Roller bearing faults classification using Artificial Neural Network
based on Servo system with Observer design
THANASAK WANGLOMKLANG, THANYABOON TUNTAVESESAK,
WINAI TUMTHONG, JIRAPHON SRISERTPOL
School of Mechanical Engineering, Suranaree University of Technology
111 University Avenue Muang, Nakhon Ratchasima 30000
THAILAND
Abstract: - The roller bearing is the main component of rotating machines, which is used to reduce friction while
the machine operation. The bearing faults are the key problem of the rotating machine because they affect the
unusual operation and caused machine downtime. This paper presented the fault detection approach based on an
Artificial Neural Network (ANN) to recognize the bearing conditions. Servo systems with observers were
designed for motor control and estimating current signal. The bearing conditions demonstrated in three cases
consisted of normal, no lubricant, and outer race defect. For ANN model training, four statistical parameters
including mean, crest factor, kurtosis, and root-mean-square (RMS) were selected to identify the causal
characteristics of the motor current from observer and observation error data. the result indicated that the fault
detection model has been displayed a classification accuracy of 94.4% which appropriate using in real operations.
Key-Words: - roller bearing, fault classification, observer design, artificial neural network, DC motor control.
Received: August 19, 2021. Revised: October 18, 2022. Accepted: November 14, 2022. Published: December 8, 2022.
1 Introduction
Nowadays, many industries utilized the rotating
machine as a power drive for several mechanisms and
shafts. The rotating machine is required a control
algorithm, for instance, a PID controller to regulate a
speed at the desired operating point and other
components to support their motion. The roller
bearing is the main part used for supporting shaft
rotation. In the real process, the machine was
operated continuously under load situations. This
condition caused the bearing element tends to failure
effect control system performance and reducing the
machine life cycle.
The roller bearing faults have various
pursuits to detect and classify. Cui Lingli, [1],
presented the fault diagnosis based on adaptive
machining. The vibration signal was used as
analytical data and the result affected to stability and
controllability of the model. The statistical
parameters consisting of RMS, crest factor, and peak
value were studied to analyze the bearing defects
scenario, [2]. P.D. McFadden, [3], has researched
vibration monitoring in a high-frequency range to
explain the envelope of signal utilized for the rolling
element of bearing. To accurately control motor
position and speed, the adaptive load torque
compensation algorithm was applied. The state
variables were estimated using the observer-based
Kaman filter technique which suggests a reasonable
result for the dynamics response, [4]. Debabrata Pal
proposed a full-state observer with state feedback
control that is simulated on motor speed control. To
obtain the appropriate parameter, MATLAB software
was implemented for designing the state feedback
and observer gain. The experimental validation
discussed in performance response and the noise
signal influent the control system, [5]. In many works
in fault detection and diagnosis, modern control
design shame is a powerful method that combines an
intelligence approach to produce fault-tolerant
control which is used in high-speed machines, [6].
The spindle bearing fault diagnosis was investigated
in multiple conditions using energy-fluctuate with
machine learning technique that is provided by
Xiaoxi Ding, [7]. Several signal processing has been
developed for failure analysis due to handling data
characteristics. Toumi Yassine, [8] worked on
predictive maintenance for bearing fault analysis.
Hilbert transform and the FFT technique were used
for feature extraction that will be transferred to create
an artificial neural network model. Pratik Phalle, [9],
presented fault identification using condition
monitoring which contributed to the bearing
diagnosis. The inter race defect was fabricated as a
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2022.21.26
Thanasak Wanglomklang, Thanyaboon Tuntavesesak,
Winai Tumthong, Jiraphon Srisertpol
E-ISSN: 2224-2678
241
Volume 21, 2022
faulty situation by using an electrical discharge
machine (EDM). The classical control as PID
controller cannot carry out the external disturbance
and feedback measuring with uncertainty suggested
by Jing Sun, [10].
This research purposed the development of
bearing fault detection and classification approaches
by using an artificial neural network. The PI servo
system with state observer design is selected to
control the motor angular velocity and it’s used for
estimating the motor current during operation. Three
conditions of roller bearing which consist of normal,
no lubricate, and outer race defect were investigated
to demonstrate the different fault situations. The state
observer variables were used as raw data for
extracting into feature data which include mean,
kurtosis, crest factor, and RMS. These parameters
will be used to train an ANN model for the specific
type of bearing fault.
