2022 International Conference on Electrical Engineering and Sustainable Technologies (ICEEST)

Advance Fault Identification of Rolling Element

Bearing
2022 International Conference on Electrical Engineering and Sustainable Technologies (ICEEST) | 978-1-6654-6209-9/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICEEST56292.2022.10077865

Ammar Naseer Dina Shona Laila Iqra Farhat

Department of Electrical, Electronics and Electrical and Electronic Engineering, Department of Electrical, Electronics and
Telecommunication Engineering, Faculty of Engineering, Telecommunication Engineering ,
University of Engineering and Technology Universiti Teknologi Brunei, University of Engineering and Technology
New Campus, Brunei Darussalam New Campus,
Lahore, Pakistan [email protected] Lahore, Pakistan
[email protected] [email protected]

Abstract — Health monitoring and fault identification of Vibration signals are usually analyzed in time domain [8],
unhealthy bearing have been of great significance in industries. and frequency domain [9]. Time domain analysis fails identify
Various techniques to identify faults have been developed, but the frequency feature of faulty data hence this technique is not
techniques to filter the vibrations generated due to synchronous preferred. Frequency-domain analyses are able to identify
noise and random noise precisely are still scarce. This work frequency component but due to nonstationary and non-linear
proposes development of a more precise technique where Morlet features of signals generated by bearing, this technique is also
Wavelet (MW) is used as a filter for pre-processing of data. Then not suitable. To obtain frequency components and due to non-
fast Kurtogram and Hilbert Transform are applied to identify stationary nature of vibration signals, the most applicable
faults on rolling elements. The proposed signal processing
technique would be a combined time-frequency domain
algorithm is then implemented using National Instruments
LabVIEW for real-time automatic bearing fault detection. The
analysis [10].
result shows that the fault diagnostics of the proposed technique Vibration signals generated due to early bearing failure are
is more precise. usually weak hence they ae easily submerged in the
environmental noise, which prevents from early detection of
Keywords — Morlet Wavelet, Rolling element bearings, the failure [11]. In this paper, a real-time advance fault
Kurtogram, Hilbert Transform, detection technique is developed where preprocessing of non-
stationary noisy signal is performed by using Morlet Wavelet,
then those signals are analyzed using Kurtogram, a time-
I. INTRODUCTION frequency domain analysis. Furthermore, by calculating the
Rolling bearing elements are key components in health signal power spectral density (PSD), the time domain
management of engineering machinery, as it withstands all vibrational data is converted to frequency domain to detect the
radial and axial loads and reduces continuous rotational type of fault. National Instrument (NI) LabVIEW is used to
friction. Mostly due to excessive load and insufficient implement the detection algorithm.
lubrication, bearing may get damaged, thus they are The remaining paper is distributed as follows. Section II
considered as the most vulnerable part in electrical machine. explains the proposed technique for fault detection in bearing.
According to a survey, bearing faults are noticed in 40 to 90% In Section III, experimental setup was laid for two bearing
of machines, based on the type and size of the machine[1]. fault data sets. In Section IV the faulty data were implemented
Due to its vast applications in motors, engines and turbines, with and without filter and a comparative analysis is
there is a crucial need to develop detection and identification presented. Conclusion and future works are provided in
techniques for bearing faults. Section V.
There are various kind of rolling bearing faults, such as
surface rub damage, wear and tear. Amongst them, the most II. VIBRATIONAL ANALYSIS
common one is the damage caused due to surface friction; The raw vibrational signal generated by an accelerometer
These faults can be distributed from surface cracks to surface gives only minimal and vague information regarding the
spalling which can lead to abnormal conditions [2]. To bearing fault. The existence of high noise in the signal also
identify the faults in a motor bearing, most commonly used often causes misleading interpretation of these data. In this
techniques are based on vibration analysis [3], [4], using section a step by step advance fault detection method is
motor current signature analysis [5] or by the detection of presented.
acoustic emission [6]. All these techniques are used to detect
As in gearbox the vibration signals produced by bearing
bearing roughness or crack formation. Acoustic emission
are very weak even if the bearing has faults hence these
sensors detect the crack extremely early even when the
vibrations are suppressed by high vibration of gear and noise
remaining useful life of the bearing is long whereas using
signal. To make data valuable certain pre-processing
vibration monitoring device, cracks are detected when the
techniques are implemented.
remaining useful life is minimal [7]. Hence the technique of
vibration analysis is preferred.

