12th

IFToMM World Congress, Besanon (France), June18-21, 2007

Machines and mechanisms design for reliability

A. Hhnel
*
M. Lemaire

F. Rieuneau

F. Petit

LaMI IFMA & UBP LaMI IFMA & UBP RENAULT s.a.s RENAULT s.a.s
Clermont-Ferrand, France Clermont-Ferrand, France Lardy, France Rueil Malmaison, France


AbstractWe propose
1
a framework for assessing and
improving the reliability of machines and mechanisms during
their design process. Actually, a multidisciplinary viewpoint is
adopted to develop a reliability approach based on flexible
representations that benefit from a comprehensive probabilistic
and physical modeling of either the performance or the failure
scenarios of the considered systems. The suggested
FORM/SORM-based analysis yields traditional reliability
measures. It also participates in the definition of chains of
sensitivity factors that detail the weight of each failure scenario
and associated design parameters with respect to the considered
failure modes or loss of performance. These results are exploited
to contribute efficiently to design for reliability efforts.

Keywords: Design for reliability, Physics of failure,
Uncertainty propagation, FORM/SORM, Metamodeling.
I. Introduction
The mechanical engineering industry is dealing with the
challenging issue to design, in an uncertain context,
highly performant and reliable machines and mechanisms.
A machine is an assemblage of mostly mechanical parts
that transmit forces, motion and energy in a predetermined
way. The term mechanism is applied to the combination of
geometrical bodies which constitute a machine. The main
difference is that the former transforms energy to do work,
while the latter does not necessarily perform this function.
Mechanical design is the process by which the needs of a
customer to transmit forces, motion and energy are
transformed into a product, e.g. a machine, satisfying
these needs. It is managed in an uncertain context since
the products environment (their real conditions of use, the
loads) and manufacturing (e.g. geometric or material
characteristics) cannot be known deterministically or
exactly, a fortiori during their definition and modeling
stages. It is thus an exercise in problem solving carried out
by people from dissimilar fields of expertise: designers,
reliability engineers, mechanical engineers, statisticians,
probabilistic method specialists and decision makers (Fig.
1). Reliability is the probability of an item (parts,
components or systems) operating for a given amount of
time without failure. In other words, it is its capability to

1

*
E-mail : Anthony.Hahnel@ifma.fr

E-mail : Maurice.Lemaire@ifma.fr

E-mail : Francois.Rieuneau@renault.fr

E-mail : Frederic.f.Petit@renault.fr

execute required functions for a desired amount of time
without failures, over an expected performance level.
Machines and mechanisms performances, in particular
their reliability, is obviously impacted by the decisions
made during the design process and the uncertain context
introduced beforehand. Therefore, to design for reliability,
tailored reliability-based approaches have to be integrated
to concurrent engineering design process.
The objective of this paper is to present an original
framework for assessing and improving the reliability of
mechanical systems during their design process while
encouraging critical synergies between its multiple actors.
Taking advantage of the considerable developments of
simulation tools and computing resources, a
multidisciplinary viewpoint is adopted to develop an
adaptive and probabilistic reliability approach based on
the sequential physical metamodeling of the failure
scenarios of mechanical systems (Fig. 1). Benefiting from
techniques used in the fields of system reliability [1,2] and
structural reliability [3,4], it pilots the diverse iterations of
an industrial design process. The availability of numerical
models (even approximated) of the behaviors of the
machines is the only prerequisite to make the most of it.
In the first part of the paper, using an usual design
process decomposition, we measure the potential of the
methodology up to classical reliability approaches.
Then, we detail the main steps of the implementation of
reliability
Models of uncertain
input data
Probabilistic reliability methods
FORM/SORM, Random variable
approximation, Polynomial Chaos,
Mechanical and physical
behavior models
EP
EP
IP
EP
IP
IP
EP
TE IP
IP
EP
EP
EP
EP = Elementary Parameter
IP = Intermediate Parameter
TE = Top Event
System failure analysis
Failure mechanism tree
System reliability model
Reliability
engineers
Specialists of physical
behavior modeling
Probabilistic
method specialists
Statisticians
Physical phenomenon / Mechanical behavior
Results
Probability of
failure

f
P

f
P
Importance
measures
Nodal Global
Chains of
importance
measures
Probability
of failure
PDF
approximation
PDF
approximation
Design optimization for reliability, by reliability Update
Designers
1
2
3
4
5

