitrarily
then asThe
obtained as residual
[1.8e4 insignal
shown2.3e4] after
by
Figure the
15b,
applying
authors.
of whichTheband-pass filter
residual increases
the kurtosis is then
signal after obtained
to applying as
208.849, five shown
band-pass in Figure 15b,
filter isthan
times greater of
thenthe which
obtained
previousthe kurtosis
as one. increases
shown in Figure 15b, to
208.849,
of which fivethetimes greater
kurtosis than the
increases to previous
208.849, fiveone.times greater than the previous one.
Figure 14. Residual signal after the AR filtering.

Figure 14.Residual
Figure 14. Residualsignal
signalafter
afterthe
theAR
ARfiltering.
filtering.
Appl. Sci. 2020, 10, 7302 16 of 23
SK filtered signal
SK filtered signal
(a) (b)
Figure
Figure 15. Demodulation
15. Demodulation (a) bandselection
band selectionfor
for data
data of
of bearing
bearing1:1:(a)(a)
spectral kurtosis,
spectral (b) and and
kurtosis, (b) vibration
(b) vibration
signal after band-pass filtering.
signalFigure
after band-pass filtering.
15. Demodulation band selection for data of bearing 1: (a) spectral kurtosis, and (b) vibration
signal after band-pass filtering.
AfterAfter typing in the frequency band, it continues to the envelope analysis (lines 50–64), and the
typing in the frequency band, it continues to the envelope analysis (lines 50–64), and the
results are shown in Figure 16. In the figure, the four fault frequencies are plotted as vertical lines, at
results areAfter
shown in Figure
typing 16. In theband,
in the frequency figure, the fourto
it continues fault frequencies
the envelope are plotted
analysis as vertical
(lines 50–64), and thelines,
which the peak is examined to identify whether the fault has occurred or not. The result, however,
results
at which are
thebepeakshown in Figure
is examined 16. In the
to identify figure, the
whether four fault
the fault frequencies
has occurred are plotted
or not. as vertical
The result, lines, at
however,
should interpreted with caution. It shows multitude of peaks at the several frequencies: the first
should which the peak
be interpreted is examined to identify whether the fault has occurred or not. The result, however,
group is at the shaftwith caution.
frequency (20 Hz It shows multitude
= 1200 rpm/60 s) andof its
peaks at the (black
harmonics several frequencies:
arrows), which canthe first
group should beshaft
interpreted with (20
caution. = It1200
shows multitude of its
peaks at the several frequencies: the first
be caused by unbalance, misalignment, or coupling lock up. Details associated with shaft harmonicscan be
is at the frequency Hz rpm/60 s) and harmonics (black arrows), which
group is at the shaft frequency (20 Hz = 1200 rpm/60 s) and its harmonics (black arrows), which can
causedcanbybeunbalance,
referred to misalignment,
the literature [34]. or coupling
Next is atlock up. Details
the BPFI (130 Hz) associated with shaft
and sidebands harmonics can
with harmonics
be caused by unbalance, misalignment, or coupling lock up. Details associated with shaft harmonics
spaced at
be referred tothe
theinterval
literatureof shaft
[34].frequency
Next is at (20the
Hz)BPFI
around(130 it (orange
Hz) andarrows).
sidebandsOncewith
the peaks like these
harmonics spaced
can be referred to the literature [34]. Next is at the BPFI (130 Hz) and sidebands with harmonics
at theare identified around the BPFI, it(20 indicates the inner race fault occurrence. Note the
spaced at the interval of shaft frequency (20 Hz) around it (orange arrows). Once the peaks like these are
interval of shaft frequency Hz) around it (orange arrows). Once that,peaks
in caselike
of inner
these
race
identified fault,
aroundsidebands
the typically
BPFI, appearthe around the BPFI, whereas it is notNote
so inthat,
the other faults [35]. race
are identified around theitBPFI,
indicates
it indicates inner race
the inner fault occurrence.
race fault occurrence. Note that,inincase
caseofofinner
inner
Note also that sometimes the peaks can overlap by chance in the same frequency range, as is found
fault, sidebands typically appear
race fault, sidebands typicallyaround
appearthe BPFI,the
around whereas it is notitsoisin
BPFI, whereas the
not soother
in thefaults
other [35].
faultsNote
[35]. also
at BPFO, which should not be mistaken as the fault.
that sometimes
Note also that thesometimes
peaks canthe overlap
peaks can by chance
overlap by in the
chancesame frequency
in the range, as
same frequency is found
range, at BPFO,
as is found
whichatshould
BPFO, notwhichbeshould
mistaken not be
as mistaken
the fault.as the fault.
Figure 16. Frequency domain of envelope analysis.
Figure
Figure
5. Diagnosis of CWRU Bearing 16.16. Frequencydomain
Frequency
Faults domain of
of envelope
envelope analysis.
analysis.
Vibration
5. Diagnosis
5. Diagnosis data
of
of CWRU provided
CWRU by
Bearing
Bearing the Case Western Reserve University (CWRU) Bearing Data Center
Faults
Faults
are employed [36,37] to demonstrate another application of the diagnosis code. In this case, artificial
Vibration
Vibration data
data in provided
provided by Case
byfrom
the the Case Western Reserve University(CWRU)
(CWRU)Bearing
Bearing Data
Data Center
faults ranging diameter 0.007Western
to 0.028Reserve
inches University
were seeded Center are
with the electro-discharge
are employed [36,37] to demonstrate another application of the diagnosis code. In this case, artificial
employed [36,37]
machining to demonstrate
(EDM) on drive-endanother application
(SKF 6205-2RS JEM)of the
and diagnosis code.