978-1- 6654-6209-9/22/$31.00 © 2022

Authorized licensed use limited to: Rajamangala Univ of978-1- 6654-6209-9/22/$31.00
Technology ©Downloaded
Srivijaya provided by UniNet. 2022 IEEEon August 31,2023 at 06:22:07 UTC from IEEE Xplore. Restrictions apply.
 2022 International Conference on Electrical Engineering and Sustainable Technologies (ICEEST)

A. Morlet Wavelet 𝐸(𝑋 4 (𝑡))

𝐾𝑢𝑟𝑡𝑜𝑠𝑖𝑠 = − 3 
The pre-processing technique can be performed using (𝐸(𝑋 2 (𝑡)))2
numerous techniques but wavelet transform is preferred as it
gives information about the signal both in frequency and time where 𝐸(𝑋 4 (𝑡)) represents the fourth moment and 𝐸(𝑋 2 (𝑡))
domains. Wavelet transform is the representation of signal in represents variance or the second moment. The subtraction of
terms of wavelets. The mother wavelet is then scaled and 3 in (4) is to avoid excess kurtosis respective to normal
translated to provide daughter wavelets. distribution [14].
The mother wavelet and it's corresponding Wavelet Spectral Kurtosis computes the kurtosis for short-term
transform, represented as "φ" and "C(a,b)" respectively are window in order to detect the frequency band where hidden
calculated as. non-stationarities are present. The techniques to compute
Spectral Kurtosis are Short Term Fourier Transform (STFT)
𝑡−𝑏 and Kurtogram.
𝜑𝑎,𝑏 (𝑡) = 𝜑 ( )  Kurtogram analysis is a frequency/frequency resolution
𝑎
(f⁄Δf dyad) that maximizes its Kurtosis. A set of low-pass and
1 ∞
𝑡−𝑏 high-pass filters are used to perform the elementary
𝐶(𝑎, 𝑏) = ∫ 𝑥(𝑡)𝜑′ ( ) 𝑑𝑡  decomposition which is iterated in a pyramidal manner to
√𝑎 −∞ 𝑎
produce the tree of filter-banks as shown in Fig 1. Kurtosis for
each signal is calculated the highest value represents
where, "t" represents the time and "φ'” represents the Fourier appropriate frequency with bandwidth where the fault is
transform of base wavelet. The mother wavelet depends on present.
two characteristic parameters namely, the scaling factor "a"
which represents to high frequency component and the
location or space "b" which controls the position of the
wavelet in time [12].
Wavelet transform lets the daughter wavelet to acts as
filter kernel which can be used to smoothen the signal. This Fig 1: Kurtogram as frequency/frequency resolution (𝒇⁄∆𝒇 dyad)
smoothening property of the signal can be used for denoising
the signal. The Morlet Wavelet is defined as. C. Envelope Detection
Due to wear and tear of the bearing any surface defect will
generate vibrations that would be amplitude modulated due to
𝛽 2 (𝑡 − 𝑏)2 𝜋(𝑡 − 𝑏) periodic changes in the forces. To demodulate the signal
𝜑𝑎,𝑏 (𝑡) = exp (− ) cos ( ) 
𝑎 2 𝑎 envelope detection techniques are performed. The power
spectrum of the envelope signal generates peaks that represent
where a and b control the scale and shifting of the wavelet, the frequency generated due to fault.
respectively. The denoising factor would be controlled by Advance signal processing techniques are applied to
adjusting the value of a and b till the number of oscillations extract envelope by using Hilbert transform, Shock filtering
becomes greater than 5 [12]. and Gaussian filtering. Amongst them Hilbert transform
Gao, Yu and Wang [13] proposed a technique whereby method detects the best envelope of the signal [15] following
using Morlet Wavelet technique, nonsynchronous noises were the peaks of the input signal.
filtered from the vibrational signal collected from broken Hilbert transform is the convolution of time domain input
tooth of gearbox. By tweaking the procedure, it is possible to signal with 1/πt and can be written as
filter the gear and shaft vibration signal out and obtaining the
bearing faulty vibration. If the nonsynchronous de-noised 1 ∞ 𝑥(𝑡) 1
signal was subtracted from the original signal it would give us 𝐻[𝑥(𝑡)] = ∫ 𝑑𝑥 = ∗ 𝑥(𝑡) 
𝜋 −∞ 𝑡 − 𝜏 𝜋𝑡
vibration signal with only bearing fault as the rest of the shaft
In frequency domain Hilbert Transform can be used to
and gear noises would be subtracted.
perform 90o phase shift. Positive frequency component
Once the data is pre-processed then Spectral Kurtosis is generates a negative phase shift whereas negative frequency
implemented to detect the faulty frequency band. component generates a positive phase shift.
B. Kurtogram The extraction of the function can be performed by adding
For fault detection, the most important part is to detect the a phase shifted signal with original signal which would create
appropriate demodulation band. For that, one of the earliest a time domain analytical signal written as
techniques is the spectral comparison between healthy and
faulty bearing. As this technique needs the information of 𝑧(𝑡) = 𝑥(𝑡) + 𝑖𝑥𝐻 (𝑡) = 𝐴(𝑡)𝑒 𝑖𝜃  
healthy bearing which is not advisable hence other techniques
known as Spectral Kurtosis was developed, with which Using the analytical signal, the envelope and the phase
Kurtosis for complete spectral are calculated. angle can be extracted respectively as