Fig. 1. Proposed framework for mechanical system design for reliability.
12th

IFToMM World Congress, Besanon (France), June18-21, 2007


the suggested probabilistic reliability analysis (Fig. 2):
1. the structural definition of an original quantitative
system reliability model using qualitative failure trees;
2. the evolutive mathematical representation of the system
physical behavior, performance and limit state functions;
3. the adaptive probabilistic modeling of the uncertainties
related to the design parameters characterization;
4. the uncertainty aggregation, through the whole tree
formalism, using specific probabilistic tools;
5. the evaluation of the resulting system reliability model
driven by an original FORM/SORM [3,4] based analysis,
which results are subsequently post-processed and
eventually used for design optimization purposes.
The main features of the methodology are illustrated
using close-ups on the example of an automotive engine
multi-perimeter system. The issues of the gas sealing
quality of a cylinder head gasket and the seizing of a
piston pin due to its deformation are isolated and
examined. There are analyzed with respect to the
sequence of phenomena illustrated in figure 2.

Combustion
chamber filling
Combustion
Cylinder head,
Gasket & Housing
Crankshaft, Connecting
rods & Pistons
Elementary
parameters
Performance functions
Limit state functions
Engineering or control limit states of performance
Gas sealing quality
2 criteria
Piston pin seizing
3 criteria
Physics-of-failure based
sequential model
P
max
P
max

Fig. 2. Central example - synoptic scheme of the engine sub-system.
II. Design process and classical reliability analyses
Considering a systematic decomposition [5], industrial
design process consists of three main stages: conceptual,
embodiment and detail design. They are usually preceded
by a statement of intent plus the definition of design
specifications, and followed by product manufacturing
and sale. Actually, the applicability and the efficiency of
standard reliability methods for design enhancing matters
strongly depend on the design stages (Fig. 3).

Time
Impact on the elementary design parameters definition
Conceptual
Design
Embodiment Design Detail Design Industrialisation
Physical life testing analysis
Reliability growth analysis
Physical and system reliability interpretations
Failure scenarios representation integrating
metamodels of the system physical behavior
Extended use ? Physical without system
reliability interpretations
Which potential for
a unified approach ?
System without physics reliability interpretations
Qualitative risk &
reliability analysis
Qualitative risk &
reliability analysis

Fig. 3. Reliability analysis and the design process of mechanical systems.

Conceptual design is the action of conceiving ideas and
working principles from a functional perspective.
Qualitative risk and reliability analyses are carried out
iteratively during this phase. Preliminary Hazard Analysis
(PHA), Failure Mode Effects and Criticity Analysis
(FMECA), qualitative Event and Fault Tree Analysis
(ETA/FTA) [1] guarantee that all the systems likely
failure modes are identified. The related failure scenarios
are exhaustively described to be further analyzed.
The embodiment design consists in defining the
preliminary layout for the selected concept. The system
architecture takes form while initial performance studies
are performed using simplified models and design rules.
The design is assessed against objective criteria (cost,
reliability, manufacturing) and refined iteratively. At
this stage, reliability interpretation and evaluation based
on time to failure approaches [2] are usually derived. They
allow the analyst to take into account the system
dimension and the logic of the item failures based either
on static or dynamic formulations of the reliability
problem. Their inputs (e.g. failure or state transition rates,
statistical failure models) stem from reliability database,
field data or experts judgment. They tend to validate the
architecture of the system giving estimation of reliability
measures such as probability of failure, component
importance, etc However, their high level of abstraction
is inadequate to link its reliability to the most influent
design variables (e.g. material properties, loads,
geometry). Therefore, we denote them as system
without physics reliability analyses.
Detail design participates in validating and refining the
preliminary design. Comprehensive test campaigns are
performed, optimized using Design Of Experiments
(DOE) techniques [6]. At this stage, system without
physics analyses are completed by life data or physical
life testing analyses [7]. Alternatively, physical reliability
approaches based on numerical simulations of the system
physics-of-failure and performance are exploited. Relying
on highly representative physical behavior models (e.g.
Finite Element Models), they encompass evaluation
techniques such as FORM/SORM analysis [3,4]. These
methods are convenient to design mechanical components
in an uncertain context modeled by random variables or
processes. They give essential insight into their response
and reliability sensitivities with respect to the design
parameters. Still, we denote these approaches as physical
without system since their use is yet too restricted to the
study of isolated parts. Although system formulations
exist for the reliability analysis of complex machines, the
identification issue of the physical sequences involved in
their failures still needs to be better addressed.
This comparison clearly emphasizes the needs to close
the gap between system without physics and physical
without system reliability interpretations. The use of
physical modeling based reliability analysis to the
preliminary design stage should be broadened as well.
12th

