Application of Artificial Intelligence For The Prediction of Plain Journal Bearings Performance

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 19

Alexandria Engineering Journal (2022) 61, 9011–9029

H O S T E D BY
Alexandria University

Alexandria Engineering Journal


www.elsevier.com/locate/aej
www.sciencedirect.com

Application of artificial intelligence for the


prediction of plain journal bearings performance
M.B. Badawi *, Taher H. Awad, I.M. El Fahham

Mechanical Engineering Department, Faculty of Engineering, Alexandria University, El-Chatby, Alexandria 21544, Egypt

Received 20 May 2021; revised 11 November 2021; accepted 13 February 2022


Available online 26 February 2022

KEYWORDS Abstract Artificial intelligence techniques are applied to predict the performance of a plain journal
Plain journal bearing; bearing instead of classical methods. AI techniques are known to be superior for prediction; they
Lubrication; are accurate and fast compared to finite difference, finite element, and finite volume methods. To
Artificial intelligence; obtain the data needed for the AI prediction, the finite difference method is used to solve the dimen-
Neural network; sionless Reynolds equation at various aspect ratios. The bearing performance characteristics, such
Fuzzy logic technique as load-carrying capacity, attitude angle, friction variable, and maximum-film-pressure ratio, are
determined considering isothermal conditions. Four aspect ratios are considered from 0.25 to 4,
with eccentricity ratios varying between 0.2 and 0.8. Three artificial neural networks (Feed-
forward, Radial basis, and Generalized regression networks) and fuzzy logic techniques were
applied to the obtained data from FDM simulation to predict the performance parameters. The
three trained neural networks and the fuzzy system were tested to obtain the performance charac-
teristics for aspect ratios and eccentricity ratios that were not included in the FDM study. The cur-
rent response of the trained ANN models and the fuzzy logic technique is found to be very fast and
precise, with a prediction computational time of less than one second and an error of less than 2.5
percent.
Ó 2022 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Engineering, Alexandria
University. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/
4.0/).

1. Introduction effects. Many methods were considered to study the behavior


of plain journal bearings, including the finite element method
The most basic type of journal bearing is plain journal bearing, (FEM), the finite difference method (FDM), and the finite vol-
in which a shaft or journal rotates within the bearing with a ume method (FDM). Many parameters, such as eccentricity
layer of lubricant separating the two parts via fluid dynamic ratio and aspect ratio (L/D), were the subject of bearing per-
formance research. One of the critical specifications that can
affect the performance of a journal bearing is the aspect ratio.
* Corresponding author. The pressure distribution and the load-carrying capacity of the
E-mail addresses: mohammed.badawi@alexu.edu.eg (M.B. Badawi), plain journal bearing have been determined using both analyt-
taher.awad@alexu.edu.eg (T.H. Awad), ibrahim.elfahham@alexu. ical and finite element methods by Nuruzzaman et al. [1]. The
edu.eg (I.M. El Fahham).
finite element results showed sound agreement when compared
Peer review under responsibility of Faculty of Engineering, Alexandria
with previous research. Khan et al. [2] used the finite difference
University.
https://doi.org/10.1016/j.aej.2022.02.041
1110-0168 Ó 2022 THE AUTHORS. Published by Elsevier BV on behalf of Faculty of Engineering, Alexandria University.
This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
9012 M.B. Badawi et al.

Nomenclature

R Journal radius [m] R 2


Symbol S Sommerfeld number, S ¼ lNLD W C [–]
C Radial clearance [m] S1 The number of neurons in the first layer of a radial
D Journal diameter [m] basis neural network [–]
f Coefficient of friction [–] S2 The number of neurons in the second layer of a ra-
ff Dimensionless friction force [–] dial basis neural network [–]
h Fluid film thickness [m] 
U Characteristics velocity, U ¼ 2pNR [m/s]
h Dimensionless film thickness, h ¼ Ch [–] W
Load carrying capacity 
[N]
L Bearing length [m] W

