Professional Documents
Culture Documents
Kha Mari 2021
Kha Mari 2021
https://doi.org/10.1007/s40430-021-02874-0
TECHNICAL PAPER
Received: 19 August 2020 / Accepted: 29 January 2021 / Published online: 26 February 2021
© The Brazilian Society of Mechanical Sciences and Engineering 2021
Abstract
Most studies on dynamic coefficients of bearings are focused on evaluation using different analytical methods. Minimal
emphasis is given to the level of influence of each geometrical variable, the corresponding range of these variables for opti-
mum stiffness and damping and the measure of performance of the analytical method used. The objective of this paper was
to study the influence and sensitivity of length-to-diameter ratio, eccentricity ratio, bearing number, whirl ratio, and bearing
compliance on the stiffness and damping of gas foil bearing. A numerical model is developed by utilizing the finite differ-
ence method to evaluate the dynamic coefficients. The results reveal that the normalized stiffness increases with the bearing
number and decreases with increased bearing compliance whereas the normalized damping shows an opposite nature. Further,
the stiffness coefficients tend to increase and the damping coefficients tend to decrease corresponding to increase in speed
up to 240 krpm. The characteristic data sets obtained from the analysis is used to train an artificial neural network (ANN).
Performance of ANN network is evaluated though computation of root-mean-square error (RSME) and regression coefficient
(R2) and Mean Absolute Error (MAE). Utilizing the neural network results, a Sobol’s sensitivity test is carried out to identify
most effective parameters which have a significant influence on the dynamic coefficients of gas foil bearing. After that, an
adaptive neurofuzzy interface system (ANFIS) is established to determine the optimum range of data for which maximum
stiffness and damping can be obtained. The results deduce that the neural network shows high efficacy in predicting the
output variables correctly with a regression of more than 95%. It is also observed that the variation of dynamic coefficients
is the highest for eccentricity ratio whereas lowest for whirl ratio. The maximum stiffness and damping coefficients are also
obtained for a wide range of geometrical variables which can help in designing the gas foil bearing.
Keywords Dynamic coefficients · Artificial neural network · Sobol’s sensitivity test · Adaptive neurofuzzy interface
system · Gas foil journal bearing
List of symbols E Young’s modulus (N/m 2)
R Radius of the journal (mm) tb Bump foil thickness (mm)
h Film thickness (mm) p Non-dimensional pressure (p/p a)
hb Bump height (mm) C Nominal bearing clearance (mm)
Rb Bump radius (mm) h Non-dimensional film thickness (h/C)
C Radial clearance (mm) t Normalized time variable (ν t)
p Aerodynamic pressure (N/m 2) L/D Length-to-diameter ratio
s Pitch of bump (mm) h0,hx,hy,h0. ,hx. ,hy. Perturbation components of h
P Atmospheric pressure (N/m 2) Δ x, Δ y Normalized perturbations
2 l Length of bump (mm) p0, px, py, p0. , px. , py. Perturbation components of p
K mn Normalized stiffness of bearing
(cK mn ∕pa R2)
Technical Editor: Daniel Onofre de Almeida Cruz. Cmn Normalized damping of bearing
(c𝜔Cmn ∕pa R2)
* Debanshu S. Khamari k Normalized stiffness of the founda-
debanshushekhar@gmail.com
tion ( kc∕pa)
1
Department of Mechanical Engineering, NIT Rourkela, u Normalized foil deflection (u/c)
Rourkela, Odisha 769008, India Fx,Fy Resultant forces in x and y directions
13
Vol.:(0123456789)
167 Page 2 of 19 Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167
W Normalized bearing load (W∕pa R2) damping. So it is a matter of importance to identify the key
z Normalized axial coordinate factors influencing the dynamic coefficients. Furthermore,
θ Circumferential coordinate the optimum range of the design parameters for reliable stiff-
ness and damping of the GFJB must be identified.