IFToMM World Congress, Besanon (France), June18-21, 2007


Consequently, the physical and system interpretation
that we suggest is more integrated into the successive
stages of the design process (Fig. 3). Thereafter, we focus
on its main features and we show how it allows us:
to evaluate machines reliability taking into account their
failure logic as well as their physics of failure sequences;
to characterize the physical phenomena and the design
parameters that impact the most the reliability;
to identify suitable design levers to improve it;
to benefit from the synergies and the expertise of the
main contributors of the design process.
III. Physical and probabilistic system reliability model
A. Qualitative & quantitative failure mechanism scenarios
First, the critical aspects of the mechanical system as
well as all its potential failure modes are pointed out by
means of PHA and FMECA. On this basis, we take
advantage of failure mechanism tree [8] to fully describe
the failure modes (also denoted as Top Events - TE) of the
system and its physical failure scenarios. Actually, they
consist in a thorough bottom-down qualitative analysis of
its failure sequences. It leads to the successive
identification of all the physical phenomena and the
elementary design parameters that have an impact on the
TE and on the reliability of the system. Hence, failure
mechanism trees set up the physical links between the
failure modes of the mechanical system, its loss of
performance and the elementary design parameters.
Once a qualitative failure mechanism tree is available, it
is used to build a quantitative formalism that is valuable to
derive original and efficient reliability estimation. The
root of the tree, i.e. the level (0), contains the vector X
(0)

of the nva
0
design parameters identified by the qualitative
analysis. Level (1) encompasses level (0) variables as well
as the Nc
1
variables resulting from the transition from
level (0) to level (1). In fact, they follow from nodal
transformations M
k
(1)
that model the physics of the related
phenomena. The vector X
(1)
of level (1) is then defined as:
( )
( )
( )
( )
( )
1
1
0
1
1
0
1
0
1 , , 1 0
nva
Nc
nva
Nc k
M
nva
k
)
`

=
=
S
X
X X
K
(1)

This incremental procedure leads to the definition of X
(TE)
the vector of the TE level. The transition from level (l) to
the next level (l+1) is given by:
( )
( )
( )
( )
( )
l
l
l
l
l
k
l
nva
Nc
l
nva
l
l Nc k
M
nva
l
)
`

=
S
X
X X
1
1
1
, , 1 1 K
(2)
where TE l , , 0 K = and

=
+ =
l
i
i l
Nc nva nva
1
0
.
Each node of the model consists in a logical as well as
physical entity clarifying a specific physical phenomenon
or a particular aspect of the behavior of the system that is
involved in its failure. Thus, performance functions can be
defined for any intermediate or terminal parameter as:
( ) ( ) ( )
( ) ( ) ( )
( )
( )
TE l Nc k M G
nva k G
l
l
k
l l
k
l
k
k k k
, , 1 , , , 1 ,
, , 1 ,
1
0
0 0 0
K K
K
= = =
= =

X
X
(3)
where
( ) l
k
is an engineering threshold. They consist in
numerical interpretations of the system changes in state or
critical losses of performance.
( )
0
l
k
G defines a failure
domain
( ) l
f
k
D while
( )
0 >
l
k
G defines a safe domain
( ) l
s
k
D .
There are derived using design and physical rules. The
reliability of the system is measured according to these
particular functions. For example (Fig. 2), a control
performance function is associated to the combustion
pressure P
max
and a limit state function describes each gas
sealing and piston pin seizing failure criterion.
Note that the detail fineness of the reliability model is
adjusted depending on what we want to demonstrate in
terms of reliability, what we really know about the
behavior of the system, which parameters and phenomena
we want to monitor [8]. Figure 4 illustrates an extract of
the model structure corresponding to the two TE related to
the gas sealing quality of the cylinder head - gasket -
cylinder housing assembly. It reveals the dependencies
between parameters that have a potential influence on the
gas sealing quality criteria and which we want to monitor
and control for reliability analysis purposes. It could have
been divided in more levels but that would not have given
more useful information from the reliability point of view.