fan-end (SKF In this case,
6203-2RS JEM) artificial
bearings.faults
faults ranging in diameter from 0.007 to 0.028 inches were seeded with the electro-discharge
During
ranging the tests, data
in diameter fromwere
0.007collected
to 0.028atinches
12 andwere
48 kHz sampling
seeded withrates for several datasets.machining
the electro-discharge Test
machining (EDM) on drive-end (SKF 6205-2RS JEM) and fan-end (SKF 6203-2RS JEM) bearings.
(EDM)During
on drive-end
the tests,(SKF
data6205-2RS JEM) and
were collected at 12fan-end
and 48 (SKF 6203-2RSrates
kHz sampling JEM)for
bearings. During the
several datasets. Testtests,
data were collected at 12 and 48 kHz sampling rates for several datasets. Test bearings were driven
with different rotation speeds ranging from 1797 to 1720 rpm. Accelerometers were installed on three
Appl. Sci. 2020, 10, 7302 17 of 23
different locations: on the drive-end bearing (DE), the fan-end bearing housing (FE), and the motor
supporting base plate (BA).
Among others, this paper considers the bearing with inner race fault on the17 of fan23
end.
Geometry information and fault frequencies of the bearings are listed in Table 4
bearings were driven with different rotation speeds ranging from 1797 to 1720 rpm. Accelerometers
(the number of rolling
elements
were and contact
installed on angles are not provided
three different locations: directly, and they bearing
on the drive-end are calculated by fan-end
(DE), the using the suggested
bearing
fault housing
frequencies),
(FE), and andthethe sampling
motor rate base
supporting and plate
rotation
(BA).speed are 12 kHz and 1796 rpm, respectively.
The data Among
are alsoothers,
available withconsiders
this paper the name the“bearing 270”inner
bearing with (therace
original data
fault on thewas resaved
fan end. with two
Geometry
information and fault frequencies of the bearings are listed in Table 4 (the number
variables: vib for vibration data, and t for measuring time in rawData) at https://www.kau-sdol.com/. of rolling elements
anddetails
Further contactregarding
angles are the not data
provided directly,
and test andbe
rig can they are calculated
found in referencesby using the suggested fault
[36,37].
frequencies), and the sampling rate and rotation speed are 12 kHz and 1796 rpm, respectively. The
data are also available with the nameTable
“bearing 270” (the
4. Bearing original data was resaved with two variables:
information.
vib for vibration data, and t for measuring time in rawData) at https://www.kau-sdol.com/. Further
Parameter
details regarding Name
the data Value
and test rig can be Parameter
found in references Name
[36,37]. Value
Pitch diameter (D) 1.122 inch BPFO 3.0530 fr
Ball diameter (d) Table 4. Bearing
0.2656 inch information. BPFI 4.9469 fr
Number of rolling element
Parameter Name (n) 13 Value FTF
Parameter Name Value0.3817 fr
Contact angle (α) 0 BSF 1.994 f
Pitch diameter (D) 1.122 inch BPFO 3.0530𝑓 r
Ball diameter (d) 0.2656 inch BPFI 4.9469𝑓
Number of rolling element (n) 13 FTF
Based on the information provided by the literature, the code to test the 0.3817𝑓
bearings is changed
as follows: Contact angle (α) 0 BSF 1.994𝑓
Based on the information provided by the literature, the code to test the bearings is changed as
- Line 2: rawData = load (‘bearing270’);
follows:
- -LineLine
3: sampRate = 12e3;
2: rawData = load (‘bearing270’);
- -LineLine
4: rpm = 1796; = 12e3;
3: sampRate
- -LineLine 4: rpm = 1796;
5: bearFreq = [3.053 4.947 0.382 1.994]*rpm/60;
- -LineLine 5:
6: maxP = 82; = [3.053 4.947 0.382 1.994]*rpm/60;
bearFreq
- Line 6: maxP = 82;
- Line 7: windLeng = [2ˆ4 2ˆ5 2ˆ6].
- Line 7: windLeng = [2^4 2^5 2^6].
In line
In 6, since
line the highest
6, since fault frequency
the highest is BPFI
fault frequency is with
BPFI 148.0802 Hz (4.947*1796/60),
with 148.0802 the maximum
Hz (4.947*1796/60), the
ordermaximum
of the ARordermodelof should
the AR be 82 (12000/148.0802),
model which is thewhich
should be 82 (12000/148.0802), number of data
is the numberwithin a fault
of data withininterval.
a fault interval. In line 7, window lengths are set based on the criteria mentioned in
In line 7, window lengths are set based on the criteria mentioned in Section 3.3. The results are shownSection 3.3. The
results
in Figure 17.areBased
shownoninthe
Figure 17.inspection
visual Based on the visual inspection
of spectral kurtosisofinspectral
Figure kurtosis in Figure
17a, [5.3e3 5.9e3]17a,Hz is
[5.3e3
selected in the5.9e3]
commandHz iswindow
selected infor
thethe
command window band
demodulation for theselection.
demodulationAfterband selection.filtering
band-pass After by
band-pass filtering by this interval, the results of the envelope analysis are shown in Figure 17b,
this interval, the results of the envelope analysis are shown in Figure 17b, which shows a dominant
which shows a dominant peak at BPFI (148.08 Hz) and sidebands with harmonics at the interval of
peak at BPFI (148.08 Hz) and sidebands with harmonics at the interval of shaft frequency (29.93 Hz),
shaft frequency (29.93 Hz), convincing that the bearing has inner race fault.
convincing that the bearing has inner race fault.
(a) (b)
Figure 17. Results obtained from the code: CWRU bearing data (bearing280.mat) (a) Spectral
Figure 17. Results obtained from the code: CWRU bearing data (bearing280.mat) (a) Spectral kurtosis
kurtosis for bearing 270, and (b) Envelope spectrum after SK filtering.
for bearing 270, and (b) Envelope spectrum after SK filtering.
Appl. Sci. 2020, 10, 7302 18 of 23

