AI-Enhanced Fault Diagnosis in Rolling

Element Bearings: A Comprehensive

Prasanta Kumar Samal
Vibration Analysis Approach
Department of Mechanical Engineering
The National Institute of Engineering This research presents a comprehensive approach for bearing fault
[NIE], Mysuru Karnataka diagnosis using artificial intelligence (AI), particularly through the
India application of artificial neural networks (ANNs). By integrating these
networks into vibration analysis, the approach aims to meet the critical
Sunil K. need for prompt fault detection. The methodology comprises three key
Department of Mechanical Engineering steps: vibration signal acquisition, feature extraction, and fault
The National Institute of Engineering
[NIE], Mysuru Karnataka classification. Experiments were conducted to acquire vibration signals for
India the test bearings on a machinery fault simulator. Six time-domain features
were extracted using MATLAB, creating a comprehensive dataset for
Imran M Jamadar training the ANN models with three algorithms: Levenberg-Marquardt
Department of Mechanical Engineering backpropagation (LMBP), scaled conjugate gradient backpropagation
The National Institute of Engineering
[NIE], Mysuru Karnataka (SCGBP), and Bayesian regularization backpropagation (BRBP). The
India BRBP algorithm achieved the highest correct classification rate (97.2%),
followed by LMBP (90%) and SCGBP (83.6%). To evaluate their efficacy
Srinidhi R. in bearing fault classification, these three networks were simulated,
Department of Mechanical Engineering
JSS Science and Technology
revealing that BRBP could predict all four classes of bearings with zero
University,Mysuru Karnataka errors.
India
Keywords: Artificial Intelligence,Artificial Neural Network, Fault
Diagnosis, Fault Classification, Rolling Element Bearing, Vibration Signal
Acquisition, Time-Domain Features.