Applied clamping stress
Stabilized
clamping stress
Clamping
stress loss
Analytical models
Stabilized clamping stress
FEM
Thermal analysis
Thermal loading
Heat transfer properties
Material properties
Geometric properties
Thermal stresses
Temperature
gradients
Maximal combustion
pressure
Cylinder head
deformation
Cylinder housing
deformation
Gasket bearing
Burring under
the screw head
Pressure on gasket
sealing elements
Bearing of gasket
sealing elements
Material properties
Geometric properties
Screw threads loaded
Cylinder head
burring
Elementary
parameters
FEM + Analytical models
Screwing/Clamping effects
Combustion
chamber filling
Combustion
FEM
Mechanical analysis
Cylinder head
deformation
Cylinder housing
deformation
Crankshaft, Connecting rods & pistons
Gas sealing
quality - 2 criteria
Plastic elongation
Parts stiffness
Temperature

Fig. 4. Idealized structure of the failure scenarios related to the gas sealing performance criteria.
12th

IFToMM World Congress, Besanon (France), June18-21, 2007


B. Adaptive modeling of the system physical behavior
The nodal transformations M
k
(l)
introduced above consist
in more or less accurate mathematical representations of
both physical phenomena and systems physical behavior.
In the best case, models of knowledge are derived
explicitly from equations provided by available theories.
Alternatively, numerical solutions to the knowledge
models are defined implicitly through finite element codes
or any other simulation analysis codes. Metamodeling
techniques [9] are sensibly used to obtain simplified and
computationally less expensive models of the model. The
common approach is to apply a numerical DOE to identify
a good set of input combinations and then use regression
analysis on the corresponding simulated response to create
a polynomial approximation of a computer analysis code.
Note that the latest applications in the field of reliability
analysis use techniques such as response surface, kriging,
bootstrap, neural networks and support vector machine
[10]. Finally, models of expertise provide a better
understanding of badly mastered phenomena giving
insight into the functional relationship between design
variables and response variables. Derived from the
available information (e.g. field data, expert judgment)
and from experimental design strategies, they are built
using statistical approaches and metamodeling.
Benefiting from metamodeling, we work with relevant
physical models at each design step: the most appropriate
regarding our knowledge of the considered physical
behaviors or phenomena and the most pertinent with
respect to the goal of the reliability analysis. No matter
how rough the resulting reliability estimation might be
since it is refined during the subsequent steps of the
design process. Moreover, the representativeness of the
knowledge and the information used to build the models
as well as the model parameters are integrated in the
proposed reliability analysis. As a result, their influence
on the system reliability and its estimation are monitored.
It allows us to point out the weakest nodes and to define
specific, not more than necessary, efficient model
enhancing actions at each design iteration.
For example, considering the combustion phenomenon
(Fig. 2), we use a simple analytical model stemming from
expertise considerations starting the embodiment design.
During the subsequent design iterations, a partial
Quadratic Response Surface is built using an adequate
DOE. It is later refined during the detail design to obtain a
complete QRS and eventually a 3D finite element model.

