This action might not be possible to undo. Are you sure you want to continue?
Design of Crankshaft Main Bearings under
Uncertainty
Zissimos P. Mourelatos
Mechanical Engineering Department
Oakland University
Rochester, MI 48309, USA
ABSTRACT
A probabilistic analysis is presented for studying the variation effects on the main bearing
performance of an I.C. engine system, under structural dynamic conditions. The analysis is
based on surrogate models (metamodels), which are developed using the kriging method.
The metamodels provide an efficient and accurate substitute to the actual engine bearing
simulation models. The bearing performance is based on a comprehensive engine system
dynamic analysis which couples the flexible crankshaft and block dynamics with a detailed
main bearing elastohydrodynamic analysis. The clearance of all main bearings and the oil
viscosity comprise the random design variables. Probabilistic analyses are performed to
calculate the mean, standard deviation and probability density function of the bearing
performance measures. A ReliabilityBased Design Optimization (RBDO) study is also
conducted for optimizing the main bearing performance under uncertainty. Results from a V6
engine are presented.
keywords: I.C. engines, crankshaft, main bearing elastohydrodynamic analysis, engine
system dynamics, surrogate modeling, design under uncertainty, reliabilitybased design
optimization
1. INTRODUCTION
This paper presents a probabilistic analysis of the main bearing lubrication
performance of an operating internal combustion engine. Surrogate models (metamodels)
are developed for critical lubrication performance measures based on a detailed dynamic
engine simulation solver which couples the structural dynamics of the crankshaft and block
with detailed main bearing elastohydrodynamic behavior. The Kriging method [1] is used to
generate the metamodels based on a limited number of sample points. Probabilistic analyses
are first performed to calculate the main bearing statistical performance in terms of the mean,
standard deviation and probability density function of defined bearing performance
measures. Subsequently, a probabilistic sensitivity analysis is described for identifying the
important random variables. Finally, a ReliabilityBased Design Optimization (RBDO) [2,3]
study is conducted for optimizing the main bearing performance under uncertainty and
results from a V6 engine are presented.
A significant amount of work in the area of elastohydrodynamic (EHD) analysis of
connecting rod bearings has been reported in the literature. An integrated system level
operating V6 engine simulation model, consisting of flexible crankshaft and engine block
dynamics model coupled by an efficient elastohydrodynamic bearing lubrication solver has
been presented in [4,5]. A detailed coupling of the crankshaft rigid and flexible body
dynamics [6] was used. The work in [4,5] is employed in this paper for calculating the
performance measures used in the metamodel generation.
An operating V6 engine represents a complicated nonlinear system which is affected
by variation in manufacturing processes, operating conditions, material properties, etc. For
this study, variability is introduced in some engine design variables and a probabilistic
analysis is preformed for the main bearing performance. The clearance at each main bearing
2
and the oil viscosity comprise the random variables. All random variables are assumed
normally distributed. The maximum oil film pressure and the percentage of time (the time
ratio) within each cycle that a bearing operates with film thickness lower than a user defined
threshold value are defined as the performance measures. Eight performance measures are
considered. A probabilistic sensitivity analysis is also performed for identifying the most
important design variables and for determining the degree to which these design variables
influence the performance.
ReliabilityBased Design Optimization (RBDO) [2,3] provides an optimum solution in
the presence of uncertainty. A mean performance measure is usually optimized subject to
the probability of satisfying a constraint being greater than a prescribed reliability level.
Deterministic optimal designs that are obtained without considering uncertainty are usually
unreliable. In contrast, input variation is fully accounted for in RBDO through probability
distributions which describe the stochastic nature of design variables and model parameters
[2,3,7,8]. In this paper, an RBDO study of the main bearing lubrication performance is
presented based on a newly developed singleloop RBDO algorithm [8].
2. DEVELOPMENT OF A SYSTEMLEVEL ENGINE BEARING SIMULATION MODEL
The integrated systemlevel engine bearing simulation model developed in [46] is
used. This model consists of a flexible crankshaft model and a flexible engine block model
connected by a detailed elastohydrodynamic lubrication model. A brief description is given in
this section.
The CraigBampton method [9] is employed for reducing the physical Degrees of
Freedom (DOF) for both the crankshaft and the engine block. The reduced dynamic
equations of motion are expressed as
{ }
i
i
r
i
i
i
r
i
i
i
r
i
i
F
x
K
x
C
x
M
α α α
; i=b,c (1)
where M , C and K are the reduced mass, damping and stiffness matrices
respectively, and
F
is the modal force vector. The superscripts “b” and “c” denote the
engine block and crankshaft, respectively. Details are given in [4,6].
The flexible crankshaft and flexible block are interacting through a set of distributed
nonlinear springs and dampers which are represented by a set of stiffness and damping
matrices for each bearing. The stiffness and damping matrices are determined through a
detailed lubrication model [5].
The lubricating oil film pressure distribution for each journal bearing is described by
the following Reynolds equation,
h
x
h
R
x
P
h
x z
P
h
z
12 6
3 3
(2)
where
) , , ( t z x P P
is the oil film hydrodynamic pressure and
) , , ( t z x h h
is the lubricant film
thickness. Figure 1 shows the used notation. Eq. (2) is discretized and solved numerically for
the hydrodynamic pressure field subject to an imposed known pressure at the two bearing
ends and along the oil grooves. The Reynolds cavitation condition is used to account for oil
cavitation.
3
Figure 1. Journal bearing notation
A linear perturbation approach is used to solve Eq. (2). At each time step, the film
thickness and squeeze film thickness distributions are assumed equal to
+ =
0
h h
and
0
h h
, respectively, where o
h
and
o
h
are the distributions from the previous time step
and and are their corresponding perturbations. In this case, the pressure distribution is
approximated as
h
P
h
P
P P
o , (3)
where the terms
h
P
K] [
and
h
P
C] [
constitute the oil film stiffness and damping matrices,
respectively.
Under the assumption of small perturbation values for and , Eq. (2) yields the
following perturbation equations [4]
= 
.

\

∂
∂
∂
∂
+ 
.

\

∂
∂
∂
∂
o o
P
x
h
x
P
z
h
z
3 3
o
o
h
x
h
R 12 6
(4)
= 
.

\

∂
∂
∂
∂
+ 
.

\

∂
∂
∂
∂
δ δ ] [ ] [
3 3
K
x
h
x
K
z
h
z
− 
.

\

∂
∂
∂
∂
−
∂
∂
o o
P
z
h
z x
R
2
3 6

.