Dimensionless load, W ¼ W
 RÞ2
[–]
pDLlNð
L/D Aspect ratio [–] Z Dimensionless axial distance, Z C¼ ZL [–]
N Rotational speed [rps] / Attitude angle [Degree]
nprod Generalized neural regression network product h Circumferential direction [Degree]
multiplication [–] e Eccentricity ratio [–]
P Characteristics pressure, P ¼ LD
W
[Pa] l Lubricant viscosity [Pa. s]
p Fluid static pressure [Pa]
  p
p Dimensionless static pressure, p ¼ R 2
[–]
2pNlðCÞ

method to investigate the static performance of a plain journal [7] investigated the variation of shear stress, power losses, and
bearing with a non-Newtonian lubricant while changing the RMS in various lubricants under different loading conditions
aspect ratio with values equal to 0.5, 1, and 1.5. According and at different rotating speeds. The experimental results were
to the results, it is clear that as the aspect ratio increases, the then used to train an artificial neural network to predict bear-
dimensionless maximum pressure, the load-carrying capacity, ing performance. Kumar et al. [8] applied ANN to predict the
and the attitude angle increase, while the non-dimensional fric- maximum pressure of a journal bearing, as a micropolar lubri-
tion variable decreases. Using the COMSOL Multiphysics pro- cated hybrid journal bearing was investigated using FEM. The
gram, Mane and Soni [3] created a three-dimensional model of results are used for training and testing the artificial neural net-
a hydrodynamic plain journal bearing. The pressure distribu- work model. The results of FEM and ANN are nearly identi-
tion was calculated using a steady-state analysis, and the cal, with an error of less than 0.5 percent. Badawi et al. [9] used
results were compared to an analytical finite element method COMSOL Multiphysics to compare the performance of tilting
that matched COMSOL results. Rajoub et al. [4] used the finite pad journal bearings to plain journal bearings at an aspect
difference method to study the steady-state characteristics of ratio of 0.5. When compared to COMSOL, it was possible
plain bearing with a textured surface, including the Sommer- to predict the performance of a tilting pad journal bearing with
feld number, friction variable, and attitude angle, and com- trivial computer time. Three artificial neural networks (ANN)
pared the results to normal plain journal bearing to were built to simulate the performance, and the feed-forward
demonstrate the effect of texture on bearing performance. network resulted in superior performance with an error of less
Sharma and Awasthi [5] investigated the effect of the aspect than 1%. ANN was also used in other engineering fields. Pele-
ratio, varying from 0.5 to 1.5, on fluid-film bearing system per- saraei et al. [10] applied an artificial neural network with feed-
formance. The eccentricity ratio, coefficient of friction, mini- forward back-propagation models based on the Levenberg-
mum film thickness, maximum pressure, attitude angle, Marquardt training algorithm to predict electricity and envi-
damping coefficients, and the critical mass of hydrodynamic ronmental factors against energy consumption in municipal
journal bearings are all calculated using the dynamic Reynolds solid waste management (MSW). A. Rosenkranz et al. [11]
governing equation for each aspect ratio. The results show that highlight and review some of the recent advancements in tribo-
the minimum fluid film thickness, attitude angle, coefficient of logical research achieved using artificial intelligence and
friction, and damping coefficient all increase as the aspect ratio machine learning, specifically artificial neural networks. A
increases. Nowadays, artificial intelligence such as artificial fuzzy logic model was also used in the prediction of bearing
neural networks (ANN) and fuzzy logic techniques are performance. Ahmed et al. [12] developed a fuzzy logic model
advanced tools for predicting the performance behavior of for predicting the highest pressures on oil films in a plain jour-
processes that represent physical phenomena and are widely nal bearing. Rotational speed, bearing load, and oil-feed pres-
used in mechanical engineering applications. Because of sure is examples of model-independent variables. The results
advances in computational technologies, the response time of demonstrated that the model could accurately predict maxi-
ANN models is very fast and precise. As the application of mum oil-film pressure. Applying fuzzy logic technique Reh-
ANN Sinanoglu et al. [6], experimental results were compared man et al. [13] demonstrated an active hydrostatic journal
with that of the feed-forward neural network model. The with a mechatronic system, which is in high demand for
FFNN was used to predict bearing performance by presenting rapidly expanding industrial needs. The goal is to improve
and analyzing the pressure variation on the steel shaft of plain the dynamics and static properties of hydrostatic journal bear-
journal bearings at various speeds and temperatures. The ings by employing two distinct control strategies. According to
study demonstrates that both experimental and artificial neural the results, intelligent control based on fuzzy logic has a faster
network models have values close to each other. Esmaeili et al. response, greater amplitude reduction, and good stiffness
Application of artificial intelligence for the prediction 9013
 2 Z L=2 Z Rheff
against load. The goal of the present work is to predict plain R
journal bearing performance using artificial neural network FR ¼ fR pLDlN ¼ RpcosðhÞdhdz ð5Þ
C L=2 0
and fuzzy logic techniques and then compare both methods
to determine which method is more accurate with the mini-  2 Z L=2 Z Rheff
R
mum errors and minimum computational time. FT ¼ fT pLDlN ¼ RpsinðhÞdhdz ð6Þ
C L=2 0