C. Stochastic modeling of the design variables
During the design process of mechanical systems, the
characterization of the vector X
(0)
of design variables is
uncertain. Common techniques to uncertainty modeling
are based e.g. on probability theory, interval arithmetic or
possibility theory [11]. Here, a consistent probabilistic
approach using the Bayesian updating rule [12] is derived.
Actually, the first information regarding the design
parameters X
(0)
comes from the initial definition of
specified values and their corresponding tolerance interval
given by the designers during the embodiment design.
These tolerance intervals can be modeled stochastically
fitting a uniform Probability Density Function (PDF)
according to the maximum entropy principle. Otherwise,
this prior PDF may be chosen among a set of usual PDFs
(e.g. normal, lognormal). Note that the proposed approach
allows the analyst to take into account the possible
correlation between the design parameters X
(0)
.
As we progress through the different stages and
iterations of the design process, the information on X
(0)
is
systematically refined based on all the available data and
sources of knowledge. The Bayesian approach allows us
to use all the available and meaningful information to
improve the PDF characterization of each design
parameter. Indeed, given a set of observations of the
design parameter of interest X
k
(0)
, the Bayesian updating
rule and the total probability theorem are jointly used to
derive the predictive PDF of X
k
(0)
updating its prior PDF.
Observations of X
k
(0)
result either from expert judgments,
test and field data or production control. They depend on
the design process stage and the related knowledge level.
As regards the elicitation of experts' opinions, systematic
approaches are exploited to get the most of them,
gathering and quantifying their expertise [13].
Subsequently, we associate a representativeness measure
to the Bayesian formalization of the uncertainties of the
design parameters. It is assessed on a 1 to 5 empirical
scale to capture the maturity of the stochastic model.
Integrated to the reliability analysis proposed thereafter,
their weight on the system reliability is monitored to point
out the most influent design parameters and probabilistic
models. In fact, it helps the definition of specific, not
more than necessary, design enhancing actions.
Figure 5 shows the iterative stochastic model and
representativeness of the Young modulus (GPa) of the
piston pin examined in the central example (Fig. 2).


Fig. 5. Piston pin Young modulus iterative stochastic models.
12th

IFToMM World Congress, Besanon (France), June18-21, 2007


D. Uncertainty aggregation
Our goal is now to monitor the propagation of the design
uncertainties through the full failure tree whatever the
type of the nodal transformations is. The issue consists in
evaluating the statistical properties and the PDF of a
random variable that result from combinations of several
random variables. Approximation methods are needed
since it has no generic solution.
We implement an original stochastic node crossing
procedure based on polynomial chaos expansions [14] and
on the Exponential of a Polynomial Method (EPM) [15].
Each random variable of the tree that results from a
transformation M
k
(l)
(Eq. 2) is first approximated as:
( ) ( ) ( )
{ }

=
=
+ +
|
.
|

\
|
=

1
0
1
1 1

P
j
r
q
l
k j j
l
nva k
l
nva k
q l l
a X X (4)
a
j
are coefficients to be determined.
j
are the
multidimensional Hermite polynomials of order less or
equal than p (the so-called polynomial chaos). r is the
number of input random variables and P=(r+p)!/(r!p!) is
the size of the chaos basis of order p.
k
(l)
is a vector of
standard normal variables that is built according to the
statistical properties of the variables of X
(l-1)
. Actually, the
a
j
coefficients are estimated using a classical regression
method that requires a limited number n of deterministic
calls to the transformation M
k
(l)
(see [16] for details). Once
the estimate of the random variable is available through
Eq. 4, an analytical PDF EPM approximation of the
form
( )
)
) (

Q
X
e c f
l
k
is derived. In fact, the constant c
and the polynomial Q are identified using the computation
of the moments of the random variable using Eq. 4 [15].
Note that its PDF can also be built point by point using
either Monte Carlo simulation or FORM/SORM analysis.
Eventually, given the polynomial expression of the
random variables in Eq. 4, the correlation coefficients of
the multiple variables of the tree model are estimated
analytically or using crude Monte Carlo simulation.
IV. Reliability evaluation and optimization
FORM/SORM analysis strategies and reliability measures
The previous procedure ends up with a fully functional
and flexible system reliability model built on the physical
and probabilistic interpretations of failure scenarios. We
now discuss estimation strategies to make the most of it.

33
1
PDF approximation
+
FORM / SORM
Monte Carlo simulation
2
FORM / SORM
Monte Carlo simulation

Fig. 6. Analysis schemes for the reliability model evaluation.
The first suggested analysis scheme (-1- in figure 6)
corresponds to the most natural approach. It follows the
construction of the probabilistic failure mechanism tree
discussed beforehand. The PDF of the nodal parameters as
well as their correlation coefficients are determined on a
level by level progression basis. Additionally, directly
from the parameters PDF approximation, we approximate
the probabilities of failure and the reliability indexes:
( )
( )
( ) ( )
( )
( )
( )
( ) 1 1
0
1
1