1. INTRODUCTION the frequency and time domains as well as acoustic

measures involving sound pressure, sound intensity, and
Rolling element bearings (REBs) find extensive acoustic emission techniques; it also explored high-
application in rotating machinery and are regarded as frequency resonance techniques.
vital components. Their reliable operation holds utmost Vibration is the most commonly used condition
significance in diverse sectors, including nuclear power monitoring of rotary machines. This technique is
plants, chemical industries, aviation industries, and extensively used in applications like material handling,
various process industries. Timely fault detection in aerospace, and power generation [11]. Comprehending
rotating machinery can mitigate the potential for the origins of vibration is crucial for grasping vibration
damage and, consequently, reduce the need for costly patterns to enable effective fault detection. Even for
emergency repairs. During regular operations, both healthy bearings, there will still be vibration, termed
mechanical and electrical systems generate distinctive variable compliance, which is considered normal [7].
signals. Any alteration in a machine's operating Different vibration analysis methods are available.
conditions will result in deviations from this Howard [8] uses accelerometers for analog signal
characteristic signal. Indeed, variations in the typical acquisition. Saruhan et al. [12] use four defect states to
signal can serve as an indicator of an impending fault diagnose faults in rolling element bearings.
[1]. Machine Condition Monitoring (CM) is the process The time-domain signal records energy history and
of continuously monitoring various parameters, which is frequently used for extracting statistical insights. It
can offer some insights into the condition of the can detect defects and assess their severity using
machine during its operation, including factors like indicators like root mean square (RMS), kurtosis (KU),
vibration and temperature. Modern condition crest factor (CF), impulse factor (IF), peak value,
monitoring systems typically consist of data acquisition energy index (EI), K-factor, and shape factor, with KU
systems with sensors and are integrated with software and CF being more sensitive to larger defects [9-10, 13-
for signal analysis. 14]. The frequency domain signals can be obtained by
Numerous researchers have reviewed bearing fault converting the time-domain vibration signals using the
diagnosis using various monitoring techniques, fast Fourier transformation[15].Calculating bearing
including vibration and acoustic measurements[2-10]. characteristic frequenciesis helpful as they can indicate
Their review covered vibration measurements in both the fault's location. Bearing defects produce periodic
vibration impulses that contain energy across a broad
frequency range, exciting the resonant frequencies of
Received: December 2023, Accepted: June 2024
individual elements, known as the fundamental vibra–
Correspondence to: Prasanta Kumar Samal
tion frequencies. When the outer race remains fixed
Department of Mechanical Engineering, The National
while the inner race rotates, four distinct defect frequ–
Institute of Engineering, Mysuru, Karnataka, India.
encies become apparent: BPFO (ball pass frequency
Email: prasantaku.samal@gmail.com
doi: 10.5937/fme2403450S
© Faculty of Mechanical Engineering, Belgrade. All rights reserved FME Transactions (2024) 52, 450-460 450
outer race), BPFI (ball pass frequency inner race), BSF predictive capabilities. For training, testing, and vali–
(ball spin frequency), and FTF (fundamental train dation, the extracted dataset has been utilized. The
frequency) [16]. details of this model development and dataset utilization
Acquiring the vibration signal, signal processing, are further explained in Section 2.2.
and diagnosis of fault utilizing machine learning can be
performed in the MATLAB environment. Seokgoo et al.
[17] used MATLAB for diagnosing faults in REBs. AI
techniques are employed for pattern recognition in
machine diagnostics, with artificial neural networks
being the prominently used techniques. Among these,
the feed-forward neural network (FFNN), specially
trained with the back-propagation (BP) algorithm, is
extensively utilized for machine fault diagnosis [18-22].
The aforementioned literature provides a compre–
hensive background on the significance of the timely
detection of faults in REBs and the utilization of vib–
ration analysis to achieve this. As the demand for en–
hanced machinery reliability grows in today's AI-driven
era, combining AI with vibration analysis greatly en–
hances REB fault diagnosis. Specifically, ANNs have
found extensive applications in fault identification and
classification. Unlike other ANN models that typically
classify bearings as either faulty or healthy, the pro– Figure 1. Methodology Flowchart
posed ANN models classify various types of faulty bea–
2.1 Experimentation and Data Acquisition
rings, such as faults in the inner race, outer race, or ball,
and distinguish them from healthy ones. Furthermore, the
The machinery fault simulator (MFS) developed by
performance of these three algorithms has been in–
Tyrannus Innovative Engineering & Research Academy
vestigated, a topic rarely addressed in previous literature,
PVT LTD, shown in Fig. 2, has been used for expe–
thereby establishing the novelty of this research.
rimentation. It comprises a 3-phase 0.25 HP motor, two
Hence, this paper endeavors to integrate AI tech–
aluminum pedestals, a flexible coupling, two bearings,
niques, particularly ANN, into fault diagnosis of rolling
and a variable frequency drive (VFD). The flexible
element bearings, presenting a comprehensive, data-
coupling connects the driven shaft with the motor output
driven methodology. The proposed approach combines
shaft, compensating for any potential misalignment. The
vibration analysis with AI, including three main steps:
two aluminum pedestals house the bearings, supporting
acquisition of vibration signal, extraction of time-do–
the shaft. A functioning bearing is always positioned at
main features, and classification of bearing faults. The
the drive end (closer to the motor), while the test
subsequent quantitative evaluation highlights the prac–
bearing is installed at the non-drive end. Motor speed is
tical applicability and effectiveness of this novel app–
regulated by the VFD.A PCB Piezotronics tri-axial
roach in the realm of industrial machinery health moni–
accelerometer is mounted onto the pedestal, as illus–
toring and predictive maintenance.
trated in Figure 2, to acquire the vibration of the test
Section 2 provides a detailed methodology, inclu–
bearing. The accelerometer, through a lightweight cable,
ding experimentation and data acquisition in sub-section
is connected to a dedicated data acquisition card, the
2.1, and outlines the procedure for extracting time-
NI9234, designed for sound and vibration acquisition.
domain features and training the ANN models in sub-
This card is mounted onto the NI cDAQ 9178 chassis.
section 2.2. The results and discussion are presented in
The chassis, in turn, is connected to a computer, with
section 3, with conclusions summarized in section 4.
LabVIEW installed, via a USB cable.To acquire the
2. METHODOLOGY
analog acceleration signal in both the time domain as
well as in the frequency domain and to save the data for
The flow chart for the approach used in this research is further analysis, a LabVIEW application has been deve–
shown in Figure 1. Using a machinery fault simulator, loped. The authors have utilized the same experimental
the vibration signals of the test bearings under various setup for fan blade fault analysis [23].
conditions - such as healthy bearings and faulty bearings In this investigation,the test bearing selected is the
with defects in the outer race, inner race, and rolling SKF 6304, a widely studied rolling element bearing that
element - have been acquired. Section 2.1 details the has been investigated using a different approach in
experimental setup and data acquisition methods. previous literature [14]. The specifications of the test
Subsequently, six key time-domain features – na– bearing are outlined in Table 1. Localized faults of
mely, RMS, kurtosis, crest factor, peak-to-peak, impulse circular shapes with a diameter of 0.5 mm have been
factor, and energy - have been extracted from the artificially created by laser engraving on the balls, inner
acquired vibration data utilizing MATLAB. Section 2.2 races, and outer races.The test bearings both with and
provides a comprehensive explanation of this process. without fault have been reinstalled separately in the
An ANN model with a feedforward (FF) backpro– setup to acquire vibration data.
pagation (BP) algorithm has been developed to enhance