2. Performance of plain journal bearing qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi


W¼ F2R þ F2T ð7Þ
2.1. Governing equations
The dimensionless force components fR and fT are defined
by equations (5) and (6). The Sommerfeld number is defined
Fig. 1 shows the geometry and coordinate system of the jour- as:
nal bearing. The steady-state dimensionless Reynolds equation  2  2
governing the hydrodynamic lubrication of the bearing [14] is: lN R lNLD R lNLDðR=CÞ2
S¼ ¼ ¼ 
!   !  ðW=LDÞ C W C W plNLDðR=CÞ2
 2 
@ 3 @ p R @ 3 @ p @h
h þ  h  ¼6 ð1Þ 1
@h @h L @z @z @h ¼  ð8Þ
Wp
The attitude angle, defined as the angle formed by the load
2.1.1. Plain journal bearing vector and the line of centers, is given by:

The dimensionless film thickness h of a plain journal bearing  
f 
can be expressed as follows: / ¼ arctan T  ð9Þ
fR

h ¼ 1 þ ecosðhÞ ð2Þ The friction force can be calculated from:
 
R
Ff ¼ ff pLDlN
2.1.2. Bearing performance parameters C
The force balance as shown in Fig. 1 can be represented as: Z L=2 Z Rheff  
h @p l
¼ R þ U dhdz ð10Þ
FR þ Wcosð/Þ ¼ 0 ð3Þ L=2 0 2R @h h
The friction coefficient is calculated by:
FT  Wsinð/Þ ¼ 0 ð4Þ    
Ff pLDlN RC ff ff C
where FR and FT represent the pressure force components f¼ ¼   ¼  ð11Þ
W pLDlN R 2 W W R
along and normal to the line of centers, respectively, and W C
represents the external load on the shaft: The friction variable will be:
ff
fðR=CÞ ¼  ð12Þ
W

2.2. Methodology

The Reynolds equation in the fluid film region is to be replaced


by its corresponding finite difference form and solved using the
successive over-relaxation method (SOR) with MATLAB soft-
ware. A program was developed to investigate the steady-state
analysis of plain journal bearings with eccentricity ratios rang-
ing from 0.2 to 0.8 and length to diameter ratio (L/D = 0.25,
0.5, 1, and 4). The flow chart in Fig. 2 depicts the steps of the
program that was used to perform the calculations. The input
data for the FDM program are the shaft diameter, bearing
length, radial clearance, angular velocity, minimum film thick-
ness, lubricant viscosity, and lubricant density. The isothermal
case is considered with the lubricant flow assumed laminar and
incompressible. An axial groove is located at the maximum
film thickness. The non-cavitation Swift-Stieber condition is
to be used [15]. Atmospheric pressure is assumed at the bearing
inlet and sides, and the viscosity of the lubricant is assumed to
be constant, with no slip at the boundary. The FDM program
outputs were the maximum pressure, the load-carrying capac-
ity, the friction variable, the Sommerfeld number, and the atti-
Fig. 1 Geometry and coordinate system for plain journal tude angle. Also, a plot for the pressure distribution in
bearings. circumferential and axial directions.
9014 M.B. Badawi et al.

Fig. 2 Flow chart for finite difference method using MATLAB.

2.3. Validation of the model non-dimensional load, and the graph shows a good agreement
between the proposed FDM and previous work [16], with an
To validate the FDM code, the results for aspect ratio 0.5 were average relative error of 3.12%.
compared with EL-Said et al. [14] results. Fig. 3 represents the
relation between eccentricity ratio and both the friction vari- 2.4. Acquiring data needed for artificial intelligence techniques
able and Sommerfeld number, respectively. The current work’s
results are in good agreement with [14], and the average rela- There is only one maximum pressure point along the circum-
tive error for the friction variable and Sommerfeld number ference of a plain journal bearing. This is due to the geometry
were 2.31% and 1.72%, respectively. Also, another compar- of the bearing and how the fluid gap expands and contracts
ison was made at an aspect ratio of 0.8 with Borse et al. [16]. once around the journal shaft circumference. The input data
Fig. 4 depicts the relationship between eccentricity ratio and listed in Table 1 and the simulating of the FDM program
Application of artificial intelligence for the prediction 9015

