This action might not be possible to undo. Are you sure you want to continue?
AND SHELLS INCLUDING UNCERTAINTIES
Von der Fakultät für Maschinenbau der
Technischen Universität Chemnitz
genehmigte
Dissertation
zur Erlangung des akademischen Grades
Doktoringenieur
(Dr.Ing.)
vorgelegt
von MSc. Thanh Ngọc Trần
geboren am 03. Februar 1975
in Nam Dinh, Vietnam
eingereicht am 12. Dezember 2007
Gutachter:
Prof. Dr.Ing. Reiner Kreißig
Prof. Dr.Ing. Manfred Staat
Prof. Dr.Ing. Christos Bisbos
Tag der Verteidigung: 12. März 2008
Trần, Thanh Ngọc
Limit and shakedown analysis of plates and shells including uncertainties
Dissertation an der Fakultät für Maschinenbau der Technischen Universität Chemnitz,
Institut für Mechanik und Thermodynamik, Chemnitz 2008
149 + vii Seiten
55 Abbildungen
28 Tabellen
162 Literaturzitate
Referat
The reliability analysis of plates and shells with respect to plastic collapse or to inadaptation is
formulated on the basis of limit and shakedown theorems. The loading, the material strength and
the shell thickness are considered as random variables. Based on a direct definition of the limit
state function, the nonlinear problems may be efficiently solved by using the First and Second
Order Reliability Methods (FORM/SORM). The sensitivity analyses in FORM/SORM can be
based on the sensitivities of the deterministic shakedown problem. The problem of the reliability
of structural systems is also handled by the application of a special barrier technique which
permits to find all the design points corresponding to all the failure modes. The direct plasticity
approach reduces considerably the necessary knowledge of uncertain input data, computing costs
and the numerical error.
Die Zuverlässigkeitsanalyse von Platten und Schalen in Bezug auf plastischen Kollaps oder
NichtAnpassung wird mit den Traglast und Einspielsätzen formuliert. Die Lasten, die
Werkstofffestigkeit und die Schalendicke werden als Zufallsvariablen betrachtet. Auf der
Grundlage einer direkten Definition der Grenzzustandsfunktion kann die Berechnung der
Versagenswahrscheinlichkeit effektiv mit den Zuverlässigkeitsmethoden erster und zweiter
Ordnung (FROM/SORM) gelöst werden. Die Sensitivitätsanalysen in FORM/SORM lassen sich
auf der Basis der Sensitivitäten des deterministischen Einspielproblems berechnen. Die
Schwierigkeiten bei der Ermittlung der Zuverlässigkeit von strukturellen Systemen werden durch
Anwendung einer speziellen Barrieremethode behoben, die es erlaubt, alle Auslegungspunkte zu
allen Versagensmoden zu finden. Die Anwendung direkter Plastizitätsmethoden führt zu einer
beträchtlichen Verringerung der notwendigen Kenntnis der unsicheren Eingangsdaten, des
Berechnungsaufwandes und der numerischen Fehler.
Schlagworte:
Limit analysis, shakedown analysis, exact Ilyushin yield surface, nonlinear programming, first
order reliability method, second order reliability method, design point
Archivierungsort:
http://archiv.tuchemnitz.de/pub/2008/0025
ACKNOWLEDGEMENTS
This work has been carried out at the Biomechanics Laboratory, Aachen University of
Applied Sciences, Campus Jülich. The author gratefully acknowledges the Deutscher
Akademischer Austausch Dienst (DAAD) for a research fellowship award under the grant
reference A/04/20207.
The author is indebted to Prof. Dr.Ing. M. Staat who has been the constant source of
caring and inspiration for his helpful guidance and encouragement. His commitment and
assistance were limitless and this is greatly appreciated.
The author would like to express his deep gratitude to Prof. Dr.Ing. R. Kreißig for giving
him the permission to complete Doctorate of Engineering at the Chemnitz University of
Technology and for kindly assistance and supervision.
The author would like to thank Prof. Dr.Ing. C. Bisbos, Aristotle University of
Thessalonoki, Greece for having kindly accepted to review this thesis.
The author is thankful to Dr.Ing. Vũ Đức Khôi for help and advice, to Ms Wierskowski
and Ms Dronia for their programming as part of their diploma theses in some parts of FEM
source code. The author’s thanks are also extended to Prof. Dr. rer. nat. Dr.Ing.
S. Sponagel and to the other colleagues at the Biomechanics Laboratory for their helpful
assistance.
The author is immensely indebted to his father Trần Thanh Xuân and his mother Nguyễn
Thị Hòa who have been the source of love and discipline for their inspiration and
encouragement throughout the course of his education including this Doctorate.
Last but not least, the author is extremely grateful to his wife Mrs. Nguyễn Thị Thu Hà
who has been the source of love, companionship and encouragement, to his daughters My
and Ly who have been the source of joy and love.
iii
iv
TABLE OF CONTENTS
INTRODUCTION........................................................................................................ 1
1. FUNDAMENTALS.................................................................................................. 3
1.1 Basic concepts of plasticity................................................................................. 3
1.1.1 Elastic and rigid perfectly plastic materials................................................. 3
1.1.2 Fundamental principles in plasticity............................................................ 4
1.1.3 Drucker’s postulate...................................................................................... 6
1.1.4 Yield criteria ................................................................................................ 7
1.1.5 Plastic dissipation function in local variables.............................................. 8
1.2 Normalized shell quantities ................................................................................ 9
1.2.1 Reference quantities..................................................................................... 9
1.2.2 Stress quantities ........................................................................................... 9
1.2.3 Strain quantities ......................................................................................... 10
1.2.4 StressStrain relation.................................................................................. 11
1.3 Exact Ilyushin yield surface.............................................................................. 12
1.3.1 Derivation .................................................................................................. 12
1.3.2 Description of the exact Ilyushin yield surface ......................................... 14
1.3.3 Reparameterization .................................................................................... 16
1.3.4 Plastic dissipation function........................................................................ 18
1.3.5 Reformulation............................................................................................ 19
2. MATHEMATICAL FORMULATIONS OF LIMIT AND SHAKEDOWN
ANALYSIS IN GENERALIZED VARIABLES ....................................................... 21
2.1 Theory of limit analysis .................................................................................... 22
2.1.1 Introduction................................................................................................ 22
2.1.2 General theorems of limit analysis ............................................................ 23
2.2 Theory of shakedown analysis.......................................................................... 24
2.2.1 Introduction................................................................................................ 24
2.2.2 Definition of load domain.......................................................................... 25
2.2.3 Fundamental of shakedown theorems........................................................ 27
2.2.4 Separated shakedown limit ........................................................................ 30
2.2.5 Unified shakedown limit............................................................................ 33
v
Table of Contents
3. DETERMINISTIC LIMIT AND SHAKEDOWN PROGRAMMING.................. 38
3.1 Finite element discretization............................................................................. 39
3.2 Kinematic algorithm......................................................................................... 41
4. PROBABILISTIC LIMIT AND SHAKEDOWN PROGRAMMING................... 49
4.1 Basic concepts of probability theory ................................................................ 50
4.1.1 Sample space.............................................................................................. 50
4.1.2 Random variables ...................................................................................... 50
4.1.3 Moments .................................................................................................... 51
4.2 Reliability analysis............................................................................................ 53
4.2.1 Failure function and probability .............................................................. 53
4.2.2 First and SecondOrder Reliability Method ............................................. 55
4.3 Calculation of design point ............................................................................... 58
4.4 Sensitivity of the limit state function................................................................ 61
4.4.1 Mathematical sensitivity............................................................................ 62
4.4.2 Definition of the limit state function.......................................................... 63
4.4.3 First derivatives of the limit state function................................................ 65
4.4.4 Second derivatives of the limit state function............................................ 66
4.4.5 Special case of probabilistic shakedown analysis...................................... 70
5. MULTIMODE FAILURE AND THE IMPROVEMENT OF FORM/SORM
RESULTS................................................................................................................... 72
5.1 Multimode failure ............................................................................................. 73
5.1.1 Bounds for the system probability of failure ............................................. 73
5.1.2 Firstorder system reliability analysis........................................................ 74
5.2 Solution technique ............................................................................................ 76
5.2.1 Basic idea of the method............................................................................ 76
5.2.2 Definition of the bulge............................................................................... 77
6. LIMIT AND SHAKEDOWN ANALYSIS OF DETERMINISTIC PROBLEMS 79
6.1 Limit analysis of a cylindrical pipe under complex loading............................. 80
6.2 Limit and shakedown analysis of a thinwalled pipe subjected to internal
pressure and axial force .......................................................................................... 82
6.3 Cylindrical shell under internal pressure and temperature change................... 84
6.4 Pipejunction subjected to varying internal pressure and temperature............. 86
6.5 Grooved rectangular plate subjected to varying tension and bending.............. 90
6.6 Square plate with a central circular hole........................................................... 92
6.7 Elbow subjected to bending moment................................................................ 97
6.8 Limit and shakedown analysis of pipeelbow subjected to complex loads .... 103
6.9 Nozzle in the knuckle region of a torispherical head...................................... 109
vi
vii
7. PROBABILISTIC LIMIT AND SHAKEDOWN ANALYSIS OF STRUCTURES
.................................................................................................................................. 115
7.1 Square plate with a central circular hole......................................................... 115
7.2 Pipejunction subjected to internal pressure ................................................... 121
7.3 Limit analysis of cylindrical pipe under complex loading ............................. 123
7.4 Folding shell subjected to horizontal and vertical loads................................. 129
8. SUMMARY.......................................................................................................... 133
REFERENCES ......................................................................................................... 136
APPENDIX: PROBABILITY DISTRIBUTIONS AND TRANSFORMATION TO
THE STANDARD GAUSSIAN SPACE................................................................. 146
1. Normal distribution........................................................................................... 147
2. LogNormal distribution................................................................................... 147
3. Exponential distribution.................................................................................... 147
4. Uniform distribution ......................................................................................... 148
5. Gamma distribution .......................................................................................... 148
6. Weibull Distribution ......................................................................................... 149
7. Extreme Type I Distribution............................................................................. 149
INTRODUCTION
The present work aims at providing an effective numerical method for the limit and
shakedown analysis (LISA) of general shell structures with the help of the finite element
method. Both deterministic and probabilistic limit and shakedown analyses are presented.
For deterministic problem, three failure modes of structure such as plastic collapse, low
cycle fatigue and ratchetting are analysed based upon an upper bound approach.
Probabilistic limit and shakedown analysis deals with uncertainties originating from the
loads, material strength and thickness of the shell. Based on a direct definition of the limit
state function, the calculation of the failure probability may be efficiently solved by using
the First and Second Order Reliability Methods (FORM/SORM). Since the deterministic
problem is a subroutine of the probabilistic one, thus, even a small error in the
deterministic model can lead to a big error in the reliability analysis because of the
sensitivity of the failure probability. To this reason, a yield criterion which is exact for
rigidperfectly plastic material behaviour and is expressed in terms of stress resultants,
namely the exact Ilyushin yield surface, will be applied instead of simplified ones (linear
or quadratic approximations). The problem of reliability of structural systems (series
systems) will also be handled by the application of a special technique which permits to
find all the design points corresponding to all the failure modes. Studies show, in this case,
that it improves considerably the FORM/SORM results.
The thesis consists of two parts: the theory part (chapters 15) and numerical part
(chapters 67). Chapter 1 introduces some basic concepts of plasticity theory, including the
fundamental principles and yield criteria. Based on the LoveKirchhoff theory, several
relations between physical and normalized values for plates and shells are presented. The
derivation and description of the exact Ilyushin yield criterion is briefly summarized.
In chapter 2, we present the two fundamental theorems of limit and shakedown
analysis, the static and kinematic theorems. Based on the original ones which were
proposed by Melan and Koiter, an extension for lower and upper bound theorems in terms
of generalized variables are proposed and formulated. A simple approach for the direct
calculation of the shakedown limit as the minimum of incremental plasticity limit and
alternating plasticity limit is also presented.
In chapter 3, a kinematic approach of limit and shakedown analysis, which is
adopted for shell structures is developed (the deterministic LISA). Starting from a finite
element discretization, a detailed kinematic algorithm in terms of generalized variables will
2
be formulated and introduced. A simple technique for overcoming numerical obstacles,
such as the nonsmooth and singular objective function, is also proposed.
Chapter 4 focuses on presenting a new algorithm of probabilistic limit and
shakedown analysis for thin plates and shells, which is based on the kinematical approach.
The loads and material strength as well as the thickness of the shell are to be considered as
random variables. Many different kinds of distribution of basic variables are taken into
consideration and performed with First and Second Order Reliability Methods
(FORM/SORM) for calculation of the failure probability of the structure. In order to get
the design point, a nonlinear optimization was implemented, which is based on the
Sequential Quadratic Programming (SQP). Nonlinear sensitivity analyses are also
performed for computing the Jacobian and the Hessian of the limit state function.
Chapter 5 presents a method to successively find the multiple design points of a
component reliability problem, when they exist on the limit state surface. Each design
point corresponds with an individual failure mode or mechanism. The FORM
approximation is, then applied at each design point followed by a series system reliability
analysis leading to improved estimates of the system failure probability.
In chapter 6, we aim at presenting various typical examples of deterministic limit
and shakedown analyses to illustrate and validate the theoretical methods. Numerical
results are tested against analytical solutions, experiments and several limit loads which
have been calculated in literature with different numerical methods using shell or volume
elements.
Numerical studies of limit and shakedown analysis for probabilistic problems are
introduced in chapter 7. Uncertainties which originate from the loads, the strength of
material and the thickness of the shell are all analyzed. For each test case, some existing
analytical and numerical solutions found in literature are briefly represented and compared.
Finally chapter 8 contains some main conclusions and future perspectives.
1 FUNDAMENTALS
In the following, some theoretical foundations are stated, which are necessary for
the developments in the subsequent chapters. We start with a brief introduction of
plasticity theory, including the fundamental principles and yield criteria. Based on the
LoveKirchhoff theory, several relations between physical and normalized values for plates
and shells are presented. The derivation and description of the exact Ilyushin yield criterion
is summarized, which is closely related to the works of Burgoyne and Brennan [1993b],
Seitzberger [2000]. For convenience, we will use only the concept of shells, instead of
plates and shells.
1.1 Basic concepts of plasticity
1.1.1 Elastic and rigid perfectly plastic materials
Mechanical behaviour of rate intensities elasticplastic, nonhardening solid body is
idealized by the elastic perfectly plastic model. In this model, the material behaves
elastically below the yield stress and will begin to yield if the stress intensity reaches the
yield stress. Stresses are not allowed to become higher than this threshold. Furthermore,
the elastic deformation can usually be disregarded when compared with the plastic
deformation. This is equivalent to the rigid plastic material model. It can be proved that
elastic characteristics do not affect the plastic collapse limit state and thus the application
of the elastic perfectly plastic material model becomes same to that of the rigid perfectly
plastic model for limit analysis.
In the geometrically linear theory the total strain
ij
ε is assumed to be decomposed
additively into an elastic or reversible part and an irreversible part . If some thermal
effects occur, a thermal strain term should be added and thus
e
ij
ε
p
ij
ε
θ
ε
ij
e p
ij ij ij ij
θ
ε ε ε ε + . (1.1) = +
The elastic part of the strain obeys Hooke’s law, its relationship with stress is linear
1 e
ij ijkl kl
C ε σ
−
= (1.2)
where , called the elastic constants, are components of a tensor of rank 4. For an
isotropic material, this tensor is expressed in the form below
ijkl
C
( ) ( ) ( )
(
1 1 2 2 1
ijkl ij kl ik jl il jk
E E
C
)
ν
δ δ δ δ
ν ν ν
= +
+ − +
δ δ + (1.3)
( , , , , ) i j k l
αβ
where denotes the Young’s modulus and E ν the Poisson ratio, δ α β = the
Kronecker delta. The inverse relationship of (1.2) can be written as
( )
2
2
1 2
e e
ij ij ij kk
G G
ν
σ ε δ ε
ν
= +
−
)
(1.4)
where
( ν +
=
1 2
E
G is the shear modulus of elasticity.
The plastic strain rate obeys an associated flow law
p
ij
ij
f
ε λ
σ
∂
=
∂
(1.5)
where λ is a nonnegative plastic multiplier and
( )
ij
f f σ = represents a timeindependent
yield surface such as
( ) 0
ij
f σ < corresponds to elastic behaviour, (1.6a)
( ) 0
ij
f σ = corresponds to appearance of plastic deformation, (1.6b)
( ) 0
ij
f σ > corresponds to a region inaccessible for the material. (1.6c)
The definition of yield function means that the stress point cannot move outside the
yield surface. Plastic flow can occur only when the stress point is on the yield surface and
for elasticperfectly plastic material the additional loading
ij
σ can only move along the
tangential direction.
1.1.2 Fundamental principles in plasticity
Consider a structure subjected to volume loads
i
f and surface loads
i
t . The stresses
ij
σ are said to be in equilibrium if they satisfy the equations of internal equilibrium
0 in
j ij i
f V ∂ σ + = (1.7)
and the conditions of equilibrium at the surface of the body
on
i j ij i
t n t V
σ
σ = = ∂ . (1.8)
Any stress field respecting conditions (1.7) and (1.8) is called a statically admissible field.
Furthermore, if this stress field nowhere violates the yield criterion, , it is called
a plastically admissible or licit stress field.