FME Transactions VOL. 52, No 3, 2024 ▪ 451

Figure 2. Experimental setup
Table 1. Specification of Test Bearing where the instantaneous magnitude (yi), the mean( y ),
Bearing number 6304 the sample length (N), and the standard deviation (σ).
Type deep groove The crest factor is defined as a ratio of peak value to
Inner diameter (ID) 20 mm RMS, as given in (3).
Outer diameter (OD) 52 mm
Race width (B) 15 mm max yi
i
Ball diameter (d) 9.525 mm CrestFactor  (3)
1 N
  yi 
Pitch diameter (D) 35.99 mm 2
No. balls (z) 7 N i 1
Contact angle(ϕ) 0 degree
Material chrome steel The impulse factor compares the peak height relative
to the signal's mean level, calculated as the peak value
2.2 Feature Extraction and ANN Model divided by the mean of the absolute value as given in
(4).
The time domain technique stands out as the most max yi
straightforward and simplest approach for analyzing ImpulseFactor  i
(4)
vibration signals. Many time domain features, including 1 N
crest factor, kurtosis, peak-to-peak value, RMS, and  yi
N i 1
others, can be utilized for condition monitoring [2]. The
RMS measures the overall level of a discrete signal, Peak-to-peak value is calculated by finding the
which can be computed as the square root of the average difference between the maximum and minimumvalues
of the values that are squared using (1), in the discrete signal data.Energy is computed as the
mean of the squared values of the discrete signal data.
1 N From the raw vibration signals acquired by follo–
 yi
2
RMS  (1) wing the procedure outlined in section 2.1, a MATLAB
N i 1 program has been employed to extract six essential time
domain features: RMS, crest factor, kurtosis, impulse
where "N" represents the number of discrete points, that
factor, peak-to-peak, and energy. These features serve
is the signal from each sampled point. These RMS
as the input nodes for the neural networks, with four
values are compared with established standards to
classes of bearings represented as output nodes.
evaluate the bearing’s condition.
For the development and simulation, the Neural
Kurtosis represents the fourth moment of a distri–
Networks Toolbox in MATLAB has been utilized.
bution normalized by the fourth power of the standard
Specifically, a Multi-Layer Feedforward Backpropa–
deviation. It offers a balanced measure that combines
gation (FFBP) ANN operating under supervised
lower and higher moments to assess the shape and tail
learning techniqueshas been employed. This is known
behavior of a dataset and can be computed using (2). It
for its efficacy in delivering highly precise results [24].
has been demonstrated that kurtosis is highly valuable in
There are three built-in training algorithms namely
fault diagnosis,
Levenberg-Marquardtbackpropagation (LMBP), scaled
N
 i 1  yi  y 
4 conjugate gradient backpropagation (SCGBP), and Ba–
Kurtosis  (2) yesian regularization backpropagation (BRBP), for trai–
N  4 ning a network. Weight and bias values are updated by
the LMBP algorithm in accordance with Levenberg-Mar

452 ▪ VOL. 52, No 3, 2024 FME Transactions

Figure 3. ANN Architecture

quardt optimization. Although it takeshigher me–mory z  d 

than other algorithms, this is frequently the tool–box's BPFI  f s  1  cos   (6)
2  D 
quickest backpropagation method and comes hig–hly
recommended as a first-choice supervised approach. z  d2 
The default algorithm is SCGBP, which uses the scaled BSF  f s  1  2 cos2   (7)
conjugate gradient approach in updating the weight and 2  D 

bias values. BRBP uses Levenberg-Marquardt optimi–
zation in updating the weight and bias variables. It 1  d 
FTF  f s  1  cos   (8)
determines the optimal blend to construct a network that 2  D 
exhibits strong generalization capabilities, initially by
Table 3. Characteristic Frequencies of Test Bearing
reducing a composite of squared errors and weights.
Figure 3 illustrates the ANN architecture, while Table 2 Characteristic Frequencies (Hz)
provides details about the four target values corres– BPFO 64
ponding to four classes of bearings.The four classes of BPFO 110
bearings are healthy bearings (target value of 1000), BSF 43,6
bearing with an outer race defect (target value of 0100), FTF 87.2
an inner race defect (target value of 0010), and a ball
defect (target value of 0001) as presented in Table 2. The experiments were conducted using the setup and
procedure explained in Section 2.1. Bearings with
Table 2. Target Values
various conditions, including both healthy ones and
Bearing Conditions Target those with defects in different elements, were tested at a
Healthy Bearing 1000 motor speed of 1490 rpm. Even for healthy bearings,
Bearing with Defect on Outer race 0100 some level of vibration, referred to as variable compli–
Bearing with Defect on Inner race 0010 ance [7], is observed. When a defect is present in the
Bearing with Defect on Ball 0001 interface between the ball and the raceways, it results in
a distinct peak in the amplitude of vibration due to the
3. RESULTS AND DISCUSSIONS impact. These defects can occur in theinner race, in the
outer race, or in the balls themselves. Figures 4-6 depict
In the field of vibration analysis, the signals are ac– the time domain plots and frequency domain plots of
quired from machine components and both the time vibration signals acquired from the outer race defect
domain data and the frequency domain data are ana– bearing, the inner racedefect bearing, andthe bearing
lyzed. A fault in a bearing element is indicated by a with balldefect, respectively.
peak observed in the vibration spectrum. However, the This can be observed from Figure 4(b) that the first
magnitude of this vibration peak depends on parameters prominent peak occurs at approximately 25 Hz,
like the size, shape, location, and type of defect. The representing the shaft frequency. Additionally, there are
theoretical bearing characteristic frequencies for the peaks at approximately 65 Hz, 130 Hz, and 196 Hz,
outer race defect, inner race defect, rolling element corresponding to the frequencies linked to a bearing
defect, and cage defectare typically labeled as BPFO, having an outer race defect (fod) and its harmonics. The
BPFI, BSF, andFTF, respectively. These theoretical theoretical bearing characteristic frequency, ball pass
bearing characteristic frequencies depend on both the frequency for an outer race defect (BPFO), is 64 Hz.
bearing's geometry and the shaft speed, as given in This observation suggests an outer racedefect.In Figure
equations (5-8) [25]. The theoretical bearing charac– 5(b), an initial prominent peak is seen at around 25 Hz,
teristic frequencies for the test bearing, with specified representing the shaft frequency. There are also
geometric details from Table 1 for a shaft speed of 1490 noticeable peaks at 112 Hz and 225 Hz. These
rpm, have been determined using the equations (5-8), frequencies are related to an inner race defect (fid) and
and are presented in Table 3. its harmonics. The expected theoretical bearing
z  d  characteristic frequency for an inner race defect (BPFI)
BPFO  f s  1  cos   (5) is 110 Hz, indicating the presence of adefect in the inner
2  D  race.In Figure 6(b), the first significant peak is observed
at around 25 Hz, representing the shaft frequency.