\

∂
∂
∂
∂
o o
P
x
h
x
2
3
(5)
. 12 ] [ ] [
3 3
C
x
h
x
C
z
h
z
(6)
Equations (4), (5) and (6) constitute the governing equations for the hydrodynamic oil
film pressure distribution, stiffness matrix, and damping matrix, respectively. They are all
solved numerically using a finite difference method. Details can be found in [4,5].
If the oil film domain is discretized with N nodes, the explicit form of the stiffness
matrix is
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
∂
=
N
N N N
N
N
h
p
h
p
h
p
h
p
h
p
h
p
h
p
h
p
h
p
K
2 1
2
2
2
1
2
1
2
1
1
1
. (7)
A similar expression holds for the damping matrix
C
. The ij element of the stiffness
(damping) matrix represents the pressure change in node i due to a unit oil film thickness
(squeeze film velocity) change at node j. The size of the fully populated oil film stiffness and
damping matrices is large in order to accurately capture the highly nonlinear oil film pressure
distribution, particularly for high journal eccentricity and journal misalignment conditions. For
4
computational efficiency reasons, a diagonalization procedure of the oil film stiffness and
damping matrices has been implemented [5]. The oil film stiffness and damping matrices are
used in the system level equations between the crankshaft and the block.
For the solution of Eqs (4) through (6), the oil film thickness distribution
t z x h , ,
is
needed. It is derived as a function of the generalized coordinates of the block b
x
and the
crankshaft c
x
as
{ }  { }
b b c c
x T x T c t z x h ) , , (
(8)
where c
T
and b
T
are appropriate transformation matrices [4]. The magnitude of the
interaction forces {Q} created between the flexible block and the flexible crankshaft, is a
function of the oil film thickness (Eq. (8)) and the stiffness (Eq. (7)) and damping matrices.
These interaction forces provide the coupling mechanism between the two components.
In general, the system of reduced dynamic equations for the combined crankshaft
and block are written as
{ }  { }  { } { } { } Q Q q K q C q M
e
(9)
where
q
are the generalized coordinates. The reduced matrices M , C and K , and the
reduced generalized external forces vector e
Q
are given in [9]. The expression for the
reduced vector
Q
of the lubrication interaction forces, is provided in [4].
The nonlinear system of reduced Eqs (9) is solved in the time domain using a
modified Newmark method for time integration, which includes a NewtonRaphson iteration
within each time step. Details for the scheme selection can be found in [4]. An operating V6
engine is analyzed by employing the outlined engine dynamic simulation model. Figure 2
shows the used finite element model of the engine block and crankshaft and the combustion
pressure.
Figure 2. Crankshaft and block FE mesh and cylinder pressure
3. METAMODEL DEVELOPMENT AND VALIDATION
Metamodels have been developed for selected bearing performance measures of an
operating V6 engine. The systemlevel engine dynamic solver of the previous section was
employed for computing the performance measures at all used sample points. The operating
V6 engine represents a typical systemlevel automotive application. It is a highly nonlinear
and complicated system, which involves a large number of degrees of freedom (DOF).
The maximum oil pressure over a cycle is crucial in selecting bearing materials with
high enough fatigue resistance for protecting the bearing against failure under an overlay
5
fatigue mode and is a critical indicator of the bearing capacity. For this reason, four of the
chosen performance measures (press
1
, press
2
, press
3
, and press
4
) are the maximum oil
pressure over a cycle at each main bearing. The percentage of time (time ratio) within each
cycle that a bearing operates with a lower film thickness than a user defined threshold
constitutes the other four performance measures ratio1 , ratio 2 , ratio3 , and ratio 4 . This time
ratio indicates the severity of the bearing working condition. Eight metamodels have been
developed for the defined performance measures. The initial clearance between the
crankshaft and each bearing (C
1
, C
2
, C
3
, C
4
) and the oil viscosity (VIS) comprise the five
random variables which are assumed to be normally distributed and uncorrelated.
The engine dynamic solver is computationally intensive. A single run of the system
level engine simulation requires 7 hours on a SUN workstation or 45 minutes on a high
speed SGI supercomputer. The required computational time therefore, makes a probabilistic
analysis practically infeasible if the engine system model is used directly. However, the
metamodels make the probabilistic analysis feasible since they significantly reduce the
computational time to evaluate the performance measures for a given set of random
variables. The Kriging method [1] is used to create the metamodels.
Table 1 shows the utilized mean values and standard deviations for all five random
variables. Their ranges are representative for industrial applications. An Optimal Symmetric
Latin Hypercube (OSLH) algorithm [10] is employed for constructing high quality metamodels
with a relatively small number of samples.
Table 1. Definition of random variables for the V6 engine application
Mean (µ) Minimum Maximum
Standard
Deviation (σ)
Coefficient
of Variation
C
1
(µm) 30 15 45 5 16.67%
C
2
(µm) 30 15 45 5 16.67%
C
3
(µm) 30 15 45 5 16.67%
C
4
(µm) 30 15 45 5 16.67%
VIS(Pa.s) 0.01 0.0058 0.0142 0.0014 14%
A data set consisting of 200 OSLH sample data points is employed. The engine
solver is used for computing the values for the performance variables at all 200 sample
points. Collecting data for the 200 sample points takes about 130 CPU hours of simulation on
an SGI supercomputer. The accuracy of the developed metamodels has been validated.
Figure 3 shows the PDF of the press3 performance measure, calculated with the
Monte Carlo Simulation (MCS) method. The PDF demonstrates a nonnormal type of
distribution. Because the V6 engine is a highly nonlinear system, the random distribution of
press
3
is not normal, although the input random variables are assumed normal and
uncorrelated. Similar nonnormal behavior is observed for the other performance measures.
The developed metamodels are used to perform probabilistic analyses for the operating V6
engine, efficiently. A sensitivity analysis based on Monte Carlo sampling and a Reliability
Based Design Optimization (RBDO) study, are presented next.
6
Figure 3. PDF of performance measure press
3
4. PROBABILISTIC SENSITIVITY ANALYSIS
Probabilistic sensitivity analysis gives the sensitivity of probability of failure f
P
with respect to
the mean and standard deviation of the input random variables. In probabilistic analysis, a
performance function (or performance measure) is defined as
y (X) = y(X1 , X2 ,…, Xd ) where X ={X1 , X2 ,…, Xd }
T
is the vector of d random variables. A
limit state function g (X) = y (X)  0
y
= 0 where 0
y
is a particular value of y, separates the
performance space into a failure region [g 0] and a safe region [g>0]. Given the joint
probability density function (PDF), f x (X), the probability of failure is
f
P
(y< 0
y
)=
Ω
...
f x (X) d X (10)
where denotes the failure region
0 g
. The probability of failure can be interpreted as the
probability of violating a performance measure.
Two dimensionless probabilistic sensitivity coefficients have been proposed [11]; the
mean sensitivity coefficient,
i
S
, and the standarddeviation sensitivity coefficient,
i
S
. They
are defined as
i
S
=
i i
f f
P P
∂
∂
(11)
i
S
=
i i
f f
P P
∂
∂
(12)
where i and i are the mean and standard deviation of random variable i
X
.
i
S
and
i
S
are the nondimensional sensitivity measures of f
P
with respect to the mean and
standard deviation respectively, for random variable i
X
. The probability sensitivity
coefficients can be positive, negative, or of zero value. A large magnitude sensitivity
coefficient indicates that the corresponding random variable is important.
The sensitivity coefficients of Eqs (11) and (12) are employed in the dynamic analysis
of the operating V6 engine to determine the sensitivity of the performance measures with
respect to the mean and standard deviation of each input random variable. The five random
variables C
1
, C
2
, C
3
, C
4
and VIS are assumed normally distributed and uncorrelated. A large
number of samples is generated and the corresponding performance measures are
Maximum Pressure (MPa)
P
D
F
Maximum Pressure (MPa)
P
D
F
7
evaluated using the metamodels. The random variables C
3
and VIS have been identified as
the most important ones.
5. RELIABILITYBASED DESIGN OPTIMIZATION OF MAIN BEARING PERFORMANCE
FOR A V6 ENGINE
In deterministic design we assume that there is no uncertainty in the design variables
and/or modeling parameters. Therefore, there is no variability in the simulation outputs.
However, there exists inherent input and parameter variation that results in output variation.
Deterministic optimization typically yields optimal designs that are pushed to the limits of
design constraint boundaries, leaving little or no room for tolerances (uncertainty) in
manufacturing imperfections, modeling and design variables. Therefore, deterministic optimal
designs that are obtained without taking into account uncertainty are usually unreliable. Input
variation is fully accounted for in ReliabilityBased Design Optimization (RBDO) [2,3].
Probability distributions describe the stochastic nature of the design variables and model
parameters. Variations are represented by standard deviations (typically assumed to be
constant) and a mean performance measure is optimized subject to probabilistic constraints.
5.1. OVERVIEW OF RELIABILITYBASED DESIGN OPTIMIZATION (RBDO) METHODS
A deterministic optimization problem is converted to a probabilistic optimization or
RBDO, problem if the inequality constraints of the former are satisfied probabilistically. In
such a case, the probability of satisfying an inequality constraint must be greater than a
prescribed reliability level which is usually very high. A typical RBDO problem is then
formulated as
X
f
s.t.
n i R G P
i
,..., 2 , 1 0 , , P X d
,
n i ,..., 2 , 1
(13)
U L
d d d
L
X
where
k
R d is the vector of deterministic design variables,
m
R X is the vector of random
design variables,
q
R P is the vector of random and deterministic design parameters,
f
is the objective function and n, k, m and q are the number of constraints, deterministic design
variables, random design variables and design parameters, respectively. According to the
used notation, a bold letter indicates a vector, an upper case letter indicates a random
variable or random parameter and a lower case letter indicates a realization of a random
variable or random parameter. The actual reliability level for the i
th
deterministic constraint is
n i p R
i
f i
,..., 2 , 1 1
, (14)
where