( )
0
ij
f σ ≤
4
The actual flow mechanism composed of the velocities and strain rate
i
u
ij
ε in the
body which satisfy the compatibility and kinematical boundary conditions
( )
1
in
2
ij j i i j
u u ε ∂ ∂ = + V
(1.9)
on
i i
u u V
u
= ∂
. (1.10)
Any mechanism
(
,
i ij
u
)
ε respecting conditions (1.9) and (1.10) is called kinematically
admissible. Furthermore, if this mechanism furnishes a non negative external power
0
E i i i i
V V
W f u dV t u dA
σ
∂
= + ≥
∫ ∫
(1.11)
then it is called a licit mechanism. A kinematically admissible strain and displacement field
can be defined in a similar manner.
Principle of virtual power
One of the main tools in the mechanics of continua is the principle of virtual power,
which states that for an arbitrary set of infinitesimal virtual velocity variation δ that are
kinematically admissible, the necessary and sufficient condition to make the stress field
i
u
ij
σ
equilibrium is to satisfy the following equation
ij ij i i i i
V V V
dV f u dV t u dS
σ
σ δε δ δ
∂
= +
∫ ∫ ∫
ij
δσ
ij
. (1.12)
Principle of complimentary virtual power
For an arbitrary set of infinitesimal virtual variations of the stress tensor that
are statically admissible, the necessary and sufficient condition to make the strain rate
tensor ε
i
and velocity vector u compatible is to satisfy the following equation
u
ij ij j ij i
V V
dV n u dS ε δσ δσ
∂
=
∫ ∫
. (1.13)
Equation of virtual power
From the two above principles, one can easily deduce that for all strain rate tensors
ij
ε
i
u
ij
σ and velocity vectors that are kinematically admissible, and for all stresses that
are statically admissible we have the following virtual work equation
u
ij
σ ε
ij i i i i j ij i
V V V V
dV f u dV t u dS n u dS
σ
σ
∂ ∂
= + +
∫ ∫ ∫ ∫
. (1.14)
5
It is well known that the above variational principles are independent of the
constitutive equation of the material.
1.1.3 Drucker’s postulate
"Over the cycle of application and removal of the set of forces, the new work
performed by the external agency on the changes in the displacements it produces is non
negative" [Martin, 1975].
This is expressed mathematically by the following inequality
( ) 0
0
≥ −
∫
ij ij ij
dε σ σ (1.15)
where
∫
is the integral taken over a cycle of applying and removing the added stress set,
( ) 0 =
ij
f
ij
σ is the stress tensor on the yield surface satisfying the yield condition σ , and
is the plastically admissible stress tensor such that ( )
0
ij
σ 0
0
≤
ij
f σ
0
. Starting from this
postulate, we have the following important consequences [Lubliner, 2005]:
Principle of maximum energy dissipation
0
( )
p
ij ij ij
σ σ ε − ≥ , (1.16a)
or
, (1.16b) 0
p
ij ij
σ ε ≥
sometimes known simply as Drucker’s inequality. It is valid for both workhardening and
perfectly plastic materials.
Normality rule
From (1.16) one can deduce that the plastic strain rates tensor must be normal to the
yield surface at a smooth point or lie between adjacent normals at a corner (nonsmooth
point), see figure 1.1. In the case of having intersected differentiable yield surfaces at a
singular point, (1.5) should be replaced by:
p
ij
ε
n
1
n
p k
ij k
k ij
f
ε λ
σ
=
∂
=
∑
. (1.17)
∂
Convexity of yield surface
It is clear form figure 1.1 that if there is any
0
ij
σ lying on the outside of the tangent,
the inequality (1.16a) is violated. In other words, the entire elastic region must lie to one
side of the tangent. As a result, the yield surface is convex.
6
The convexity of the yield surface has a very important role in plasticity. It permits
the use of convex programming tools in limit and shakedown analysis. It should be noted
that Drucker’s postulate is quite independent of the basic laws of thermodynamics. It does
not hold if internal structural changes occur or for temperature dependent behaviour
[Kalisky, 1985]. Furthermore, the yield surface fails to be convex if there is an interaction
between elastic and plastic deformations, i.e. if the elastic properties depend on the plastic
deformation [Panagiotopoulos, 1985].
ε
p
elastic domain
ε
p
inaccessible domain
yield surface
σ
σ−σ
0
σ
0
Figure 1.1 Normality rule
1.1.4 Yield criteria
The yield criterion defines the elastic limits of a material under a combined state of
stress. The yield function f in stress space may be written with no loss of generality in
terms of the stress deviator and the first invariant of stress, that is
( ) ( )
1
, , f f I = σ , ξ s ξ (1.18)
where
1 kk ij ij
I σ δ σ = = is the trace of
ij
σ , ξ are internal variables which are determined
experimentally and is the deviator defined as
ij
s
1
1 1
3 3
kl kl kl ik jl ij kl ij
s I σ δ δ δ δ δ σ
⎛ ⎞
= − = −
⎜ ⎟
⎝ ⎠
1
. (1.19)
Since the concept of plasticity was first applied to metals, in which the influence of
mean stress on yielding is generally negligible [Bridgman, 1923 and 1952], the oldest and
most commonly used yield criteria are those that are independent of I . They are therefore
formulated with using .We present here two most wellknown yield criteria
in plasticity.
ij
s
1
0
kk
J s = =
7
Tresca criterion
The Tresca yield criterion is historically the oldest one; it embodies the assumption
that plastic deformation occurs when the maximum shear stress over all planes attains a
critical value, namely, the value of the current yield stress in shear, denoted . This
criterion may be represented by the yield function
T
k
( ) ( )
0 k − =
1 2 3 1
2
T
f Max σ σ σ σ = − − σ
2 3
, , σ σ − (1.20)
where
2
y
T
k
σ
= ,
y
σ is yield stress. It can also be expressed explicitly in terms of the
invariants and of the stress deviator tensor as J
2
J
3
3 2 2 2 4 6
2 3 2 3 2 2
( , ) 4 27 36 96 64 0
T T T
f J J J J k J k J k = − − + − = . (1.21)
Von Mises criterion
The von Mises criterion is known as the maximumoctahedralshearstress or
maximumdistortionalenergy criterion, which states, that yielding begins when the
octahedral shearing stress reaches a critical value such as
v
k
2
v
k
2 2
( ) 0 f J J = − = (1.22)
where
3
y
v
k
σ
= . We may also formulate the von Mises yield criterion in the form of
principal stresses
. (1.23)
2 2 2
1 2 2 3 3 1
( ) ( ) ( ) 6
v
k σ σ σ σ σ σ − + − + − =
2
The projection of the Tresca yield surface in the ( ) ( )
1 3 2 3
σ σ σ σ − − plane takes the
form of an irregular hexagon, where the von Mises criterion is an ellipse. In the next
section the von Mises criterion will be used to derive the socalled exact Ilyushin yield
surface for the elasticplastic analysis of shelllike structures.
1.1.5 Plastic dissipation function in local variables
The plastic dissipation function is defined by
*
max( )
p p
ij ij ij ij
D
p
σ ε σ ε = = (1.24)
where is a plastically admissible stress tensor, i.e. satisfying
*
ij
σ ( ) 0
*
≤
ij
f σ ,
ij
σ is the
stress tensor satisfying yield condition ( ) 0
ij
f σ = . The first equality of (1.24) is the
definition, while the second one is due to the Drucker stability postulate:
( )
*
0
p
ij ij ij
σ σ ε − ≥ .
The plastic dissipation for the Mises criterion and associated flow rule is given by
[Lubliner, 2005]
8
( ) 2
p p p p
ij v ij ij
D k ε ε ε = . (1.25)
1.2 Normalized shell quantities
1.2.1 Reference quantities
The representation of the exact Ilyushin yield surface and the description of its
applications for the limit and shakedown analysis of shell structures are given in terms of
normalized generalized stresses and strains. The membrane forces and bending moments
are normalized with respect to the plastic limit loads in uniaxial tension and bending,
respectively
2
0 0
, M
4
y y
h
N h σ σ = = (1.26)
where
y
σ and are yield stress and shell thickness, respectively. Physical strain values
are referred to the reference strain
h
0
ε . A convenient measure for
0
ε is given by
2
0
(1 )
y
E
ν
ε σ
−
= . (1.27)
0
ε corresponds to the elastic strain of an uniaxially stretched plate at the yield stress
y
σ .
For the sake of simplicity, the reference “curvature”
0
κ is defined as follows
0
0
4
h
ε
κ = . (1.28)
so that
0 0 0 0
N M ε κ = , [Burgoyne and Brennan 1993b], [Seitzberger, 2000]. This definition
does not represent a kinematic relation.
A dimensionless thickness coordinate z is used for throughthickness integrations.
The relation between z
3
s and the thickness coordinate is as follows
3
s
z
h
= . (1.29)
1.2.2 Stress quantities
Physical and normalized stress vectors for a state of plane stress are given by,
respectively
11 11
22 22
12 12
1
and
y
σ σ
σ σ
σ
σ σ
⎛ ⎞ ⎛ ⎞
⎜ ⎟ ⎜ ⎟
= =
⎜ ⎟ ⎜ ⎟
⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠
σ σ . (1.30)
9
From these, the normalized membrane force and bending moment vectors and can be
obtained as follows
n m
11
/ 2 1/ 2 1/ 2
22 3
0 0 / 2 1/ 2 1/ 2
12
11
/ 2 1/ 2 1/ 2
2
22 3 3 2
0 0 / 2 1/ 2 1/ 2
12
1 1 1
1 1 4
4
h
y h
h
y h
N
N ds hdz dz
N N h
N
M
M s ds h zdz zdz
M M h
M
σ
σ
− − −
− − −
⎛ ⎞
⎜ ⎟
= = = =
⎜ ⎟
⎜ ⎟
⎝ ⎠
⎛ ⎞
⎜ ⎟
= = = =
⎜ ⎟
⎜ ⎟
⎝ ⎠
∫ ∫ ∫
∫ ∫
n σ σ σ
m σ σ
∫
σ
(1.31)
where the terms N
αβ
and { } , ( , 1, 2 ) M
αβ
α β ∈ are the physical membrane force and
bending moment components of the shell, respectively.
s
N
11
N
22
M
12
M
11
N
12
12
N
12
M
M
22
q
1
3
s
2
s
Figure 1.2 Static shell quantities
1.2.3 Strain quantities
The physical strain and curvature vectors
11 11
22 22
12 12
,
2 2
ε κ
ε κ
ε κ
⎛ ⎞ ⎛
⎜ ⎟ ⎜
= =
⎜ ⎟ ⎜
⎜ ⎟ ⎜
⎝ ⎠ ⎝
ε κ
⎞
⎟
⎟
⎟
⎠
(1.32)
are normalized with respect to the
0
ε and
0
κ , which gives
11 11
22 22
0 0
12 12
1 1
,
2 2
ε κ
ε κ
ε κ
ε κ
⎛ ⎞ ⎛
⎜ ⎟ ⎜
= =
⎜ ⎟ ⎜
⎜ ⎟ ⎜
⎝ ⎠ ⎝
e k
⎞
⎟
⎟
⎟
⎠
. (1.33)
The LoveKirchhoff assumptions state that the normals (i.e. the lines perpendicular
to the shell’s midplane) remain straight, unstretched and normal (i.e. they always make a
10
right angle to the midplane) after loading. These mean that the transversal shear strains in
the thickness coordinate are negligible which is only valid for thin shells with small
displacement ( u ). Based on the LoveKirchhoff hypothesis the physical strain
vector can be written in the standard form as follows
3
s
3
u
3
h
3
( ) s
3
s = + ε ε κ (1.34)
where ε and are physical midplane strain and curvature vectors ( ). The relation
between them and the displacements delivers
κ
3
0 s =
1 2 2
11 22 12
1 2 1
2 1 2
11 22 12
1 2 2
, , 2
, , 2
u u u
s s s
s s s
∂ ∂ ∂ ∂
ε ε ε
∂ ∂ ∂ ∂
1
2
1
1
u
s
s
∂θ ∂θ ∂θ
κ κ κ
∂θ
∂ ∂ ∂
= = = +
= = − = −
∂
(1.35)
with { } , ( 1,2 ) u
α α
θ α ∈ being translations and rotations of the midplane, respectively. By
the introduction of z and
0
ε , (1.34) becomes
0 0 0
1 1 4
z
ε ε κ
= + ε ε κ . (1.36)
By using of (1.28), the Kirchhoff hypothesis may be written in normalized form as
( ) 4 z z = + e e k . (1.37)
1.2.4 StressStrain relation
As will be shown in the subsequent chapters, it is convenient to group both the
strain vectors and the curvature vectors as well as the section force and moment vectors
into “engineering” vectors, as follows
, =
⎛ ⎞ ⎛ ⎞
=
⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠
e n
ε σ
k m
. (1.38)
For an elastoplastic material behaviour, the tangential stressstrain relations for a
Kirchhoff shell may be written as
=
d d
d d
d d
⎛ ⎞ ⎛ ⎞⎛ ⎞ ⎛ ⎞
= =
⎜ ⎟ ⎜ ⎟⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠⎝ ⎠ ⎝ ⎠
n C B e C B
σ ε
m B D k B D
(1.39)
where are submatrices of the physical tangential stiffness matrix. For purely
elastic material behaviour we have
B, C, D
11
1 0
4
0, 1 0 ,
3
1
0 0
2
ν
ν
ν
⎛ ⎞
⎜ ⎟
⎜ ⎟
= = =
⎜ ⎟
⎜ ⎟
−
⎜ ⎟
⎝ ⎠
B C D C. (1.40)
1.3 Exact Ilyushin yield surface
In order to calculate the maximum collapse load of a shell, a criterion is needed to
assess when the shell reaches a situation where the behaviour is governed by plasticity.
Two approaches have been identified. It is possible to work either in terms of stresses
(Moxham [1971], Little [1977]) which vary through the thickness, in which case a yield
criterion such as the von Mises criterion is used, or in terms of stress resultants (Crisfield
[1973], Frieze [1975]), when a more complex fully plasticity yield surface is needed. When
dealing with stress resultants, it is of course important to identify a yield surface, which
marks the limiting values of the stress resultants, beyond which the shell may not be
loaded. In 1948 Ilyushin published the derivation of a stress resultant yield surface
describing the case where a crosssection of a shell is fully plastified and thus reaches its
load capacity. This yield surface, however, has not been used because the parametric form
in which it was described by Ilyushin was not amenable to calculation. Some
approximations have been used, e.g. a linear approximation proposed by Ilyushin himself,
the Ivanov yield surface. A reparameterization of the exact Ilyushin yield surface for thin
shells which produces a simpler (though still exact) form was presented by Burgoyne and
Brennan [1993b]. Their work opens the way for the practical use of the exact Ilyushin yield
surface in structural calculations.
1.3.1 Derivation of the exact Ilyushin yield surface
The derivation of the exact Ilyushin yield surface is based on the following
assumptions
• perfectly plastic isotropic material behaviour obeying the von Mises yield criterion,
• validity of the normality rule for the plastic deformations,
• plane stress conditions in each material point,
• validity of the Kirchhoff hypothesis for both total and plastic strains.
At the limit state, each material point through the thickness has a plastic material
behaviour. Thus both, the von Mises yield condition and the normality rule, are valid
( )
1 1/ 2 0
1 0, 1/ 2 1 0
0 0
T
f
−
⎛ ⎞
⎜
= − = = −
⎜ ⎟
⎜ ⎟
⎝ ⎠
σ σ Pσ P
3
⎟
(1.41)
2
p
f
d d d ξ ξ
∂
= =
∂
e σ
σ
P . (1.42)
12
From Eqs. (1.41) and (1.42) σ and dξ can be expressed as functions of d
p
e
( )
( )
1
1
2
p p
p
d d
d d ξ
−
= σ e P e
e
(1.43)
( ) ( )
1
1
4
T
p p
d d d d ξ
−
= e e P
p
e . (1.44)
It is assumed that the plastic strain increment resultants obey the Kirchhoff
hypothesis. From (1.37) we have
( ) 4
p p
d z d z d = + e e
p
k . (1.45)
Substituting (1.45) into (1.44), which gives the consistency parameter
2
1
2
3
d P P z P
ε εκ κ
ξ = + + z (1.46)
with the incremental plastic strain resultant intensities
( ) ( )
( ) ( )
( ) ( )
1
1
1
2
1
3
3 3
( 0)
4 4
3 3
12 12 ( 0)
T T
p p p p
T T
p p p p
T T
p p p p
P d d d d
P d d d d
P d d d d
ε
εκ
κ
−
−
−
= =
= =
= =
e P e ε P ε
e P k ε P ε
k P k ε P ε
≥
≥
(1.47)
where
1 1
1 2 3 1 1
/ 2
, ,
/ 2
− −
− −
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
= = =
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠ ⎝ ⎠
0 0 P 0 0 P
P P P
0 P 0 0 P 0
. (1.48)
These incremental plastic strain resultant intensities are subject to the condition (by
the Schwarz inequality)
2
P P P
ε κ εκ
≥ (1.49)
Substitution of Eqs. (1.45) and (1.46) in Eq. (1.43) gives
(
1
2
3 1
4
2
2
p
d z d
P P z P z
ε εκ κ
−
=
+ +
σ P e k
)
p
+ . (1.50)
From Eqs. (1.31) and (1.50), the stress resultants may finally be written as
1 1 1 1
0 1 0 1
1 1 1 1
1 2 1 2
4 4 3 3
2 2 4 16 4 16
p
p
p
d J J J J
d
J J J J d
− − − −
− − − −
⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
= = =
⎜ ⎟ ⎜ ⎟ ⎜ ⎜ ⎟
⎝ ⎠ ⎝ ⎠ ⎝ ⎠ ⎝ ⎠
n e P P P P
σ ε
m P P P P k
⎟
(1.51)
where the integrals (not to confuse with the invariants of the deviator) can be calculated
as follows
i
J
13
1/ 2
2
1/ 2
1
3
2
i
i
z
J
P P z P z
ε εκ κ
−
=
+ +
∫
dz
i
J
. (1.52)
Equation (1.51) can be regarded as a sixdimensional stress resultant yield surface
for the limit that the shell is wholly plastic and thus in each point over the thickness the
von Mises yield criterion and the normality rule are satisfied. If the direction of the plastic
strain increment resultants is given, the stress resultants can be obtained from Eq. (1.51),
provided the integrals can be evaluated numerically.