2 Dc motor with rigid shaft modeling
There is a wide application of DC motors for driving
a rotating machine, one of which coupling with the
rotation shaft. The mathematical modeling of a DC
motor with a shaft is important because of its use for
system response analysis and controller design. This
work discusses on modeling of a motor with a rigid
shaft. Two roller bearings are supported on both sides
of the shaft and it’s also coupled with a loaded disc.
The physical diagram illustrates the position of the
DC motor and mechanical part in figure 1.
Fig.1: Physical Model of DC motor with Shaft
The model was considered through electrical and
mechanical relations. The electrical part can be
derived from basics series electrical circuit which
follows as:
(1)
The model parameter is armature resistance,
is armature inductance is back emf constant, and
is motor angular velocity in rad/s. the mechanical
shaft is connected to the load disk which can explain
by newton’s second law which refers to (2).
(2)
Where is motor inertia, is load inertia, is
motor viscous damping, and the motor torque and
load torque are shown in (3) and (4) respectively.
(3)
(4)
The shaft speed illustrates by , however, we
assume that the motor and shaft speed are equal that
can arrange into (5) as follow
(5)
Where and
From (1) and (5) the state space model is represented
as
(6)
system identification is a common technique based
on measuring system input and output data to
approximate the system parameters. It was used for
estimating the motor with shaft parameters as shown
in Table 1.
Table 1. Motor with Shaft parameter
3 State feedback with observer design
architecture
The purpose of the servo system as shown in figure 2
is one of the modern control arts which is a design
using the pole placement method. This control
system needs to be done by feedback state variable to
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2022.21.26
Thanasak Wanglomklang, Thanyaboon Tuntavesesak,
Winai Tumthong, Jiraphon Srisertpol
E-ISSN: 2224-2678
242
Volume 21, 2022
compute optimal input but normally some variables
are difficulted to measure directly therefore, the
observer is utilized for its estimation. The functional
Fig.2: Servo system with observer architecture
block diagram exposed a vital parameter which
relevant in each state. The actual speed is measured
for the feedback loop and compared with their
estimation from the observer. The design of response
specification is defined based on the second order
overdamped response with suppose 8 sec of settling
time. Therefore, the closed-loop pole location is
placed at -0.5, -20, and -50. In addition, the observer
pole must be faster than 10 times of close-loop pole
in practice. The result of designing is feedback gain
K is [2.95, 0.46], observer gain L is [45.1, 150.56]
and Ki = 0.36 respectively. The state observer
estimated the current and shaft rotational speed
through the matrix L which can reduce the error
between actual and estimate output. To validate the
control system performance, reference input tracking
was implemented. The set point is set as a step test by
changing the profile from 500 to 600 RPM as
indicated in figure 3.
Fig.3: Reference input tracking test and state variables
From the response, the graph is clear that the
control system can control the motor speed nearly by
the desired operating point with properly transient
and steady-state behavior. Moreover, the observer
presented a good estimation of output speed and
provided the motor current from its computation.
3 The purpose of feature extraction
for time domain data.
Signal processing is the standard approach for
handling measurement data. there is various method
which is applied in the time and frequency domain
that depends on signal characteristics. Signal noise is
a usual term that comes with raw data, and it needs to
be curtailed. This paper employed both low-pass and
band-pass filters to reduce the influent of an
unexpected signal. After the data pre-processing, the
statistical features including mean, crest factor,
kurtosis, and root mean square (RMS) was declared
to extract the crucial manner of data. The significant
feature that will be implemented to fabricate the
artificial neural network model for classifying the
bearing fault conditions. the mathematical formulas
are following (7) to (10).
Mean (7)
Crest factor (8)
Kurtosis (9)
Where is the standard deviation.
Root Mean Square (10)
4 Experimental setups
The functional diagram of testing and data collection
is shown in figure 4. The procedure started with
defining of reference speed input profile. The bearing
conditions are prepared for three different situations.