Greek symbols There are many methods dealing with the dynamic coef-
Α Compliance number ( pa ∕ck) ficients of GFJB. Perturbation analysis and Simpson’s rule
γ Whirl frequency ratio (ν/ω) of integration are vastly utilized to decide the dynamic coef-
Λ Bearing number (6𝜇𝜔R2 ∕pa c2) ficients of GFJB [1]. Equations of fluid model are generally
Ε Eccentricity ratio paired with the equations of structural model. The equations
μ Lubricant viscosity of the model are finally resolved by using the SOR approach
ν Whirl frequency (Successive Over Relaxation) [2]. Hydrodynamic behavior
ω Rotor angular velocity of GFJB is estimated by utilizing this model because it takes
Acronyms in to account both compressibility as well as compliance.
MLP Multi-layer perceptron The performance is predicted by evaluating the pressure
MF Membership function and load [3]. One of the earliest attempt toward solving the
MAE Mean absolute error GFJB problem is done by Heshmat et al. [4]. The finite dif-
RMSE Root-mean-squared error ference formulations are evaluated using Newton Raphson
approach for calculating both collinear as well as cross cou-
pled dynamic coefficients. Both single-pad and multipad
1 Introduction structure are used for the analysis to calculate the effect of
operating variables on bearing behavior. Later, Kim et al.
Gas Foil Journal Bearings (GFJBs) are self-acting hydrody- [5] considered the air foil bearing of four different types and
namic bearings which are used to overcome the limitations analyzed their stiffness and damping. The investigation leads
of oil lubricated bearings like contamination of oil particles to the conclusion that stiffness variation of bump foil does
and low stiffness and damping characteristics. There are also not significantly affect the load capacity. After that, Paulsen
several merits of GFJBs over conventional rigid bearings et al. [6] determined the dynamic coefficients and load by
such as high load carrying capacity, low friction at steady solving Reynold’s equation. The dynamic coefficients are
operation, and endurance to foreign matter. GFJBs basically also evaluated.
contains a bump foil and top foil over it which creates an Later, Howard et al. [7] tested foil bearing steady state
elastic structure and behaves as a spring bed responsible for stiffness with increasing temperature and observed that
increasing bearing stiffness as shown in Fig. 1. The dynamic the stiffness is decreased by a factor of two with respect to
coefficients of GFJB rely on behavior of the gas film as well increasing temperature. Carpino et al. [8] modelled a finite
as foil structure. Proper prediction of dynamic coefficients element approach in which the film thickness, foil deflection
of the GFJB accounts for the rotor stability and performance. and gas pressure are perturbed to evaluate the dynamic coef-
Both its structure and used material makes the prediction ficients. Further, Lamhar et al. [9] determined the dynamic
process of these coefficients more complicated as there coefficients for certain frequency of excitation by using the
are various parameters which influences the stiffness and first-order perturbation. Outcomes reveal dependence of
excitation frequency rises with rise in dynamic coefficients.
13
Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167 Page 3 of 19 167
Later, Rao et al. [10] developed expressions for more sophis- network to predict the performance of the model. Numeri-
ticated calculation of stiffness and damping and studied the cal models used for gas bearings always require an efficient
effect of dynamic coefficients and speed for different ratios iterative method to satisfy the order of accuracy and fast
of length and diameter. Feng et al. [11] presented numerical convergence rate. As a result of which the analysis requires
model of a GFJB by taking in to account the dynamic forces a large quantity of accurate data which can never be deter-
in the structure of the foil. Springs are used to replicate every mined with complete accuracy. Furthermore, the computa-
bump. Results show that the frequency and friction factor tional resources can hardly entertain the increasing refine-
are the major factors influencing the dynamic characteristics ment and complexity of the numerical analysis. For the
of the GFJB. Recently, Larsen et al. [12] investigated the studies on neural networks for gas bearings, the literature
mechanical bump foil behavior at different friction coeffi- is mainly focused on speed as output parameter. The cur-
cients. The numerical model also showed that stiffness is rent work using ANN and ANFIS is focused on length-to-
underestimated in analytical model. Moreover, Larsen et al. diameter ratio, eccentricity ratio, bearing number, whirl ratio
[13] solved an equation of zeroth-order using Successive and bearing compliance as input parameters and stiffness
under relaxation approach. The pressure profiles are ana- and damping as output parameters. As neural networks cali-
lyzed by implicitly considering the deformations in the lay- brate a given random vector of inputs with corresponding
ers of the compliant foil. Further, Rubio et al. [14] deduced output, it is more of a data driven technique without neces-
that the structural stiffness obtained from two identical bear- sarily involving physics of the entire process. But such an
ings and found an increasing nature of the structural stiff- approach alone based entirely on data-driven pattern has
ness with the increasing foil deflection. Later, Le Lez et al. not received wide acknowledgement from the community
[15] developed the foil bearing model corresponding to set who traditionally believes in the physics-based approach. So
of interacting bumps of multi-degree of freedom. Further, we have therefore tried to combine both numerical as well
effect of structure of the foil bearing on dynamic coefficients as data driven techniques in this work for prediction of gas
is observed. Feng et al. [16] examined a number of param- bearing parameters for a wide range of variables. The range
eters such as load capacity, flow rate, mid-plane pressure, of these parameters not only helps in improvement of design
dynamic coefficients corresponding to the eccentricity ratio, of the bearing but also for fabrication aspects of the bearing
and speed. It is found that with increase in speed, the stiff- to carry out experimental investigation. This paper further
ness increases whereas change in the damping is very mar- aims to clinch the literature gap by conducting a Sobol’s
ginal. Guo et al. [17] calculated the dynamic coefficients sensitivity test to identify major non-dimensional variables
of GFJB using the perturbation method corresponding to which influence the stiffness and damping of the gas bearing.