f
Pp
(15)
is the target probability of violating the i
th
deterministic constraint which is usually very small.
Note that in the RBDO formulation of Eq. (13) the design variables include only the means of
the random variables. The target probability of failure f
p
is usually approximated by the
following firstorder relation
t f
p − Φ ≈
(16)
where t is the target reliability index and is the standard normal cumulative distribution
function.
Problem (13) can be solved using two nested optimization loops (doubleloop RBDO
method); the design optimization loop (outer) and the reliability assessment loop (inner). The
latter is needed for the evaluation of each probabilistic constraint in Eq. (13). For this reason,
the doubleloop RBDO method is computationally very expensive and therefore, almost
impractical for largescale design problems. There are two different methods for the reliability
assessment; the Reliability Index Approach (RIA) [2] and the Performance Measure
8
Approach (PMA) [3]. Although either approach can be used, PMA is in general more efficient,
especially for high reliability problems [12]. Every time the design optimization loop calls for a
constraint evaluation, a reliability assessment loop is executed which searches for the MPP
in the standard normal space.
If the PMAbased approach is used, the general RBDO problem of Eq. (13) is stated
as [12]
X
f
s.t.
0
i
p
G
,
n i ,..., 2 , 1
(17)
U L
d d d
L
X
.
where each probabilistic constraint is transformed to an equivalent inequality constraint
involving the performance measure p
G
which is calculated from the following reliability
minimization problem
p
(18)
s.t t
= U
where the vector U represents the random variables in the standard normal space.
As shown from Eq. (17), the PMAbased RBDO formulation involves nested
optimization loops, which may hinder on its computational efficiency and convergence
properties. The same holds for the RIAbased RBDO formulation. To improve the
computational efficiency, two new classes of RBDO formulations have been recently
proposed. The first class decouples the RBDO process into a sequence of a deterministic
design optimization followed by a set of reliability assessment loops [7]. The series of
deterministic and reliability loops is repeated until convergence. The second class of RBDO
methods converts the problem into an equivalent, singleloop deterministic optimization [8]
providing therefore, a substantial computational advantage. The singleloop RBDO method
of [8] is used in the case study of the next section.
5.2. APPLICATION
An RBDO study of the main bearing lubrication performance of an operating V6
engine is presented here. The maximum oil film pressure for each bearing is minimized
subject to each maximum pressure being below a specified value and the oil film time ratio
for each bearing being less than a specified value. The oil film time ratio is defined as the
percentage of time within each cycle that a bearing operates with a film thickness less than a
threshold. Two random variables are considered; the radial clearance C of each bearing and
the oil viscosity VIS. The objective function is simply the sum of the means of maximum oil
pressure in each bearing. The RBDO problem is stated as
4 3 2 1
,
min
press press press press
f
VIS C
u u
+ + + =
s.t.
8 ..., 2 , 1 ) 0 ) ( ( j R G P
j j
X
I
4 ,..., 1 0
134
1 ) ( i
press
G
i
i
X
4 ,..., 1 0
27 . 0
1 ) (
4 +
i
ratio
G
i
i
X
;
s Pa s Pa
VIS
* 0142 . 0 * 0058 . 0 u
m
C
5
,
sP
IS
σ
28 . 1
j or
8 ,..., 2 , 1 52 . 1 j for
j
9
where VIS C
,
and VIS C
,
are the mean values and standard deviations respectively, of
the two design random variables C and VIS. The bounds for VIS C
,
and VIS C
,
are given
in Table 1. For demonstration purposes, the same target reliability index
28 . 1 β
or
52 . 1 β
is used for all eight constraints. In general, a different target reliability index may be
used for each constraint. A reliability index of 1.28 or 1.52 corresponds to a reliability level R
of 89.97% or 93.57%, respectively. The first four inequality constraints indicate that the
maximum oil pressure in all four bearings must be less than 134 MPa. The last four
inequality constraints impose the requirement that the time ratio for each bearing must be
less than 0.27.
Figure 4 shows the progress of the singleloop RBDO process in the X space for
28 . 1 β
. Only the constraints corresponding to press
2
, press
3
, ratio
2
and ratio
3
are plotted.
The other constraints are outside the design space of Figure 4. The algorithm starts with an
6
7
8
9
10
11
12
13
14
15
16
17
1.5 2 2.5 3 3.5 4 4.5 5
C*10
5
(m)
V
I
S
*
1
0