Fig. 3 Comparison between EL-Said et al. [14] and FDM to validate the model at an aspect ratio of 0.5 (a) eccentricity ratio versus
friction variable, (b) eccentricity ratio versus Sommerfeld number.

for eccentricity ratios varying from 0.2 to 0.8 and aspect ratio angle, and maximum-film-pressure ratio. The suggested net-
constant with a value of 0.5 are considered. Figs. 5 and 6 show works will be trained using the input and output data obtained
that the increase in maximum pressure appears to be increasing from the FDM code. The trained networks help the designer to
with the increase in eccentricity ratio. To investigate the effect predict output performance based on input values that were
of changing the aspect ratio on bearing performance, the not included in the training process in a short period that does
model was run with aspect ratios of 0.25, 0.5, 1, and 4, and var- not exceed one second. The following three neural network
ious performance parameters (Sommerfeld number, friction models were designed, trained, and tested to fit the data
variable, maximum-film-pressure ratio, and attitude angle) obtained from FDM (Feed-forward neural network, Radial
were calculated. Fig. 7-a and -b show a significant decrease basis neural network, and Generalized regression neural
in Sommerfeld number and friction variable with the increase network).
in eccentricity ratio. Also, as the aspect ratio increases, the
Sommerfeld number, and friction variable decrease at a certain 3.1. The feed-forward neural network
value of eccentricity ratio. Fig. 7-c and -d represent a decrease
in the maximum-film-pressure ratio and attitude angle with the As shown in Fig. 8, the feed-forward neural network (FFNN)
increase in eccentricity ratio. Furthermore, calculating the has one input and one output layer, as well as several hidden
maximum-film-pressure ratio aspect ratio has a minor effect layers in between. The number of hidden layers, the neurons
on attitude angle. in them, and the types of neurons were determined based on
the designer’s experience and some trials. The FDM data were
3. Artificial neural network divided into three groups: 70% for training the FFNN, 15%
for validation, and 15% for testing the proposed FFNN. To
The neural network is a field of artificial intelligence in which achieve the weight and bias of the FFNN, the Levenberg-
data structures and algorithms for the study and classification Marquardt optimization technique introduced by Ibrahimy
of information are found inspired by the human brain. As out- et al. [17] was used. Finally, after many trials with different
puts to the input data aspect ratio and eccentricity ratio, arti- numbers and types of neurons, the error between the FFNN
ficial neural networks with various structures were proposed to and FDM values was minimized. The optimized neural net-
obtain the Sommerfeld number, friction variable, attitude work consists of three hidden layers, each with 50, 10, and 5
9016 M.B. Badawi et al.

Fig. 4 Comparison between Borse et al. [16] and FDM to validate the model at an aspect ratio of 0.8.

Simulating the suggested GRNN with a spreading factor of


Table 1 Dimensions of plain jour-
0.001 using the same data applied to both feed-forward and
nal bearings and oil properties.
radial basis networks.
Parameters
Length [m] 0.025 3.4. The results from the suggested neural networks
Journal diameter [m] 0.05
Aspect ratio 0.5 The data obtained from FDM in Table 2 with different aspect
Radial clearance [m] 2.5  10-5
ratios and eccentricity ratios are employed in the training of
Eccentricity [m] 1.25  10-5
the suggested three neural networks. After 86 iterations, the
Journal speed [rpm] 1000
Dynamic viscosity [Pa.s] 0.19 optimum feed-forward neural network optimized using the
Levenberg-Marquardt technique [16] has a minimum mean
square error of 0.0149 between FDM and trained FFNN data.
Fig. 11 depicts the regression of the trained data, the validation
data, the test data, and the overall three data. The same input
and output variables listed in Table 2 are used to train radial
neurons made up of purelin, purelin, and tansig neurons, basis and generalized regression neural networks. Figs. 12-13
respectively. show the comparison between the FDM and the trained three
neural networks. From the figures, it can be noticed that the
3.2. The radial basis neural network Sommerfeld number, friction variable, attitude angle, and
maximum-film-pressure ratio fit well with FDM data. The
The radial basis neural network (RBNN) is made up of two trained neural networks are superior because they can obtain
layers, input, and output layers as depicted in Fig. 9. A spread performance data for input aspect ratios and eccentricity ratios
factor of 0.04 is used to simulate the RBNN with the same that were not included in the training process. Table 3 shows
input and output data as the feed-forward neural network. the prediction of the three suggested neural networks to get
The training process generated 19 and 4 hidden layer neurons bearing performance for data not trained on as aspect ratio
S1 and output layer neurons S2, respectively. 0.3 and eccentricity ratio 0.5 and shows that the FFNN gives
a more realistic response with minimum errors. Figs. 14-17
3.3. The generalized regression neural network show the prediction of the three neural networks at aspect
ratio 0.8 at eccentricity ratio varying from 0.2 to 0.8. From
Figs. 14 and 15, it is clear that FFNN predicts the performance
The structure of the generalized regression neural network
of Sommerfeld number and friction variable with minimum
(GRNN) shown in Fig. 10 is similar to that of the RBNN,
error while GRNN has bad prediction with maximum errors.
but it differs in the second layer with the nprod block shown.
Application of artificial intelligence for the prediction 9017