1.3.2 Description of the exact Ilyushin yield surface
Corresponding to the quadratic strain resultant intensities, quadratic stress resultant
intensities can also be defined
( 0)
( 0)
T
t
T
tm
T
m
Q
Q
Q
= ≥
=
= ≥
n Pn
n Pm
m Pm
. (1.53)
From these, the surface (1.51) can be reduced to a surface in the threedimensional
Qspace as follows
( )
2 2
0 0 1 1
2
0 1 0 2 1 1 2
2 2
1 1 2 2
/ 4 3 / 2 2
/16
t
tm
m
J J J J
Q P
Q J J J J J J J
Q P
J J J J
ε
P
εκ
κ
⎛ ⎞
⎛ ⎞ ⎛
⎜ ⎟
⎜ ⎟ ⎜
= +
⎜
⎜ ⎟ ⎜
⎜ ⎟ ⎜ ⎟ ⎜
⎜ ⎟
⎝ ⎠ ⎝
⎝ ⎠
⎞
⎟
⎟
⎟
⎟
⎠
. (1.54)
The surface is bounded by the condition that
2
t m tm
QQ Q ≥ (1.55)
which corresponds to
2
P P P
ε κ εκ
≥
( , F F Q
. Equation (1.54) describes a surface, which can be
represented in parameter form as a function of two independent parameters. An implicit
form of Eq. (1.54), i.e. , however, can not be obtained. Ilyushin
[1948] represented the surface in parameterized form, by introducing the two following
parameters
, ) 0
t tm m
Q Q = =
( )
1/ 2
1/ 2
2
/ 4 P P P
ε εκ κ
⎛ ⎞ − +
,
/ 4
.
/ 4
P P P
P P P
P P P P
ε εκ κ
ε κ εκ
κ ε εκ κ
μ
⎜ ⎟
+ +
⎝ ⎠
⎛ ⎞
−
=
⎜ ⎟
+ +
⎝ ⎠
ζ =
(1.56)
The resulting equations of the exact Ilyushin yield surface are
14
( )
( )
( ) ( ) ( )
2 2 2
2
1
2 2 2 2
3
1
2 2 2 2 2 2 2 2 2 2
4
1
1
2
4
4 2 2 2
t
tm
m
Q
Q
Q
μ ψ ϕ
μ ψ ϕ μ ϕψ ϕχ
μ ψ μ ϕ μ μ ϕψ μ ψχ ϕχ χ
= +
Δ
= Δ + Δ + +
Δ
= + Δ + + Δ + Δ − + Δ
Δ
+
(1.57)
where
( ) ( )
2 2
2 2 2
2 2 2
1
2
1
1
1 1
ln ln
1
1
1
2
ϕ ζ
μ ζ ζ μ
ψ
μ μ
χ μ ζ ζ μ
μ ζ μ
ζ
= −
+ − + −
= ±
= − ± −
Δ = − ± −
−
Δ =
Δ
(1.58)
subject to the conditions
0 1, 1 μ μ ζ ≤ ≤ ≤ ≤ . (1.59)
The boundary is given by 0 μ = . Ilyushin’s original parameterization makes it
necessary to divide the surface into four regions, known as the “inplane dominant” and
“bending dominant” regions, which are governed by different equations (due to the
alternative signs). Since the lines of constant ζ and μ are virtually parallel in many cases
and, a numerical algorithm based on these parameters will be illconditioned and
numerically unstable [Burgoyne and Brennan, 1993b].
In his original paper Ilyushin proposed a linear approximation to his exact surface,
which is usually referred to as Ilyushin yield surface
1
1
1
3
t tm m
F Q Q Q = + + = . (1.60)
This crude approximation consists of two planes in the Qspace. It introduces a
discontinuity at the line Q . Figure 1.3 shows a graphical representation of the exact
and the linear approximation of the yield surface in the Qspace. As can be seen, the
surface is symmetric with respect to the
0
tm
=
t m
Q Q − plane. Thus, it can also be plotted in two
dimensional form as Q over Q
tm t m
Q − without loss of clarity [Burgoyne and Brennan,
1993b], see figure 1.3.
15
Figure 1.3 Exact and linear approximation of Ilyushin yield surfaces (from Burgoyne and
Brennan, 1993b)
Ivanov [1967] proposed a quadratic approximation of the exact Ilyushin yield
surface
2
2 2
2
1
/ 4 1
2 4 0.48
m t m
t m tm
t m
Q QQ
F Q Q Q
Q Q
⎛ ⎞ −
= + + + − =
⎜
+
⎝ ⎠
tm
Q
⎟
1
t
. (1.61)
The Ivanov yield surface overcomes many of the difficulties associated with the
approximate Ilyushin yield surface, it has no discontinuities in slope except one at Q = ,
where the exact surface also has a slope discontinuity, and always lies within 1% of the
exact Ilyushin yield surface.
Further suggestions of approximate full plasticity yield surfaces, partly including
the effect of transverse shear as well as hardening effects, can be found e.g. in [Robinson,
1971]. According to Robinson, the maximum error of the linear approximation is 6% on
the safe side and 3,5% on the unsafe side. However, the error can increase up to
approximately 10% according to [Preiss, 2000]. In structural reliability analyses such
errors in the deterministic model are not acceptable because of the sensitivity of the failure
probability.
1.3.3 Reparameterization
In order to avoid the difficulties arising with the parameterization of Ilyushin and
open a possibility using the exact yield surface in practical computations, Burgoyne and
Brennan [1993b] introduced the parameters
16
2
, and =
P P
P P
ε εκ
κ κ
υ β γ = = − − υ β (1.62)
where β and γ are proposed as two independent parameters for the description of the
yield surface. β has the physical meaning of being the position within the thickness of the
shell, where the consistency parameter dξ in Eq. (1.46) is a minimum. With these
parameters, the yield surface assumes the form
( )
( ) ( )
( )
2
2
0 1 0
0 1 1 2 0
2
2
1 2 1
4
16 16
t
tm
m
Q K K K
Q K K K K K
Q K K K
β β γ
β γ
= −
= − − +
= − +
β γ +
1
4 K (1.63)
where the integrals are given by
i
K
1/ 2
2
1/ 2
3
2
i
i i
z
K P J
P P z P z
κ
ε εκ κ
−
= =
+ +
∫
dz (1.64)
Figure 1.4 One half of the exact Ilyushin yield surface ( ) 0
tm
Q ≤ constructed in terms of β
and γ (from Burgoyne and Brennan, 1993b)
This yield surface is subject to the conditions
2
0 ,
0 .
β υ
γ
≤ ≤ ≤ ∞
≤ ≤ ∞
(1.65)
The integrals
i
K can be evaluated analytically giving
17
( ) ( )
( ) ( )
( ) ( )
( ) ( ) ( ) ( )
2
0
2
2 2
1 0
2 2
2 1 0
0.5 0.5
ln ,
0.5 0.5
0.5 0.5 ,
2 0.5 0.5 0.5 0.5 2 .
K
K K
K K
β γ β
β γ β
β γ β γ β
β β γ β β γ β γ
⎛ ⎞
− + + −
⎜ ⎟
=
⎜ ⎟
+ + − +
⎝ ⎠
= − + − + + +
= + − + + − + + + − K
( ) 0
tm
Q ≤
(1.66)
Figure 1.4 shows a twodimensional representation of one half of the yield surface
in terms of the two parameters β and γ . This representation is the most
convenient for calculations, since nowhere on the yield surface do lines of constant values
of the two parameters become parallel.
1.3.4 Plastic dissipation function
The derivation and description of the exact Ilyushin yield surface presented above
was performed with incremental strain quantities. For the evaluation of the power of
internal forces, however, a description in terms of strain rate quantities is more convenient.
It is to be noted that the state relations are not affected, if the strain rate quantities are used
throughout in stead of the incremental strain quantities, provided the reference values
0
ε
and are replaced by reference strain and curvature rates
0
κ
0
ε
0
κ
ε
and , respectively.
According to Eq. (1.34) a dimensionless generalized strain rate vector is defined
( )
11 22 12 11 22 12
0 0 0 0 0 0
1 1 1 1 1 1
2 2
T
ε ε ε κ κ κ
ε ε ε κ κ κ
= =
⎜ ⎟
⎝ ⎠
ε e k
⎛ ⎞
(1.67)
replacing the generalized strain increment vector . dε
The plastic dissipation function for a shell structure may be written in the form
0 0
T
p
p
p
p
D
d
N ε
⎛ ⎞
⎛ ⎞
= = =
⎜ ⎟
⎜ ⎟
⎝ ⎠
⎝ ⎠
n e
σ ε
m
k
T p
p
d
p
(1.68)
where and are the physical and normalized plastic dissipations per unit area of the
midplane of the shell. With the sixdimensional representation of the exact Ilyushin yield
surface Eq. (1.51) (written in rate form), may be expressed as a function of the strain
rate resultant quantities ε
p
D
p
d
( ) ( )
1 1
0 1
1 1
1 2
4 3
2 4 16
T
p p p p
J J
d
J J
− −
− −
⎛ ⎞
=
⎜ ⎟
⎝ ⎠
P P
ε ε ε
P P
. (1.69)
Analytical evaluation of this relation gives
( )
1/ 2
2
1/ 2
2
2
3
p p
d P P z
ε εκ κ
−
= + +
∫
ε
P z dz (1.70)
18
which finally may be written as [Seitzberger, 2000]
( )
2 2
1 1 2 2 0
2 for 0 ( )
3
for 0 ( )
3
p
P
P a
d
P
K P
ε
κ
κ
κ
β β γ β β γ γ
⎧
=
⎪
⎪
=
⎨
⎪
+ + + + >
⎪
⎩
b
(1.71)
where
1
β and
2
β are
1 2
0.5 and 0.5 β β β = − = + β . (1.72)
It is to be noted that the value of will become indefinite if both conditions
0
K
0.5 β ≤ 0 and = are fulfilled. However, as long as γ γ is not exactly equal to zero, but to
some small positive numbers, a “regularized” evaluation of may be obtained
[Seitzberger, 2000]. Otherwise, in general, is convex but not everywhere differentiable
[Capsoni and Corradi, 1997]. In order to allow a direct nonlinear, nonsmooth, constrained
optimization problem, as will be discussed later, a “smooth regularization method” will be
used for overcoming the nondifferentiability of the objective function. On the other hand,
without the loss of generality, it is supposed that
0
K
P
p
d
κ
is always positive in order to have the
expression of the plastic dissipation function as described in (1.71b). To this end it is
necessary to add to γ and to a small positive number. Thus, in this case, Eq. (1.71) is
amenable to a numerical evaluation for all values of ε .
P
κ
p
y
1.3.5 Reformulation
For our general purpose to deal with probabilistic problems, as will be discussed
later, the yield limit σ and thickness h might be no longer constant but random variables,
and then the reference quantities are also changed together with the different ‘realizations’
of random variables. Additionally, in reliability analysis, the sensitivities are required
which contain first and second derivatives versus loading, material strength and thickness
random variables. Thus some expressions of quantities which are necessary for analysis
algorithm should be reformulated. Let us restrict ourselves to the case of homogeneous
material and shells of constant thickness in which the yield limit
y
σ and thickness are
the same at every Gaussian point of the structure. So we always can write
h
0 0 y
Zh , Y h σ σ = = (1.73)
where
0 0
, h σ are constant reference values and are random variables. By that way the
normalized quantities in Eqs. (1.26), (1.27) and (1.28) assume the new form as follows
, Y Z
0
0 0 0
y
h
N
N h
YZ YZ
σ
σ = = =
(1.74a)
19
2 2
0
0 0 2 2
4 4
y
h h
M
Y Z YZ
σ
σ = = =
0
M
(1.74b)
( ) ( )
2 2
0
0
0
1 1
y
E YE
σ ν σ ν
Y
ε
ε
− −
= = = (1.74c)
and
0 0
0
0
4 4
0
Z Z
h Yh Y
ε ε
κ κ = = =
. (1.74d)
With the new normalized quantities, the new “engineering” strain and stress vectors
are obtained
2
ˆ ˆ , =
Y
YZ
Y
YZ
Z
⎛ ⎞
⎛ ⎞
⎜ ⎟
=
⎜ ⎟
⎜ ⎟
⎝ ⎠
⎝ ⎠
e
n
ε σ
k m
(1.75)
and the incremental plastic strain resultant intensities have the form
( ) ( ) ( )
( ) ( )
( )
2
1 1
2
2 2
2
3 3 2
3 3
ˆ
ˆ ˆ ( 0)
4 4
ˆ
ˆ ˆ 3 3
ˆ
ˆ ˆ 12 12 = ( 0)
T T
p p p p
T T
p p p p
T
T
p p p p
P Y Y Y P
Y Y
P Y P
Z Z
Y Y Y
P P
Z Z Z
ε ε
εκ εκ
κ κ
= = = ≥
⎛ ⎞
= = =
⎜ ⎟
⎝ ⎠
⎛ ⎞ ⎛ ⎞
= =
⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠
ε Pε ε P ε
ε P ε ε P ε
ε Pε ε P ε
≥
(1.76)
Adapting the parameters , , υ β γ which were introduced in (1.62) with these new
strain resultant intensities and then substitute them into (1.71), we obtain the new
expression of the plastic dissipation function
( )
0 0
2 2
0 0 1 1 2 2 0
ˆ
ˆ
2 for 0 (a)
3
ˆ
ˆ
for 0 (b)
3
p
P
YZN P
D
P
YN K P
ε
κ
κ
κ
ε
ε β β γ β β γ γ
⎧
= ⎪
⎪
=
⎨
⎪
+ + + + >
⎪
⎩
(1.77)
with the new
1 2
, β β and are
0
K
2
1
1 2 0
2
2 2
, , ln
2 2
Z Z
K
1
β γ β
β β β β
β γ β
⎛ ⎞
+ +
⎜ = − = + =
⎜
+ −
⎝ ⎠
⎟
⎟
. (1.78)
20
2 MATHEMATICAL FORMULATIONS OF LIMIT AND
SHAKEDOWN ANALYSIS IN GENERALIZED VARIABLES
It is the objective of structural analysis to determine the load carrying capacity. In
the early 20
th
century, it has been relatively easily defined by forcing the stress intensity at
a certain point of the structure to attain the yield stress of the material. This implies that
structural failure occurs before yielding. However, many materials, for example the
majority of metals, exhibit distinct, plastic properties. Such materials can deform
considerably without breaking, even after the stress intensity attains the yield stress. This
implies that if the stress intensity reaches the critical (yield) value, the structure does not
necessarily fail or deform extensively. To this case, in order to permit higher loads, elastic
plastic structural analysis becomes more general than the classical elastic one. Among the
plasticity methods, Limit and Shakedown Analysis (LISA) seems to be the most powerful
one. In Europe LISA has been developed as a direct plasticity method for the design and
the safety analysis of severely loaded engineering structures, such as nuclear power plants
and chemical plants, offshore structures etc. [Staat, 2002], [Staat and Heitzer, 2003a].
Annex B of the new European pressure vessel standard EN 134453 is based on LISA
[European standard, 200506], [Taylor et al., 1999] thus indicating the industrial need for
LISA software. All design codes are based on perfectly plastic models. The extension of
LISA to hardening materials is no problem [Staat and Heitzer, 2002].
Based on the elasticperfectly plastic or rigidperfectly plastic models of material
and considering the loads as monotonic and proportional, limit analysis evaluates the
plastic collapse load or the largest load to which the structure would be subjected during its
lifetime. Beyond this limit, the structure will fail due to global plastic flow. Limit analysis
was pioneered approximately from the works of Kazincky in 1914 and Kist in 1917. The
first complete formulations of both upper bound and lower bound theorems were
established later by Drucker, Greenberg and Prager and the alternative formulation using
rigidplastic material was given by Hill in 1951. Since then, the applications of limit
analysis theory in engineering have been widely reported, such as the works of Hodge
[1959, 1961, 1963], Maier [1970], Prager [1972], Martin [1975], Lubliner [2005] and
Capsoni and Corradi [1997].
However, in practice, the loads are generally timedependent or may vary
independently. Practical experience showed that in the case of variable repeated loads, not
only can lowcycle fatigue cause structural failure below the plastic collapse load
calculated with limit analysis but also an accumulation of plastic deformations may occur,
resulting in excessive deflections of the structure [König, 1987]. It may also happen that
the structure comes back to its elastic behaviour after a certain time period. In this way, a
new branch of plasticity, the theory of shakedown came to existence. The first static
shakedown theorem was formulated by Bleich [1932] for a system of beam of ideal I
crosssections. This static theorem was then extended by Melan [1936] to more the general
case of a continuum. In an alternative way, Koiter [1960] developed a general kinematic
shakedown theorem based on an analogy to limit analysis. He stated and proved the plastic
analysis theorems, i.e. the limit analysis and the shakedown ones in the form used
nowadays. Since then, large amount of work were reported in the literature, cf. Maier
[1969, 1973], König [1966, 1969, 1972], Polizzotto et al. [1993a, b], Ponter et al. [1997a,
1997b, 2000], Nguyen Dang et al. [1976, 1990], Weichert et al. [1988, 2002], Staat et al.