3
(
P
a
.
s
)
Prob. Opt. History
Det. Opt. History
Press 3
Press 2
Ratio 2
Ratio 3
Reliable Det. Optimum
1
st
Iteration
2
nd
Iteration
Initial pt
Initial β "circle"
A B
MPP
β =1.28
Figure 4. RBDO progress for
28 . 1 β
initial point
I X
which is infeasible. The corresponding initial –“circle” is also shown. It should be noted that the –“circle” does not have a
circular shape because the two random variables have different standard deviations. Figure
4 shows the position of the –“circle” for the two iterations the singleloop algorithm needed
toconverge.Thecenterofeach –“circle”indicatesthevalueofthedesignvector
I X
.Asshown,theoptimum(pointB)withXhasbeenreachedwithonlytwoiterations.Constraint 7
G
correspondingtoratio
3
,istheonlyactiveconstraint.Itmustbenotedthatphysicallythe –“circle”moveswithinthefeasibledomaindefinedbyallconstraints,untiliteitherbecomestangenttoatleastoneconstraint(activeconstraint)withoutpenetratingthefixedboundariesofthefeasibledomain(constrainedoptimum)ortheobjectivefunctionisminimizedwithoutanyconstraintsbeingactive
(unconstrained optimum). In the latter case, the –“circle” is within the feasible domain
without touching its boundaries.
For the
28 . 1 β
case, there is an 89.97% chance that a design realization will fall
within the –“circle”. We have therefore, an 89.97% reliable design; all constraints will be
satisfied 89.97% of the time. Figure 5 shows the RBDO results for
52 . 1 β
. In this case,
10
both constraint 7
G
corresponding to ratio
3
and constraint 3
G
corresponding to press
3
, are
active. The –“circle” is tangent to both 7
G
and 3
G
. With the given variability of C and VIS,
indicated by their standard deviations, there is no feasible solution for greater than 1.52.
The maximum reliability we can achieve for this problem is therefore, 93.57%.
6
7
8
9
10
11
12
13
14
15
16
17
1.5 2 2.5 3 3.5 4 4.5 5
C*10
5
(m)
V
I
S
*
1
0