=
l l
G
l
f
d f P
l l
k
l
k
X x
x
X
(5)
( ) ( )
( )
l
f
l
k
k
P
1
= (6)
related to each performance function G
k
(l)
defined in the
model. FORM implementation following this strategy also
leads to the evaluation of the parameters X
(l-1)
influence on
the nodal transformations M
k
(l)
or G
k
(l)
and their associated
reliability measures. In fact, we evaluate there sensitivities
to small variations of the input parameters of interest X
(l-1)
.
The second scheme (-2- in figure 6) consists in
performing upwards shortcut FORM/SORM analyses
considering each node of the intermediate levels with
respect to the parameters of level (0). It aims at evaluating
the influence of the elementary parameters X
(0)
on the
intermediate parameters X
(l)
and on each associated
performance function and probability of failure.
The third scheme (-3- in figure 6) deals with backwards
FORM/SORM analyses considering each node of the TE
level with respect to each lower level: the global influence
of the transitional parameters on the TE level is captured.
On the one hand, coupling these schemes we are able to
evaluate the linearized correlation coefficients of the limit
states defined in the model. By extension, we derive [17]:
first-order approximation of the probabilities of
occurrence of unions and intersections of their related
states and associated events;
first-order approximation of the conditional probability
of any potential failure or loss of performance identified
in the failure tree given a sequence of occurred events.
Table 1 gives the correlation coefficient of the limit states
for the TE related to the piston pin seizing performance
criteria defined in the introduction. Table 2 shows
reliability indexes measured for each performance of
interest taken separately (on the diagonal) or assuming
that another performance is not satisfied (outside of the
diagonal). Actually, they are normalized by a non-given
target reliability index
(ref)
. /
(ref)
1 means that the
design meets the reliability objectives. /
(ref)
< 1 denotes
that the proposed design is not satisfactory. /
(ref)
>1
corresponds to designs which exceed their reliability
specifications. In this testbed case, the evaluation is made
at an intermediate iteration of the embodiment design
phase. The piston pin Out-of-Round (O-o-R) wear
criterion does not meet the reliability target while the
others criteria oversatisfy it. Thus, the piston definition
needs to be improved during subsequent design iterations.
In the following, we show how the proposed analysis and
its outputs efficiently participate in enhancing the design.
specifications.
12th

IFToMM World Congress, Besanon (France), June18-21, 2007


Bending
Bosses
loading
O-o-R
wear
Bending 1 0.617 0.947
Bosses loading 0.617 1 0.778
Out-of-Round wear 0.947 0.778 1

TABLE 1. Correlation of the limit state and performance functions.

/
(ref)
of
knowing occurrence of
Bending
Bosses
loading
O-o-R
wear
Bending 2.12 -0.12 -1.36
Bosses loading 1.53 1.26 -1.36
Out-of-Round wear 2.12 1.26 -0.87

TABLE 2. Normalized probabilities of failure occurrence.

On the other hand, coupling the strategies in the order 1,
3, 2, we construct original chains of importance factors
using classical chain rules of differentiation. They capture:
the global influence of any parameter of the model on a
higher level parameter or performance function and its
associated reliability measures;
the partial influence of any given parameter through
each of the failure mechanisms it potentially impacts.
Consequently, we identify the most critical parameters as
well as the most likely failure scenarios considering the
examined system failure mode. Figure 7 represents the
simplified chains of importance measures relatively to the
minimum pressure allowed on the gasket body (one of the
TE gas sealing criteria implicitly defined in figure 2). It
shows that P
max
is the most influent parameter on the
reliability of the system. Moreover, it is involved in more
than one failure scenario. Figure 7 stresses that its partial
contribution in the upper branch is insignificant compared
to its impact through the lower branch. For reliability
improvement purposes, if P
max
can be controlled, then the
chain shows that the injection advance and the combustion
filling parameters should be monitored and modified first
and foremost. If not, it reveals that the branches ending
with the applied clamping stress
10.3 Stabilized
clamping
stress
89.6
Applied
clamping
stress
10.4 Applied clamping
stress loss
67.4 Screw stiffness
3.8 Cylinder head stiffness
1.5 Cylinder housing stiffness
27.3 Screw plastic
elongation
71.6
5.7 Cylinder
housing stiffness
22.7
Temperature
P
max
Top Event
Gas sealing quality
89.7 P
max
0.4
0.3
4.1
65.9
0.0
8.9
0.0
Air intake adjustment
Injection advance
Fuel jet nappe angle
Distance
Piston-Cylinder head
Injector protusion
Common rail pressure
Quantity of injected fuel
Swirl
0.0
0.1
14.1
0.0
76.7
1.0
1.8
0.1
Exhaust adjustment
Compression efficiency
Turbocharger efficiency
Turbocharger permeability
Low Pressure charge loss
High Pressure charge loss
Exhaust charge loss
Supercharging Air
Cooler thermal efficiency
6.2 20.4 Combustion chamber filling
Minimum pressure
criteria on the gasket

Fig. 7. Pressure criteria on the gasket: chain of importance factors (%).
with the applied clamping stress and the distance from the
piston to the cylinder head should then be examined.