[1997, 2001, 2003b].
The theory of limit and shakedown analysis were established long time ago.
However their numerical applications encountered some difficulties and were limited to
some relatively simple structures. In the case of more complex structures, e.g. shelllike
structures, a direct application of the limit and shakedown theorems is cumbersome if
possible at all. The mathematical difficulties arising from the need to use a three
dimensional analysis are difficult to be overcome. In this case it is convenient to introduce
some generalized, integrated variables, which are used to describe the static and kinematic
quantities. Such an approach allows one to reduce complicated threedimensional problems
to simple plane or onedimensional ones.
In this chapter, two fundamental theorems of limit and shakedown analysis are
introduced, the static and kinematic theorems. Based on the original ones which were
proposed by Melan and Koiter, an extension for lower and upper bound theorems in terms
of generalized variables are proposed and formulated. A simple approach for direct
calculation of the shakedown limit as the minimum of incremental plasticity limit and
alternating plasticity limit, namely the unified shakedown limit method, is also presented.
2.1. Theory of limit analysis
2.1.1 Introduction
Let us consider a structure of volume V made of elasticperfectly plastic or rigid
perfectly plastic material and subjected to external loading . The external loading
consists of general body force
P P
f in V and surface traction t V on
σ
∂ . We assume that all
loads are applied in a monotonic and proportional way
0
α = P P (2.1)
where
( )
0 0 0
, = P f t denotes the nominal or initial load. If the value of α is sufficiently
small, the body behaves elastically. As α increases and reaches a special value, the first
point in the body reaches the plastic state. This state of stress is called elastic limit. Further
22
increase of α will lead to the expansion of plastic region in the body. The structure
gradually forms a collapse mechanism. At limit state, the structure fails to support the
applied load and collapses. If represents the applied load, the value
0
P
l
α corresponding to
the plastic collapse state is called the safety factor of the structure or the limit load
multiplier.
The theory of limit analysis offers a way to solve directly the problem of estimating
the plastic collapse load, bypassing the spreading process of the plastic flow. The limit
value of the load is estimated and at the same time the limit state of stress in the whole
structure can be evaluated. The limit load and stresses so obtained are of great interest in
practical engineering whenever the perfectly plastic model and small deformation
assumption constitute a good approximation of the material.
2.1.2 General theorems of limit analysis
Lower bound theorem
Based on the variational principle of Hill for perfectly plastic material, also known
as the principle of maximum plastic work, the lower bound theorem of limit analysis can be
stated as follows
The exact limit load factor
l
α is the largest one among all possible static solutions
corresponding to the set of all licit stress fields , that is
−
l
α σ
l l
α α
−
≤ . (2.2)
To prove this theorem the principle of virtual work and the property of convexity of
the yield surface are used, [Hodge, 1959], [Prager, 1959], [Lubliner, 2005]. Then, the task
of computing the limit load factor becomes a nonlinear optimization problem
[ ]
( )
0
max
0 (a)
s.t.:
0 (b)
l l
l
L
f
α α
α
−
−
=
+ =
≤
σ P
σ
⎧
⎪
⎨
⎪
⎩
(2.3)
where Eq. (a) is equilibrium condition, denotes a linear operator (usually differential
one) and is the yield function. For the continuum we have .
L
( ) σ f [ ] iv σ σ L d =
Upper bound theorem
The upper bound theorem can be demonstrated as
The actual limit load multiplier
l
α is the smallest one of the set of all multipliers
corresponding to the set of all licit velocity fields u
+
l
α
l l
α
+
≤ α (2.4)
where
23
0 0
(a)
( ) (b)
0 (
on (d)
l in ex
p p
in
V
T T
ex
V V
u
W W
W D dV
W dV dS
V
σ
α
+
∂
=
=
= + >
= ∂
∫
∫ ∫
ε
f u t u
u 0
c)
n
(2.5)
with and the total power of the internal deformation and the power of the external
loads of the structure. The upper bound theorem permits us to estimate the limit load factor
by solving the following optimization problem (written in a normalized form)
in
W
ex
W
min
s.t.: 1
l i
ex
W
W
α =
=
. (2.6)
2.2. Theory of shakedown analysis
2.2.1 Introduction
It has been understood in limit analysis considered above that the loading is simple,
namely, monotonic and proportional mechanical load (selfequilibrium thermal load has no
effect on classical limit analysis). In practice, however, structures are often subjected to the
action of varying mechanical and thermal loading. These loads may be repeated (cyclic) or
varying arbitrarily in certain domain. In this case, loads which are less than plastic collapse
limit may cause the failure of the structure due to an excessive deformation or to a local
break after a finite number of loading cycles.
Inelastic structures such as for example pressure vessels and pipelines subjected to
variable repeated or cyclic loading may work in four different regimes, which are
presented in the Breediagram (figure 2.1, [Bree, 1967]) together with the evolution of the
structural response: elastic, shakedown (adaptation), inadaptation (nonshakedown), and
limit (ultimate) state. Since for the elastic regime there are no plastic effects at all, whereas
for the adaptation regime the plastic effects are restricted to the initial loading cycles and
then they are followed by asymptotically elastic behaviour, both regimes are considered as
safe working ones and they constitute a foundation for the structural design. We do not
consider elastic failure such as buckling or high cycle fatigue here. The inadaptation
phenomena such as low cycle fatigue and or ratchetting should be avoided since they lead
to a rapid structural failure. At the limit load the structure looses instantaneously its load
bearing capacity. Limit and shakedown analyses deal directly with the calculation of the
load capacity or the maximum load intensities that the structure is able to support. The
structural shakedown takes place due to development of permanent residual stresses which,
imposed on the actual stresses shift them towards purely elastic behaviour. Residual
stresses are a result of kinematically inadmissible plastic strains introduced to the structure
by overloads. They clear out effects of all preceding smaller loads. They also avoid any
plastic effects in the future provided that the loads are smaller than the initial overload.
24
Therefore, in limit and shakedown analyses the knowledge of the exact load history is not
necessary. Only the maximum loads (limits) count and the envelopes should be taken into
consideration.
σ
mech.
0
σ
ther.
σ
σ
0
1
1
2
0
pure elastic
behaviour
collapse
σ
ε
σ
ε
σ
σ
ε
ratchetting
low cycle fatigue
elastic shakedown
ε
Figure 2.1 Breediagram of a pressurized thin wall tube under thermal and
mechanical loads
Viewing the situations above, one can see that the first and second situations may
not become dangerous but maximum exploitation of materials can only be attained in the
adaptation or shakedown case. The maximum safe load is defined as the shakedown load
avoiding low cycle fatigue and ratchetting. Thus, the main problem of shakedown theory is
to investigate whether or not a structure made of certain material will shake down under
the prescribed loads.
2.2.2 Definition of load domain
We study here the shakedown problem of a structure subjected to n time
dependent (thermal and mechanical) loads ( ) t P
k
0
with time is denoted by t , each of them
can vary independently within a given range
( )
0 0 0
, , ,
k k k
1,
k k k
P t I P P P k n μ μ
− + − +
⎡ ⎤ ∈ = =
⎣ ⎦
k
⎡ ⎤ =
⎣ ⎦
. (2.7)
25
These loads form a convex polyhedral domain of n dimensions with
vertices in the load space as shown in figure 2.2 for two variable loads. This load domain
can be represented in the following linear form
L
n
m 2 =
( ) ( )
∑
=
=
n
k
k k
P t t P
1
0
μ (2.8)
where
( ) , 1,
k k k
t k μ μ μ
− +
≤ ≤ = n . (2.9)
1
μ
+
μ

1
μ
+
2
2
μ

P
1
1
P
P
2
2
P P
3
4
P
L
αL
Figure 2.2 Two dimensional load domain and L '=α L L
In the case of shell structures, it is useful to describe this load domain in the
generalized stress space. To this end, we use here the notion of a fictitious infinitely elastic
structure which has the same geometry and elastic properties as the actual one. Let the
crosssections of the shell be identified by a vectorvariable . The actual ‘engineering’
stress field in Eq. (1.75) in the elasticplastic shell can be expressed in the following way
x
( )
ˆ ˆ ( , ) ( , )
E
t t = + σ x σ x ρ x (2.10)
with the fictitious elastic generalized stress vector ( )
ˆ ,
E
t σ x is defined as that would appear
in the fictitious infinitely elastic structure if this structure was subjected to the same loads
as the actual one. This fictitious elastic generalized stress vector may be written in a form
similar to (2.8)
( ) ( ) (
1
ˆ ,
n
E
k
k
t t μ
=
=
∑
σ x σ x)
ˆ
Ek
(2.11)
26
where denotes the generalized stress vector in the infinitely elastic (fictitious)
structure when subjected to the unit load mode .
( )
ˆ
Ek
σ x
(
0
k
P ) ρ x
0 P =
denotes a timeindependent
residual generalized stress field. This residual generalized stress field satisfies the
homogeneous static equilibrium and boundary conditions (2.3.a) for .
Following (2.11), let us define a load domain '=α L L
(
ˆ ,
E
t
such that when subjected to
the fictitious elastic generalized stress vector ' L )
′
σ x of the structure under
consideration is equal to σ multiplied by a load factor ( )
ˆ ,
E
t x α
( )
′
σ x ( )
1
ˆ ˆ , ,
n
E E E
k
t t α α
=
= =
∑
σ x ( )
k
t μ σ (
ˆ x)
k
. (2.12)
In shakedown analysis, the problem is to find the largest value of α which still
guarantees elastic shakedown. This situation means that after some time t or some cycles
of loading, plastic strain ceases to develop and the structure returns to elastic behaviour.
One criterion for an elastic, perfectly plastic material to shake down elastically is that the
plastic generalized strains and therefore the residual generalized stresses become stationary
for given loads ( ) P t
( )
( )
ˆ lim , 0,
lim 0,
p
t
t
t
S
→∞
→∞
=
= ∀ ∈
ε x
ρ x x
(2.13)
where denotes the middle surface of the shell. It is shown in this case, the total amount
of plastic energy dissipated over any possible load path within the domain must be
finite
S
' L
0 0
0
ˆ ˆ
t
T p
in
W N dt ε = < ∞
∫
σ ε
. (2.14)
The inequality (2.14) may be considered as an intuitive examination which verifies
if a given structure is going to shake down. Indeed, most of the authors have formulated
shakedown criteria in this way. However, it should be noted that such a concept leads to an
approximate description: the total energy dissipated may be finite even if plastic strain
increments appear at every load cycle but comprise a convergent series. Furthermore,
boundedness of the total plastic work, without its maximum value being specified, seems
to be sometimes too weak a requirement, for example when low cycle fatigue is
considered.
2.2.3 Fundamental of shakedown theorems
Static shakedown theorem
The decomposition (2.10) is an extension of the classical formulae, which expresses
27
in terms of stresses. By comparing with the classical static shakedown criterion, which
states for stress variables, one can easily see that the shakedown of a shell structure is
equivalent to the existence of a timeindependent residual generalized stress field ( ) ρ x in
structure such that it does not anywhere violate the yield criterion
( ) ( ) ( )
ˆ ˆ ( , ) ( , ) 0
E
f t f t = + ≤ σ x σ x ρ x (2.15)
where f denotes the yield function in term of generalized variables. The following
theorem shows that this is the necessary and sufficient condition for a structure to shake
down
Theorem II.1:
1. Shakedown occurs if there exists a timeindependent residual generalized stress
field ( ) ρ x , statically admissible, such that
( ) ( ) ( )
ˆ ˆ ( , ) ( , ) 0
E
f t f t = + < σ x σ x ρ x . (2.16)
2. Shakedown will not occur if no ( ) ρ x exists such that
( ) ( ) ( )
ˆ ˆ ( , ) ( , ) 0
E
f t f t = + ≤ σ x σ x ρ x . (2.17)
Based on the above static theorem, we can find a permanent statically admissible
residual generalized stress field in order to obtain a maximum load domain αL
−
α
that
guarantees (2.17). The obtained shakedown load multiplier is generally a lower bound.
From the above static theorem, the shakedown problem can be seen as a mathematical
maximization problem in nonlinear programming
[ ]
( ) ( ) ( )
max
0 (a)
s.t.:
ˆ , 0 (b)
E
L
f t t
α α
α
−
=
= ⎧
⎪
⎨
+ ≤ ∀
⎪
⎩
ρ
σ x ρ x
(2.18)
Kinematic shakedown theorem
Using plastic strain field to formulate shakedown criterion, kinematic shakedown
theorem is the counterpart of the static one. The theorem was given by Koiter [1960] and
some of its applications in analysis of incremental collapse were derived by Gokhfeld
[1980], Sawczuk [1969a, b]. The same as proposed by Koiter for plastic strain field, we
introduce here an admissible cycle of plastic generalized strain field . The plastic
generalized strain rate may not necessarily be compatible at each instant during the
time cycle T but the plastic generalized strain accumulation over the cycle
ˆ
p
Δε
ˆ
p
ε
28
0
ˆ ˆ
T
p p
dt Δ =
∫
ε ε
(2.19)
is required to be kinematically compatible such that
[ ]
ˆ ( )
p
R Δ = ε u x (2.20)
and
( )
0 0
0
ˆ ˆ 0
T
T
E p
S
N dtd ε >
∫ ∫
σ ε
S (2.21)
where R is a linear, differential or algebraic operator, denotes the vector of
displacements on the middle surface. The virtual power principle (1.14) permits us to write
the external power in a more general form which contains the generalized variables
( ) u x
( ) ( ) ( )
0 0
0 0
ˆ ˆ ( , ) ,
T T
T
p E
i i ij ij i i
V V S
p
f u t dV t u dS dt N t dSdt
σ
θ
σ
σ ε ε
∂
⎡ ⎤
+ + =
⎢ ⎥
⎢ ⎥
⎣ ⎦
∫ ∫ ∫ ∫ ∫
x σ x ε
(2.22)
( ) t where is a selfequilibrium thermal stress due to the temperature field ( , )
ij
t
θ
σ x θ . From
Eq. (2.22), the following extension of Koiter theorem holds
Theorem II.2:
1. Shakedown may happen if the following inequality is satisfied
( ) ( )
0 0
0 0
ˆ ˆ ˆ ,
T T
T
E p p p
S S
N t dSdt D dSdt ε ≤
∫ ∫ ∫ ∫
σ x ε ε
( )
( )
+
α
. (2.23)
2. Shakedown can not happen when the following inequality holds
( ) ( )
0 0
0 0
ˆ ˆ ˆ ,
T T
T
E p p p
S S
N t dSdt D dSdt ε >
∫ ∫ ∫ ∫
σ x ε ε
. (2.24)
Based on the kinematic theorem, an upper bound of the shakedown limit load
multiplier can be computed. The shakedown problem can be seen as a mathematical
minimization problem in nonlinear programming
( ) ( )
[ ]
0
0 0
0
0
ˆ ( )
min
ˆ ˆ ,
ˆ ˆ
s.t:
ˆ ( )
T
p p
S
T
T
E p
S
T
p p
p
D dSdt
N t dSdt
dt
R
α
ε
+
=
⎧
Δ =
⎪
⎨
⎪
Δ =
⎩
∫ ∫
∫ ∫
∫
ε
σ x ε
ε ε
ε u x
(2.25)
29
In order to calculate the shakedown limit load multiplier, the two following methods can
be applied: separated and unified methods. While the former analyses separately two
different failure modes: incremental plasticity (ratchetting) and alternating plasticity, the
latter analyses them simultaneously. Both methods deserve special attention due to their
role in structural computation. In the following sections, from the original ones which state
for local variables [König, 1987], an extension for generalized variables is presented.
2.2.4 Separated shakedown limit
As was mentioned above, incremental collapse and alternating plasticity may
happen to combine. These inadaptation modes can be defined precisely in the following
way
Theorem II.3:
1. A perfect incremental collapse process (over a certain time interval ) is a
process of plastic deformation
( ) T , 0
( ) , t ε x in which a kinematically admissible plastic
generalized strain increment ( ) ( ) ( ) , , , 0 t T Δ = − ε x ε x ε x is attained in a
proportional and monotonic way, namely
[ ]
( ) ( )
( )
( )
( )
,
, 0
, 0 0
, 1
t
t
T
= Λ ⋅ Δ
Λ ≥
Λ =
Λ =
ε
ε x ε x
x
x
x
( ) R Δ = u x
(2.26)
( ) , t ε x 2. An alternating plasticity process is any process of plastic deformation
within a certain time interval ( ) T , 0
( ) ε x
such that the total increment of the plastic
generalized strain Δ
over this period is zero,
( )
T
( )
0
, t dt 0 Δ = =
∫
ε x x ε
. (2.27)
The criteria of safety with respect to alternating plasticity or incremental collapse
may be obtained by substituting the plastic strain history (2.26) or (2.27) into the
shakedown condition (2.23).