As a supplementary explanation, the effects of each step in the signal preprocessing are also
demonstrated As in Figure 18, in which
a supplementary the envelope
explanation, the effects analysis
of each stepareincarried outpreprocessing
the signal to the raw signal
are also(a) and
to the residual signalin that
demonstrated Figure went
18, inthrough
which thethe AR filtering
envelope analysis are(b),carried
respectively. In Figure
out to the raw signal (a)18a,
and the
to peak
the the
at BPFI in residual signal that
spectrum went through
is apparent buttheisAR filteringthan
smaller (b), respectively.
the others,InwhichFigure 18a, the peak atof
is doubtful BPFI
the fault.
Anotherinpeak
the spectrum
appeared is apparent but is Hz)
at BSF (59.69 smaller than
also doesthe not
others,
givewhich is doubtful
us the confidenceof thebecause
fault. Another
it coincides
peak appeared at BSF (59.69 Hz) also does not give us the confidence because it coincides with the
with the harmonic of shaft frequency, which are the multiples of 29.93 Hz. After the AR filtering,
harmonic of shaft frequency, which are the multiples of 29.93 Hz. After the AR filtering, the peak at
the peakthe
at the
BPFIBPFI is increased
is increased greatlygreatly along
along with with the sidebands
the sidebands with the
with the intervals intervals
of shaft of shaft
frequency frequency
in Figure
in Figure18b.
18b. This
This confirms
confirms theatfault
the fault at that frequency.
that frequency. On the otherOn thethe
hand, other hand,
height at thethe
BSFheight at the BSF is
is not changed
not changed relative
relative to the neighbors,
to the neighbors, whichthat
which suggests suggests
it may notthatbeitthemay
faultnot
butbe
thethe fault but
harmonic theshaft
of the harmonic
frequency.
of the shaft frequency.In this case, case,
In this the fault at BPFI
the fault is already
at BPFI identified
is already at thisatstep
identified this without furtherfurther
step without
preprocessing, i.e., the demodulation band selection, although the result may be better interpreted as
preprocessing, i.e., the demodulation band selection, although the result may be better interpreted as
shown in Figure 17b. As such, not all bearings have to resort to the full preprocessing steps for the
shown in Figure 17b. As such, not all bearings have to resort to the full preprocessing steps for the
fault identification.
fault identification.
(a) (b)
Figure 18. Envelope spectrum reflecting the raw and AR filtered signals: (a) envelope spectrum of the
Figure 18. Envelope spectrum reflecting the raw and AR filtered signals: (a) envelope spectrum of the
raw signal, and (b) envelope spectrum after AR filter.
raw signal, and (b) envelope spectrum after AR filter.
Refer to the following code to obtain the results in Figure 18.
Refer toaX
the= following code to obtain%===
hilbert(rawData.vib); the results
(a) Rawinsignal
Figure 18.
envel = abs(aX);
aX = hilbert(rawData.vib); %=== (a) Raw signal
envel = envel-mean(envel);
envel = abs(aX);
fftEnvel = abs(fft(envel))/Ne*2;
fftEnvel = fftEnvel(1:ceil(N/2));
freq = (0:Ne-1)/Ne*sampRate;
fftEnvel = freq
abs(fft(envel))/Ne*2;
= freq(1:ceil(N/2));
fftEnvel = figure(3)
fftEnvel(1:ceil(N/2));
stem(freq,fftEnvel,’LineWidth’,1.5); hold on;
freq = (0:Ne-1)/Ne*sampRate;
[xx,yy] = meshgrid(bearFreq,ylim);
freq = freq(1:ceil(N/2));
plot(xx(:,1),yy(:,1),’*-’,xx(:,2),yy(:,2),’x-’,xx(:,3),yy(:,3),’d-
figure(3) ’,xx(:,4),yy(:,4),’^-’)
legend(‘Envelope spectrum’,’BPFO’,’BPFI’,’FTF’,’BSF’);
xlabel(‘Frequency [Hz]’); ylabel(‘Amplitude[g]’); xlim([0 max(bearFreq)*1.8])
aX = hilbert(e); %=== (b) Residual signal
envel = abs(aX);
’,xx(:,4),yy(:,4),’ˆ-’)
xlabel(‘Frequency
fftEnvel [Hz]’); ylabel(‘Amplitude[g]’); xlim([0 max(bearFreq)*1.8])
= fftEnvel(1:ceil(N/2));
figure(4)
aX = hilbert(e); %=== (b) Residual signal
envel = abs(aX);
fftEnvel = fftEnvel(1:ceil(N/2));
Appl. Sci. 2020, 10, 7302 19 of 23
figure(4)
’,xx(:,4),yy(:,4),’ˆ-’)
xlabel(‘Frequency [Hz]’); ylabel(‘Amplitude [g]’); xlim([0 max(bearFreq)*1.8])
6. Conclusions
Signal processing for feature extraction is an important step in the fault diagnosis and prognosis,
but the process in detail are not fully understood for the beginners of PHM to implement by themselves.
With this motivation, this paper has presented a tutorial for how to identify the fault using the
envelope analysis along with the preprocessing steps such as AR filtering and SK filtering to remove
the unnecessary signals and noises, which are widely used in the area of bearing fault diagnostics.
MATLAB codes for the algorithms are provided, which consists of 64 lines and can be simply modified
for the other applications. Two sets of bearing vibration data (data made by the authors and CWRU
bearing data) are used to help the user to implement and understand the algorithms. The code and
bearing data are available at https://www.kau-sdol.com/ and https://sites.google.com/view/sg-kim.
Apart from the datasets in this tutorial, there are several more datasets open to the public, and some of
those are listed in Table 5 to help the user practice diagnosis themselves.
Table 5. Open Data for Fault Diagnosis and Prognosis of Bearings.
Data Set Name Diagnosis Prognosis Failure Type Operating Condition