3
(
P
a
.
s
)
Prob. Opt. History
Det. Opt. History
Press 3
Press 2
Ratio 2
Ratio 3
Reliable
O ti
Det. Optimum
1
st
Iteration
2
nd
Iteration
Initial pt
Initial β "circle"
A B
MPP
β=1.52
MPP
Figure 5. RBDO progress for
52 . 1 β
For comparison purposes, Figures 4 and 5 also show the progress of a deterministic
optimization problem with the same initial point. The deterministic optimization converged to
the constrained deterministic optimum (point A) in two iterations. Point A has low reliability
(around 50%) because uncertainty in the design variables C and VIS will result in design
realizations which will violate the ratio
3
constraint.
It should be noted that for both the deterministic and probabilistic optima, the mean oil
viscosity value was pushed to its upper limit of 0.0142 Pa*s (see Table 2 and RBDO problem
statement). This upper limit was responsible for achieving a maximum reliability of only
93.57%. A higher reliability level can be achieved if the viscosity upper limit is raised.
Table 2 illustrates the efficiency of the singleloop RBDO method by comparing it
against deterministic optimization. Both need two iterations to converge. The final value of
the objective function is 352.09, 366.07 and 370.56 for the deterministic and singleloop
methods with
28 . 1 β
and
52 . 1 β
, respectively. Both the deterministic and the singleloop
RBDO methods required 88 function evaluations. A function evaluation represents a
calculation of the objective function or any of the eight constraints. The efficiency of the
singleloop RBDO method is the same with that of deterministic optimization. In general, the
efficiency of the singleloop method is comparable to the deterministic optimization. It should
be noted that the efficiency of any RBDO method can not theoretically exceed the efficiency
of the deterministic optimization. For comparison purposes, Table 2 shows the initial point
used in both the deterministic and probabilistic optimizations.
Table 2. Summary of results for deterministic and singleloop RBDO methods
11
Initial
Point
Det. Opt. Single Loop
Design Variables β=1.28 β=1.52
1
C (µm) 23 24.055 31.57 32.64
VIS 0.01 0.0142 0.0142 0.0142
Objective
f(X) 397.6581 352.0961 366.0703 370.5696
Constraints
0 134 / 1
1 1
press G 0.2391 0.3211 0.2369 0.2217
0 134 / 1
2 2
press G 0.0293 0.1782 0.2152 0.0982
0 134 / 1
3 3
press G 0.1200 0.1492 0.0441 0.0015
0 134 / 1
4 4
press G 0.6440 0.7238 0.5994 0.5733
0 27 . 0 / 1
1 5
ratio G 0.8725 0.9127 0.8894 0.8858
0 27 . 0 / 1
2 6
ratio G 0.0146 0.2405 0.2620 0.2633
0 27 . 0 / 1
3 7
ratio G 0.3571 0.0101 0.0001 0.0529
0 27 . 0 / 1
4 8
ratio G 0.6734 0.9907 0.8284 0.7302
No. of Iterations 2 2 2
No. of F. E. 88 88 88
REFERENCES
[1] Sacks, J., Welch, W. J., Mitchell, T. J., and Wynn, H. P., “Design and Analysis of
Computer Experiments,” Statistical Science, 4(4), 409435, 1989.
[2] Lee, J. O., Yang, Y. O. and Ruy, W. S., “A Comparative Study on Reliability Index
and Target Performance Based Probabilistic Structural Design Optimization,”
Computers and Structures, 80, 257269, 2002.
[3] Tu, J., Choi, K. K. and Park, Y. H., “A New Study on ReliabilityBased Design
Optimization”, ASME Journal of Mechanical Design, 121, 557564, 1999.
[4] Ebrat, O., Mourelatos, Z. P., Hu, K., Vlahopoulos, N. and Vaidyanathan, K., “An
Elastohydrodynamic Coupling of a Rotating Crankshaft and a Flexible Engine Block,”
Journal of Tribology, 126, 19, 2004.
[5] Ebrat, O., Mourelatos, Z. P., Vlahopoulos, N. and Vaidyanathan, K., “Calculation of
Journal Bearing Dynamic Characteristics Including Journal Misalignment and Bearing
Structural Deformation,” Tribology Transactions, 47, 19, 2004.
[6] Hu, K., Mourelatos, Z. P. and Vlahopoulos, N., “A Finite Element Formulation for
Coupling Rigid and Flexible Body Dynamics of Rotating Beams,” Journal of Sound
and Vibration, 253(3), 603630, 2002.
[7] Du, X. and Chen, W., “Sequential Optimization and Reliability Assessment Method for
Efficient Probabilistic Design,” Proceedings of ASME Design Engineering Technical
Conferences (DETC), Paper # DETC2002/DAC34127, 2002.
[8] Liang, J., Mourelatos, Z. P. and Tu, J., “A SingleLoop Method for ReliabilityBased
Design Optimization,” Proceedings of ASME Design Engineering Technical
Conferences (DETC), 2004.
[9] Craig, R. R. and Bampton, M. C. C., “Coupling of Substructures for Dynamic
Analysis,” AIAA Journal, 6(7), 13131319, 1968.
[10] Ye, K. Q., Li, W. and Sudjianto, A., “Algorithm Construction of Optimal Symmetric
Latin Hypercube Designs,” Journal of Statistical Planning and Inference, 90, 145159,
2000.
12
[11] Mohanty, S. and Wu, YT., “CDF Sensitivity Analysis Technique for Ranking
Influential Parameters in the Performance Assessment of the Proposed HighLevel
Waste Repository at Yucca Mountain, Nevada, USA,” Reliability Engineering and
System Safety, 167176, 2001.
[12] Tu, J., Choi, K. K. and Park, Y. H., “Design Potential Method for Robust System
Parameter Design”, AIAA Journal, 39(4), 667677, 2001.
The maximum oil film pressure and the percentage of time (the time ratio) within each cycle that a bearing operates with film thickness lower than a user defined threshold value are defined as the performance measures. The CraigBampton method [9] is employed for reducing the physical Degrees of Freedom (DOF) for both the crankshaft and the engine block.c i (1) where C and K are the reduced mass.3] provides an optimum solution in the presence of uncertainty. damping and stiffness matrices respectively. DEVELOPMENT OF A SYSTEMLEVEL ENGINE BEARING SIMULATION MODEL The integrated systemlevel engine bearing simulation model developed in [46] is used.3.7. input variation is fully accounted for in RBDO through probability distributions which describe the stochastic nature of design variables and model parameters [2. A probabilistic sensitivity analysis is also performed for identifying the most important design variables and for determining the degree to which these design variables influence the performance. ReliabilityBased Design Optimization (RBDO) [2.and the oil viscosity comprise the random variables. A mean performance measure is usually optimized subject to the probability of satisfying a constraint being greater than a prescribed reliability level. z . (2) is discretized and solved numerically for the hydrodynamic pressure field subject to an imposed known pressure at the two bearing ends and along the oil grooves. i=b. t ) is the oil film hydrodynamic pressure and h h 3 P x 6 R h 12 h x (2) 2 . The flexible crankshaft and flexible block are interacting through a set of distributed nonlinear springs and dampers which are represented by a set of stiffness and damping matrices for each bearing. t ) is the lubricant film thickness. P z z where P h 3 h( x. In contrast. Eq. z . engine block and crankshaft. an RBDO study of the main bearing lubrication performance is presented based on a newly developed singleloop RBDO algorithm [8]. respectively. In this paper. This model consists of a flexible crankshaft model and a flexible engine block model connected by a detailed elastohydrodynamic lubrication model. 2. Figure 1 shows the used notation. The Reynolds cavitation condition is used to account for oil cavitation. The stiffness and damping matrices are determined through a detailed lubrication model [5].6].8]. All random variables are assumed normally distributed. and F is the modal force vector. x P ( x. The superscripts “b” and “c” denote the M . A brief description is given in this section. Deterministic optimal designs that are obtained without considering uncertainty are usually unreliable. Details are given in [4. Eight performance measures are considered. The reduced dynamic equations of motion are expressed as M i αi x i r C i αi x i r K i αi x i r {F }. The lubricating oil film pressure distribution for each journal bearing is described by the following Reynolds equation.