From the above definitions (2.26) and (2.27), it is easy to see that any plastic
generalized strain history ( )
ˆ ,
p
t ε x , which leads to a kinematically admissible plastic
generalized strain increment within a periodic interval ( ) T , 0 , can be decomposed into two
components of perfectly incremental collapse and alternating plasticity
( ) ( ) ( ) , + x ˆ
p
ε , , t t t = x ε x ε
. (2.28)
30
2.2.4.1 Incremental collapse criterion
If the safety condition against any form of perfectly incremental collapse is
considered, the plastic strain field is assumed by (2.26). Substituting (2.26) into (2.23), one
obtains
0 0
0 0
ˆ ( , ) ( , ) ( ) ( )
T T
E p
ex in
S S
W N t t dSdt W D dSdt ε = Λ Δ ≤ = ΛΔ
∫ ∫ ∫ ∫
σ x x ε x ε
. (2.29)
By taking into account the properties of the dissipation function and the plastic
strain history (2.26), we can write
0 0
( ) ( ) ( )
T T
p p
in
S S S
W D dSdt D dSdt D dS = ΛΔ = Λ Δ = Δ
∫ ∫ ∫ ∫ ∫
ε ε
p
ε
ex
in
) , ( t x Λ
0 ) , ( ≠ Λ t x
. (2.30)
From the shakedown condition (2.29), the smallest upper bound of incremental
limit could be attained when the external power W assumes its maximum and the internal
dissipation W takes its minimum. To this end, the function is selected in such a
way that only when the product ˆ ( , ) ( )
E
t Δ x ε x
L
ex
σ takes its maximum possible
value for a given load domain . In this case, the external power W can be written as
0 0 0 0
0
ˆ
E E
ˆ ( , ) ( , ) ( ) ( ) ( )
ex
S S
W N t t dSdt N dS ε ε = Λ Δ = Δ
∫ ∫ ∫
σ x x ε x σ x ε x
T
(2.31)
in which
( )
ˆ
ˆ ( ) ( ) max ( , ) ( )
E E
t Δ = Δ σ x ε x σ x ε x . (2.32)
By this way, the safety condition against any form of perfectly incremental collapse
thus has the form
( )
0 0
ˆ
( ) ( )
E p
S S
N dS D α ε Δ ≤ Δ
∫ ∫
σ x ε x ε
dS
n
. (2.33)
If the load variation domain is prescribed by (2.8), (2.9) and (2.11), namely
independently varying loads, the formulation (2.33) becomes
( )
0 0
1
ˆ
( ) ( )
n
Ek p
k
k
S S
N dS D α ε μ
=
Δ ≤ Δ
∑
∫
σ x ε x ε
dS
∫
(2.34)
in which
ˆ if ( ) ( ) 0
ˆ if ( ) ( ) 0
Ek
k
k
Ek
k
μ
μ
μ
+
−
⎧ Δ ≥
=
⎨
Δ <
⎩
σ x ε x
σ x ε x
(2.35)
From condition (2.34), the shakedown load multiplier against incremental collapse
can be formulated as a nonlinear minimization problem
+
α
31
⎟
⎟
⎠
⎞
⎜
⎜
⎝
⎛
=
+
ex
in
W
W
min α (2.36)
or in normalized form
1 : .
min
=
=
+
ex
in
W t s
W α
. (2.37)
2.2.4.2. Alternating plasticity criterion
If the safety condition against alternating plasticity is considered, the plastic strain
field must be satisfied (2.27). The shakedown condition (2.24) in this case has the form
(2.38) ( ) ( )
0 0
0 0
ˆ , , ( )
T T
E
S S
N t t dSdt D dSdt α ε ≤
∫ ∫ ∫ ∫
σ x ε x ε
p
∈ x
x
( )
with
( )
0
1, , 0 for all S
T
t dt α > =
∫
ε x
. (2.39)
Starting from the kinematic theorem and the last constraint in (2.39), the
optimization problem leading to the most stringent limit condition can be established at
each point separately. The global safety factor against alternating plasticity limit will be
the minimum of local x defined as α
( ) ( )
0 0
0
0
0
1
ˆ max , ,
( )
( ) 1
. :
( , ) 0
ij
T
E
T
p
T
N t t
D dt
s t
t dt
ε
ε
α
=
⎧
=
⎪
⎪
⎨
⎪
=
⎪
⎩
∫
∫
∫
σ x ε x
x
ε
ε x
dt
(2.40)
By solving this problem, the static shakedown condition against any form of
alternating plasticity can be obtained
A given structure is safe against alternating plasticity if there exists a time
independent generalized stress field ρ which, if superimposed on the envelope of elastic
generalized stresses, does not violate the yield condition
ˆ ( (
E
, ) ) 0 f t + ≤ x ρ σ . (2.41)
It should be noted that the stress field ρ in (2.41) is an arbitrary timeindependent
generalized stress field and not necessarily selfequilibrated as that in Melan’s theorem
(2.16) and (2.17). If we define a general stress response
32
*
1
ˆ ˆ ( ) ( ) ( )
n
Ek
k k
k
μ μ
=
= +
∑
σ x σ x (2.42)
where is the elastic generalized stress field in the reference structure when
subjected to the kth load and
ˆ ( )
Ek
σ x
,
2 2
k k k
k k
k
μ μ μ
μ μ
μ
+ − +
+ −
= =
−
)
. (2.43)
In view of (2.40), the plastic shakedown load multiplier (lower bound) may be
calculated as
(
*
1
min
ˆ ( ) ( ) F
α =
+
x
σ x ρ x
(2.44)
where
1 f F = − . (2.45)
The sign of
k
μ must be chosen so that the value of function is maximal. By
considering the alternating characteristic of the stress corresponding to an alternating strain
rate, the optimal timeindependent generalized stress field ρ can be defined by
F
1
ˆ
n
Ek
k
k
μ
=
= −
∑
ρ σ . (2.46)
Then the plastic shakedown limit load multiplier can be finally represented as
1
1
min
ˆ ( )
n
Ek
k
k
F
α
μ
=
=
⎛ ⎞
⎜ ⎟
⎝ ⎠
∑
x
σ x
(2.47)
where all the combinations of the signs ± must be accounted for.
From equation (2.47) it can be shown that the constant loads have no influence on
the plastic fatigue limit if these constant loads do not change the geometry and material
properties. Kinematical strain hardening has also no influence on the alternating plasticity
limit because it does not change the allowed stress variation. On the other hand, the
alternating plasticity limit is proportional to the current material strength for an isotropic
strain hardening material [Yan et al., 2003].
As counterparts of the static condition (2.40), a kinematic condition as well as an
upper bound of the plastic shakedown load multiplier can be found, see Polizzotto [1993a,
b] for further details.
2.2.5 Unified shakedown limit
In practical computation, in most cases it is impossible to apply lower and upper
theorems to find directly the shakedown limit defined by the minimum of incremental
33
plasticity limit and alternating plasticity limit. The difficulty here is the presence of the
timedependent generalized stress field ( )
ˆ ,
E
t x
L
L
L
σ in (2.16)÷(2.17) or the time integration in
formulation (2.25). These obstacles can be overcome with the help of the following two
convexcycle theorems, introduced by König and Kleiber [1978].
Theorem II.4:
“Shakedown will happen over a given load domain if and only if it happens over the
convex envelope of ”.
L
Theorem II.5:
“Shakedown will happen over any load path within a given load domain if it happens
over a cyclic load path containing all vertices of ”.
P
P P
4
P
3
1
1
P
2
P
2
P
P
4
P
1
P
3
P
1
2
P
2
a) A cycle connecting all vertices of
with the origin
L b) A cycle covering the vertices of on
the entire boundary
L
Figure 2.3 Critical cycles of load for shakedown analysis
These theorems, which hold for convex load domains and convex yield surfaces,
permit us to consider one cyclic load path instead of all loading history. They allow us to
examine only the stress and strain rate fields at every vertex of the given load domain
instead of computing an integration over the time cycle. Based on these theorems, König
and Kleiber suggested a load scheme as shown in figure 2.3.a for two independently
varying loads. This scheme was applied in a simple stepbystep shakedown analysis by
Borkowski and Kleiber [1980]. Another scheme (figure 2.3.b) was adopted later by
Morelle [1984]. Extensions and implementations of these theorems can also be found in
the works of Morelle [1989], Nguyen and Morelle [1990], Polizzotto [1991], Jospin
[1992], Yan [1997].
34
Let us restrict ourselves to the case of a convex polyhedral load domain . The
question is how to apply the above theorems to eliminate timedependent elastic
generalized stress field σ and time integrations in the lower and upper shakedown
theorems. In order to do so, let us consider a special load cycle ( passing through all
vertices of the load domain such as
L
( )
ˆ ,
E
t x
0, ) T
L
) (
k
t
1
ˆ
( , ) ( ) ( )
m
k k
k
t t t μ
=
= −
∑
P x P x (2.48)
where is the total number of vertices of , is the total number of varying loads,
n
m 2 = L n
δ is the Dirac distribution with the property that
( ) ( ) ( )
k k
f t f x t t δ
∞
−∞
= −
∫
dx
f
.
Intuitively it is understood by
and ( ) 1 if
( )
0 i
k
k
k
x dx t t
t t
t t
δ
δ
∞
−∞
⎧
∞ = =
⎪
− =
⎨
⎪
≠
⎩
∫
(2.49)
We define for eq. (2.48)
. [ ( ) ( )
k
t t x t t dx μ δ
∞
−∞
− = − −
∫
]
k
ˆ
Over this load path, the generalized strain at any instant is represented by t
ˆ( ) ( )
k k
k
t t t μ = −
∑
ε
ε
ˆ
. (2.50)
At each instant (or at each load vertex), the kinematical condition may not be
satisfied, however the accumulated generalized strain over a load cycle
1
ˆ
m
k
k=
Δ =
∑
ε
ε (2.51)
must be kinematically compatible.
Obviously, the Melan condition required in the whole load domain will be satisfied
if and only if it is satisfied at all vertices (or the above special loading cycle) of the domain
due to the convex property of load domain and yield function. This remark permits us to
replace the timedependent generalized stress field ( )
ˆ ,
E
t x σ by its values calculated only at
load vertices. We have the following static shakedown theorem
35
Theorem II.6:
The necessary and sufficient condition for shakedown to occur is that there exists a
permanent residual generalized stress field ( ) ρ x , statically admissible, such that
( )
( )
ˆ
ˆ ( , ) 0 1,
E
k
f P + ≤ ∀ = σ x ρ x k m
ˆ
k
( )
p
dS
∫
. (2.52)
The application of loading cycle (2.48) also leads to the elimination of time
integration in kinematic shakedown condition as stated in the theorem hereafter
Theorem II.7:
The necessary and sufficient condition for shakedown to occur is that there exists a plastic
accumulation mechanism ε such that
( ) ( )
[ ]
0 0
1 1
1
ˆ
ˆ ˆ ˆ ,
ˆ ˆ
ˆ ( )
m m
T
E
k k k
k k
S S
m
k
k
N P dS D
R
ε
= =
=
⎧
≤
⎪
⎪
⎪
⎨
Δ =
⎪
⎪
Δ =
⎪
⎩
∑ ∑
∫
∑
σ x ε ε
ε ε
ε u x
(2.53)
The bounds of shakedown limit load multiplier (2.18) and (2.25) now can be
reformulated in simpler forms corresponding to the static theorem (2.52) and the kinematic
theorem (2.53)
1. The lower bound
[ ]
( )
( )
( )
max (a)
0 (b)
s.t.:
ˆ
ˆ , 0 1, (c)
E
k
L
f P k m
α α
α
−
=
= ⎧
⎪
⎨
+ ≤ ∀ =
⎪
⎩
ρ
σ x ρ x
(2.54)
2. The upper bound (in normalized form)
[ ]
( ) ( )
1
1
0 0
1
ˆ min ( ) (a)
ˆ ˆ (b)
ˆ s.t.: ( ) (c)
ˆ
ˆ ˆ , 1 (d)
m
p
k
k
S
m
k
k
m
T
E
k k
k
S
D dS
R
N P dS
α
ε
+
=
=
=
=
⎧
Δ =
⎪
⎪
⎪
Δ =
⎨
⎪
⎪
=
⎪
⎩
∑
∫
∑
∑
∫
ε
ε ε
ε u x
σ x ε
(2.55)
36
Let us note that if there is only one load and this load does not vary then according
to load domain definition (2.9) one has
+ −
=
1 1
μ μ . (2.56)
In this case it is easy to see that the above upper bound and lower bound reduce to
the formulations of upper and lower bounds of limit load factor (2.3) and (2.6). This fact
means that limit analysis can be considered as a special case of shakedown analysis.
37
3 DETERMINISTIC LIMIT AND SHAKEDOWN PROGRAMMING
Based on two theorems, the lower bound and upper bound theorems, different
numerical methods for LISA which deal with both linear and nonlinear programming
were built to analyse complicated structures which analytical tools fail to deal with. In fact
it is very cumbersome to use “stepbystep” procedures in solving the problem of limit and
shakedown analysis. Direct procedures are thus necessary. With the help of the finite
element method, the problem of finding the limit and shakedown load factors can be
discretized and transformed into a problem of mathematical programming. In lower bound
shakedown analysis, we calculate the interior approximation of the load domain, i.e. we
determine the maximum load factor for shakedown. Alternatively, the exterior
approximation of the load domain will be sought in upper bound analysis, i.e. we
determine the minimum load factor for nonshakedown. Displacement finite elements
work best with the upper bound approach in which kinematic conditions are satisfied.
Users of the upper bound method sometimes state that the lower bound approach should be
used with stressbased finite elements because otherwise the bounding character is lost.
Actually, the minimum and maximum problems resulting from the lower bound and upper
bound theorems are dual. Studies showed that they converge to the same load factor [Vu et
al., 2004, 2007].
Being considered as a nonlinear programming technique, the elastic compensation
method modifies the Young's modulus of each element during an iterative linearelastic
finite element calculation in order to obtain an optimized statically admissible stress field.
An upper bound and a pseudolower bound can be obtained after each iteration. In [Taylor
et al., 1999], elastic compensation and the deviatoric map have been used for LISA of
perfectly plastic shell structures by the finite element method. Recently many largescale
optimization methods have been developed for LISA with continuum finite elements in an
European research project [Staat et al., 2003a]. The research showed that today LISA can
achieve exact values of the carrying capacity under monotone and under cyclic loading for
bounded linearly kinematic hardening material with the potential inclusion of continuum
damage and moderately large deformations.
The elastoplastic analysis of failure loads and failure modes for shell structures has
been a problem of great interest to many designers. This can be achieved by performing a
multilayer analysis which requires the knowledge of the stress distribution through the
thickness and thus requires extensive computing resources both of time and storage. The
more effective method is to deal with a single layer analysis which relates to stress
resultants. However, in this case wellformulated finite element methods for the direct
evaluation of shakedown limits are currently still rare due to the complexity of geometrical
shape and the mathematical difficulties in the stress resultant yield criterion. Franco and
Ponter [1997a, b] developed an approximate technique which is based on the reformulation
of the kinematic shakedown theorem for an axisymmetric shell with piecewise linear yield
conditions. Yan and Nguyen Dang [2000] proposed an approximate stress distribution
through the thickness of the shell and simplified the von Mises criterion for generalized
variables to get a yield condition in stress resultants. Bisbos and Papaioannou [2006],
Bisbos and Pardalos [2007] formulated a lower bound shakedown analysis of steel shells
as a specific mathematical optimization problem known as secondorder cone
programming by using the linear approximation of the exact Ilyushin yield surface (usually
referred to as Ilyushin’s generalized yield model)
In this chapter, we present a kinematic approach of limit and shakedown analysis,
which is adopted for shell structures [Tran et al., 2007a, c, e]. The approach was initially
proposed by Zhang [1995b] and further developed by Liu et al. [1997], Carvelli et al.
[1999]. Vu [2001] applied this in conjunction with Newton’s method to establish a
numerical algorithm for dual shakedown limits. Starting from a finite element
discretization, a detailed kinematic algorithm in terms of generalized variables will be
formulated and introduced. A simple technique for overcoming numerical obstacles, such
as nonsmooth and singular objective functions, is also proposed.
3.1 Finite element discretization
By subdividing the whole middle surface into elements , the integration
(2.33.a) becomes
S
e
S
1 1
ˆ ( )
e
m ne
p
k
k e
S
D d
= =
∑∑
∫
ε
S
0
(3.1)
where is the number of vertices of the load domain and is the total number of
elements. The two major numerical obstacles that appear here in dissipation function are
m L ne
• the value of in the expression of the dissipation function may become
infinite in the case of
0
K ˆ ( )
p
k
D ε
γ = (as mentioned in the first chapter).
• if the rigid or elastic perfectly plastic material model is adopted with the Ilyushin
yield criterion, the objective function (3.1) is only differentiable in the plastisfied
region of structures while powerful optimization methods require its gradient to
be available everywhere.
Being considered as a special case of shakedown analysis, limit analysis based on
kinematic formulation suffers from the same difficulty. Dealing with singular dissipation
function in limit analysis, Andersen et al. [1995, 1998a, 1998b, 2000] introduced a
39
“smooth regularization method” by replacing the original dissipation function by
its disturbed one
( )
p
k
D ε
( )
2
0
,
p
k
D η ε
0
. In this function η is a very small number and by that way
all elements are seen as plastified or on the plastified verge. The technique is actually the
wholeregion regularization method mentioned in [Yan, 1997] and will be adopted here in
our analysis.