FEMTO [38] X O Natural Different load and speed
MFPT [39] O X Artificial Different load
IMS [40] X O Natural Single condition
WT [41] X O Natural Single Condition
Paderborn [42] O X Artificial, Natural Different operating condition
Author Contributions: Conceptualization, S.K. and D.A.; writing-original draft preparation, S.K.; validation, S.K.;
MATLAB coding, S.K.; review and editing, D.A. and J.-H.C.; supervision, D.A. and J.-H.C. All authors have read
and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Acknowledgments: This work was supported by the National Research Foundation of Korea (NRF) grant funded
by the Korea government (MSIT). (No. 2020R1A4A4079904).
Conflicts of Interest: The authors declare no conflict of interest.
Appendix A. Bearing Fault Simulation

%% Bearing fault simulation signal
% Parameter setting ===================================================
fr = 600; % Carrier signal
fd = 13; % discrete signal
ff = 10; % Characteristic frequency(Modulating signal)
a = 0.02; % Damping ratio
T = 1/ff; % Cyclic period
fs = 3e3; % Sampling rate
K = 50; % Number of impulse signal
t = 1/fs:1/fs:2; % Time
A=5; % Maximum amplitude
noise = 0.5;
%=====================================================================
Appl. Sci. 2020, 10, 7302 20 of 23
for k = 0 : K-1
for i = 1 : length(t)
if t(i)-k*T>=0
x1(i) = A*exp(-a*2*pi*fr.*(t(i)-k*T));
x2(i) = sin(2*pi*fr.*(t(i)-k*T));
x3(i) = x1(i).*x2(i);
end;end;end
x5 = normrnd(0,noise,1,length(x3));
x4 = 2*sin(2*pi.*t.*fd);
vib = x3 + x4 + x5;
save(‘Simulation’,’vib’,’t’)
Appendix B. Bearing Fault Diagnosis
1 %======================PROBLEM DEFINITION ===========================

2 rawData = load(‘Simulation’); % Data load
3 sampRate = 3e3; % Sampling rate (Hz)
4 rpm = 60; % Shaft rotating speed
5 bearFreq = [10]*rpm/60; % BPFO, BPFI, FTF, BSF
6 maxP = 300; % Maximum order of AR model
7 windLeng = [2ˆ4 2ˆ5 2ˆ6 2ˆ7]; % Window length of STFT
8 %==============Discrete signal separation (AR model) ==================
9 x=rawData.vib(:); N=length(x);
10 for p = 1 : maxP
11 if rem(p,50)==0; disp([‘p=‘ num2str(p)]); end
12 a = aryule(x,p); % aryule returns the AR model parameter, a(k)
13 X = zeros(N,p);
14 for i = 1 : p; X(i+1:end,i) = x(1:N-i); end
15 xp = -X*a(2:end)’;
16 e = x-xp;
17 tempKurt(p,1) = kurtosis(e(p+1:end));
18 end
19 optP = find(tempKurt==max(tempKurt)); %==== Optimum solution
20 optA = aryule(x,optP);
21 xp = filter([0 -optA(2:end)],1,x);
22 e = x(optP+1:end) - xp(optP+1:end); % residual signal
23 %============Demodulation band selection (STFT & SK) =================
24 Ne = length(e);
25 numFreq = max(windLeng)+1;
26 for i = 1 : length(windLeng)
27 windFunc = hann(windLeng(i )); %==== Short Time Fourier Transform
28 numOverlap = fix(windLeng(i)/2);
29 numWind = fix((Ne-numOverlap)/(windLeng(i)-numOverlap));
30 n = 1:windLeng(i);
31 STFT=zeros(numWind,numFreq);
32 for t = 1 : numWind
33 stft = fft(e(n).*windFunc, 2*(numFreq-1));
34 stft = abs(stft(1:numFreq))/windLeng(i)/sqrt(mean(windFunc.ˆ2))*2;
35 STFT(t,:) = stft’;
36 n = n + (windLeng(i)-numOverlap);
37 end
Appl. Sci. 2020, 10, 7302 21 of 23
38 for j = 1 : numFreq %==== Spectral Kurtosis