certain nominal clearances. Nonlinear rotordynamic behav- In this paper, the dynamic coefficients of the gas foil bear-
ior is also predicted by using a loss factor model. Recently, ing are predicted using a finite difference method. Behavior
Lai et al. [18] reviewed the performance of a high-speed of normalized stiffness and damping is observed with the
turboexpander. Influence of dynamic coefficients on the change in bearing number and bearing compliance. Further,
rotor behavior are also presented for reliable operation of the effect on changing stiffness and damping coefficients
the turbomachinery. is estimated up to a speed of 240krpm. A neural network
Good amount of work is done based upon various numer- is simulated to evaluated the performance of the model.
ical models but little importance is given to the use of neural After that, a sensitivity test is conducted to estimate the
network in the field of bearings. Hasan et al. [19] reported most sensitive parameters among the parameters length-
the training of a neural network to analyze the overall numer- to-diameter ratio, eccentricity ratio, bearing number, whirl
ical GFJB model. The model is subjected to different input ratio and bearing compliance. Then, a fuzzy interface system
data sets over a wide range of speeds. It is found that the is designed to determine the range of these parameters for
neural network successfully identifies the numerical model maximum stiffness and damping.
of the GFJB through different validation studies. Qin et al.
[20] demonstrated the training of neural network model and
studied the motion aspect of a rigid rotor. From the compu- 2 Artificial neural network (ANN)
tational results, motion behavior of neural network is found
to be identical to that of numerical model. Manoj et al. [21] ANN is a simulation and predicting tool designed on the
recently have reported the use of ANN and ANFIS model basis of biological neural network of human brain. The net-
based on the geometrical data sets to predict and maximize work comprises of many interconnecting entities known as
the efficiency of the turbine for a cryogenic turboexpander. neurons (nodes or processing units) connecting linear and
A detailed literature review reveals that there is huge nonlinear functions. ANN is a type of modelling tool which
amount of work available on numerical methods to pre- converts nonlinear functions into input and output during
dict the dynamic coefficients, but none of them use neural training. The feed forward types multi-layer perceptron
13
167 Page 4 of 19 Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167
13
Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167 Page 5 of 19 167
( ( ) )
h0 = 1 + 𝜀cos𝜃 + 𝛼p0 (6) 𝜕 3 𝜕 p0 py 𝜕p0
h + 3p0 h0 hy +
𝜕𝜃 0 𝜕𝜃 𝜕𝜃
First-order equations ( ( ) )
( ( ) ) ( )2 3 𝜕 p0 py 𝜕p0
R 𝜕
𝜕 3 𝜕 p0 px 𝜕p0 + h + 3p0 h0 hy +
h + 3p0 h0 hx + L 𝜕z 0 𝜕z 𝜕z
𝜕𝜃 0 𝜕𝜃 𝜕𝜃 ( )
( ( ) ) 𝜕 p0 hy + py h0 ( )
( )2 3 𝜕 p0 px 𝜕p0
R 𝜕 =Λ − 2Λ𝛾 p0 hẏ + pẏ h0