FME Transactions VOL. 52, No 3, 2024 ▪ 453

 (a) (b)
Figure 4. Vibration Signal for Bearing with Outer Race Defect (a) Time Domain and (b) Frequency Domain

(a) (b)
Figure 5. Vibration Signal for Bearing with Inner Defect (a) Time Domain and (b) Frequency Domain

(a) (b)
Figure 6. Vibration Signal for Bearing with Ball Defect (a) Time Domain and (b) Frequency Domain

454 ▪ VOL. 52, No 3, 2024 FME Transactions

 Additionally, there are peaks at approximately 87 Hz From Table 5, it is evident that the highest accuracy
and 131 Hz.These are the harmonics of a ball defect is achieved with 12 hidden layers. Consequently, 12
frequency (fbd). The theoretical bearing characteristic hidden layer nodes were selected for the ANN model. A
frequency, ball spin frequency (BSF), is 87.2 Hz. This total of 360 samples (90 samples each from the four
observation indicates a ball defect. classes of bearings) were provided as input. The scatter
plots of the input data are shown in Figures 7-9.
3.1 Fault Classification using ANN

The six essential time-domain features - RMS,crest fac–

tor, kurtosis, impulse factor, peak-to-peak, and energy -
have been extracted from the vibration signals using a
MATLAB program. Table 4 presents the statistical
parameters, namely range (minimum and maximum
values), mean, and standard deviation, of the six time-
domain features for four distinct bearing conditions.
The next step in ANN modeling is selecting the
number of hidden neurons. A back-propagation neural
network's predictionaccuracy is greatly influenced by
the quantity of hidden layers. An inadequate number of
nodes can hinder learning and lengthen the training
process, which will impact accuracy. On the other hand,
using too many layers can result in over-fitting and Figure 7. Scatter Plot (RMS Vs Peak)
prolonged training times. The quantity of hidden layers
can be calculated using the formula provided in
equations (9,10) [26].

1 m  n   a (9)

1  log2 n (10)

With the number of output nodes (m), the number of

input nodes (n), 'a' being a constant, that varies from one
to ten.In this case, the value of ‘m’ and ‘n’ are
respectively four and six. These values have been
substituted in equation (9) resulting in 'l' ranging from
four to thirteen. However, as per equation (10), 'l'
should be greater than six. Table 5 presents the
classification accuracy varying the number of hidden
layer nodes from 7 to 13. Figure 8. Scatter Plot (Kurtosis Vs Crest Factor)
Table 4. Time Domain Features of Test Bearings

Bearing Condition Statistical Crest

RMS Kurtosis Impulse Factor Peak to Peak Energy
Parameters Factor
Minimum Value 1.54 1.90 2.15 2.53 7.10 11.18
Maximum Value 1.99 2.66 3.22 3.97 9.23 13.86
Good Bearing
Mean 1.76 2.20 2.55 3.04 8.06 12.65
Standard Deviation 0.09 0.17 0.28 0.35 0.58 0.60
Minimum Value 1.45 2.03 2.32 2.72 8.21 10.12
Bearing with
Maximum Value 1.95 3.75 4.69 5.74 11.68 14.72
Outer Race
Mean 1.67 2.74 3.45 4.20 10.04 11.67
Defect
Standard Deviation 0.11 0.37 0.49 0.64 0.99 0.97
Minimum Value 1.96 4.51 4.20 5.57 20.05 12.88
Bearing with Maximum Value 2.93 9.06 7.30 9.83 28.95 17.19
Inner Race Defect Mean 2.46 5.94 5.65 7.48 24.49 15.28
Standard Deviation 0.17 0.94 0.72 0.98 2.46 0.95
Minimum Value 1.33 2.00 2.17 2.58 6.39 9.25
Bearing with Ball Maximum Value 1.86 3.93 4.37 5.23 10.04 12.85
Defect Mean 1.55 2.58 3.08 3.73 8.14 10.92
Standard Deviation 0.11 0.38 0.47 0.62 0.94 0.86
Table 5.Classification Accuracy for Different Hidden Layer Nodes