Kurtosis is the fourth order of moment. It is a

dimensionless quantity that defines the peaked-ness of a 𝐴(𝑡) = √𝑥(𝑡)2 + 𝑥𝐻 (𝑡)2  
distribution relative to a normal distribution. Kurtosis of a
function is defined as

Authorized licensed use limited to: Rajamangala Univ of Technology Srivijaya provided by UniNet. Downloaded on August 31,2023 at 06:22:07 UTC from IEEE Xplore. Restrictions apply.
 2022 International Conference on Electrical Engineering and Sustainable Technologies (ICEEST)

𝑥𝐻 (𝑡) Start
𝜃 = tan−1  
𝑥(𝑡)
Vibration Signal to be analyzed
D. Bearing Fault Frequency
When the rolling element passes any part that has defect
an impact shock is produced which excites high-frequency Perform Morlet Wavelet
resonances of the whole structure. The value of the faulty (MW) on the signal
frequencies depends on the rotational speed of the shaft and Pre-processing
on the geometrical parameters of the bearing. The geometric
parameters of the bearing are ball diameter, the pitch diameter,
the contact angle of the balls on the race and the number of -
balls Fig 2 shows the geometric parameters of the ball bearing
that are involved to generate the fault frequencies.
Compute Fast Kurtogram
using 1/3-binary tree
structure

Select the Bandwidth

with maximum Kurtosis

Demodulate the signal using Hilbert

transform to obtain the envelope

Calculate power spectral density of the

Fig 2: Structure and fault frequency in bearing envelope signal and evaluate whether it
causes fault
The major faulty frequencies a ball bearing can generate
are named as fundamental train frequency (FTF), also known
as cage pass frequency, ball pass frequency of inner race Detect the type of fault
(BPFI), ball pass frequency of outer race (BPFO), and ball using geometric parameters
spin frequency (BSF) [16].
The vibration frequencies generated are mathematically End
evaluated using the following equations:
Fig 3: Flowchart of the proposed technique

The outer race bearing fault frequency:

𝑁𝑏 × 𝑆𝐹 𝐵𝑑
𝑓𝑜𝑢𝑡 = (1 − cos 𝛽)
2 𝑃𝑑  III. EXPERIMENTAL SETUP
The absolute vibration of the bearing was calculated by
The inner race bearing fault frequency:
mounting an accelerometer ADXL377 in perpendicular axis
𝑁𝑏 × 𝑆𝐹 𝐵𝑑
𝑓𝑖𝑛 = (1 + cos 𝛽) to the mounting surface. The test rig for the experiment was
2 𝑃𝑑  assembled as shown in Fig 4.
The fundamental train frequency:
𝑆𝐹 𝐵𝑑
𝑓𝐹𝑇𝐹 = (1 − cos 𝛽)
2 𝑃𝑑 

The ball defect frequency:

2
𝑆𝐹 𝐵𝑑 Bearing Accelerometer
𝑓𝑏 = (1 − ( cos 𝛽) )  Motor
2 𝑃𝑑 Fig 4: Test Rig Installation

where 𝐵𝑑 is the ball diameter, 𝑃𝑑 the pitch diameter, 𝛽 the The dataset was collected from a test rig which was
contact angle of the balls on the race, 𝑁𝑏 the number of balls equipped with NTN UCP 202 bearing with the value of the
and 𝑆𝐹 the shaft frequency. parameters listed in Table 1. The shaft rotation frequency used
for experiments was 1500 RPM.
The step by step procedure to detect faults in a noisy Table 1: Geometric Parameter for bearing
environment is summarized in the flowchart shown in Fig 3. Geometric Parameter
Ball diameter 𝐵𝑑 0.3125
Pitch diameter 𝑃𝑑 1.319
Number of balls 𝑁𝑏 8
Contact angle 𝛽 0

Authorized licensed use limited to: Rajamangala Univ of Technology Srivijaya provided by UniNet. Downloaded on August 31,2023 at 06:22:07 UTC from IEEE Xplore. Restrictions apply.
 2022 International Conference on Electrical Engineering and Sustainable Technologies (ICEEST)

Two different bearings were used each with fault only on IV. ANALYSIS AND RESULTS
one side of race either inner race or outer race. To validate our Once the faulty vibration signals are generated, the
experiment a healthy bearing was also used. Fig 5 show the procedure, described on flowchart would be followed but
faults on bearings. A low-level fault was visible at inner race prior to that faulty frequencies known as Bearing
which was generated as the surface on inner race was worn Characteristic Frequencies (BCF) should be calculated. These
out where as faults on outer race were clearly visible due to frequencies are calculated by using the parameters defined in
crack. Table 1 and placing them in Equations (9) to (12). These
defective frequencies of the bearing are as shown in Table 2.
(a) (b)

As the shaft was rotating at 1500RPM hence the

Inner race
fault frequency generated due to shaft would be calculated as

1500𝑅𝑃𝑀⁄ = 25𝐻𝑧
60
Table 2: Bearing Characteristic Frequencies
Outer race Types of defects Frequency
(c) fault Shaft frequency SF 25 Hz
Ball Pass Frequency on the 81.12 Hz
outer race 𝑓𝑜𝑢𝑡
Ball Pass Frequency on the 118.81 Hz
inner race 𝑓𝑖𝑛
Cage pass frequency 𝑓𝐹𝑇𝐹 10.14 Hz
Ball Spin rotation frequency 𝑓𝑏 127.73 Hz

Once the BCF were calculated then the vibration signal

Fig 5: Sideview of (a) Healthy bearing (b) Worn out inner race (c) Worn
was de-noised using Morlet Wavelet and to calculate the
out outer race
nonsynchronous noise the de-noised signal was subtracting
from the original signal.
Experimental data for bearing was collected at a sampling
rate of 18,828 samples per second for a duration of 3s. The The appropriate frequency band where faults were present
load at the shaft was maintained at 30lbs. The vibrational data was calculated using Fast Kurtogram. The frequency range
for different bearings are shown in Fig 6. where Kurtosis is evaluated is large hence a 4th level
Kurtogram is calculated The Kurtosis calculated was plotted
(a) 8 in a tabular manner as shown in fig 8 to get the appropriate
(a) 6

4
Kurtogram. The Kurtogram is represented using grey scale
2 hence the highest value for Kurtosis is represented using black
0
colour.
-2
(a)
-4

-6

-8
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3
Time (s)
(a) 40

30
(b) 20

10

-10 (b)
-20

-30

-40

-50
0 0.25 0.5 0.75 1 1.25 1.5 1.75 2 2.25 2.5 2.75 3
Time (s)

Fig 6: Vibration data for bearing when the defect was generated due to
contact of ball with (a)damaged outer race (b)damaged inner race

Usually, noise in the data represents imbalance and Fig 8: Kurtogram representation for bearing with a defect located on (a)
mechanical looseness in bearing or improper installation of outer race (b) inner race
accelerometer. To test the Advance fault detection system a From the Kurtogram we can notice that maximum value
medium noise is generated and added to the vibration signal. of Kurtosis for the outer race fault is at 4th level node 12 and
The noise signal generated is a white Gaussian noise with inner race fault is at level 1 node 2. The faulty frequency
mean 0 and standard deviation of 0.8 as shown in fig 7. ranges for each type of bearing fault was calculated as shown
in Table 3
Waveform Graph
4