+ h + 3p0 h0 hx + 𝜕𝜃 (9)
L 𝜕z 0 𝜕z 𝜕z
( )
𝜕 p0 hx + px h0 ( ) hy = 𝛼py − cos𝜃 (10)
=Λ − 2Λ𝛾 p0 hẋ + pẋ h0
𝜕𝜃 (7)
13
167 Page 6 of 19 Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167
∫ ∫
Fx sin𝜃
Parameter Value =− p Rd𝜃dz (18)
Fy −cos𝜃
0 0
Network type MLP
Training function TRAINLM The stiffness and damping are given as:
Number of hidden layers 1 { } ( )
Number of hidden neurons 10 K xx K xy 1 Kxx Kxy
=
Training algorithm Back propagation K yx K yy pa L Kyx Kyy
Number of epochs 1000 ( ) (19)
1 2𝜋
R ∫0 ∫ 0
L px sin𝜃 py sin𝜃
=− d𝜃dz
−px cos𝜃 −py cos𝜃
( ( ) )
𝜕 3 𝜕 p0 pẋ 𝜕p0
h + 3p0 h0 hẋ + { } ( )
𝜕𝜃 0 𝜕𝜃 𝜕𝜃 Cxx Cxy 𝜔 Cxx Cxy
( ( ) ) =
( )2 3 𝜕 p0 pẋ 𝜕p0 Cyx Cyy pa L Cyx Cyy
R 𝜕
+ h + 3p0 h0 hẋ + 1 2𝜋 ( )
L 𝜕z 0 𝜕z
R ∫0 ∫0
𝜕z L pẋ sin𝜃 pẏ sin𝜃
( ) =− d𝜃dz
−pẋ cos𝜃 −pẏ cos𝜃 (20)
𝜕 p0 hẋ + pẋ h0 ( )
=Λ
𝜕𝜃
+ 2Λ𝛾 p0 hẋ + pẋ h0
(11) The numerical procedure begins after dividing the bear-
ing surface into number of grids (Δ θ × Δ z). MATLAB
programming tool is used to calculate the pressure at each
hẋ = 𝛼pẋ (12) point. The solution starts with solving the equations first by
( ) discretization then calculation begins with the use of itera-
( )
𝜕 3 𝜕 p0 pẏ 𝜕p0 tive method know as successive over relaxation method. The
h + 3p0 h0 hẏ + convergence found in order of accuracy is maintaining by
𝜕𝜃 0 𝜕𝜃 𝜕𝜃
( ( ) ) the 0.0001% of error. The perturbed pressures are found out
( )2
R 𝜕 3 𝜕 p0 pẏ 𝜕p0 at each grid point, the perturbed pressures are integrated
+ h + 3p0 h0 hẏ +
L 𝜕z 0 𝜕z 𝜕z using Simpson’s rule. The integration gives the values of
( ) dynamic coefficients.
𝜕 p0 hẏ + pẏ h0 ( ) The output data from the numerical model is given as
=Λ + 2Λ𝛾 p0 hẏ + pẏ h0 input to the ANN. A target space is generated for predicting
𝜕𝜃 (13)
the new dynamic coefficients.
The ANN model is developed using five input param-
hẏ = 𝛼pẏ (14) eters, i.e., L/D (length-to-diameter ratio), ε, Λ, γ & α and
The boundary conditions are given as: different output variables, i.e., normalized bearing stiffness
( K xx,K xy,K yx,K yy) and damping ( Cxx,Cxy,Cyx,Cyy). The results
( ) ( )
p0 z = 0 = p0 z = 1 = 1 obtained from the finite difference model is used to develop
px = py = pẋ = pẏ = 1 (15) the combined model. The neural network is created with the
training function as: Trainlm. The literature review suggests
( ) that single hidden layer is enough for ’universal approxima-
p0 𝜃 = 𝜃1 = 1
( ) tion property’, but requires a higher number of neurons [23].
h0 𝜃 = 𝜃1 = 1 + 𝜀cos𝜃1 So one hidden layer and ten neurons are taken. Typically,
( ) ( ) ( ) ( )
px 𝜃 = 𝜃1 = py 𝜃 = 𝜃1 = pẋ 𝜃 = 𝜃1 = pẏ 𝜃 = 𝜃1 = 1 many epochs in order of thousands at a time, are required
( ) ( ) to train the neural network efficiently. Also, three transfer
hx 𝜃 = 𝜃1 = sin𝜃1 hy 𝜃 = 𝜃1 = cos𝜃1 functions are used to compare predicted results, i.e., Losgig,
( ) ( )
hẋ 𝜃 = 𝜃1 = hẏ 𝜃 = 𝜃1 = 0 Purelin, and Tansig. The other details regarding training of
(16) neural network is shown in Table 1.