No. of Hidden Layers 7 8 9 10 11 12 13

Accuracy (%) 47 60 63 65 68 87 80

FME Transactions VOL. 52, No 3, 2024 ▪ 455

 actual target values. In each confusion matrix, correct
classifications are represented by green background
squares along the diagonal, while misclassifications are
indicated by red background squares off the diagonal.
The last row of each matrix displays the percentages
of correctly classified samples in green text and
misclassified samples in red text, out of the actual
samples of a particular class. The square in the last row
and last column of each matrix uses a gray background:
the percentage of all the correctly classified samples are
in green text, and the percentage of all misclassified
samples are in red text.
As observed in Figures 10-11,252 samples (70%)
were utilized for training, and 54 samples (15%) were
utilized for validation and testing each, for the networks
Figure 9. Scatter Plot (Impulse Factor Vs Energy)
trained with LMBP and SCGBP algorithms. However,
This can be observed from the scatter plots as shown in for the network, trained with BRBP algorithm, 306
Figures 7-9 that the data points are clustered for good samples (85%) and 54 samples (15%) were used for
bearings and bearings with defects in the outer race and in training and testing, respectively. In this algorithm, the
rolling element. However, the data points for inner race generalization took place within the training process
defects are distinct from the other three classes of bearings. without validation. This can be observed fromFigure 12’s
These samples are categorized into three types: trai– validation confusion matrix that all values are zeros. It is
ning samples, validation samples, and testing samples. evident from Figure 10's all-confusion matrix that each
During training, the network is exposed to training sam– row gives the predicted output values. Out of 90 samples
ples and adjusted based on its error. Validation samples of class 1, 86 were classified correctly, however, 4 were
are utilized to evaluate the network's generalization misclassified as fourth class.In the second column, out of
ability and to cease training when generalization ceases to 90 samples belonging to class 2, 5 samples were
improve. Testing samples remain unaffected by training misclassified as class 1 (first row), 75 samples were
and thus offer an independent assessment of network correctly classified (diagonal element), no sample was
performance both during and after training. The ANN misclassified as class 3 (third row), however fourth row
model randomly allocated 70% of samples for training shows 10 samples were misclassified as class 4. From the
and reserved 15% for both testing and vali–dation each third column, it is evident that al OF THE 90 samples of
for LMBP and SCGBP algorithms. How–ever, in the class 3 wwre classified correctly. And out of 90 samples
BRBP algorithm, for training 85% samples have been of class 4, 73 samples were classified correctly, whereas
utilized, and the remaining 15% samples have beenused 17 samples were misclassified.
for testing. The ANN model underwent training with Overall, 90% of the samples were correctly iden–
these three algorithms, and the respective confusion tified and 10% of the samples were misclassified when
matrices are displayed in Figures10-12. The cross- the network was trained with the lmbp algorithm. Simi–
entropy (CE) and the mean squared error (MSE) values larly, the network, when trained with SCGBP algorithm,
are presented in Table 6 to compare the per–formance of gives 83.6% correct classification and 16.4% miscla–
these three algorithms. Cross-entropy measures the ssifications. However, the BRBP algorithm improves
performance of a network by comparing the targets and the capability of the network in correctly classifying
outputs, with the option to include performance weights 97.2% samples.The same conclusion can be observedin
and other parameters. A lower cross-entropy value Table 6; the BRBP algorithm exhibits the highest per–
indicates better classifier performance. For each output- formance in terms of CE and MSE, followed by the
target pair, cross-entropy is calculated as (11), LMBP and SCGBP algorithms. Unlike the other three
CE  t  log  y  (11) classes, the class 3 samples were 100% correctly classi–
fied, without confusion, by all three algorithms. This
where ‘t’ is the target value and ‘y’ is the output value. result is due to the fact that the class 3 data points are
The overall cross-entropy performance is the average of distinct from the other three classes of bearings, as seen
these individual values. MSE is the average of the in the scatter plots.
squared differences between the outputs and the targets.
Lower MSE values indicate better performance, with a 3.2 Prediction of Bearing Condition by Simulating the
Trained ANN
value of zero signifying no error.For the arrays of
output-target pair, MSE is calculated as (12),
To further evaluate the capability of the network trained
N with the three algorithms in bearing fault classification,
1
  yi  ti 
2
MSE  (12) they have been simulated to predict the bearing conditions
N i 1
using a separate dataset of the six time-domain features.
where 'N' is the number of data points, ‘yi' is the output The trained ANN models were fed these data for each
value, and 't' is the target value. bearing and the respective predicted values were noted.
In the confusion matrices, the rows represent Table 7 presents a comparison of the predicted values by
predicted output values, while the columns display the three algorithms for the four classes of bearings.