39 specKurt(i,j) = mean(abs(STFT(:,j)).ˆ4)./mean(abs(STFT(:,j)).ˆ2).ˆ2-2;
40 end
41 lgd{i} = [‘window size = ‘,num2str(windLeng(i))];
42 end
43 figure(1) %==== Results
44 freq = (0:numFreq-1)/(2*(numFreq-1))*sampRate;
45 plot(freq,specKurt); legend(lgd,’location’,’best’)
46 xlabel(‘Frequency[Hz]’); ylabel(‘Spectral kurtosis’); xlim([0 sampRate/2]);
47 [freqRang] = input(‘Range of bandpass filtering, [freq1,freq2] = ‘);
48 [b,a] = butter(2,[freqRang(1) freqRang(2)]/(sampRate/2),’bandpass’);
49 X = filter(b,a,e); % band-passed signal
50 %=======================Envelope analysis ============================
51 aX = hilbert(X); % hilbert(x) returns an analytic signal of x
52 envel = abs(aX);
53 envel=envel-mean(envel); % envelope signal
54 fftEnvel = abs(fft(envel))/Ne*2;
55 fftEnvel = fftEnvel(1:ceil(N/2));
56 figure(2) %==== Result plot
57 freq = (0:Ne-1)/Ne*sampRate;
58 freq = freq(1:ceil(N/2));
59 stem(freq,fftEnvel,’LineWidth’,1.5); hold on;
60 [xx,yy]=meshgrid(bearFreq,ylim);
61 plot(xx(:,1),yy(:,1),’*-’)
62 % ,xx(:,2),yy(:,2),’x-’,xx(:,3),yy(:,3),’d-’,xx(:,4),yy(:,4),’ˆ-’)
63 legend(‘Envelope spectrum’,’BPFO’,’BPFI’,’FTF’,’BSF’);
64 xlabel(‘Frequency [Hz]’); ylabel(‘Amplitude [g]’); xlim([0 max(bearFreq)*1.8])
References
1. He, D.; Bechhoefer, E.; Ma, J.; Zhu, J. A particle filtering based approach for gear prognostics. In Diagnostics and
Prognostics of Engineering Systems: Methods and Techniques; IGI Global: Hershey, PA, USA, 2013; pp. 257–266.
2. Kim, S.; Choi, J.H. Convolutional neural network for gear fault diagnosis based on signal segmentation
approach. Struct. Health Monit. 2019, 18, 1401–1415. [CrossRef]
3. An, D.; Choi, J.H.; Kim, N.H. Remaining useful life prediction of rolling element bearings using degradation
feature based on amplitude decrease at specific frequencies. Struct. Health Monit. 2018, 17, 1095–1109.
[CrossRef]
4. Siegel, D.; Ly, C.; Lee, J. Methodology and framework for predicting helicopter rolling element bearing
failure. IEEE Trans. Reliab. 2012, 61, 846–857. [CrossRef]
5. Kim, S.; Park, S.; Kim, J.W.; Han, J.; An, D.; Kim, N.H.; Choi, J.H. A new prognostics approach for bearing
based on entropy decrease and comparison with existing methods. In Proceedings of the Annual Conference
of the Prognostics and Health management Society, Denver, CO, USA, 3–6 October 2016.
6. Lim, C.; Kim, S.; Seo, Y.H.; Choi, J.H. Feature extraction for bearing prognostics using weighted correlation
of fault frequencies over cycles. Struct. Health Monit. 2020, 22, 1475921719900917. [CrossRef]
7. Lee, J.; Wu, F.; Zhao, W.; Ghaffari, M.; Liao, L.; Siegel, D. Prognostics and health management design for
rotary machinery systems—Reviews, methodology and applications. Mech. Syst. Signal Process. 2014, 42,
314–334. [CrossRef]
8. Wang, D.; Tsui, K.L.; Miao, Q. Prognostics and health management: A review of vibration based bearing and
gear health indicators. IEEE Access. 2017, 6, 665–676. [CrossRef]
9. De Azevedo, H.D.M.; Araújo, A.M.; Bouchonneau, N. A review of wind turbine bearing condition monitoring:
State of the art and challenges. Renew. Sustain. Energy Rev. 2016, 56, 368–379. [CrossRef]
Appl. Sci. 2020, 10, 7302 22 of 23
10. Rai, A.; Upadhyay, S.H. A review on signal processing techniques utilized in the fault diagnosis of rolling
element bearings. Tribol. Int. 2016, 96, 289–306. [CrossRef]
11. Cerrada, M.; Sánchez, R.V.; Li, C.; Pachecod, F.; Cabreraa, D.; de Oliveirae, J.V.; Vásquezf, R.E. A review
on data-driven fault severity assessment in rolling bearings. Mech. Syst. Signal Process. 2018, 99, 169–196.
[CrossRef]
12. Hamadache, M.; Jung, J.H.; Park, J.; Youn, B.D. A comprehensive review of artificial intelligence-based