In order to overcome the two obstacles above, we replace γ and the bending strain
resultant intensity in the expression of the dissipation function by the new ones
ˆ
P
κ
2 2
0 0
ˆ
and P
κ
γ η η + +
ˆ
P
. By that way, the dissipation function is finite and differentiable
everywhere, even in the nonplastisfied region. Furthermore, since the bending strain
resultant intensity
κ
is nonnull, thus, the dissipation function now is
( ) ( )
2 2 2
0 0 0 1 1 2 2
ˆ
ˆ ,
3
0 k
P
D YN K
κ
η ε β β γ β β γ γ = + + + + ε
p
(3.2)
with
1 2 0
, , , K β β γ calculated by Eqs. (1.78) and (1.62). It is to be noted that, the quadratic
strain resultant intensities now become
( )
( )
( )
1
2
2
3 0
3
ˆ
ˆ ˆ ( 0)
4
ˆ
ˆ ˆ 3
ˆ
ˆ ˆ 12 + (>0)
T
k k
T
k k
T
k k
P
P
P
ε
εκ
κ
η
= ≥
=
=
ε Pε
ε P ε
ε Pε
(3.3)
where is the generalized strain rate vector corresponding to load vertex , ˆ
k
ε
k
P
ˆ
11 22 12 11 22 12
0 0 0 0 0 0
1 1 1 1 1 1
ˆ 2 2
T
k k k k k k
k
ε ε ε κ κ κ
ε ε ε κ κ κ
⎡ ⎤
=
⎢ ⎥
⎣ ⎦
ε
. (3.4)
It should be noted that the incompressibility condition, although it is true for plastic
deformation of metal, introduces some numerical difficulties. Fortunately, by using plane
stress or shelltype finite elements, this condition can be naturally achieved by adopting the
Kirchhoff’s hypothesis. This can reduce the computational cost in the optimization
process.
The integrations in objective function (3.1) are calculated by Gauss
Legendre integration technique. From equations (3.1) and (3.2) one has
ˆ ( )
e
p
k
S
D d
∫
ε
S
( )
( )
2 2
0 0 1 1 2 2 0
1 1 1 1
2 2
0 0 1 1 2 2 0
1 1
ˆ
ˆ ( )
3
ˆ
3
e e
m ne m ne
p
k
k e k e
S S
m NG
i
k i
P
D dS YN K d
P
wYN K
κ
κ
ε β β γ β β γ γ
ε β β γ β β γ γ
= = = =
= =
= + + +
= + +
∑∑ ∑∑
∫ ∫
∑∑
ε
S +
+ +
(3.5)
40
where is the weighting factor of the Gauss point . The same integration technique is
applied to evaluate external power (2.33.d)
i
w i
( ) ( ) ( ) ( ) 0 0 0 0
1 1 1
0 0
1 1
ˆ ˆ
ˆ ˆ ˆ ˆ , ,
ˆ ˆ
e
m m ne
T T
E E
k k k k
k k e
S S
m NG
T E
i ik ik
k i
N P dS N P dS
w N
ε ε
ε
= = =
= =
=
=
∑ ∑∑
∫ ∫
∑∑
σ x ε σ x ε
ε σ
(3.6)
where denotes the plastic generalized strain rate vector corresponding to vertex k of
the load domain and calculated at the Gauss point i in the structure. The compatible
condition (2.33.c) is verified at each Gauss point
ˆ
ik
ε
i
1
ˆ
m
ik i
k=
=
∑
B u
ε (3.7)
with denotes the deformation matrix
i
B ( ) x B at Gauss point i .
From (3.5÷3.7) the shakedown limit load multiplier may be formulated as
( )
2 2
0 0 1 1 2 2 0
1 1
1
0 0
1 1
ˆ
min
3
ˆ 1,
s.t:
ˆ ˆ 1
m NG
i
k i
m
ik i
k
m NG
T E
i ik ik
k i
P
wYN K
i NG
w N
κ
α ε β β γ β β
ε
+
= =
=
= =
= + +
⎧
= ∀ =
⎪
⎪
⎨
⎪
=
⎪
⎩
∑∑
∑
∑∑
B u
ε
ε σ
γ γ + +
ng ne NG × =
(3.8)
In the above formulation, for simplicity, the sums over all elements and over all
Gauss points of one element (in the expressions of internal and external powers) are
unified in the sum over all Gauss points of discretized structure. The total number of Gauss
points in structure is , where ng is the number of Gauss point on an element.
3.2 Kinematic algorithm
By restricting ourselves to polyhedral form of load domain, in this section we
introduce a kinematic algorithm for limit and shakedown analysis of shell structures. To
begin with, for the sake of simplicity, let us rewrite the upper bound limit (3.8) in a simpler
form by introducing some new notations
• The new strain rate vector
ik
e
ˆ
ik i ik
w = e
ε . (3.9)
• The new fictitious elastic generalized stress field
ik
t
0 0
ˆ
E
ik ik
N ε = t
σ . (3.10)
41
• The new deformation matrix
i
B
ˆ
ˆ
i i
w = B B
i
. (3.11)
Following these definitions, the objective function in (3.8) becomes
( )
2 2
0 0 1 1 2 2 0
1 1
ˆ
min
3
m NG
k i
P
YN K
κ
α ε β β γ β β γ
+
= =
= + +
∑∑
γ + +
ˆ
ik
ik
1 2 0
, , , K
(3.12)
where are calculated by Eqs. (1.72), (1.62), (1.66) and (3.3) in which the
generalized strain vector ε is replaced by the rate e . It is to be noted that this replacing
does not affect the value of
1 2 0
ˆ
, , , , K P
κ
β β γ
β γ
ˆ
P , but
κ
. β
By substituting (3.9)÷(3.12) into (3.8) one obtains a simplified version for upper
bound of shakedown limit load
( )
2 2
0 0 1 1 2 2 0
1 1
1
1 1
ˆ
min
3
ˆ
1,
s.t:
1
m NG
k i
m
ik i
k
m NG
T
ik ik
k i
P
YN K
i NG
κ
α ε β β γ β β γ
+
= =
=
= =
= + +
⎧
= ∀ =
⎪
⎪
⎨
⎪
=
⎪
⎩
∑∑
∑
∑∑
e B u
e t
γ + +
(3.13)
Dealing with the nonlinear constrained optimization problem (3.13), the penalty
method is used for satisfying the compatibility condition. To this purpose, let us write the
penalty function as
( )
2 2
0 0 1 1 2 2 0
1 1 1 1
ˆ
ˆ ˆ
3 2
T
NG m m m
P ik i ik i
i k k k
P c
F YN K
κ
ε β β γ β β γ γ
= = = =
⎧ ⎫
⎛ ⎞
⎛ ⎞ ⎛ ⎞ ⎪ ⎪
⎜ ⎟ = + + + + + −
⎨ ⎬
⎜ ⎟ ⎜ ⎟
⎜ ⎟
⎝ ⎠ ⎝ ⎠
⎪ ⎪
⎝ ⎠ ⎩ ⎭
∑ ∑ ∑ ∑
e B u e B u
−
(3.14)
where is penalty parameter such that . This parameter c may be dependent on
integration points or load vertices and c should be adjusted to fit different compatibility
criteria. However, at this stage, for the sake of simplicity, c is let to be constant
everywhere. Theoretically, when c goes to infinity we will recover related conditions.
c 1 >> c
Following (3.14) the modified kinematic formulation (3.13) becomes
1 1
min (a)
s.t:
1 (b)
P
NG m
T
ik ik
i k
F α
+
= =
=
=
∑∑
e t
(3.15)
The corresponding Lagrange function of (3.15) is
42
1 1
1
NG m
T
PL P ik ik
i k
F F λ
= =
⎛ ⎞
= − −
⎜ ⎟
⎝ ⎠
∑∑
e t . (3.16)
Obviously, the normalization can be also treated by a penalty method, which may
help to remove the Langrangian multiplier λ from our problem. However, the application
of a penalty method to (3.15.b) leads to a system of equations which requires more effort
and computer memory to solve. For this reason that λ is retained as variable and will be
handled later on. The stationarity condition for the Lagrange function states that
PL
F
( )
( )
( )
( ) ( )
2 2
1 1 2 2 0 3
2 2
0 0 1 2 2 3
2
0 1 2 3
1
4 3
ˆ
4 3
4
ˆ
3
8 16
ˆ
2
ˆ
+ i, k (
PL
ik
ik
m
ik i ik
k
K
P
F
YN
P
K
P
c
κ
κ
κ
β β γ β β γ γ
ε β γ β γ β
β γ β
λ
=
⎛ ⎞
⎜ ⎟ + + + +
⎜ ⎟
⎜ ⎟
∂ ⎜ ⎟
= + + − + +
⎜ ⎟
∂
⎜ ⎟
⎜ ⎟
⎜ ⎟
+ + − −
⎜ ⎟
⎝ ⎠
⎛ ⎞
− − = ∀
⎜ ⎟
⎝ ⎠
∑
P
P P e
e
P P P
e B u t 0
1 1
1 1
a)
ˆ ˆ
(b)
1 0 (c)
NG m
T PL
i ik i
i k
NG m
T PL
ik ik
i k
F
c
F
λ
= =
= =
⎧
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
∂ ⎛ ⎞
= − − =
⎜ ⎟
∂
⎝ ⎠
∂ ⎛ ⎞
= − =
⎜ ⎟
∂
⎝ ⎠ ⎩
∑ ∑
∑∑
B e B u 0
u
e t
⎪
⎪
⎪
⎪
⎪
(3.17)
The first equation of (3.17) can be rewritten in the following form
( )
( )
( )
( ) ( )
2 2
1 1 2 2 0 3
2 2
0 0 1 2 2 3
1
2
0 1 2 3
4 3
ˆ ˆ ˆ
4 3 4 +
3
8 16
2
m
ik ik i ik
k
K
YN c P P
K
κ κ
β β γ β β γ γ
ε β γ β γ β λ
β γ β
=
⎛ ⎞
+ + + +
⎜ ⎟
⎜ ⎟
⎛ ⎞
⎜ ⎟
+ + − + + − −
⎜ ⎟
⎜ ⎟
⎝ ⎠
⎜ ⎟
⎜ ⎟
+ + − −
⎜ ⎟
⎝ ⎠
∑
P
P P e e B u t
P P P
= 0 .
(3.18)
This form of stationary condition will help us reduce numerical difficulty in
singular situation when
ˆ
0 P
κ
→
0
. This means that (3.18) may be satisfied even when there
is no plastic deformation and η is set equal to zero while in the same conditions the
equation (3.17.a) is indefinite. For convenience, let us rewrite Eqs. (3.18), (3.17.b) and
(3.17.c) as follows
43
1
1 1 1
1 1
ˆ ˆ ˆ
(a)
ˆ ˆ
(b)
1 0
m
ik ik ik i ik ik
k
NG m NG
T
i ik i i
i k i
NG m
T
ik ik
i k
c P P
κ κ
λ
=
= = =
= =
⎛ ⎞
+ − − = =
⎜ ⎟
⎝ ⎠
⎛ ⎞
− = =
⎜ ⎟
⎝ ⎠
− =
∑
∑ ∑ ∑
∑∑
H e e B u t 0 f
B e B u 0 h
e t
(c)
⎧
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎩
(3.19)
where
( )
( )
( )
( ) ( )
2 2
1 1 2 2 0 3
2 2
0 0 1 2 2 3
2
0 1 2 3
4 3
4 3 4
3
8 16
2
ik
K
YN
K
β β γ β β γ γ
ε β γ β γ β
β γ β
⎛ ⎞
+ + + +
⎜ ⎟
⎜ ⎟
⎜ ⎟
= + + − + +
⎜ ⎟
⎜ ⎟
⎜ ⎟
+ + − −
⎜ ⎟
⎝ ⎠
P
H P
P P P
P (3.20)
By applying NewtonRaphson’s method to solve the modified system (3.19) one
gets
1
1 1 1
1 1 1 1
ˆ ˆ ˆ
(a)
ˆ ˆ
(b)
1
m
ik ik ik i ik ik
k
NG m NG
T
i ik i i
i k i
NG m NG m
T T
ik ik ik ik
i k i k
d c P d d d P
d d
d
κ κ
λ
=
= = =
= = = =
⎛ ⎞
+ − − = −
⎜ ⎟
⎝ ⎠
⎛ ⎞
− = −
⎜ ⎟
⎝ ⎠
= − +
∑
∑ ∑ ∑
∑∑ ∑∑
M e e B u t f
B e B u h
e t e t
(c)
⎧
⎪
⎪
⎪
⎪
⎨
⎪
⎪
⎪
⎪
⎩
(3.21)
where
3
1
12
ˆ
ˆ
m
T
ik ik ik ik i ik ik
k
c
P
κ
λ
=
⎛ ⎛ ⎞
= + + − −
⎜ ⎜ ⎟
⎝ ⎠ ⎝ ⎠
∑
M H D e B u t e
⎞
⎟
P (3.22)
with
44
( )
( ) ( )
2 2 1 2
1 2 3
2 2
1 2
0 2
2 2
1 2
2 2
0 0
1 2
2 2 2 2
1 1 1 2 2 2
48 3 24 3
ˆ ˆ
3 3 1 1 24 3
ˆ ˆ
3 3 1 1
ˆ
3 3 1 1
ˆ
8
T
ik ik
ik
P P
K
P P
YN P
P
κ κ
κ κ
κ
κ
β β
β γ β γ
β γ β γ
β γ β γ
ε
β γ β γ
β γ β γ β β γ β γ β
⎛ ⎞ ⎛ ⎞
⎜ ⎟ ⎜ ⎟ − + − + + +
⎜ ⎟
⎜ ⎟
+ +
⎝ ⎠
⎜ ⎟
⎜ ⎟ ⎛ ⎞
⎜ ⎟ ⎜ ⎟ + − −
⎜ ⎟ ⎜ ⎟
+ +
⎝ ⎠ ⎝ ⎠
⎛ ⎞
⎜ ⎟ −
⎜ ⎟
=
+ +
⎝ ⎠
+
⎛ ⎞
⎜ ⎟
+ −
⎜
+ + + + + − ⎜
⎝ ⎠
P A
e e A
B P
A
D
( )
( )
0 3
2 3
2
1 2 3
2
1 2 3
6 3
ˆ
4
8 16
24 16 5
T
ik ik
T
ik ik
K
P
κ
β
β γ β
β γ β
⎛ ⎞
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎛ ⎞
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎜ ⎟ ⎟
⎟ ⎜ ⎟
⎜ ⎟
⎝ ⎠
⎜ ⎟
⎜ ⎟
−
⎜ ⎟
⎜ ⎟
⎜ ⎟
⎝ ⎠
= +
= + − −
= + − −
e e B
B
P e e C
A P P
B P P P
C P P P
(3.23)
In the system (3.21) denote the incremental vectors of generalized strain
rate and displacement respectively while
,
ik
d d e u
dλ denotes the incremental value of λ . By
moving the last two terms in (3.21.a) to lefthand side and then, multiplying both sides
with we get
1
ik
−
M
1 1
1
ˆ ˆ ˆ
1
ik ik ik ik ik i ik ik
k
d d d P
κ κ
λ
− − −
=
⎛ ⎞
= − − +
⎜ ⎟
⎝ ⎠
e M e B u M t
m
P
∑
M d c − f . (3.24)
Writing (3.24) for m k , 1 = and summing them up, after some manipulations one
has
1 1 1 1 1
ˆ ˆ ˆ
m m m m
ik i ik ik i ik i i ik ik
k k k k
d c P d d
κ κ
λ
− − − − −
= − + +
∑ ∑ ∑ ∑
e E M f E M B u E M t
1
P
−
(3.25)
where
1
ˆ
m
i i ik
k
c P
κ
−
= +
∑
E I M (3.26)
and is the identity matrix.
i
I
The substitution of (3.25) in (3.21.b) leads to
45
1 1 1 1
1 1 1 1
ˆ ˆ ˆ
NG NG NG m NG m
T T T
i i i i i ik ik i i ik i
i i i k i k
d d
κ
λ
− − − −
= = = =
= − +
∑ ∑ ∑ ∑ ∑ ∑
B KB u h B E M f B E M t
ˆ
k
P (3.27)
where
1
ˆ
m
i
k
c P
κ
1
ik
− −
= −
∑
K I E M . (3.28)
Starting from the definition of , Eq. (3.26), we can further simplify (3.28) as
i
E
( )
1 1 1
ˆ
m
i ik i i
k
c P
κ
− − −
= − = − − =
∑
K I E M I E E I E
1
i
−
. (3.29)
By introducing (3.29) in (3.27) and take into account Eqs. (3.21.a), (3.21.b), one
has the following system
( )
1 2
d d λ λ = − + + + S u Su f f (3.30)
where
( )
1
1
1 1
1
1 1
1 1
2
1
ˆ ˆ
(a)
ˆ
(b)
ˆ ˆ
(c)
NG
T
i i i
i
NG m
T
i i ik ik ik ik
i k
NG m
T
i i ik ik
i k
P
κ
−
=
− −
= =
− −
=
⎧
=
⎪
⎪
⎪
= −
⎨
⎪
⎪
=
⎪
⎩
∑
∑ ∑
∑ ∑
S B E B
f B E M M H e
f B E M t
(3.31)
The system (3.30) with the last two terms on the righthand side may be interpreted
as the linear system arising in purely elastic computations with the global stiffness matrix
and applied force vector ( )
1 2
d S λ λ + + ⎡ ⎤
⎣ ⎦
f f
1
i
. The matrix
−
E
i
plays the role of the elastic
matrix at the Gauss point while the vector
( ) d
1 2
λ = + f f f (3.32) λ +
can be considered as the vector of nodal forces applied on the structure. A difficulty arises
here in the system (3.30) due to a nonsymmetry of the matrix S. In order to reduce the
computational costs, one can make the matrix
1
i
−
E symmetric by
(
( )
)
1 1 1
1
2
T
i i i
− − −
E E E = + . (3.33)
From (3.30), (3.25), and (3.24) and after some manipulations we get the
incremental vectors of displacement and strain rate as
( )
( ) ( ) ( )
1 2
1 2
(a)
(b)
ik ik ik
d d d d
d d d d
λ λ
λ λ
= + + ⎧
⎪
⎨
= + +
⎪
⎩
u u u
e e e
(3.34)
where
46
( )
( )
1
1 1
1
2 2
1 1 1 1
1
1
1
1 1 1 1
2
2
ˆ ˆ ˆ
ˆ ˆ ˆ ˆ
m m
ik ik i ik ik ik i ik i ik ik ik
k k
m
ik ik i i ik ik ik ik
k
d
d
d c P d
d c P d P P
κ
κ κ κ
−
−
− − − −
=
− − − −
⎧ = − +
⎪
=
⎪
⎪
⎪ ⎛ ⎞ ⎛ ⎞
= + − − ⎨
⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠ ⎪
⎪
⎛ ⎞
⎪
= − − + +
⎜ ⎟
⎪
⎝ ⎠ ⎩
∑ ∑
∑
u u S f
u S f
e M E M H e B u e B u M H e
e M E B u M t M t
− (3.35)
The vectors are actually Newton’s directions which assure that a suitable
step along them will lead to a decrease of the objective function.