456 ▪ VOL. 52, No 3, 2024 FME Transactions

Figure 10. Confusion Matrices for the ANN trained with LMBP Algorithm

Figure 11. Confusion Matrices for the ANN trained with SCGBP Algorithm

FME Transactions VOL. 52, No 3, 2024 ▪ 457

Fig. 12. Confusion Matrices for the ANN trained with BRBP Algorithm
Table 6. Comparison of 3 Algorithms’ Performance

Performance Cross Entropy (CE) Mean Squared Error (MSE)

Indicator
Training Validation Testing Training Validation Testing
Algorithm
LMBP 0.0500 0.0427 0.0246 0.0511 0.0372 0.0496
SCGBP 1.3682 3.6727 3.7072 0.0529 0.0637 0.0493
BRBP 0.0189 0.0000 0.0408 0.0269 0.0000 0.0565
Table 7. Comparison of Predicted Values with the Target Values for 3 Algorithms in Bearing Fault Classification
Algorithm
BearingCondition LMBP SCGBP BRBP
Values
Target 1 0 0 0 1 0 0 0 1 0 0 0
Good Predicted 0.9891 0.0052 0.0001 0.0056 0.9424 0.0337 0.0002 0.0237 1.0000 0.0000 0.0000 0.0000
Error 0.0109 0.0052 0.0001 0.0056 0.0576 0.0337 0.0002 0.0237 0.0000 0.0000 0.0000 0.0000
Target 0 1 0 0 0 1 0 0 0 1 0 0
OuterRaceDefect Predicted 0.0085 0.9722 0.0040 0.0154 0.0028 0.9957 0.0012 0.0004 0.0000 1.0000 0.0000 0.0000
Error 0.0085 0.0278 0.0040 0.0154 0.0028 0.0043 0.0012 0.0004 0.0000 0.0000 0.0000 0.0000
Target 0 0 1 0 0 0 1 0 0 0 1 0
InnerRaceDefect Predicted 0.0000 0.0006 0.9994 0.0000 0.0000 0.0025 0.9975 0.0000 0.0000 0.0000 1.0000 0.0000
Error 0.0000 0.0006 0.0006 0.0000 0.0000 0.0025 0.0025 0.0000 0.0000 0.0000 0.0000 0.0000
Target 0 0 0 1 0 0 0 1 0 0 0 1
Ball Defect
Predicted 0.0000 0.0288 0.0000 0.9711 0.0046 0.0143 0.0000 0.9811 0.0000 0.0000 0.0000 1.0000
Error 0.0000 0.0288 0.0000 0.0289 0.0046 0.0143 0.0000 0.0189 0.0000 0.0000 0.0000 0.0000

As observed from Table 7, the network with the 4. CONCLUSION

BRBP algorithm predicts all four classes of bearings
with zero errors. The LMBP algorithm performs with Timely fault detection in rotating machinery is crucial
higher accuracy than the SCG BP algorithm. Even the to minimize the costs associated with downtime,
networks trained with LMBP and SCG BP algorithms unexpected failures, and potential casualties. This
predictminimal errors for inner race defect bearings paper aims at fault diagnosis of rolling element
(class 3). bearings using artificial intelligence techniques,