D. Design enhancing using reliability
In the scope of design refinement for reliability, we
apply an original optimization procedure coordinated by
the proposed approach. It takes advantage of the formerly
mentioned reliability measures and chains of importance
factors to identify dominant parameters that improve the
system performance if modified in a prescribe way.
Using the example of the piston pin seizing (Fig. 2), we
show how it can guide the design process of complex
mechanical systems. For illustration matters, focusing on
reliability aspects, we consider that design iterations only
consist in one change in one identified design parameter.
In real applications, several parameters are notified for
modifications that are compatible with reliability
objectives and other criteria such as manufacturing or
cost. In addition, the sensitivities of the reliability analysis
with respect to the parameters of the nodal behavior
models are not taken into account in the following.
Figure 8 illustrates the evolution of the piston pin
normalized reliability with respect to the three seizing
criteria, according to successive design modifications.
Note that they depend on the same set of design variables,
namely: the width of the connecting rod head B, the
distance between the two pin supports on the piston head
(the bosses) E
b
, the Young modulus of the piston pin E, its
length L, its outer and inner diameters D
ext
and D
int
, the
maximum combustion pressure P
max
, and the cylinder
diameter . The first points match the results obtained
while performing the analysis onto a preliminary design of
the system during its embodiment phase. For example, at
this stage the Young modulus is stochastically modeled by
a uniform distribution defined over its tolerance interval
and which representativeness degree is 1. These results
are far from being acceptable regarding the reliability
objectives. Figure 9 shows the elasticities (i.e. the
normalized version of the importance factors) of the
piston pin bending reliability index
bending
with respect to
the design parameters mean and standard deviation.
Obviously, it is highly sensitive to several parameters and
thus there is more than one lever for reliability
improvement:
increasing the mean of D
ext
or decreasing the mean of L,
E
b
, Phi, P
max
would have a positive impact on the
reliability of the system with respect bending failure;
checking and controlling the standard deviation of E,
P
max
and E
b
would also enhance its reliability.
Adopting a multi-failure mode approach, the chains of
importance factors of the different failure modes are used
jointly with the estimation of the limit states correlation
coefficients to pilot effective design modifications. Here,
the first resulting action consists in reducing the mean of
D
int
. The reliability results still do not meet the target (Fig.
8). Consecutively, P
max
standard deviation is reduced, the
stochastic model of the Young modulus E is improved
12th

IFToMM World Congress, Besanon (France), June18-21, 2007


1 2 3 4 5 6 7
-1
-0.5
0
0.5
1
1.5
2
Design modification

(
r
e
f
)
1 2 3 4 5 6 7
0
1
5
10
15
20
25
30
Design modification
P
f

/

P
f (
r
e
f
)

Bending
Bosses loading
Out-of-Round wear

Fig. 8. Piston pin reliability evolution.
1 2 3 4 5 6 7 8
-30
-20
-10
0
10
20
30
40
50
60
Design parameters
E
l
a
s
t
i
c
i
t
i
e
s

t
o

t
h
e

m
e
a
n

1 2 3 4 5 6 7 8
-0.7
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
Design parameters
E
l
a
s
t
i
c
i
t
i
e
s

t
o

t
h
e

s
t
a
n
d
a
r
d

d
e
v
i
a
t
i
o
n

1 : B
2 : E
b
3 : E
4 : D
ext
5 : D
int
6 : P
max
7 :
8 : L

Fig. 9. Elasticities of
bending
at the first design iteration.