1
Table 3: Cut-off frequencies for bandpass filter.
0

-1

-2
Lower Upper
-3

-4
frequency (Hz) frequency (Hz)
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3
Time (s)
Outer race fault 13375 14500
Fig 7: Noise signal
Inner race fault 10000 19000

Authorized licensed use limited to: Rajamangala Univ of Technology Srivijaya provided by UniNet. Downloaded on August 31,2023 at 06:22:07 UTC from IEEE Xplore. Restrictions apply.
 2022 International Conference on Electrical Engineering and Sustainable Technologies (ICEEST)
Waveform Graph 2

Once the faulty frequency range was detected then the 0.09
0.08
filtered signal was demodulated using envelope detection.
0.06
Hilbert transform was used to demodulate the signal. After the
transformation, the Power Spectral Density (PSD) of the data 0.04

was calculated to get the faulty frequency. 0.02

By calculating PSD the vibrations in time domain were 0

10 25 50 75 100 125 150 175 200 225 250 275 300 325 350 375 400 425 450 475 500
converted into frequency domain. The PSD calculated for Frequency )Hz)

bearings are shown in fig 9 The point where the main peak is Fig 10: Power Spectral density without filtering noisy data (a) Outer
race fault (b) Inner race fault
achieved represents the frequency where maximum power is
dissipated. As the proposed technique can easily identify the types of
From fig 9(a), it is visible that for health bearing the fault for both inner and outer race with noisy vibrations hence
maximum value of power dissipated is extremely small which the proposed technique is better.
proves the system as healthy. Whereas for figure 9(b) the V. CONCLUSION
highest peak marked as fout is at 81Hz which represent as outer
race faulty bearing. The multiples of fout represent higher This paper presents an improved fault detection technique
harmonics. Similarly, for fig 9(c) the highest peak marked as where bearing faults for noisy vibration signal were detected.
fin is at 118Hz which represents inner race faulty bearing. The In the proposed technique a new step was added where an
multiples of f𝑖𝑛 represent higher harmonics and the SB additional filter Morlet wavelet was used to filter
represents side bands which are visible due to amplitude synchronous noise. The vibration signals collected from two
modulating effect kinds of rolling element bearings are used with additional
added noise to test the improved performance of the proposed
(a) (a)
0.0004

0.00035

0.0003
method. The results obtained validate the effectiveness and
0.00025 preciseness of the proposed method.
0.0002