458 ▪ VOL. 52, No 3, 2024 FME Transactions

specifically artificial neural networks. The following Today: Proceedings, Vol. 4, No. 2, pp. 2085–2094,
conclusions can be drawn: 2017, https://doi.org/10.1016/ j.matpr.2017.02.054
(1) Experiments were conducted and vibration signals [8] I. Howard, A Review of Rolling Element Bearing
were acquired for the test bearings using NI DAQ and Vibration “Detection, Diagnosis and Prognosis”,
LABVIEW. Six time-domain features were extracted Department of Defence, Melbourne, 1994.
using MATLAB to create a comprehensive dataset for
[9] A. Rai, S. H. Upadhyay, A review on signal
training, validation, and testing of the ANN models.
processing techniques utilized in the fault
(2) The networks were trained with three algorithms:
diagnosis of rolling element bearings, Tribology
Levenberg-Marquardt backpropagation, scaled
Interna–tional, Vol. 96, pp. 289–306, Apr. 2016,
conjugate gradient backpropagation, and Bayesian
https://doi. org/10.1016/j.trib oint.2015.12.037
regularization backpropagation. It was observed that
the BRBP algorithm achieved the highest percentage of [10] M. N. Jagdale, G. Diwakar, A critical review of
correct classification (97.2%), followed by LMBP condition monitoring parameters for fault
(90%) and SCGBP (83.6%). diagnosis of rolling element bearing, in IOP
(3) To evaluate their efficacy in bearing fault classi– Conference Se–ries: Materials Science and
fication, these three networks were simulated for pre– Engineering, Vol. 455, p. 012090, Dec. 2018,
diction. It was observed that BRBP could predict all https://doi.org/10.1088/1757-899x/455/ 1/012090
four classes of bearings with zero errors from the target [11] Demetgul, K. Yildiz, S. Taskin, I. N. Tansel, O.
values. Yazicioglu, Fault diagnosis on material handling
The proposed model’s accuracy in predicting bea– system using feature selection and data mining
ring conditions enhances maintenance strategies, pre– techniques, Measurement, Volume 55, pp. 15–24,
venting failures, and extending machinery lifespan. September 2014, doi.org/10.1016/j.measurement.
This investigation advances condition monitoring and 2014.04.037
predictive maintenance, laying the groundwork for [12] H. Saruhan, S. Sandemir, A. Çiçek, I. Uygur, Vib-
future fault diagnosis in mechanical systems. ration analysis of rolling element bearings defects,
Journal of Applied Research and Technology, Vol.
REFERENCES 12, No. 3, pp. 384–395, Jun. 2014, https://doi.org/
[1] D. Abboud, M. Elbadaoui, W. A. Smith, and R. B. 10.1016/s1665-6423(14)71620-7
Randall, Advanced bearing diagnostics: A compa– [13] Q. Xu, S. Lu, W. Jia, C. Jiang, Imbalanced fault
rative study of two powerful approaches, Mecha– diagnosis of rotating machinery via multi-domain
nical Systems and Signal Processing, Vol. 114, pp. feature extraction and cost-sensitive learning,
604–627, 2019, https://doi.org/10.1016/j.ymssp. Journal of Intelligent Manufacturing, Vol. 31, No.
2018.05.011 6, pp. 1467–1481, 2020, https://doi.org/10.1007/
[2] Althubaiti, Adnan, Faris Elasha, Joao Amaral s10845-019-01522-8
Teixeira. Fault diagnosis and health management [14] Mohammed, Jawad Saja, Jaber Alaa Abdulhady.
of bearings in rotating equipment based on Rolling bearing fault detection based on vibration
vibration analysis–a review, Journal of signal analysis and cumulative sum control chart,
Vibroengineering, Vol. 24, no. 1, pp. 46-74, 2022, FME Transactions, Vol. 49, No. 3, pp. 684-695,
https://doi.org/10. 21595/jve.2021.22100 (2021), doi:10.5937/fme2103684M.
[3] P.Y. Kim, I.R.G. Lowe, A review of rolling [15] R. Pidl, P. Böröcz, Discrete Fourier transform and
element bearing health monitoring, In: cepstrum analysis of vibration events on
Proceedings of Machinery Vibration Monitoring semitrailer truck, FME Transactions, vol. 47, pp.
and Analysis Meeting, Vibration Institute, 177- 182, 2019, doi: 10.5937/fmet1901177P.
Houston, TX, 19–21 April 1983, pp.145–54, [16] J. R. Stack et al, Fault classification and fault
[4] P.D. McFadden, J.D. Smith, Vibration monitoring signature production for rolling element bearings
of rolling element bearings by the high-frequency in electric machines, IEEE Transactions on
resonance technique- a review, Tribology Interna– Industry Applications, Vol. 40, No. 3, pp. 735–
tional, Vol. 17, No. 1, pp. 3-10, 1984. 739, May 2004, https://doi.org/10.1109/
[5] N. Tandon, B.C. Nakra, Vibration and acoustic tia.2004.827454
monitoring techniques for the detection of defects [17] Kim, Seokgoo An, Dawn Choi, Joo-Ho, Diagnos–
in rolling element bearings - a review The Shock tics 101: A Tutorial for Fault Diagnostics of
and Vibration Digest, Vol. 24, No. 3, pp. 3-11, Rolling Element Bearing Using Envelope Analysis
1992. in MATLAB, Applied Sciences, Vol. 10. No. 20,
[6] N. Tandon, A. Choudhury, A Review of Vibration 7302, pp. 1-23, 2020, https://doi.org/10.3390/app
and Acoustic Measurement Methods for the 10207302
Detection of Defects in Rolling Element Bearings, [18] B. Li, M.-Y. Chow, Y. Tipsuwan, J. C. Hung,
Tribology International, Vol. 32, pp.469–480, Neural-network-based motor rolling bearing fault
1999. diagnosis, IEEE Transactions on Industrial Elect–
[7] P. Gupta, M. K. Pradhan, Fault detection analysis ronics, Vol. 47, No. 5, pp. 1060–1069, 2000, https:
in rolling element bearing: A review, Materials //doi.org/10.1109/41. 873214