(2) yields the (4) 12 h o ∂ ∂ 2 − 3h o ∂x ∂z ∂ ∂ 2 Po − 3h o ∂z ∂x ∂ 3 ∂ ∂ 3 ∂ [ K ]δ + [ K ]δ = 6 h h ∂z ∂ z ∂x ∂ x z h 3 ∂ Po ∂x (5) (6) z [ C] x h 3 x [C ] 12 . (5) and (6) constitute the governing equations for the hydrodynamic oil film pressure distribution. Details can be found in [4.5]. h P [C ] h and (3) P h constitute the oil film stiffness and damping matrices. The size of the fully populated oil film stiffness and damping matrices is large in order to accurately capture the highly nonlinear oil film pressure distribution.Figure 1. (2). The ij element of the stiffness (damping) matrix represents the pressure change in node i due to a unit oil film thickness (squeeze film velocity) change at node j. particularly for high journal eccentricity and journal misalignment conditions. where h o and h o are the distributions from the previous time step and and are their corresponding perturbations. and damping matrix. For 3 . Journal bearing notation A linear perturbation approach is used to solve Eq. At each time step. stiffness matrix. ∂p N ∂hN (7) A similar expression holds for the damping matrix C . In this case. Equations (4). where the terms [ K ] respectively. Under the assumption of small perturbation values for following perturbation equations [4] ∂ 3 ∂ ∂ 3 ∂ Po + Po = 6 h h ∂z ∂ z ∂x ∂x R ho x R and . the explicit form of the stiffness matrix is ∂p1 ∂h 1 ∂p 2 K = ∂h 1 ∂p N ∂ h1 ∂p1 ∂ h2 ∂p 2 ∂ h2 ∂pN ∂ h2 ∂ p1 ∂hN ∂p2 ∂hN . They are all solved numerically using a finite difference method. the film thickness and squeeze film thickness distributions are assumed equal to h = h 0 + and h h0 . respectively. If the oil film domain is discretized with N nodes. respectively. the pressure distribution is approximated as P Po P h P . Eq.
and the are given in [9]. It is a highly nonlinear and complicated system. which includes a NewtonRaphson iteration within each time step. The reduced matrices M . which involves a large number of degrees of freedom (DOF).computational efficiency reasons. Details for the scheme selection can be found in [4]. C reduced generalized external forces vector Qe reduced vector Q of the lubrication interaction forces. z . is a function of the oil film thickness (Eq. METAMODEL DEVELOPMENT AND VALIDATION Metamodels have been developed for selected bearing performance measures of an operating V6 engine. (7)) and damping matrices. It is derived as a function of the generalized coordinates of the block xb crankshaft xc as and the (8) h ( x. The nonlinear system of reduced Eqs (9) is solved in the time domain using a modified Newmark method for time integration. For the solution of Eqs (4) through (6). An operating V6 engine is analyzed by employing the outlined engine dynamic simulation model. The operating V6 engine represents a typical systemlevel automotive application. The expression for the where q are the generalized coordinates. Figure 2. the system of reduced dynamic equations for the combined crankshaft and block are written as M {q} [C ]{q} [K ]{q} {Q } {Q } e (9) and K . The magnitude of the interaction forces {Q} created between the flexible block and the flexible crankshaft. the oil film thickness distribution h x. The oil film stiffness and damping matrices are used in the system level equations between the crankshaft and the block. Figure 2 shows the used finite element model of the engine block and crankshaft and the combustion pressure. (8)) and the stiffness (Eq. t ) c Tc {xc } [Tb ]{xb } where Tc and Tb are appropriate transformation matrices [4]. z . These interaction forces provide the coupling mechanism between the two components. In general. The systemlevel engine dynamic solver of the previous section was employed for computing the performance measures at all used sample points. The maximum oil pressure over a cycle is crucial in selecting bearing materials with high enough fatigue resistance for protecting the bearing against failure under an overlay 4 . is provided in [4]. a diagonalization procedure of the oil film stiffness and damping matrices has been implemented [5]. t is needed. Crankshaft and block FE mesh and cylinder pressure 3.
and press4) are the maximum oil pressure over a cycle at each main bearing. For this reason. The developed metamodels are used to perform probabilistic analyses for the operating V6 engine. Definition of random variables for the V6 engine application Mean (µ) C1(µm) C2(µm) C3(µm) C4(µm) VIS(Pa. The engine dynamic solver is computationally intensive. Collecting data for the 200 sample points takes about 130 CPU hours of simulation on an SGI supercomputer.fatigue mode and is a critical indicator of the bearing capacity. press3.0014 Coefficient of Variation 16.01 Minimum 15 15 15 15 0.67% 14% A data set consisting of 200 OSLH sample data points is employed.0058 Maximum 45 45 45 45 0. The percentage of time (time ratio) within each cycle that a bearing operates with a lower film thickness than a user defined threshold constitutes the other four performance measures ratio 1 . Table 1. The required computational time therefore. press2. This time ratio indicates the severity of the bearing working condition.67% 16. ratio 3 . A single run of the systemlevel engine simulation requires 7 hours on a SUN workstation or 45 minutes on a high speed SGI supercomputer. The engine solver is used for computing the values for the performance variables at all 200 sample points. C4) and the oil viscosity (VIS) comprise the five random variables which are assumed to be normally distributed and uncorrelated. makes a probabilistic analysis practically infeasible if the engine system model is used directly.67% 16. C3. Table 1 shows the utilized mean values and standard deviations for all five random variables. A sensitivity analysis based on Monte Carlo sampling and a ReliabilityBased Design Optimization (RBDO) study. 5 . Eight metamodels have been developed for the defined performance measures. The Kriging method [1] is used to create the metamodels. ratio 2 . the random distribution of press3 is not normal.67% 16. and ratio 4 . four of the chosen performance measures (press1. are presented next. Because the V6 engine is a highly nonlinear system. However.s) 30 30 30 30 0. efficiently. Figure 3 shows the PDF of the press 3 performance measure. calculated with the Monte Carlo Simulation (MCS) method. Similar nonnormal behavior is observed for the other performance measures. although the input random variables are assumed normal and uncorrelated.0142 Standard Deviation (σ) 5 5 5 5 0. Their ranges are representative for industrial applications. the metamodels make the probabilistic analysis feasible since they significantly reduce the computational time to evaluate the performance measures for a given set of random variables. The PDF demonstrates a nonnormal type of distribution. C2. The accuracy of the developed metamodels has been validated. The initial clearance between the crankshaft and each bearing (C1. An Optimal Symmetric Latin Hypercube (OSLH) algorithm [10] is employed for constructing high quality metamodels with a relatively small number of samples.