( ) ( )
p0 𝜃 = 𝜃1 + 2𝜋 = 1px 𝜃 = 𝜃1 + 2𝜋 = py 3.2 Preliminary testing of the numerical model
( ) ( ) ( )
𝜃 = 𝜃1 + 2𝜋 = pẋ 𝜃 = 𝜃1 + 2𝜋 = pẏ 𝜃 = 𝜃1 + 2𝜋 = 1
(17) The data presented in Table 2 is used to test the numerical
The bearing forces is found by integrating the pressure model and predict the behavior of the fluid film.
over the bearing surface and is given as:
13
Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167 Page 7 of 19 167
13
167 Page 8 of 19 Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167
compliance. This nature corresponds to the increase in the is negative, i.e., the journal velocity in the y direction always
deflection of the foil with the compliance. It is also seen that supplies energy, rather than dissipating energy, to the motion
the stiffness increase with the bearing number and attains a in the x (loading) direction. The other three components of
steady value at higher bearing number. This trend is because the damping are positive, so they dissipate energy from the
at low speeds or small bearing numbers, the compliance of system.
the bearing depends primarily on the lubricant film which is Figure 10 shows the change in stiffness coefficients cor-
relatively soft compared to the stiffness of the foil. So when responding to increase in speed up to 240 krpm. Here it
the bearing number is low, the stiffness tends to approach is observed that with the increase in speed, the stiffness
the same value independent of the value of bearing compli- increases promptly owing to the larger film thickness. While,
ance, whereas, at high speeds or large bearing numbers, the rise of stiffness with the increase in speed decreases gradu-
stiffness of the gas film is large compared to the stiffness ally. It may stabilize when the speed attains a higher value.
of the foil and the compliance of the bearing is primarily It can be described as with the increase in speed, the film
dependent on the foil. This effect is observed as the values of of the lubricant deliberately gets stiffer. Further, the bearing
the bearing stiffness are relatively constant for larger bearing compliance significantly depends on the bump foil acting as
compliances. the elastic foundation.
Figure 9 depicts that there is a decreasing trend in the The damping coefficients are also calculated for the same
normalized damping as the bearing number rises. The trend varying speeds as shown in Fig. 11. It depicts that as the
becomes constant for larger bearing number. Further the speed increases, all the coefficients of damping tends to
damping values tend to increase with the increasing bear- decline except Cxy . Stiffer lubricant restricts the dissipation
ing compliance. The results depict that for the cross-coupled of energy in the film of the gas, and hence decreasing coef-
coefficients of damping are contradictory in nature. Cxy term ficients of damping at greater speed of rotation.
Fig. 8 Effect of bearing compliance on normalized bearing stiffness a K xx,b K yy,c K xy,d K yx for L/D = 1, ε = 0.5
13
Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167 Page 9 of 19 167
Fig. 9 Effect of bearing compliance on normalized bearing damping a Cxx,b Cyy,c Cxy,d Cyx for L/D = 1, ε = 0.5
Fig. 10 Variation of stiffness coefficients with rotational speed Fig. 11 Variation of damping coefficients with rotational speed
13
167 Page 10 of 19 Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167
The present paper deals with Sobol’s sensitivity analysis where D is the total variance of the output functions. Hence,
for GFJB to distinguish the variables having larger influence the sensitivity indices are defined as:
on stiffness and damping. Di1,… is
With a view to interpret how the output variance be Si1,… is = (26)
D
attributed to individual input variables, sensitivity indi-
ces are calculated to accurately reflect the influence of the For instance, Si = Di is the first-order sensitivity index of
D
13
Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167 Page 11 of 19 167
Table 3 ANFIS network design indices are calculated by dividing each with D as given in
Eq. 26. This concludes the sensitivity test for one output
MFs type of input parameter Gbell,
Gaussmf and K xx . Utilizing the same procedure, sensitivity analysis for
Gauss2mf stiffness ( K xy,K yx,K yy ) and damping ( Cxx,Cxy,Cyx,Cyy ) are
also evaluated.