Fig. 5 Comparison of pressure distribution versus bearing angle at different eccentricity ratios and aspect ratio equals 0.5.

Fig. 6 Comparison of pressure distribution versus bearing length at different eccentricity ratios and aspect ratio equals 0.5.
9018 M.B. Badawi et al.

Fig. 7 The effect of changing the aspect ratio on the performance of a plain journal bearing as the eccentricity ratio is changed versus (a)
Sommerfeld number, (b) Friction variable, (c) Maximum-film-pressure ratio, and (d) Attitude angle.

Fig. 8 The feed-forward NN structure.

While Figs. 16 and 17, all three neural networks predict the 4. Fuzzy logic technique
attitude angle and maximum-film-pressure ratio with the same
response and average equal errors. Table 4 shows the RMS Fuzzy logic is a kind of multivalued logic, enabling intermedi-
errors for the three neural networks. The feed-forward, radial ate values between conventional assessments, such as true/-
basis, and generalized regression neural networks had com- false, yes/no, high/low, etc. The fuzzy programs were
puter prediction times of 0.01328, 0.00753, and 0.005172 s, developed to predict the performance parameters such as Som-
respectively.
Application of artificial intelligence for the prediction 9019

Fig. 9 Radial basis NN structure.

Fig. 10 Generalized regression NN structure.

Table 2 The data used to train the three neural networks and fuzzy logic technique.
Input Data Eccentricity ratio 0.2 0.2 0.2 0.2 0.4 0.4 0.4 0.4
L/D 0.25 0.5 1 4 0.25 0.5 1 4
Output Data S 7.6107 2.042 0.6383 0.1766 2.85 0.796 0.2623 0.0853
f(R/C) 153.41 41.23 12.94 3.64 61.55 17.32 5.818 2.005
Attitude angle (Degree) 75.33 75.06 73.7 68.64 61.16 61.43 62.44 61.93
P/Pmax 0.4895 0.506 0.5265 0.6140 0.4161 0.4386 0.4804 0.5888
Input Data Eccentricity ratio 0.6 0.6 0.6 0.6 0.8 0.8 0.8 0.8
L/D 0.25 0.5 1 4 0.25 0.5 1 4
Output Data S 1.085 0.320 0.122 0.0496 0.2646 0.0919 0.0446 0.0252
f(R/C) 26.97 8.133 3.2357 1.457 8.91 3.238 1.7021 1.084
Attitude angle (Degree) 46.73 47.87 50.23 53.6 31.2 32.85 36.13 40.49
P/Pmax 0.332 0.364 0.4122 0.533 0.239 0.271 0.316 0.41882