In each state will be recorded the estimation current
and observation error which contribute to the
MATLAB environment. Finally, feature extraction is
a minor step for generating learning data set that is
deployed to ANN model creation.
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2022.21.26
Thanasak Wanglomklang, Thanyaboon Tuntavesesak,
Winai Tumthong, Jiraphon Srisertpol
E-ISSN: 2224-2678
243
Volume 21, 2022
Fig.4: Experimental scheme.
4.1 Bearing fault conditions
In certain situations, the rotating machine has several
problems that occur with roller bearings. Whereas
there are some faults which marginal arise before the
huge damage. Thus, two fault conditions of bearing
were investigated. The starved lubricant condition
was considered as one of all faults which are
exploded to clean the grease to 0% level. Another
fault is an inside surface outer race defect It was
created by a computer numerical control (CNC)
machine. This status explained the bearing that
initiated a deep groove caused by fatigue that is run
for a long time. Figure 5 illustrated the difference in
bearing conditions which use for testing.
Fig.5: The bearing condition (a) Normal (b) No
lubricant and (C) Outer race defect
4.2 Data collection
The estimated motor current and observation error
from the control system was considered the
fundamental data for the feature extraction process.
The MATLAB/Simulink software is applied for data
collection and interface with the ARM Cortex-a72
microcontroller to run as hardware-in-loop testing.
The configuration consists of setting up the desire
speed accounting for 500 to 600 RPM. Figure 6 is
showed the comparison of current and error for each
bearing condition. The data were recorded in 30
samples per condition.
Fig.6: Fundamental data of each bearing condition.
It was found that the current between normal and
other faults is distinguishable, on the other hand, two
types of faults are quite tough to separate. As a result,
the artificial neural network (ANN) will be an
appropriate approach to classify these fault
symptoms.
4.3 Feature data and ANN model training
The maps as shown in figure 7 and figure 8 are
represented by essential feature current and error
data. it contains 30 data points for each condition
thus, there are 8 different inputs to supervise the
model. From the graph, it can be classified as normal
and other cases of fault. On the other hand, all
features are difficult to distinguish between no
lubricant and outer race defect condition.
Fig.7: Feature map of current
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2022.21.26
Thanasak Wanglomklang, Thanyaboon Tuntavesesak,
Winai Tumthong, Jiraphon Srisertpol
E-ISSN: 2224-2678
244
Volume 21, 2022
Fig.8: Feature map of error
Consequently, the ANN will be used to learn the
featured trait that is ambiguous in each fault. The
structure of the ANN model has multiple parameters
that affect the prediction accuracy such as the amount
hidden layer, activating function, learning algorithm,
and the quality of the input data set. The observation
of some neural in hidden layer change is presented to
examine the highest model accuracy. In this studying,
we have compared five different neural representing
5, 10, 15, 20, and 25 layers. The training data set was
split to training 70% for supervising and the network
is adjusted according to its error. the validation
process is used 15% of data for measuring network
generalization, and to halt training when
generalization stops improving. The final step is
model testing, it utilized the data by 15% that is for
measure the model classification after training. In
pattern recognition, the neural network is used for
classifying inputs into a set of target categories it’s
which means three bearing conditions. The network
will be trained with scaled conjugate gradient
backpropagation and set the SoftMax function as
activating function. The configuration of the model
structure for 10 neural in the hidden layer is
illustrated in figure 9.
Fig.9: The example configuration of the ANN model
The result is revealed in table 2. It is clear from Table
2 that increasing the hidden layer number implied the
optimal value representing 15 numbers which
generates the model accuracy of 94.4%.
Table 2: Model accuracy with various neural number
Many loss functions are used for the optimization of
model weight in the training process. This work has
used the Cross-Entropy loss which is one of the
powerful functions for model validation. The result
can discuss the performance curve as shown in figure
10. All curves are demonstrated the optimal point to
stop learning the highest accuracy model at epoch 30.
Minimizing cross-entropy is given better results as
can be seen from the result of the graph.