MF type of output parameter Linear
While this test indicates the most influencing parameters,
Number of MFs 3
it is also important to calibrate the optimum range of these
Optimization method Hybrid
parameters for better performance of GFJB. The optimised
Error tolerance 0.0001
range of these parameters is evaluated by using artificial
Number of epochs 100
13
167 Page 12 of 19 Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167
13
Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167 Page 13 of 19 167
0.75 0.6 0.9 0.4 1 1.135313 0.157581 0.231555 0.974566 0.559593 − 0.288124 0.274568 0.574568
0.75 0.7 1.3 0.4 1 1.393524 0.131859 0.301246 1.084512 0.523298 − 0.275124 0.254685 0.551243
1 0.625 2.3 0.5 0.7 1.679026 0.269772 0.456489 1.594561 0.482398 − 0.254123 0.236598 0.478123
1 0.725 2.7 0.5 0.7 1.819124 0.247282 0.521345 1.591245 0.458581 − 0.214556 0.201325 0.484512
1.25 0.525 3.2 0.6 0.5 1.739304 0.427183 0.594623 1.645712 0.479065 − 0.224513 0.1912457 0.451236
1.25 0.8 4.3 0.6 0.5 2.136474 0.319473 0.631458 1.801246 0.418252 − 0.2014578 0.184523 0.443135
1.5 0.675 5.1 0.7 0.3 2.230287 0.462737 0.684663 1.858711 0.407005 − 0.175896 0.1336545 0.351247
1.5 0.7 5.2 0.7 0.3 2.261400 0.441717 0.6813466 1.869746 0.401848 − 0.165239 0.135213 0.39124
1.75 0.5 5.7 1 0.1 2.239615 0.619462 0.711253 2.17322 0.416539 − 0.155789 0.117268 0.354711
1.75 0.575 6 1 0.1 2.430642 0.591673 0.721345 2.214567 0.387169 − 0.145236 0.124623 0.294578
13
167 Page 14 of 19 Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167
13
Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167 Page 15 of 19 167
0.75 0.6 0.9 0.4 1 1.14 0.16 0.259 0.903 0.552 −0.284 0.272 0.546
0.75 0.7 1.3 0.4 1 1.23 0.145 0.301 1.08 0.554 −0.294 0.264 0.559
1 0.625 2.3 0.5 0.7 1.55 0.272 0.463 1.6 0.488 −0.244 0.224 0.484
1 0.725 2.7 0.5 0.7 1.94 0.277 0.516 1.64 0.458 −0.241 0.21 0.466
1.25 0.525 3.2 0.6 0.5 1.81 0.381 0.577 1.7 0.442 −0.214 0.199 0.442
1.25 0.8 4.3 0.6 0.5 2.22 0.297 0.664 1.73 0.43 −0.221 0.189 0.454
1.5 0.675 5.1 0.7 0.3 2.19 0.492 0.697 1.91 0.416 −0.184 0.138 0.386
1.5 0.7 5.2 0.7 0.3 2.22 0.488 0.7 1.94 0.405 −0.183 0.136 0.379
1.75 0.5 5.7 1 0.1 2.39 0.591 0.748 2.1 0.41 −0.169 0.109 0.321
1.75 0.575 6 1 0.1 2.48 0.594 0.751 2.29 0.354 −0.158 0.11 0.291
gaussmf
0.75 0.6 0.9 0.4 1 1.11 0.158 0.255 0.9 0.557 −0.282 0.275 0.538
0.75 0.7 1.3 0.4 1 1.18 0.144 0.3 1.086 0.546 −0.296 0.266 0.553
1 0.625 2.3 0.5 0.7 1.53 0.27 0.465 1.62 0.484 −0.247 0.222 0.487
1 0.725 2.7 0.5 0.7 1.89 0.274 0.516 1.641 0.476 −0.231 0.215 0.463
1.25 0.525 3.2 0.6 0.5 1.82 0.38 0.574 1.701 0.454 −0.211 0.192 0.439
1.25 0.8 4.3 0.6 0.5 2.17 0.295 0.665 1.73 0.415 −0.228 0.175 0.448
1.5 0.675 5.1 0.7 0.3 2.2 0.496 0.692 1.912 0.408 −0.186 0.127 0.382
1.5 0.7 5.2 0.7 0.3 2.23 0.484 0.7 1.941 0.399 −0.173 0.135 0.377
1.75 0.5 5.7 1 0.1 2.39 0.59 0.741 2.15 0.408 −0.166 0.104 0.327