merfeld number, friction variable, attitude angle, and the outputs, various membership functions such as triangular,
maximum-film-pressure ratio as the fuzzy logic program has gauss, and gbell were used. The minimum and maximum limits
the eccentricity ratio and aspect ratio as inputs. According to obtained from the FDM program are used to define the space
Fig. 18, a fuzzy inference system (FIS) consists primarily of of each input and output. Each input and output space is
three parts: part one is called inputs, and in this part, all input divided into several membership functions (MF) according
parameters are represented in a membership function MF to data obtained from the FDM program. When the number
based on data obtained from the FDM program. Part two is of MF goes up, the accuracy of the system’s output goes up.
about the outputs, which can be handled similarly to the Figs. 19 and 20 depict the MF of the inputs and outputs after
inputs. The final section is fuzzy logic rules, which are used applying triangular membership functions to the linguistic
to create relationships between inputs and outputs based on terms defined in Table 5. As the fuzzy rule-base was developed,
database and experience. fuzzy reasoning was linked between the inputs and the outputs.
For each fuzzy rate, twenty-one rules were used to link
4.1. Fuzzy logic structure between input and output as represented in Table 6. During
the defuzzification process, the center of gravity method was
Using the data obtained from the FDM program, as shown used, with all parameters having a weight of one and if then
early in Table 2, the non-additive Mamdani model was imple- criteria were applied for all rules (i.e., for example, rule one:
mented as a fuzzy rule-based model. For both the inputs and if eccentricity ratio ZE and aspect ratio ZE, then Sommerfeld
9020 M.B. Badawi et al.

Fig. 11 For the FFNN, regression of trained data, validation data, testing data, and all three data together.

Fig. 12 Sommerfeld number and friction variable obtained by FDM and the three neural networks.
Application of artificial intelligence for the prediction 9021

Fig. 13 The maximum-film-pressure ratio and attitude angle obtained by FDM and the three neural networks.

Table 3 The results from the three neural networks for the input data aspect ratio 0.3 and eccentricity ratio 0.5.
Performance parameters FDM FFNN RBNN GRNN
Sommerfeld number 1.283 1.32719 1.367385 1.9675
Friction variable 29.44 30.5168 31.97082 44.26
Attitude angle (Degree) 54.03 54.1435 48.15552 53.945
Maximum-film-pressure ratio 0.379 0.377 0.338631 0.374472

Fig. 14 The relation between eccentricity ratio versus Sommerfeld number for FDM and suggested three neural networks at aspect ratio
0.8.
9022 M.B. Badawi et al.

Fig. 15 The relation between eccentricity ratio versus friction variable for FDM and suggested three neural networks at aspect ratio 0.8.

Fig. 16 The relation between eccentricity ratio versus attitude angle for FDM and suggested three neural networks at aspect ratio 0.8.
Application of artificial intelligence for the prediction 9023

Fig. 17 The relation between eccentricity ratio versus maximum-film-ratio for FDM and suggested three neural networks at aspect ratio
0.8.

capability of the fuzzy logic program to predict the bearing


Table 4 RMS error for the predicted performance data using performance, the program was used to predict the output data
NN at aspect ratio 0.8. at values of input as aspect ratio 0.3 and eccentricity ratio 0.5.
RMS error Table 7 shows the output performance parameters from fuzzy
logic and the original values from FDM. Comparing the
Performance parameters FFNN RBNN GRNN
results shows that the fuzzy logic can predict the bearing per-
Sommerfeld number 0.1304 0.325 0.142 formance with minimum errors. Finally, the fuzzy logic pro-
Friction variable 2.53 6.28 2.9 gram predicts the performance at an aspect ratio of 0.8 and
Attitude angle (Degree) 0.965 1.8 0.82
an eccentricity ratio varying from 0.2 to 0.8. Figs. 21-24 show
Maximum-film-pressure ratio 0.0092 0.012 0.0162
the relationship between the eccentricity ratios versus perfor-
mance parameters: Sommerfeld number, friction variable, atti-
tude angle, and maximum-film-pressure ratio for both FDM
number PB, friction variable PVB, attitude angle PB, and and fuzzy logic technique, the RMS errors are represented in
maximum-film-pressure ratio PB). Table 8. The results show the capability of the fuzzy logic tech-
niques to predict the outputs at a computer prediction time of
4.2. Fuzzy logic results 0.1691 s.

To study the effects of changing the MF shape and to obtain 5. Conclusion


the most accurate shape with the fewest errors in predicting
the output performance parameter, three different types of The main aim of the present study is to simulate the perfor-
MF were applied at an aspect ratio of 0.5 and changing eccen- mance of the plain journal bearing for the isothermal case.
tricity ratio from 0.2 to 0.8. The RMS error for predicting the First, a MATLAB program using the finite difference method
Sommerfeld number was 0.179 when the triangular member- (FDM) was built and run at different aspect ratios (0.25, 0.5, 1,
ship function was used for inputs and outputs. By changing and 4) and different eccentricity ratios varying from 0.2 to 0.8.
the triangular membership functions to Gauss and Gbell ones, From FDM, all performance parameters (Sommerfeld num-
RMS errors became 0.182 and 0.181 for Gauss and Gbell, ber, friction variable, attitude angle, and maximum-film-
respectively. Because the triangular membership functions pro- pressure ratio) were calculated. To minimize the computing
duce the most accurate results in predicting the performance of time for bearing performance calculations, two methods of
plain journal bearing, they are used in this study. To study the artificial intelligence were applied. Three neural networks were
9024 M.B. Badawi et al.