REFERENCES
0.00015

0.0001

5E-5

0
[1] Bouchikhi, V. Choqueuse and M. Benbouzid, "Current Frequency
10 20 40 60 80 100 120 140 160 180 200 220 240 260 280
frequency (Hz)
300 320 340 360 380 400 420 440 460 480 500
Spectral Subtraction and Its Contribution to Induction Machines’
(a) (b)
0.018 Bearings Condition Monitoring", IEEE Trans. Energy Conversion, vol.
0.016
0.014 fout 28, no. 1, pp. 135-144, 2013.
0.012
0.01
[2] W. Mao, L. Ding, Y. Liu, S. Afshari and X. Liang, “A new deep domain
0.008
2×fout
adaptation method with joint adversarial training for online detection
0.006
3×fout of bearing early fault”. ISA Transactions, vol 122, pp.444-458, 2022.
0.004
0.002 [3] D. Kumar, S.Mehran, M. Shaikh, M. Hussain, B. Chowdhry and T.
0
10 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360 380 400 420 440 460 480 500 Hussain, “Triaxial bearing vibration dataset of induction motor under
frequency (Hz)
varying load conditions”. Data in Brief, vol. 42, p.108315, 2022.
(a) (c) 0.45
0.5
fin
0.4 [4] W. Zhao, J. Liu, W. Zhao and Y. Zheng, “An investigation on vibration
0.35
0.3 SF
features of a gear bearing system involved pitting faults considering
0.25
2×fin
effect of eccentricity and friction.” Engineering Failure Analysis, vol
0.2
0.15
SB 131, pp.105837, 2022.
SB 3×fin
0.1
0.05
[5] L.E. Rikam, L. Bitjoka, and A. Nketsa, “Quaternion Fourier Transform
0
10 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 360 380 400 420 440 460 480 500
spectral analysis of electrical currents for bearing faults detection and
frequency (Hz) diagnosis”, Mechanical Systems and Signal Processing, vol. 168,
Fig 9: Power Spectral Density (PSD) by proposed technique for (a) pp.108656, 2022.
Healthy Bearing (b) Outer Race faulty bearing (c) Inner Race faulty bearing
[6] D. Hou, H. Qi, C. Wang and D. Han, “High-speed train wheel set
bearing fault diagnosis and prognostics: Fingerprint feature recognition
To compare the advantage of Morlet Wavelet the same method based on acoustic emission”. Mechanical Systems and Signal
Processing, vol 171, pp.108947, 2022.
technique is performed for noisy data without denoising the
[7] N. Nirwan and H. Ramani, “Condition monitoring and fault detection
signal. The results for faulty bearing data are as shown in fig in roller bearing used in rolling mill by acoustic emission and vibration
10. analysis”. Materials Today: Proceedings, vol 51, pp.344-354, 2022.
Fig 10(a) shows the results for outer race faults. As we can [8] M. Motahari-Nezhad and S. Jafar, “Bearing remaining useful life
see that the vibration signals are fused in the noise signal and prediction under starved lubricating condition using time domain
the type of fault can’t be identified. For fig 10(b) due to the acoustic emission signal processing”. Expert Systems with
Applications, vol 168, pp.114391, 2021.
rotating property of inner race there are perodic change in
vibrations hence the vibrational data is amplitude modulated. [9] Q. Zhang, J. Ding and W. Zhao, “An adaptive boundary determination
method for empirical wavelet transform and its application in wheelset
As the signal is demodulated using Hilbert transform hence bearing fault detection in high speed trains”. Measurement, vol 171,
the noise signal act less dominant and the type of fault is still pp.108746, 2021.
detectable. [10] S. He, D. Hu, G. Yu, A. Liao and W. Shi, “Trackside acoustic detection
Waveform Graph 2
0.02
of axle bearing fault using wavelet domain moving beamforming
0.0175 method”. Applied Acoustics, vol. 195, 2022.
0.015
[11] D. Hou, H. Qi, H. Luo, C. Wang and J. Yang, “Comparative study on
0.0125
0.01
the use of acoustic emission and vibration analyses for the bearing fault
0.0075
diagnosis of high-speed trains”. Structural Health Monitoring, vol
0.005 21(4), pp.1518-1540, 2021.
0.0025 [12] L. Yang, R. Duan, T. Kang, J. Li and Y. Liao, “Iterative Morlet wavelet
0
10 25 50 75 100 125 150 175 200 225 250 275 300 325 350 375 400 425 450 475 500 with SOSO boosting strategy for impulsive feature extraction”.
Frequency )Hz) Measurement, vol 193, pp.110965, 2022.

Authorized licensed use limited to: Rajamangala Univ of Technology Srivijaya provided by UniNet. Downloaded on August 31,2023 at 06:22:07 UTC from IEEE Xplore. Restrictions apply.
 2022 International Conference on Electrical Engineering and Sustainable Technologies (ICEEST)

[13] M. Gao, G. Yu and T. Wang, “Impulsive Gear Fault Diagnosis Using [15] B. Betea, P. Dobra, M. Gherman and L. Tomesc, "Comparison between
Adaptive Morlet Wavelet Filter Based on Alpha-Stable Distribution envelope detection methods for bearing defects diagnose", IFAC
and Kurtogram”. IEEE Access, vol 7, pp.72283-72296, 2019. Proceedings vol. 46, no. 6, pp. 137-142, 2013.
[14] C. Yi, Y. Li, X. Huo and K. Tsui “A promising new tool for fault [16] M. Tahir, A. Khan, N. Iqbal, A. Hussain and S. Badshah, "Enhancing
diagnosis of railway wheelset bearings: SSO-based Kurtogram”. ISA Fault Classification Accuracy of Ball Bearing Using Central Tendency
Transactions, 2021 Based Time Domain Features", IEEE Access, vol. 5, pp. 72-83, 2017.

Authorized licensed use limited to: Rajamangala Univ of Technology Srivijaya provided by UniNet. Downloaded on August 31,2023 at 06:22:07 UTC from IEEE Xplore. Restrictions apply.