1.75 0.575 6 1 0.1 2.48 0.594 0.75 2.289 0.363 −0.154 0.115 0.295
gauss2mf
0.75 0.6 0.9 0.4 1 1.28 0.149 0.247 0.9 0.569 −0.278 0.288 0.55
0.75 0.7 1.3 0.4 1 1.43 0.141 0.3 1.078 0.52 −0.286 0.26 0.553
1 0.625 2.3 0.5 0.7 1.67 0.27 0.464 1.602 0.461 −0.245 0.23 0.487
1 0.725 2.7 0.5 0.7 1.87 0.276 0.515 1.641 0.405 −0.24 0.217 0.464
1.25 0.525 3.2 0.6 0.5 1.42 0.38 0.572 1.71 0.485 −0.213 0.196 0.44
1.25 0.8 4.3 0.6 0.5 1.95 0.297 0.664 1.729 0.464 −0.226 0.185 0.457
1.5 0.675 5.1 0.7 0.3 2.24 0.489 0.692 1.907 0.401 −0.175 0.133 0.386
1.5 0.7 5.2 0.7 0.3 2.26 0.482 0.701 1.942 0.393 −0.18 0.136 0.373
1.75 0.5 5.7 1 0.1 2.17 0.59 0.742 2.11 0.396 −0.165 0.109 0.32
1.75 0.575 6 1 0.1 2.48 0.596 0.749 2.288 0.369 −0.157 0.11 0.29
13
167 Page 16 of 19 Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167
Fig. 17 Effect of a L/D and ε, (b)L/D and Λ, (c)L/D and γ, (d)L/D and α
the combined model for stiffness and damping as shown in The performance of the model is estimated from Table 5.
Fig. 14. In all the transfer functions used, the R 2 is found to be more
It is observed that the data set of the combined model than 95% which concludes the performance of the combined
shows a deviation when there is clear spike or decline in the model is satisfactory. Further the MAE and RMSE values
data set of the numerical model. Otherwise the combined are found to be less than 0.1 which is generally considered
model is well in agreement with the numerical model. to be acceptable. It is also clear that the ‘Purelin’ transfer
function has the least amount of error.
6.2 Performance of the combined model
6.3 Significant variables influencing stiffness
Hereafter, taking in to account different combinations, a and damping
large amount of data sets is calibrated to get a clear picture
of the performance of the combined model. The results are The sensitivity indices of all the input variables are calcu-
also compared with the numerical model. The regression lated and shown in Fig. 16.
curves are drawn to estimate the curve fit between the input It is found that L/D, ε, Λ and α have significant effect on
and the output as shown in Fig. 15. The numerical model the bearing stiffness and damping whereas γ has minimal
results are given in Table 4. The data predicted through the effect as compared to other parameters. Hence, it can be
combined model using transfer functions Logsig, Purelin concluded that change in γ does not result much change in
and Tansig are shown in Table 5. Mean square error (RSME) the dynamic coefficients. Further, it is observed that ε is the
and regression coefficient (R2) and Mean Absolute Error most sensitive parameter among these. Change in a small
(MAE) are calibrated to estimate the performance of the fraction of ε results in a significant change in the value of
model. dynamic coefficients.