Fig.10: Cross-Entropy validation curve
4.4 Fault classification result
The confusion matrix is commonly used to interpret
classification accuracy. As shown in figure 11
combined the detail of the neural network model for
classifying the bearing conditions. the numbers 1, 2,
and 3 are represented the normal, no lubricant, and
outer race defect classes respectively. Training,
validation, and test matrix appeared a correction of
more than 90%, and all matrices denoted with 100%
for detecting between normal and no lubricant class
and overall indicated 94.4% of accuracy.
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2022.21.26
Thanasak Wanglomklang, Thanyaboon Tuntavesesak,
Winai Tumthong, Jiraphon Srisertpol
E-ISSN: 2224-2678
245
Volume 21, 2022
Fig.11: Confusion matrix of fault classify model
5 Conclusion
This article proposes a classification of roller bearing
fault based on an artificial neural network (ANN) and
servo system with stat observer design. the estimated
motor current and observation error are utilized as
preprocessing data. The parameters consisting of
mean, kurtosis, crest factor, and root mean square are
used to extract the crucial feature from current and
error which is used to supervise the neural network
model. Three conditions of bearing include normal,
no lubricant, and outer race defect are investigated as
the target of the model. The number of neural in the
hidden layer was changed to examine a suitable
model in the training process. using state variable
data, signal processing, and ANN algorithm can be
applied to fabricate the bearing fault detection model.
The result revealed that the model has an accuracy of
94.4% for classifying demonstration roller bearing
faults.
Acknowledgment:
The authors would like to thank Suranaree University
of Technology for laboratory testing and financial
support for this research project.
References:
[1] Cui, Lingli, Jing Wang, and Seungchul Lee.
Matching pursuit of an adaptive impulse
dictionary for bearing fault diagnosis. Journal of
Sound and Vibration 333.10, 2014, 2840-2862.
[2] Tandon, N. A comparison of some vibration
parameters for the condition monitoring of
rolling element bearings. Measurement 12.3,
1994, 285-289.
[3] McFadden, P. D., and J. D. Smith. Vibration
monitoring of rolling element bearings by the
high-frequency resonance technique a review.
Tribology international 17.1, 1984, 3-10.
[4] Khajorntraidet, C., J. Srisertpol, and C.
Prattanarak. Alternative technique for DC servo
motor control using adaptive load torque
compensator. Advanced science letters 19.8,
2013, 2139-2143.
[5] Talole, Sanjay E., Jayawant P. Kolhe, and
Srivijay B. Phadke. Extended-state-observer-
based control of flexible-joint system with
experimental validation. IEEE Transactions on
Industrial Electronics 57.4, 2009, 1411-1419.
[6] Chommuangpuck, Prathan, et al. Fault Tolerant
Control Based on an Observer on PI Servo
Design for a High-Speed Automation Machine.
Machines 8.2, 2020, 22.
[7] Ding, Xiaoxi, and Qingbo He. Energy-
fluctuated multiscale feature learning with deep
convnet for intelligent spindle bearing fault
diagnosis. IEEE Transactions on
Instrumentation and Measurement 66.8, 2017,
1926-1935.
[8] Yassine, Toumi, Lachenani Sidahmed, and Ould
Zmirli Mohamed. Bearing fault classification
based on envelope analysis and artificial neural
network. 2019 International Conference on
Advanced Electrical Engineering (ICAEE).
IEEE, 2019.
[9] Phalle, Pratik, and Sangram Patil. Fault
Diagnosis of Rolling Element Bearing Using
Artificial Neural Networks. 2021 4th Biennial
International Conference on Nascent
Technologies in Engineering (ICNTE). IEEE,
2021.
[10] Sun, Jing, Chunyang Wang, and Ruihao Xin. On
disturbance rejection control of servo system
based on the improved disturbance observer.
2018 37th Chinese Control Conference (CCC).
IEEE, 2018.
WSEAS TRANSACTIONS on SYSTEMS
DOI: 10.37394/23202.2022.21.26
Thanasak Wanglomklang, Thanyaboon Tuntavesesak,
Winai Tumthong, Jiraphon Srisertpol
E-ISSN: 2224-2678
246
Volume 21, 2022
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the Creative
Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en_US