FME Transactions VOL. 52, No 3, 2024 ▪ 459

[19] T. Tung, B.-S. Yang, Machine fault diagnosis and AI Artificial Intelligence
prognosis: the state of the art, International Journal ANN Artificial Neural Network
of Fluid Machinery and Systems, Vol. 2, No. 1, BP Back-Propagation
pp. 61–71, Mar. 2009, https://doi.org/10.5293/ Ball Pass Frequency for the Inner
BPFI
ijfms.2009.2.1.061 Race
[20] B. A. Paya, I. I. Esat, M. N. M. Badi, Artificial BPFO Ball Pass Frequency for the Outer race
neural network-based fault diagnostics of rotating Bayesian regularization
BRBP
machinery using wavelet transforms as a prepro– backpropagation
cessor, Mechanical Systems and Signal BSF Ball Spin Frequency
Processing, Vol. 11, No. 5, pp. 751–765, Sep. CF Crest Factor
1997, https://doi. org/10.1006/mssp.1997.0090 CM Condition Monitoring
EI Energy Index
[21] P.K. Kankar, S.C. Sharma, S.P. Harsha, Fault
FFBP Feed-Forward Back-Propagation
Diagnosis of Ball Bearings using Machine
FFNN Feed-Forward Neural Network
Learning Methods, Expert Systems with
FFT Fast Fourier Transform
Applications, Vol. 38, pp. 1876–1886, 2011,
FTF Fundamental Train Frequency
https://doi.org/10. 1016 /j.eswa.2010.07.119
Levenberg-Marquardt
[22] J. P. Patel, S. H. Upadhyay, “Comparison between LMBP
backpropagation
artificial neural network and support vector MFS Machinery Fault Simulator
method for a fault diagnostics in rolling element OD Outer Diameter
bearings,” Procedia Engineering, Vol. 144, pp. REB Rolling Element Bearing
390– 397, 2016, https://doi.org/10.1016/j.proeng. RMS Root Mean Square
2016.05.148 scaled conjugate gradient
SCGBP
[23] Jamadar, Imran M., Ajit Kumar Patil, Prasanta backpropagation
Kumar Samal, B. Suresha, An empirical model VFD Variable Frequency Drive
integrating dimensional analysis and Box-Behnken
design for crack detection in rotor fan blades, FME
Transactions, Vol. 52, No. 1, pp. 45-56, 2024, doi: ДИЈАГНОСТИКА КВАРОВА У ЛЕЖАЈЕВИМА
10.5937/fme2401045J КОТРЉАЈУЋИХ ЕЛЕМЕНАТА
[24] Gunerkar, R.S., Jalan, A.K., Belgamwar, S.U. Fault ПОБОЉШАНА АИ: СВЕОБУХВАТАН
diagnosis of rolling element bearing based on artifi– ПРИСТУП АНАЛИЗИ ВИБРАЦИЈА
cial neural network. J Mech Sci Technol 33, 505–
511, 2019, https://doi.org/10.1007/s12206-019- П.К. Самал, К. Сунил, И.М. Јамадар, Р. Сриниди
0103-x
[25] Harris T. A, Rolling Bearing Analysis, 4th ed., Ово истраживање представља свеобухватан
John Wiley & Sons, New York, 2001. приступ за дијагностику кварова лежајева кориш–
ћењем вештачке интелигенције (АИ), посебно кроз
[26] Bishop, Christopher M., and Nasser M. Nasrabadi. примену вештачких неуронских мрежа (АНН).
Pattern recognition and machine learning, Vol. 4. Интеграцијом ових мрежа у анализу вибрација,
No. 4. New York: Springer, 2006. приступ има за циљ да задовољи критичну потребу
за брзим откривањем квара. Методологија се
NOMENCLATURE
састоји од три кључна корака: аквизиција сигнала
B Race Width вибрације, издвајање карактеристика и класифи–
d Ball Diameter кација грешке. Експерименти су спроведени за
D Pitch Diameter добијање сигнала вибрација за испитне лежајеве на
fbd Ball Defect Frequency симулатору грешке машине. Шест карактеристика
fid Inner Race Defect Frequency временског домена екстраховано је коришћењем
fod Outer Race Defect Frequency МАТЛАБ-а, стварајући свеобухватан скуп података
fs Shaft Frequency за обуку АНН модела са три алгоритма: Левенберг-
ID Inner Diameter Маркуардт бацкпропагатион (ЛМБП), скалиран
IF Impulse Factor коњуговани градијент бацкпропагатион (СЦГБП) и
KU Kurtosis Баиесиан регуларизатион бацкпропагатион (БРБП).
l Number of Hidden Layers БРБП алгоритам је постигао највећу тачну стопу
m Number of Output Nodes класификације (97,2%), а следе ЛМБП (90%) и
n Number of Input Nodes СЦГБП (83,6%). Да би се проценила њихова
z Number of Balls ефикасност у класификацији кварова лежаја, ове
ϕ Contact Angle три мреже су симулиране, откривајући да БРБП
може предвидети све четири класе лежајева са нула
ABBREVIATIONS грешака.

460 ▪ VOL. 52, No 3, 2024 FME Transactions