approaches for rolling element bearing PHM: Shallow and deep learning. JMST Adv. 2019, 1, 125–151.
[CrossRef]
13. Caesarendra, W.; Tjahjowidodo, T. A review of feature extraction methods in vibration-based condition
monitoring and its application for degradation trend estimation of low-speed slew bearing. Machines 2017,
5, 21. [CrossRef]
14. Randall, R.B.; Antoni, J. Rolling element bearing diagnostics—A tutorial. Mech. Syst. Signal Process. 2011,
25, 485–520. [CrossRef]
15. Bechhoefer, E.; Kingsley, M.; Menon, P. Bearing envelope analysis window selection using spectral kurtosis
techniques. In Proceedings of the 2011 IEEE International Conference on Prognostics and Health Management,
PHM 2011, Denver, CO, USA, 20–23 June 2011. [CrossRef]
16. Nikolaou, N.G.; Antoniadis, I.A. Rolling element bearing fault diagnosis using wavelet packets. NDT E Int.
2002, 35, 197–205. [CrossRef]
17. McInerny, S.A.; Dai, Y. Basic vibration signal processing for bearing fault detection. IEEE Trans. Educ. 2003,
46, 149–156. [CrossRef]
18. Shin, K.; Hammond, J. Fundamentals of Signal Processing for Sound and Vibration Engineers; John Wiley & Sons:
Hoboken, NJ, USA, 2008.
19. Kedadouche, M.; Liu, Z.; Vu, V.H. A new approach based on OMA-empirical wavelet transforms for bearing
fault diagnosis. Measurement 2016, 90, 292–308. [CrossRef]
20. Zhang, X.; Kang, J.; Zhao, J.; Teng, H. Bearing run-to-failure data simulation for condition based maintenance.
TELKOMNIKA Indonezian J. Electr. Eng. 2014, 12, 514–519. [CrossRef]
21. Bechhoefer, E. A Quick Introduction to Bearing Envelope Analysis; Green Power Monitoring Systems: Cornwall,
VT, USA, 2016.
22. Randall, R.B. Vibration-Based Condition Monitoring: Industrial, Aerospace and Automotive Applications; Wiley:
Hoboken, HJ, USA, 2010. [CrossRef]
23. Sawalhi, N.; Randall, R.B.; Endo, H. The enhancement of fault detection and diagnosis in rolling element
bearings using minimum entropy deconvolution combined with spectral kurtosis. Mech. Syst. Signal Process.
2007, 21, 2616–2633. [CrossRef]
24. Sawalhi, N.; Randall, R.B. The application of spectral kurtosis to bearing diagnostics. In Proceedings of the
ACOUSTICS, Gold Coast, Australia, 3–5 November 2004.
25. Ayaz, E. Autoregressive modeling based bearing fault detection in induction motors. In Proceedings of the
9th International Symposium on Industrial Electronics, Banja Luka, Bosnia and Herzegovina, 1–3 November
2012; pp. 99–102.
26. Akaike, H. Information theory and an extension of the maximum likelihood principle. In Selected Papers of
Hirotugu Akaike; Springer: New York, NY, USA, 1998; pp. 199–213.
27. AR Order Selection with Partial Autocorrelation Sequence. Available online: https://www.mathworks.com/
help/signal/ug/ar-order-selection-with-partial-autocorrelation-sequence.html (accessed on 9 October 2020).
28. Vrieze, S.I. Model selection and psychological theory: A discussion of the differences between the Akaike
information criterion (AIC) and the Bayesian information criterion (BIC). Psychol. Methods 2012, 17, 228.
[CrossRef] [PubMed]
29. Schwarz, G. Estimating the dimension of a model. Ann. Stat. 1978, 6, 461–464. [CrossRef]
30. Grünwald, P.D.; Grunwald, A. The Minimum Description Length Principle; MIT Press: Cambridge, MA,
USA, 2007.
31. Ramsey, F.L. Characterization of the partial autocorrelation function. Ann. Stat. 1974, 2, 1296–1301. [CrossRef]
32. Vrabie, V.; Granjon, P.; Serviere, C. Spectral kurtosis: From definition to application. In Proceedings of
the 6th IEEE International Workshop on Nonlinear Signal and Image Processing (NSIP 2003), Grado, Italy,
8–11 June 2003.
Appl. Sci. 2020, 10, 7302 23 of 23
33. NSK Technical Calculations. Available online: http://www.jp.nsk.com/app02/BearingGuide/m/html/en/TopS.html