PDF Maximum Pressure (MPa) Figure 3. X d } T is the vector of d random variables. a performance function (or performance measure) is defined as y (X) = y(X 1 .…. C3. PDF of performance measure press3 4. The five random variables C1. C2. separates the performance space into a failure region [g 0] and a safe region [g>0]. In probabilistic analysis. the probability of failure is Pf (y< y 0 )= . A large number of samples is generated and the corresponding performance measures are 6 .y 0 = 0 where y 0 is a particular value of y.. The probability of failure can be interpreted as the probability of violating a performance measure. the mean sensitivity coefficient.…. Ω f x (X) d X (10) where denotes the failure region g 0 . C4 and VIS are assumed normally distributed and uncorrelated. S i and S i are the nondimensional sensitivity measures of Pf with respect to the mean and standard deviation respectively. X 2 . X d ) where X ={X 1 . S i . The probability sensitivity coefficients can be positive. The sensitivity coefficients of Eqs (11) and (12) are employed in the dynamic analysis of the operating V6 engine to determine the sensitivity of the performance measures with respect to the mean and standard deviation of each input random variable. Given the joint probability density function (PDF). PROBABILISTIC SENSITIVITY ANALYSIS Probabilistic sensitivity analysis gives the sensitivity of probability of failure Pf with respect to the mean and standard deviation of the input random variables. A large magnitude sensitivity coefficient indicates that the corresponding random variable is important. f x (X). or of zero value. S i . and the standarddeviation sensitivity coefficient. They are defined as S i= S i= ∂Pf Pf ∂ ∂ i i (11) (12) i ∂Pf Pf i i where i and are the mean and standard deviation of random variable X i . Two dimensionless probabilistic sensitivity coefficients have been proposed [11]. X 2 . A limit state function g (X) = y (X) . for random variable X i .. negative.
X R m is the vector of random design variables. there exists inherent input and parameter variation that results in output variation. k. n . (13). n .. deterministic optimal designs that are obtained without taking into account uncertainty are usually unreliable. Input variation is fully accounted for in ReliabilityBased Design Optimization (RBDO) [2. Deterministic optimization typically yields optimal designs that are pushed to the limits of design constraint boundaries.. P Gi d. Probability distributions describe the stochastic nature of the design variables and model parameters. X. Note that in the RBDO formulation of Eq. the design optimization loop (outer) and the reliability assessment loop (inner)... there is no variability in the simulation outputs.. m and q are the number of constraints.. f is the objective function and n. The target probability of failure p f is usually approximated by the following firstorder relation Ri 1 p fi i 1. A typical RBDO problem is then formulated as f X s. 5. Therefore. P 0 R i 1. 5. the probability of satisfying an inequality constraint must be greater than a prescribed reliability level which is usually very high.. However. almost impractical for largescale design problems. an upper case letter indicates a random variable or random parameter and a lower case letter indicates a realization of a random variable or random parameter.2. The actual reliability level for the ith deterministic constraint is p [P (15) f where is the target probability of violating the ith deterministic constraint which is usually very small. Therefore.. modeling and design variables. the doubleloop RBDO method is computationally very expensive and therefore. The latter is needed for the evaluation of each probabilistic constraint in Eq. i 1..evaluated using the metamodels. problem if the inequality constraints of the former are satisfied probabilistically. For this reason. RELIABILITYBASED DESIGN OPTIMIZATION OF MAIN BEARING PERFORMANCE FOR A V6 ENGINE In deterministic design we assume that there is no uncertainty in the design variables and/or modeling parameters.2.1...t. respectively. OVERVIEW OF RELIABILITYBASED DESIGN OPTIMIZATION (RBDO) METHODS A deterministic optimization problem is converted to a probabilistic optimization or RBDO. the Reliability Index Approach (RIA) [2] and the Performance Measure 7 .2. n (13) d L X L d d U where d R k is the vector of deterministic design variables. In such a case. According to the used notation. a bold letter indicates a vector. deterministic design variables. (14) pf ≈ Φ − t (16) where t is the target reliability index and is the standard normal cumulative distribution function. leaving little or no room for tolerances (uncertainty) in manufacturing imperfections.. The random variables C3 and VIS have been identified as the most important ones. Problem (13) can be solved using two nested optimization loops (doubleloop RBDO method). Variations are represented by standard deviations (typically assumed to be constant) and a mean performance measure is optimized subject to probabilistic constraints. random design variables and design parameters. P R q is the vector of random and deterministic design parameters.3]. There are two different methods for the reliability assessment. (13) the design variables include only the means of the random variables.
The oil film time ratio is defined as the percentage of time within each cycle that a bearing operates with a film thickness less than a threshold. To improve the computational efficiency.. (13) is stated as [12] f X s.. 0.. The first class decouples the RBDO process into a sequence of a deterministic design optimization followed by a set of reliability assessment loops [7].. X L d d where each probabilistic constraint is transformed to an equivalent inequality constraint involving the performance measure G p which is calculated from the following reliability minimization problem p (18) s.Approach (PMA) [3].. (17).0142 Pa * s Gi ( X) C j 5 m.0058 Pa * s µVIS 0.t. Although either approach can be used. 5..27 . The second class of RBDO methods converts the problem into an equivalent. APPLICATION An RBDO study of the main bearing lubrication performance of an operating V6 engine is presented here. singleloop deterministic optimization [8] providing therefore. Every time the design optimization loop calls for a constraint evaluation. The singleloop RBDO method of [8] is used in the case study of the next section. a substantial computational advantage.28 or j 1. n U (17) d L ..8 I press i 0 i 1.2. The maximum oil film pressure for each bearing is minimized subject to each maximum pressure being below a specified value and the oil film time ratio for each bearing being less than a specified value.. the general RBDO problem of Eq. P (G j ( X) 0) 1 Rj j 1. σ P s S I 1.2. 4 0. especially for high reliability problems [12].2. the radial clearance C of each bearing and the oil viscosity VIS.t U = t where the vector U represents the random variables in the standard normal space. If the PMAbased approach is used..4 134 ratio i G4 + i ( X) 1 0 i 1. the PMAbased RBDO formulation involves nested optimization loops. a reliability assessment loop is executed which searches for the MPP in the standard normal space.. PMA is in general more efficient...2. i 1. two new classes of RBDO formulations have been recently proposed. The series of deterministic and reliability loops is repeated until convergence. µ VIS min f = press1 + press 2 + press 3 + press 4 s.. G pi 0.. As shown from Eq. The objective function is simply the sum of the means of maximum oil pressure in each bearing. The same holds for the RIAbased RBDO formulation.t. The RBDO problem is stated as µ C ... which may hinder on its computational efficiency and convergence properties.8 8 ...52 for j 1.. Two random variables are considered.