Fig. 18 Components of the fuzzy logic program used in predicting the performance parameters.

Fig. 19 Membership functions for inputs to fuzzy logic techniques.

suggested to simulate the performance parameters of the plain 2. The feed-forward only gave a good prediction of the perfor-
journal bearing namely the Feed-forward, Generalized regres- mance for data not included in the training with RMS error
sion, and Radial basis. Fuzzy Logic Techniques were also varying from 0.0092 to 2.53.
applied to get performance parameters. It is concluded here 3. In Fuzzy logic techniques, we can say that increasing the
that: number of membership functions for both inputs and out-
put data improves fuzzy logic program prediction response
1. The three neural networks were able to predict the perfor- and minimizes errors. From the fuzzy logic computer time,
mance parameters (Sommerfeld number, friction variable, it was about 0.2 s, i.e., larger than the neural networks.
attitude angle, and maximum-film-pressure ratio) in a very Also, the RMS errors for predicting the performance vary
small computer time that does not exceed 0.1 s for data from 0.054 to 1.87.
included in the training.
Application of artificial intelligence for the prediction 9025

Fig. 20 Membership functions for outputs to fuzzy logic techniques.

Table 5 Linguistic terms used in MF for the fuzzy logic program.


ZE Zero
PS Positive Small
PM Positive Medium
PB Positive Big
PVB Positive Very Big

Table 6 Rule base for the fuzzy logic program.


Inputs Outputs
Ecc. Ratio Aspect Ratio Sommerfeld No. Friction Variable Attitude angle Pressure Ratio
ZE ZE PB PVB PB PB
ZE PS PM PB PB PM
ZE PS PM PB PB PB
(continued on next page)
9026 M.B. Badawi et al.

Table 6 (continued)
Inputs Outputs
Ecc. Ratio Aspect Ratio Sommerfeld No. Friction Variable Attitude angle Pressure Ratio
ZE PM PS PM PB PB
ZE PB PS PS PB PB
PS ZE PM PB PM PM
PS PS PS PM PM PM
PS PM PS PS PM PM
PS PB ZE ZE PM PB
PM ZE PS PM PS PS
PM ZE PM PM PS PS
PM PS PS PS PS PS
PM PM PS PS PM PM
PM PB ZE ZE PS PM
PM PB ZE ZE PM PB
PB ZE PS PS ZE ZE
PB PS ZE PS ZE ZE
PB PS ZE PS ZE PS
PB PM ZE ZE ZE PS
PB PM ZE ZE PS PS
PB PB ZE ZE PS PM

Table 7 The results from the fuzzy logic technique for the input data aspect ratio 0.3 and eccentricity ratio 0.5.
Performance parameters FDM Fuzzy logic
Sommerfeld number 1.283 1.61
Friction variable 29.44 33.2
Attitude angle (Degree) 54.03 53.8
Maximum-film-pressure ratio 0.379 0.395

Fig. 21 The relation between eccentricity ratio versus Sommerfeld number for FDM and fuzzy logic technique at aspect ratio 0.8.
Application of artificial intelligence for the prediction 9027

Fig. 22 The relation between eccentricity ratio versus friction variable for FDM and fuzzy logic technique at aspect ratio 0.8.

Fig. 23 The relation between eccentricity ratio versus attitude angle (Degree) for FDM and fuzzy logic technique at aspect ratio 0.8.
9028 M.B. Badawi et al.

Fig. 24 The relation between eccentricity ratio versus maximum-film-pressure ratio for FDM and fuzzy logic technique at aspect ratio
0.8.