(accessed on 9 October 2020).
34. Ehrich, F.F. Handbook of Rotordynamics; Krieger: Malabar, FL, USA, 1999.
35. Wang, Y.F.; Kootsookos, P.J. Modeling of low shaft speed bearing faults for condition monitoring. Mech. Syst.
Signal Process. 1998, 12, 415–426. [CrossRef]
36. Bearing Data Center. Case Western Reserve University Seeded Fault Test Data; Case Western Reserve University:
Cleveland, OH, USA, 2015.
37. Smith, W.A.; Randall, R.B. Rolling element bearing diagnostics using the Case Western Reserve University
data: A benchmark study. Mech. Syst. Signal Process. 2015, 64–65, 100–131. [CrossRef]
38. FEMTO-ST, IEEE PHM Data Challenge. 2012. Available online: http://www.femto-st.fr/en/Research-
departments/AS2M/Researchgroups/PHM/IEEE-PHM-2012-Data-challenge (accessed on 9 October 2020).
39. Mechanical Failures Prevention Group (MFPT) Society (a Division of the Vibration Institute), Brook, O., IL.
Available online: http://www.mfpt.org/FaultData/FaultData.htm (accessed on 9 October 2020).
40. Lee, J.; Qiu, H.; Yu, G.; Lin, J. Rexnord Technical Services: Bearing Data Set, Moffett Field, CA: IMS, Univ.
Cincinnati; NASA Ames Prognostics Data Repository, NASA Ames: Washington, WA, USA, 2007.
41. Bechhoefer, E.; Van Hecke, B.; He, D. Processing for improved spectral analysis. In Proceedings of the Annual
Conference of the Prognostics and Health Management Society, New Orleans, LA, USA, 14–17 October 2013;
pp. 14–17.
42. Lessmeier, C.; Kimotho, J.K.; Zimmer, D.; Sextro, W. Condition monitoring of bearing damage in
electromechanical drive systems by using motor current signals of electric motors: A Benchmark data set
for data-driven classification. In Proceedings of the European Conference of the Prognostics and Health
Management Society, Bilbao, Spain, 5–8 July 2016.
Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional
affiliations.
© 2020 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/).