,
ik
d d u e
In order to compute ( ) dλ λ + we substitute (3.34.b) into (3.21.c) and have
( )
( ) ( )
( )
1
1 1
2
1 1
1
NG m
T
ik ik ik
i k
NG m
T
ik ik
i k
d
d
d
λ λ
= =
= =
⎡ ⎤
− +
⎢ ⎥
+ = ⎢ ⎥
⎢ ⎥
⎢ ⎥
⎣ ⎦
∑∑
∑∑
t e e
t e
. (3.36)
Based on (3.34) we can update the displacement vector and the global
generalized strain rate vector e
u
11 ,
... ...
T
ik NG m
= ⎡ ⎤
⎣ ⎦
e e e e . (3.37)
The new vectors of and e tend to satisfy (3.17.b) and (3.18) simultaneously. By
forcing them to fulfil (3.17.c) we get Lagrange multiplier
u
λ updated as ( ) dλ λ +
u e
in (3.36).
Iterating these steps may drive us to a stable set of , , λ satisfying all conditions
(3.17.b), and (3.17.c) and (3.18). Details of the iterative algorithm are presented hereafter.
• Step 1: Initialize displacement and strain rate vectors: and such that the
normalized condition (3.17.c) is satisfied
0 0
e u
0
1 1
1
NG m
T
ik ik
i k = =
=
∑∑
t e . (3.38)
Normally the fictitious solution must be computed first in order to define the load
domain in terms of the fictitious elastic generalized stress σ . Hence and may
assume fictitious values (after being normalized) for their initialization. Set up initial
values for the penalty parameter and for
L ˆ
E
ik
0
u
0
e
c
0
η . Set up convergence criteria, maximum
number of iterations.
• Step 2: Calculate , and from (3.31) at current values of and S
1
f
2
f u e
• Step 3: Calculate ( ) dλ λ + , from Eqs. (3.34) and (3.36) ,
ik
d d u e
47
• Step 4: Perform a line search to find
s
Λ such that
( ) ,
P s s
F d d + Λ + Λ → u u e e min . (3.39)
Update displacement, strain rate and λ as
( )
( )
( )
s
ik ik s ik
d a
d
d c λ λ λ
b
= + Λ
= + Λ
= +
u u u
e e e
(3.40)
As was shown above, it is costly to get the analytical form of the derivatives of the
objective function. Thus, a line search without using the derivatives is recommended here,
e.g. the Golden section method.
• Step 5: Check convergence criteria: if they are all satisfied go to step 6 otherwise repeat
steps 2, 3 and 4.
• Step 6: Stop
The algorithm can fail due to some reasons: failure in computing the inverse matrix
or unsuccessful initialization step, which after some iterations results in an unexpected
form
1 −
S
0
1 1
0
NG m
T
ik ik
i k = =
=
∑∑
t e
or the maximum number of iterations is insufficient to get a convergent solution. If those
obstacles do not exist, a solution set
( )
, , λ u e
can be found as proved in [Vu, 2001].
48
4 PROBABILISTIC LIMIT AND SHAKEDOWN PROGRAMMING
The traditional approach to safety assessment and design is based on a deterministic
model which invariably involves a large safety factor usually assigned from heuristic and
somewhat arbitrary decisions. This approach has almost certainly been reinforced by the
very large extent to which structural engineering design is codified and the lack of
feedback about the actual performance of the structure. Use of large safety factors can lead
to the view that “absolute” safety can be achieved. Absolute safety is, of course,
undesirable if not unobtainable, since it could only be approached by deploying infinite
resources. Moreover, these safety factors do not provide any information regarding the
probability that would lead to the loss of structural integrity.
The realistic evaluation of structural performance can be conducted only if the
uncertainty of the actual loadcarrying capacity of the structure is taken into consideration.
Uncertainty may originate from random fluctuations of significant physical properties,
from limited information and from model idealizations of unknown credibility. Structural
reliability analysis deals with all these uncertainties in a rational way. Reliability
assessment of structures requires on the one hand mechanical models and analysis
procedures that are capable of modeling limit states accurately. On the other hand, full
coverage of the present random variables is also necessary for a meaningful reliability
assessment.
Present structural reliability analysis is typically based on the limit state of initial or
local failure. This may be defined by first yield or by some member failure if the structure
can be designed on an element basis. However, this gives quite pessimistic reliability
estimates, because virtually all structures are redundant or statically undetermined.
Progressive member failures of such systems reduce redundancy until finally the statically
determined system fails. This system approach is not defined in an obvious way for a finite
element representation of a structure. The more effective method of structural reliability
analysis is probabilistic limit and shakedown analyses, which is based on the direct
computation of the loadcarrying capacity or the safety margin.
The aim of this chapter is to present a new algorithm of probabilistic limit and
shakedown analysis for thin plates and shells, which is based on the kinematic approach
[Tran et al., 2007b, d, 2008]. The loading, material as well as the thickness of the shell are
to be considered as random variables. Many different kinds of distribution of basic
variables are taken into consideration and performed with First and Second Order
Reliability Methods (FORM/SORM) for calculation of the failure probability of the
structure. In order to get the design point, a nonlinear optimization was implemented,
which is based on the Sequential Quadratic Programming (SQP). Nonlinear sensitivity
analyses are also performed for computing the Jacobian and the Hessian of the limit state
function.
4.1 Basic concepts of probability theory
4.1.1 Sample space
A standard way of determining the yield stress of a material such as steel is to
perform a number of simple tensile tests with specimens made from the material in
question. By each test a value for the yield stress is determined but this value will probably
be different from test to test. Therefore, in this connection, the yield stress must be taken as
an uncertain quantity and it is in accordance with this point of view said to be random
quantity. The set of all possible outcomes of such tests is called the sample space. The
sample space for the yield stress is the set of all positive real numbers and is continuous.
A subset of sample space is called an event. An event is the set of sample points. If
it contains no sample points, it is called an impossible event.
4.1.2 Random variables
An event can be identified through the value of a function called a random
variable. A random variable is a function which maps events ω in the sample space Ω
into the real line R . Usually a random variable is denoted by a capital letter X
: X R Ω→ . (4.1)
( ) X x A random variable can be continuous or discrete. The values of ω = are
called realizations of X and denoted by x . The space of event Ω is characterized by a
probability measure , which satisfies the properties of a normed, nonnegative and P σ 
additive measure
1) 0 ( P A) 1, ≤ ≤
2) ( ) 1, ( ) 0, P P Ω = ∅ =
3) . ( ) ( ) ( ) if P A B P A P B A B ∪ = + ∩ = ∅
A family { }
i
i
X of random variables with :
i
X R Ω→ is called stochastically
independent or independent for short, if for every choice of a subset the event
i
A R ⊂
{ } X
i i
i
A ∈ are independent of each other.
The cumulative probability distribution function (CDF) of the random X is defined
as follows
( ) ( ) F x P X x = < . (4.2)
50
This means, that the value of at F x is the probability of the event, that the random
variable X has a realization lower than x . A random variable is characterized by its CDF.
We summarize the most important characteristics of distribution functions
1) is nondecreasing and continuous ( ) F x
2) . lim ( ) 1, lim ( ) 0, ( ) ( )  ( )
x x
F x F x P a X b F b F a
→∞ →−∞
= = < < =
For the continuous random variable, it is often useful to use the derivative
probability function. This function is called the probability density function (PDF)
f : R R →
)
. We have
. (4.3) ( ) ( )
x
F x f t dt
−∞
=
∫
The probability density function has the properties
1) , ( ) (
b
a
P a X b f t dt ≤ ≤ =
∫
2) , ( ) 0 f x ≥
3) ( ) 1 f t dt
∞
−∞
=
∫
.
The appendix summarizes the most important distribution functions, densities and further
details. The following distributions play a special role in structural reliability analysis.
Normal distribution and standard normal distribution ( 1, 0) σ μ = = with the respective
densities
2 2
( ) / 2 0.5
2
1 1
( ) and ( )
2
2
2
x x
f x e f x e
μ σ
π
πσ
− − −
= = . (4.4)
4.1.3 Moments
Let X be a continuous random variable. Then its probabilistic characteristics are
described by the cumulative distribution function . However, in many applications, the
form of is not known in all details. It is therefore useful to have an approximate
discription of a random variable by its most important features. In the following it is
assumed that all the random variables are continuous.
F
F
The expected value [ ] E X of X with the density f is defined as
. (4.5) [ ] ( ) E X xf x dx
∞
−∞
=
∫
51
The expected value is also called the ensemble average, the mean or the first moment of X
and the symbol μ is often used for it. The following simple rules for the expectation hold
1) [ ] , E a a a R = ∈ ,
2) [ ] [ ] [ ] [ ] [ ] , , E X Y E X E Y E X E X R λ λ λ + = + = ∈ ,
3) [ ] [ ] [ ] ( . ) . if and only if X and Y are independent E X Y E X E Y = .
The central moment of
th
n X is defined by ( )
n
E X μ ⎡ ⎤ −
⎣ ⎦
. The second central
moment of X is called the variance of [ ] X and is denoted by
2
σ or Var X
[ ] ( )
2
2
Var X X
2 2
E E X σ μ μ ⎡ ⎤
⎡ ⎤
= = −
⎣ ⎦
= −
⎣ ⎦
. (4.6)
The variance reflects the expected deviation of a realization x from the expected value
[ ] E X . The positive square root of the variance, [ ] X σ , is called standard deviation of X .
The concept of random variable has been used only in onedimensional sense. This
definition can easily be extended to a vector valued random variable :
n
Ω→ X
n
t
R called a
random vector. The joint cumulative distribution function for the random vector X is
defined as
1
1 1
1
( ) ( ) ... ( ,..., ) ...
n
x x
n
i i n
i
F P X x f t t dt d
=
−∞ −∞
⎛ ⎞
= ≤ =
⎜ ⎟
⎝ ⎠
∫ ∫
X ∩ . (4.7)
The mixed central moment of random variables is defined as
( ) ( )
,
ij i j i i j j i j i j
Cov X X E X X E X X μ μ μ
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ Σ = = − − = −
⎣ ⎦ ⎣ ⎦ ⎣ ⎦
μ
i
(4.8)
is called covariance of random variables (written with random vector components X and
j
X ). The covariance matrix Σ is symmetric semidefinite. The ratio
ij
,
i j
ij
i j
Cov X X
ρ
σ σ
⎡ ⎤
⎣ ⎦
= (4.9)
where ,
i j
σ σ are the standard deviations of the random variables
i
X and
j
X , is called the
correlation coefficient. It can be used as a measure of mutual linear dependence between a
pair of random variables. It can be shown that 1 1
ij
ρ − ≤ ≤
i
. Two random variables X and
, 0 (or 0)
i j ij
X X ρ ⎡ ⎤
j
X are said to be uncorrelated if Cov = =
⎣ ⎦
holds. Therefore,
independent random variables are uncorrelated but not vice versa. In the appendix the
expectations and variances of the most important distribution functions for structural
reliability analysis are listed.
52
4.2 Reliability analysis
The behavior of a structure is influenced by various typically uncertain parameters
(loading type, loading magnitude, dimensions, or material data, ...). Data with random
fluctuations in time and space is adequately described by stochastic processes and fields.
Typical examples of engineering interest are earthquake ground motion, sea waves, wind
turbulence, imperfections. The probabilistic characteristics of the processes are known
from various available measurements and investigation in the past. In engineering
mechanics, the available probabilistic characteristics of random quantities affecting the
loading of the mechanical system often cannot be utilized directly to account for the
randomness of the structural response due to its complexity. In structural response
calculations a distinction is made between the involved structural model properties which
are either considered as being deterministic or stochastic.
The principles of reliability analysis have been applied to a very large class of
problems, ranging from the design of control systems for complex nuclear and chemical
plants to the design of specific mechanical and structural components, as well as more
generally in the field of aerospace industries. Reliability analysis should not, however, be
thought of as an isolated discipline as it is closely related to the theory of statistics and
probability and to such fields as operations research, systems engineering, quality control
engineering and statistical acceptance testing.
4.2.1 Failure function and probability
In reliability based structural analysis all probabilistic characteristics are modelled
as random variables. These socalled physical basic random variables are collected in a
vector ( )
1 1
, ,...,
n
X X X = X where n is the number of stochastic variables. The
deterministic safety margin M is based on the comparison of a structural resistance
(threshold) R and loading (which is usually an invariant measure of local stress at a hot
spot or in a representative crosssection). With
S
, R S X
g X
functions of the structure fails for
any realization with nonpositive failure function or limit state function ( ) , i.e.
0 for failure
( ) ( ) ( ) 0 for limit state
0 for safe structure
g R S
< ⎧
⎪
= − =
⎨
⎪
>
⎩
X X X (4.10)
The limit state defines the limit state hypersurface ( ) 0 g = X F ∂ which separates the
failure region
{ }
( ) F g = x X 0 <
,
from safe region. Figure 4.1 shows the densities of two
random variables R S which are generally unknown or difficult to establish. The failure
probability
f
P ( ) g X
X
is the probability that is nonpositive, i.e.
( ) 0 ( )
F
P P f d = ≤ =
∫
X x ( )
f
g x (4.11)
53
where is the ndimensional joint probability density function. Usually, it is not
possible to calculate
( )
X
f x
f
P analytically. In special cases, where the basic variables are jointly
Gaussian distributed and the failure surface is a hyperplane, it can be shown that
( )
f c
P β = Φ − (4.12)
where
c
β is the second moment reliability index defined as the mean of M divided by the
standard deviation of M. Φ is the standard Gaussian distribution function. ( ) .
F
The concept of a reliability index which is invariant with respect to the formulation
of the failure function was proposed by Hasofer and Lind [1974]. It is defined as the
shortest distance from origin to the failure surface ∂ in the standardised and uncorrelated
space (figure 4.2), i.e. u
( ) 0
min
T
HL
g
β
=
=
u
u u . (4.13)
For the general cases, there are several approximate methods to compute the failure
probability
f
P
4
. Direct Monte Carlo Simulation (MCS) becomes increasingly expensive
with the increase of the structural reliability but it is independent on the dimension n of
the space of basis variables. Acceptable failure probabilities might be in the range of 10
−
to 10 . They are even much lower in nuclear reactor technology. For a validation that the
failure probability
6 −
f
P
c
P
/10
c
P
5
is less than an accepted limit , the sample size required for direct
MCS must be at least leading to a minimum sample size in the range of 10
−
to
. Such a large number exceeds particularly for complex FEmodels, available
resources by far. The numerical effort can be reduced considerably by variance reduction
methods like Importance Sampling and by Response Surface Methods (RSM). However,
the most effective analysis is based on First and SecondOrder Reliability Methods
(FORM/SORM) if gradient information is available [Gollwitzer et al., 1988]. However the
numerical effort of FORM/SORM increases with the dimension n .
7
10
−
r,s
f (r)
R
S
f (s)
Loading S
Resistance R
Figure 4.1 Density functions of , R S presenting on one axis
54
4.2.2 First and SecondOrder Reliability Method
First and SecondOrder Reliability Methods (FORM/SORM) are efficient means
and standard methods of numerical probability calculations for multiple random variables.
These methods have only become popular since the advent of inexpensive computing.
They are often confused with other socalled first order methods that may rely on
assumptions of normality and/or small numbers of random variables to achieve closed
form solutions. FORM/SORM, on the other hand, can deal with random variables having
arbitrary probability distributions and can estimate reliability results for large numbers of
random variables.
u space
failure region
g(u) < 0
safe region
g(u) > 0
β
HL
g (u) = 0
L
Q
g (u) = 0
constant probability
g(u) = 0
Figure 4.2 Domain based on linear and quadratic approximations in u space
FORM/SORM are analytical probability integration methods. Therefore, the
defined problem has to fulfill the necessary analytical requirements (e.g. FORM/SORM
apply to problems, where the set of basic variables are continuous). Because of the large
computational effort of MCS due to small failure probabilities, any effective analysis is
based on FORM/SORM [Hohenbichler et al., 1987], in which the numerical effort depends
on the number of stochastic variables but not on
f
P
u
U ( ) 0, 1 μ σ = =
(contrary to MCS). Practical
experience with FORM/SORM algorithm indicates that their estimates usually provide
satisfactory reliability measure. Especially in the case of small failure probability (large
reliability), FORM/SORM are extremely efficient compared with the MCS method
regarding the requirement of computer time, such as the Central Processing Unit (CPU).