–“circle” is within the feasible domain For the β 1. In general. the without touching its boundaries.97% reliable design. Figure 4 shows the position of the –“circle” for the two iterations the singleloop algorithm needed o m A i a sntsy rm epy r wee i cruh aoo r ier pr i i hb cr i teiei i ia c e ov n al ava ax e n a m vo an nie av tceTeoh –”iaeeeinrswo u(it)iXbrcwltoooitG7 cpig3eaoit untase –msiiteild afeas. Opt. It should be noted that the –“circle” does not have a Iwhi circular shape because the two random variables have different standard deviations. Figure 4 shows the progress of the singleloop RBDO process in the X space for β 1.27. For demonstration purposes.28 . The last four inequality constraints impose the requirement that the time ratio for each bearing must be less than 0.hnfa “lidsvodvXhnp mBtheadio irnn ootr.97% chance that a design realization will fall within the –“circle”. 9 . s r noitotctn tedyh c vhfsmebnsee tntlsctiaonhethdnsfsmoi o obunn dhnrsg nec c iethl t sc t t o h e ehwts tn eda nesI bthia il e hb n ytnitbm ne on swt ti f oatee ntnp r e ts mwt oni t g e rne uhgt.hlcn. a different target reliability index may be the two design random variables C and VIS.28 or β 1. Opt. We have therefore. Optimum A st 2 Reliable nd Iteration Prob."circle" Initial pt 10 9 8 7 6 1. hi n a e yeCr I (unconstrained optimum).28 case. The algorithm starts with an 17 16 15 14 1 Iteration Det.28 initial point X ch is infeasible. all constraints will be satisfied 89. RBDO progress for β 1.where C . respectively. A reliability index of 1.5 5 Ratio 3 Press 2 Ratio 2 C*105 (m) Figure 4.e pwsnh tnta.5 3 3. there is an 89. Figure 5 shows the RBDO results for β 1.28 Press 3 VIS*103 (Pa.97% or 93.52 .5 2 2. press3. the same target reliability index β 1. VIS are given used for each constraint. The bounds for . Only the constraints corresponding to press2. VIS are the mean values and standard deviations respectively. In the latter case. VIS and C .52 corresponds to a reliability level R of 89.52 is used for all eight constraints. sothlt “” o tnao iddloittirc setaes(tcti)tuetebdfeilo i(sd motfc ii z tuctitec or etre cctaf ee o. of C in Table 1.57%. ratio2 and ratio3 are plotted. s) 13 12 11 Initial β.neoegatnntcntiongeuroad anet u)hjcnim ieioysba ov.28 or 1. The first four inequality constraints indicate that the maximum oil pressure in all four bearings must be less than 134 MPa. VIS and C . History MPP B β =1. History Det. an 89.5 4 4. In this case.97% of the time. The other constraints are outside the design space of Figure 4. The corresponding initial –“circle” is also shown.
Opt.56 for the deterministic and singleloop methods with β 1. Optimum A Initial β. the mean oil viscosity value was pushed to its upper limit of 0. History Det. The deterministic optimization converged to the constrained deterministic optimum (point A) in two iterations.57%. Both need two iterations to converge. greater than 1. This upper limit was responsible for achieving a maximum reliability of only 93. The –“circle” is tangent to both G7 and G3 .both constraint G7 corresponding to ratio3 and constraint G3 corresponding to press3. there is no feasible solution for The maximum reliability we can achieve for this problem is therefore. Table 2 illustrates the efficiency of the singleloop RBDO method by comparing it against deterministic optimization. The final value of the objective function is 352.52 .07 and 370. It should be noted that for both the deterministic and probabilistic optima.5 5 C*10 (m) Figure 5. s) 13 12 11 10 9 8 7 6 1. Point A has low reliability (around 50%) because uncertainty in the design variables C and VIS will result in design realizations which will violate the ratio3 constraint. Table 2 shows the initial point used in both the deterministic and probabilistic optimizations. are active. A higher reliability level can be achieved if the viscosity upper limit is raised.5 4 4.28 and β 1."circle" MPP Reliable O ti B MPP st nd Iteration Prob. In general. 93.52. RBDO progress for β 1. Summary of results for deterministic and singleloop RBDO methods 10 .52 For comparison purposes.57%.09. the efficiency of the singleloop method is comparable to the deterministic optimization. History β=1. For comparison purposes.5 Press 3 Initial pt Ratio 3 Press 2 Ratio 2 2 2.5 3 5 3.52 1 Iteration VIS*103 (Pa. With the given variability of C and VIS. 17 2 16 15 14 Det. respectively. The efficiency of the singleloop RBDO method is the same with that of deterministic optimization. 366. Table 2. Opt. Both the deterministic and the singleloop RBDO methods required 88 function evaluations. Figures 4 and 5 also show the progress of a deterministic optimization problem with the same initial point. A function evaluation represents a calculation of the objective function or any of the eight constraints. indicated by their standard deviations.0142 Pa*s (see Table 2 and RBDO problem statement). It should be noted that the efficiency of any RBDO method can not theoretically exceed the efficiency of the deterministic optimization.
0146 0..57 0. 126.27 0 0 0 0 0 0 1 press 3 / 134 0 No.1782 0.2633 0. J. Liang. H. 603630. R. 1968. Welch. Q. W. “A New Study on ReliabilityBased Design Optimization”. Du.2620 0.” Journal of Sound and Vibration. A.52 32.0961 0. X.. Hu. K. 2004..8894 0. C. and Vaidyanathan. 409435.. R.” Tribology Transactions.5994 0.. Mourelatos. 47. “A Comparative Study on Reliability Index and Target Performance Based Probabilistic Structural Design Optimization.27 1 ratio 4 / 0. Opt.6734 Det.0101 0. Z. 11 . J.6581 0.2391 0. P.64 0.0142 352.7302 2 88 Constraints G1 G2 G3 G4 G5 G6 G7 G8 1 press1 / 134 0 1 press 2 / 134 1 press 4 / 134 1 ratio1 / 0. E.. S. Mourelatos. N.6440 0. 121. 2004. K. Z.” Proceedings of ASME Design Engineering Technical Conferences (DETC).27 1 ratio 2 / 0. Z.2217 0. Ebrat. “Algorithm Construction of Optimal Symmetric Latin Hypercube Designs.8725 0.” Journal of Tribology. P. K. W. and Ruy. 80. of Iterations No.. 19. Choi. T.01 397. REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] Sacks. 257269. 4(4).0703 0.. 2002. 2002.055 0.. 2002.. O. O.9127 0..2152 0. 1989. J.0001 0.1200 0. C.7238 0.. “A SingleLoop Method for ReliabilityBased Design Optimization.0982 0. and Wynn. “Calculation of Journal Bearing Dynamic Characteristics Including Journal Misalignment and Bearing Structural Deformation.. and Vlahopoulos. Lee.0529 0. 13131319. and Vaidyanathan. W.2369 0.. 557564.1492 0. Mourelatos.3211 0. 90.. O. K. O.0441 0. J.” Statistical Science.. Vlahopoulos. Tu.” Journal of Statistical Planning and Inference. Ye.. Yang.28 31. “An Elastohydrodynamic Coupling of a Rotating Crankshaft and a Flexible Engine Block. P. Z.” AIAA Journal. Y. 6(7). K.8858 0. and Park. N. 1999.” Proceedings of ASME Design Engineering Technical Conferences (DETC). W. 19. H.0142 370. 253(3). K.5696 0.0142 366. K. Y. Craig. “Design and Analysis of Computer Experiments.. “A Finite Element Formulation for Coupling Rigid and Flexible Body Dynamics of Rotating Beams. Mourelatos..0293 0.27 1 ratio 3 / 0. J.. 2004.Initial Point Design Variables C1 (µm) VIS Objective f(X) 23 0..2405 0. Mitchell. 2000. Paper # DETC2002/DAC34127. “Coupling of Substructures for Dynamic Analysis. P. Vlahopoulos. of F. Hu. “Sequential Optimization and Reliability Assessment Method for Efficient Probabilistic Design. J. 24. and Chen.” Computers and Structures.5733 0. and Tu.8284 2 88 β=1. 145159. and Bampton.0015 0. Ebrat. ASME Journal of Mechanical Design.9907 2 88 Single Loop β=1.3571 0. P.. M. N. J. and Sudjianto.. Li.
USA. 2001. K. AIAA Journal. S. 667677. Nevada. “CDF Sensitivity Analysis Technique for Ranking Influential Parameters in the Performance Assessment of the Proposed HighLevel Waste Repository at Yucca Mountain. and Wu. 2001. 12 .. Tu. and Park. H. K. J.” Reliability Engineering and System Safety. YT.. “Design Potential Method for Robust System Parameter Design”.. 39(4).[11] [12] Mohanty. Choi. 167176. Y.