Applied Sciences

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Applied Sciences

Uploaded by

Copyright:

Available Formats

applied

Appl. Sci. 2020, 10, 7302; doi:10.3390/app10207302 www.mdpi.com/journal/applsci

2.1. Fault Signal Given by Amplitude Modulation

2.2. Simulation of Fault Signal

Appl. Sci. 2020, 10,

𝑥 (𝑛) = − ∑X p 𝑎(𝑘)𝑥(𝑛 − 𝑘) or in a matrix form 𝐱 = −𝐗𝐚 (7)

rxx (−1) = [x(1)x(2) + x(2)x(3) + · · · + x(N − 1)x(N ) + x(N )x(N + 1)]/N

3.3. Demodulation Band Selection (Lines 23–49)

3.3. Demodulation Band Selection (Lines 23–49)

𝑎 (𝑡) = 𝑥(𝑡) + 𝑗𝑥 (𝑡) = 𝑥 (𝑡) + 𝑥 (𝑡)𝑒 ( ) = 𝐴 (𝑡)𝑒 ( ) (14)

Appl. Sci. 2020, 10, x FOR PEER REVIEW 13 of 23

Figure 12. Korea Aerospace University (KAU) bearing test rig.

Table 3. Bearing fault frequencies.

Figure 14. Residual signal after the AR filtering.

Appl. Sci. 2020, 10, x FOR PEER REVIEW 16 of 23

Figure 16. Frequency domain of envelope analysis.

Appl. Sci. 2020, 10, x FOR PEER REVIEW 18 of 23

Table 5. Open Data for Fault Diagnosis and Prognosis of Bearings.

Data Set Name Diagnosis Prognosis Failure Type Operating Condition

Appendix A. Bearing Fault Simulation

Appendix B. Bearing Fault Diagnosis

1 %======================PROBLEM DEFINITION ===========================

38 for j = 1 : numFreq %==== Spectral Kurtosis

33. NSK Technical Calculations. Available online: http://www.jp.nsk.com/app02/BearingGuide/m/html/en/TopS.html

You might also like