stochastic model of the Young modulus E is improved
while its mean and standard deviation are modified, the
piston pin length L and the distance E
b
between its bosses
are reduced. After seven iterations, the reliability of the
system with respect to the three limit states exceeds the
targeted reliability level (Fig. 8). The design is even too
reliable regarding the bending failure. For economical
matters, revisions must be decided to get a less expensive
design, which reliability equals the specified objective.
V. Conclusion and perspectives
Efficient approaches to machine design for reliability
require the implementation of methodologies that interact
one with an other. In this paper, a multidisciplinary
standpoint is adopted to derive an adaptive probabilistic
reliability approach based on the physical modeling of
machines and mechanisms failure scenarios. It has the
aptitude to catalyze the synergies between the contributors
of an industrial design process. Besides, it contributes to
design optimization essentially providing them with:
the characterization of the physical phenomena and the
design parameters that impact the most the reliability;
the identification of the suitable design levers and the
way of activating them to improve the reliability.
The suggested system and physical reliability approach
rapidly leads to the numerical evaluation of complex
models that are prohibitive to manipulate and to evaluate
manually. It also requires an adapted management of data
and engineering knowledge. RELAB, a software platform
we implement in the Matlab environment is a first answer
to that specific issue. It orchestrates the definition and the
automatic evaluation of such reliability models [8].
The proposed framework for machines and mechanisms
design for reliability is flexible enough to integrate, e.g.:
more efficient methods to model the uncertainties of the
design parameters and to assess the uncertainties inherent
to the definition of models of expertise and metamodels;
more accurate and less time-consuming techniques to
approximate random variables that result from complex
transformation of numerous random variables.
a computerized procedure to design optimization which
yet relies on a fastidious expert approach.
Acknowledgements
This paper is supported by funding from Renault s.a.s, the
ANRT (CIFRE program) and the LaMI UBP/IFMA
laboratory. These supports are gratefully acknowledged.
References
[1] Villemeur A. Reliability, availability, maintainability and safety
assessment. John Wiley & Sons, New York, 1991.
[2] Rausand M. and Hyland A. System reliability theory. John Wiley
& Sons, New York, 2004.
[3] Ditlevsen O. and Madsen H.O. Structural reliability methods. John
Wiley & Sons, New York, 1996.
[4] Lemaire M. Fiabilit des structures : Couplage mcano-fiabiliste
statique. Hermes Science Publication, Paris, 2005.
[5] Pahl G. and Beitz W. Engineering Design: A Systematic Approach,
Second Edition. Springer-Verlag, Berlin, 1996.
[6] Mason R.L., Gunst R.F. and Hess J.L. Statistical Design and
Analysis of Experiments. John Wiley & Sons, New York, 2003.
[7] Bagdonavicius V. and Nikulin M. Accelerated life models -
Modeling and Statistical analysis. Chapman & Hall, 2002.
[8] Hhnel A. Approche mcano-probabiliste systme en conception
pour la fiabilit, Ph.D. Thesis, In revision, LaMI IFMA/UBP, 2006.
[9] Simpson T.W., Peplinski J., Koch P. and Allen J. Metamodels for
computer-based engineering design: survey and recommendations.
Engineering with computers 17(2):129-150, 2001.
[10] Lemaire M. Implementation of parametric reliability methods in
industrial applications. In Proc. of the ASRANet Colloquium, July,
10-12, Glasgow,2006.
[11] Helton J.C., Johnson J.D. and Oberkampf W.L. An exploration of
alternative approaches to the representation of uncertainty in model
predictions. Rel. Eng. & Syst. Safety, 85(1-3):39-71, 2004.
[12] Box G.E.P. and Tiao G.C. Bayesian inference in statistical
analysis. John Wiley & Sons, New York, 1992.
[13] Ayyub B.M. Elicitation of expert opinions for uncertainty and
risks. CRC press, Boca Raton, 2001.
[14] Ghanem R. and Spanos P. Stochastic finite elements: A spectral
approach. Springer-Verlag, New York, 1991.
[15] Er G.K. A method for multi-parameter PDF estimation of random
variables. Structural Safety 20(4):341-355, 1998.
[16] Sudret B., Berveiller M. and Lemaire M. Stochastic finite element:
A non intrusive approach by regression. European Journal of
Computational Mechanics 15(1-3), 81-92, 2006.
[17] Hhnel A., Lemaire M., Rieuneau F. and Petit F. A probabilistic
and physical-based framework for mechanical system design for
reliability. In Proc. of the ESREL Safety and Reliability Conference,
September, 18-22, Estoril, 2006.