13
Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167 Page 17 of 19 167
13
167 Page 18 of 19 Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167
7 Conclusions
13
Journal of the Brazilian Society of Mechanical Sciences and Engineering (2021) 43:167 Page 19 of 19 167
Compliance with ethical standards 13. Larsen JS, Santos IF (2015) Efficient solution of the non-linear
Reynolds equation for compressible fluid using the finite element
method. J Braz Soc Mech Sci Eng 37(3):945–957
Conflict of interest The authors declared that there is no conflict of
14. Rubio D, Andrés LS (2006) Bump-type foil bearing structural
interest to any person or organisation.
stiffness: experiments and predictions. J Eng Gas Turbines Power
128(3):653–660
15. Lez Le Sb,Arghir M and Frene J (2007) A new bump-type foil
References bearing structure analytical model. in Turbo Expo: Power for
Land, Sea, and Air
16. Feng H, Jiang S, Ji A (2019) Investigations of the static and
1. Peng J-P, Carpino M (1993) Calculation of stiffness and damping
dynamic characteristics of water-lubricated hydrodynamic jour-
coefficients for elastically supported gas foil bearings. ASME J
nal bearing considering turbulent, thermohydrodynamic and mis-
Tribol 115(1):20–27
aligned effects. Tribol Int 130:245–260
2. Pletcher RH, Tannehill JC, Anderson D (2012) Computational
17. Guo Z et al (2018) Measurement and prediction of nonlinear
fluid mechanics and heat transfer. CRC press, pp 155–156
dynamics of a gas foil bearing supported rigid rotor system. Meas-
3. Peng Z-C, Khonsari M (2004) Hydrodynamic analysis of
urement 121:205–217
compliant foil bearings with compressible air flow. J Trib
18. Lai T et al (2018) Numerical and experimental studies on stability
126(3):542–546
of cryogenic turbo-expander with protuberant foil gas bearings.
4. Heshmat H, Walowit J, Pinkus O (1983) Analysis of gas-lubri-
Cryogenics 96:62–74
cated foil journal bearings. J Lubr Technol 105(4):647–655
19. Hassan MFB, Bonello P (2017) A neural network identifica-
5. Kim D (2007) Parametric studies on static and dynamic per-
tion technique for a foil-air bearing under variable speed condi-
formance of air foil bearings with different top foil geom-
tions and its application to unbalance response analysis. J Tribol
etries and bump stiffness distributions. J Trobol Trans ASME
139(2):021501
129(2):354–364
20. Qin P et al (2005) Dynamic analysis of hydrodynamic bear-
6. Paulsen BT, Morosi S, Santos IF (2011) Static, dynamic, and ther-
ing–rotor system based on neural network. Int J Eng Sci
mal properties of compressible fluid film journal bearings. Tribol
43(5–6):520–531
Trans 54(2):282–299
21. Kumar M et al (2019) Experimental investigation and perfor-
7. Howard SA et al (2001) Steady-state stiffness of foil air journal
mance prediction of a cryogenic turboexpander using artificial
bearings at elevated temperatures. Tribol Trans 44(3):489–493
intelligence techniques. Appl Therm Eng 162:114273
8. Carpino M, Talmage G (2006) Prediction of rotor dynamic coef-
22. Haykin S (2010) Neural networks and learning machines. Pearson
ficients in gas lubricated foil journal bearings with corrugated
Education, India
sub-foils. Tribol Trans 49(3):400–409
23. Kumar M et al (2019) Preliminary design, flow field, and ther-
9. Lahmar M, Ellagoune S, Bou-Saïd B (2010) Elastohydrodynamic
mal performance analysis of a helium turboexpander: a numerical
lubrication analysis of a compliant journal bearing considering
approach. SN Appl Sci 1(11):1482
static and dynamic deformations of the bearing liner. Tribol Trans
24. Zhang XY et al (2015) Sobol sensitivity analysis: a tool to guide
53(3):349–368
the development and evaluation of systems pharmacology models.
10. Rao T et al (2000) An analytical approach to evaluate dynamic
CPT Pharmacomet Syst Pharmacol 4(2):69–79
coefficients and nonlinear transient analysis of a hydrodynamic
25. Sobol IM (2001) Global sensitivity indices for nonlinear math-
journal bearing. Tribol Trans 43(1):109–115
ematical models and their monte carlo estimates. Math Comput
11. Feng K, Guo Z (2014) Prediction of dynamic characteristics of a
Simul 55(1–3):271–280
bump-type foil bearing structure with consideration of dynamic
friction. Tribol Trans 57(2):230–241
12. Larsen JS, Varela AC, Santos IF (2014) Numerical and experi- Publisher’s Note Springer Nature remains neutral with regard to
mental investigation of bump foil mechanical behaviour. Tribol jurisdictional claims in published maps and institutional affiliations.
Int 74:46–56
13