[3] R.M. Mane, S. Soni, Analysis of hydrodynamic plain journal


Table 8 RMS error for the predicted bearing, Excerpt from the Proceedings of the COMSOL
performance data using fuzzy logic at Conference, Bangalore, India, 2013.
aspect ratio 0.8. [4] M.B. Rajoub, H.A. El-Gamal, E.M. Attia, The performance of
a plain journal textured bearing, Int. J. Eng. Res. Technol.
Performance parameters RMS error (IJERT) 12 (2020) 657–662.
Sommerfeld number 0.15 [5] S. Sharma, R.K. Awasthi, Effect of aspect ratio on the
Friction variable 0.998 performance and stability of hydrodynamic Journal Bearings,
Attitude angle (Degree) 1.87 Int. J. Adv. Res. Innov. 4 (2016) 96–105.
Maximum-film-pressure ratio 0.054 [6] C. Sinanoglu, A.O. Kurban, S.A. Yildirim, Analysis of pressure
variations on journal bearing system using artificial neural
network, Ind. Lubric. Tribol. 56 (2004) 74–87.
[7] M. Esmaeili, A.R. Oskouei, S.A. Mirhadizadeh, W.T. Tse, N.
Hoshyar, Prediction of Hydrodynamic bearing performance
Declaration of Competing Interest based on effective parameters by Neural Network, Int. J. Eng.
Manage. Sci. 7 (2016) 92–99.
The authors declare that they have no known competing [8] S. Kumar, V. Kumar, A.K. Singh, Prediction of maximum
financial interests or personal relationships that could have pressure of journal bearing using ANN with multiple input
appeared to influence the work reported in this paper. parameters, Aust. J. Mech. Eng. 13 (2020) 1–10.
[9] M.B. Badawi, W.A. Crosby, I.M. El Fahham, M.H. Alkomy,
Performance analysis of tilting pad journal bearing using
References COMSOL Multiphysics and Neural Networks, Alexandria
Eng. J. 59 (2) (2020) 865–881.
[1] D.M. Nuruzzaman, M.K. Khalil, M.A. Chowdhury, M.L. [10] A.N. Pelesaraei, R. Bayat, H.H. Bandbafha, H. Afrasyabi, K.
Rahaman, Study on pressure distribution and load capacity of W. Chau, Modeling of energy consumption and environmental
a journal bearing using finite element method and analytical life cycle assessment for incineration and landfill systems of
method, Int. J. Mech. Mechatron. Eng. (IJMME-IJENS) 10 municipal solid waste management-A case study in Tehran
(2010) 1–8. Metropolis of Iran, J. Cleaner Prod. 148 (2017) 427–440.
[2] P. Khan, A. Dhanola, H.C. Garg, A. Khatkar, Influence of [11] A. Rosenkranz, M. Marian, F.J. Profito, N. Aragon, R. Shah,
aspect ratio on the performance characteristics of plain journal The use of artificial intelligence in tribology—a perspective,
bearing lubricating with non-newtonian lubricant, Int. J. Eng. Lubricants 9 (2021) 1–11.
Adv. Technol. (IJEAT) 8 (2019) 1022–1026.
Application of artificial intelligence for the prediction 9029

[12] D.I. Ahmed, S. Kasolang, B.A. Khidhir, B.F. Yousif, Fuzzy [15] H.W. Swift, The stability of lubricating films in journal bearings,
logic based model to predict maximum oil-film pressure in in: Minutes of the Proceedings of the Institution of Civil
journal bearing, Res. J. Appl. Sci., Eng., Technol. 6 (2013) 3871– Engineers, Thomas Telford-ICE Virtual Library 233 (1932) 267–
3878. 288.
[13] W.UR. Rehman, Y. Luo, Y. Wang, G. Jiang, N. Iqbal, S.UR. [16] N.V. Borse, A.M. Parkar, S.P. Chippa, Analysis of
Rehman, S. Bibi, Fuzzy logic–based intelligent control for Hydrodynamic Journal Bearing Including Thermal Effect, in:
hydrostatic journal bearing, Meas. Control 52 (3-4) (2019) 229– Proceedings of TRIBOINDIA-2018 An International
243. Conference on Tribology, Mumbai, India, 2018.
[14] A.KH. EL-Said, B.M. EL-Souhily, W.A. Crosby, H.A. EL- [17] M. Ibn Ibrahimy, R. Ahsan, O.O. Khalifa, Design and
Gamal, The performance and stability of three-lobe journal optimization of levenberg- marquardt based neural network
bearing textured with micro protrusions, Alexandria Eng. J. 56 classifier for EMG signals to identify hand motions, Meas. Sci.
(4) (2017) 423–432. Rev. 13 (3) (2013) 142–151.

You might also like