The failure probability is computed in three steps
• Transformation: The physical space of uncertain parameters, , is transformed into
a new ndimensional space, , consisting of independent standard Gaussian variables
. Such a transformation is always possible for continuous random
variables. The appendix provides a description of some independent random variables,
x X
55
equations to map them into the standard Gaussian space. By this transformation, the
original limit state function ( ) 0 g = x
( ) 0 g = u u
is mapped into the new limit state function
in the space. This transformation is exact and not an approximation
[Bjerager, 1991].
• Computation of design point: The design point or point
HL
β , which is the point on the
limit state surface having the shortest distance to the origin in the space, is
determined by an appropriate nonlinear optimization algorithm. More details of this
algorithm will be discussed in next section.
u
( ) 0 = u
( ) 0
L
g = u
g
• Approximating the limit state surface and calculation of the failure probability: The
limit state surface is approximated by a tangential hypersurface at the design
point. This corresponds to an approximating hyperplane (linear or first
order) or hyperparaboloid
g
( ) 0
Q
= u (quadratic or secondorder), respectively. The
failure probability
( ( ) )
0
L
P g < u in FORM or
f
P is thus approximated by
( ) ( )
0 <
g
HL
Q
P g u
( ) 0 = u
in SORM.
4.2.2.1 FirstOrder Reliability Method (FORM)
Consider a tangential linearization at the design point on the limit state surface
, which is given by
0
u
0
( ) )
T T
g (
L
= − u a u β = + u a u (4.14)
where a is the vector of direction cosines and can be obtained from
0
0
( )
( )
g
g
∇
=
∇
u
a
u
. (4.15)
Here is the Hesse normal form of the tangential hypersurface to the limit state
function at the design point. If the failure function is not strictly nonlinear, the probability
of failure
( )
L
g = u 0
f
P can be determined with good accuracy by
2
0.5
,
1
( )
2
HL
z
f I HL
P
β
β
π
−
−
−∞
= Φ − =
∫
e dz (4.16)
where
HL
β is calculated from (4.13). The firstorder reliability index can be defined as
following
1
,
( )
I HL f I
P β β
−
= = −Φ . (4.17)
Consider a small vector which can be interpreted as an error appearing in the
calculation of the design point (i.e. is replaced by
ε
0
u
0
+ u ε
( )
HL
). The resulting disturbance of
the reliability index ε can be calculated as follows [Rackwitz, 2004] β
( ) ( ( )
T
)
1
, 0 L f
P
H I HL
β β
−
Φ + + ε u ε = − ≈ ε a , (4.18)
56
and therefore
( )
( )
( )
0
,
0
,
.
i
i
HL
i
i
f I
HL i
i
a
P
a
ε
ε
β
ε
ϕ β
ε
→
→
∂
=
∂
∂
= − −
∂
ε
ε
(4.19)
Obviously, represents the sensitivities of
i
a
HL
β and
, f I
P versus the changes of the mean
value of u .
0,
1 −
j
κ
3
S
i
4.2.2.2 SecondOrder Reliability Method (SORM)
Various methods have been suggested to improve the accuracy of FORM
calculations and to give a rough estimate of the approximation quality [Madsen et al.,
1986], [Fiessler et al., 1979]. The general idea is to approximate the limit state surface by a
quadratic hypersurface rather than by a hyperplane. The main curvatures of the quadratic
hypersurface at the design point are equal to those of the limit state surface. This leads to
the socalled SecondOrder Reliability Methods (SORM). Using all n main curvatures
, the failure probability is calculated as a three term approximation suggested by Tvedt
[1983] as
, 1 2 f II
P S S = + + (4.20a)
( )
1
1/ 2
1
1
( ) 1
n
HL HL j
j
S β β κ
−
−
=
= Φ − −
∏
(4.20b)
[ ] ( ) ( )
1 1
1/ 2 1/ 2
2
1 1
( ) ( ) 1 1 ( 1)
n n
HL HL HL HL j HL j
j j
S β β φ β β κ β κ
− −
− −
= =
⎧ ⎫
= Φ − − − − − +
⎨ ⎬
⎩ ⎭
∏ ∏
(4.20c)
[ ] ( ) ( )
1 1
1/ 2 1/ 2
3
1 1
( 1) ( ) ( ) 1 Re 1 ( )
n n
HL HL HL HL HL j HL j
j j
S i β β β φ β β κ β κ
− −
− −
= =
⎧ ⎫
⎡ ⎤
⎪ ⎪
= + Φ − − − − − +
⎨ ⎬ ⎢ ⎥
⎪ ⎪ ⎣ ⎦ ⎩ ⎭
∏ ∏
(4.20d)
where 1 i = − , [ ] Re . represents the real part of the complex argument and (.) φ is the
standard Gaussian PDF. is an asymptotic approximation of
1
S
f
P
HL
being exact for
→∞
2
S , and are correction terms.
3
S β
The curvatures are obtained in the following way. First, the coordinate system
of the standard normalized u space is rotated, so that the design point becomes part of the
last axis of the rotated normal
j
κ
y space. In other words, the last basic vector of the y space
coincides with the design point position vector in the normalized u space. This is achieved
through the following transformation
57
T
= Y D U (4.21)
and, provided that
1 T −
= D D
= U DY. (4.22)
Defining as the matrix of second derivatives of the limit state function in
0
y
H y space at
the design point
0
y , the curvatures are obtained as solutions of a characteristic equation
1
0
( )
det 0
y
n
g
y
κ
−
⎛ ⎞
⎛ ⎞ ∂
⎜ ⎟ − =
⎜ ⎟
⎜
∂
⎝ ⎠
⎝ ⎠
y
H
⎟
I
)
(4.23)
where is the identity matrix. The secondorder reliability index can be defined as
following
I
1
,
(
II f II
P β
−
= −Φ (4.24)
It should be noted that
I
β has a geometric meaning but
II
β has no such interpretation. The
open source code of FORM/SORM is available in [Wierskowski, 2005].
4.3 Calculation of design point
In order to apply FORM/SORM, the design point must be identified. This leads to a
nonlinear constrained optimization problem as follows [Heitzer, 1999]
1
minimize: ( )
2
s.t. ( ) 0
T
f
g
=
≤
u u
u
u
(4.25)
where a coefficient
1
2
is added for technical reasons. Many algorithms have been
suggested to deal with this problem. Staat et al. [2003] and Bjerager [1989] got good
results in probabilistic limit analyses with Rackwitz’s simple gradient search algorithm,
which is based on a linearization of the limit state function at each step. However, this
algorithm is only guaranteed to converge towards a locally most likely failure point in each
sequence of points on the failure surface if the safe region is quasiconvex or concave (see
section 7.4).
A more general algorithm is the Sequential Quadratic Programming (SQP). This
method has proved to be suitable for tasks in the area of the reliability theory [Rackwitz,
2004]. Due to the complex shape of the limit state surface, SQP becomes an efficient tool
for solving nonlinear optimization problems and shall be introduced for this reason here in
short statements. More details can be found in the literature, e.g Schittkowski [1983a,
1983b, 1988, 2004], Bazaraa et al. [1993]. Example of applying SQP in reliability analysis
can be found in [Dronia, 2005].
The SQP method, also known as successive or recursive quadratic programming,
employs Newton’s method (or quasiNewton methods) to solve the KKT conditions for the
58
original problem directly. As a result, the accompanying subproblem turns out to be the
minimization of a quadratic approximation to the Lagrangian function optimized over a
linear approximation to the constraints. Consider the inequality constrained nonlinear
optimization problem
minimize: ( )
s.t. ( ) 0
i
f
g i I ≤ ∈
x
x
(4.26)
where is an ndimensional parameter vector containing the design variables. For this
case, given an iterate , where is the Lagrange multiplier estimating for the
inequality constraints. Define ∇ =
x
( ) ,
k k
x λ 0
k
≥ λ
2 2 2
( ) ( ) ( )
k k ki i k
i I
L f g λ
∈
∇ + ∇
∑
x x x
k
to be the usual Hessian
of the Lagrangian at x . Then a quadratic programming subproblem of the form
2
1
min ) ( )
2
s.t. ( ) ( ) 0
T T
k
T
i k i k
f f L
g g i
+ ∇ + ∇ imize: ( ( )
k k
I + ∇ ≤
d d x d
x x d
x x
(4.27)
∈
1 k
is formulated and must be solved in each iteration. Let be the optimal solution,
k
d
+
λ
k
⎞
⎟
⎠
the
corresponding multiplier of this subproblem, then the new iterate is obtained by
1
1 1
k k
k
k k k
α
+
+ +
⎛ ⎞ ⎛ ⎞ ⎛
= +
⎜ ⎟ ⎜ ⎟ ⎜
−
⎝ ⎠ ⎝ ⎠ ⎝
x x
λ λ λ λ
k
d
(4.28)
where ( ] 0,1
k
α ∈ is a suitable step length parameter. Since might not be positive
definite, it can be replaced by a positive definite approximation , which is known as
BFGS update [Powell, 1978]
2
( )
k
L ∇ x
k
B
1
T T
k k
s
k k k k
k k T T
k k k k k
+
= − +
B p p B s
B B
p B p s p
(4.29)
with
( )
1
1 1 1
1
( , ) ( , )
1 when 0.2
0.8 otherwise
k k k
k k k k
k k k k k
T T
k k k k k
T
k k k
T T
k k k k k
L L
θ θ
θ
+
+ + +
= −
= + −
= ∇ − ∇
⎧ ≥
⎪
=
⎨
⎪
−
⎩
p x x
s q B p
q x λ x λ
p q p B p
p B p
p B p p q
(4.30)
To ensure global convergence of the SQP method and avoid the Maratos effect, a
certain irregularity leading to infinite cycles, we have to select a suitable steplength
k
ξ
subject to the augmented Lagrangian merit function ( ) ξ [Schittkowski, 1983a] ψ
59
1 2
2
2
1
( ) ( ) ( ) ( )
2 2
i
i i i i
i I i I i
f g r g
r
1 λ
ψ ξ ξ λ ξ ξ
∈ ∈
⎛ ⎞
= + + + + + +
⎜ ⎟
⎝ ⎠
∑ ∑
x d x d x d (4.31)
where { }
1
: ( ) /
i i i
I i I g r λ = ∈ ≤ x and
2 1
\ I I I = , is the penalty parameter. In order to
get the value of penalty parameter, two positive constants
1
i
r ≥
1, 1 r ε > <
k
should be chosen.
Then at each iteration, together with the search direction d , two additional variables
2
1
1 2 1
, when
min , and
, otherwise
k T
k k k k k
k k k
k k
k
λ λ
δ δ ε
ε
+
−
+
⎧ ⎫
≠
⎪ ⎪ ⎪
= =
−
⎨ ⎬ ⎨
⎪ ⎪ ⎪ ⎩ ⎭
⎩
d B d
λ λ
d
⎧
d
(4.32)
with
0
1 δ = and a small positive number
k
j , which satisfies the inequality
( )
16
ln
4 1 1 1
1 , i.e. min ,
4 4 ln
k k k
k k k k j
j j N j
r r
ε δ δ
ε δ δ
⎧ ⎫
⎪ ⎪
−
⎪ ⎪ ⎛ ⎞
< − = ∈ >
⎨ ⎬
⎜ ⎟
⎝ ⎠
⎪ ⎪
⎪ ⎪
⎩ ⎭
(4.33)
must be identified. The penalty parameter at iteration can be obtained from
k
i
r k
{ }
1
max ,
k
j k k
i i
r r
−
= r (4.34)
with
0
r r = . The step length
k
ξ can be chosen by a separate algorithm, which should take
the curvature of the merit function into account. If
k
ξ is too small, the resulting step sizes
might become too small leading to a higher number of outer iteration. On the other hand, a
large value close to one requires too many function calls during the line search. Thus we
need some kind of compromise which is obtained by applying first a polynomial
interpolation. Rackwitz [2004] proposed a simple line search, which is based on a
quadratic interpolation minimization algorithm
( )
(0)
0
2 (0) (0) (1)
k
ψ
ξ
ψ ψ ψ
γ
∇
=
+ ∇ −
≥ > (4.35)
where
( )
1
(0) ( ) ( ) ( ) ( )
T T
i i i i i
i I
f g r g g ψ λ
∈
∇ = ∇ + ∇ + ∇
T
∑
x d x d x x d (4.36)
and γ is a small positive number, which ensures that the value of
k
ξ is not too small. This
algorithm is very useful in reliability analysis, because it allows to reduce the number of
deterministic loop in an iteration to two, for 0 = and 1 ξ ξ = , and thus considerably reduce
the computational cost.
Applying SQP method for solving the optimization problem (4.25), the quadratic
programming subproblem (4.27) at iteration becomes k
60
2
1
minimize: ( ) ( ) ( )
2
s.t. ( ) ( ) 0
T T
k k k k k
T
k k k
f f f
g g
+ ∇ + ∇
+ ∇ ≤
u u d d u
u u d
k
d
k
= I
=
k
k
⎞
⎟
⎠
u
(4.37)
with and . The KarushKuhnTucker conditions for this problem
are
( )
k
f ∇ = u u
2
( )
k
f ∇ u
( , ) ( ) ( ) 0
( ) ( ) 0
k k k k k k
T
k k k
L f g
g g
λ λ ∇ = ∇ + + ∇
⎧
⎨
+ ∇ =
⎩
u u d u
u u d
. (4.38)
Applying Newton’s method for solving these equations leads to the system
1
( ) ( )
( ) 0 ( )
k k k
T
k k
g f
g g λ
+
∇ ∇
⎡ ⎤ ⎛ ⎞ ⎛
= −
⎜ ⎟ ⎜ ⎢ ⎥
∇
⎣ ⎦ ⎝ ⎠ ⎝
B u d
u u
(4.39)
with the solutions
1
1 1
1 1
1
( ) ( ) ( )
( ) ( )
( ) ( )
T
k k k k
k T
k k k
k k k k k k
g f g
g g
g f
λ
λ
−
+ −
− −
+
∇ ∇ −
= −
∇ ∇
= − ∇ − ∇
u B u u
u B u
d B u B u
(4.40)
as the Lagrange multiplier and search direction for the next iteration. If then
together with the Lagrange multiplier
0
k
= d
k
u
1 k
λ
+
yields the optimal solution for the problem
(4.25), i.e. the design point is actually found. The calculation of the necessary derivatives
are considered as the sensitivity analysis and will be discussed in the next section.
4.4 Sensitivity of the limit state function
The reliability analysis described above can be carried out now with the help of a
probabilistic limit and shakedown analysis. From the results of the finite element analysis,
the necessary derivatives of the limit state function based shakedown analysis can be
determined analytically. This represents a considerable reduction of computing time
comparing with the other methods, e.g. the difference approximation, and makes such an
efficient and costsaving calculation of the reliability of the structure. Contrary to the
numerical calculation, the analytical calculation is faster and more exact. The necessary
data for the calculation of the derivatives are available after the execution of the
deterministic shakedown analysis since they are based on the limit load factor
lim
α .
The derivatives must be calculated firstly at each iteration in the physical x space.
Then it is transferred into the standard Gaussian u space by using the chain rule
( ) ( )
2 2
( ) ( ) ( )
( ) ( ) ( ) ( )
u u x u
T
u u x u u x u x
g g g
g g g g
∇ = ∇ = ∇ ∇
∇ = ∇ ∇ ∇ = ∇ ∇ ∇ + ∇ ∇
u x x x
u x x x x x x x
2
u
(4.41)
The calculation of the derivatives in the physical x space is based on a sensitivity
analysis. Sensitivity arises prominently in microeconomic theory, where optimality
61
condition provide the basic for analysis. It is interested primarily not in finding an optimal
solution, but rather in how the solution is affected by changes in the problem data. By this
way, sensitivity analysis becomes widely in engineering, especially in reliability analisis of
the structures, where the data defining problem are random variables.
4.4.1 Mathematical sensitivity
Consider a constraint optimization problem is defined as follows P
minimize: ( )
s.t. ( ) 0
i
f
h i I = ∈
x
x
. (4.42)
Suppose that , :
n
i
f h R R → are twice continuously differentiable and I is some
index set. In many applications the objective function f as well as the constraint functions
may depend on other parameters. Consider the following pertubation problem of
the original problem
i
h ( ) P ε
R
*
( ) x ε ( P
*
λ
=
( ) P 0
minimize: f ( , )
s.t. ( , ) 0 ,
m
i
h i I = ∈ ∈
x ε
x ε ε
. (4.43)
The perturbations ε can be interpreted in two ways: as a random error or as
aspecific change in the parameters defining the problem functions. The optimal solution
of the problem with the Lagrange multipliers fulfills the following first
order KarushKuhnTucker conditions
) ε
,
* * * * *
*
, ) ( , ) ( 0
( , ) 0
i i
i I
i
L f
h i I
λ
∈
∇ = ∇ − ∇
= ∈
( , ) h
∑
x λ ε x ε x
x ε
ε
. (4.44)
If f has a continuous and nonsingular gradient matrix ( , ) f ∇
ε
x ε
( ) ε
( ) P 0
*
( ) L ε
in an open set
containing , the implicit function theorem implies the existence of a unique
differentiable local solution
(
of the problem . Let then the
conditions are fulfilled in a local solution of the problem . We defined the
optimal value function and the optimal value Lagrangian as follows
*
( , ) x ε
)
* *
( ), ( ) x ε λ ε P = ε 0
*
( ) x 0
*
( ) f ε
( )
( ), ε ε
* *
( ) f f ≡ ε x
( )
* * *
( ) ( ), ( ), L L ≡ ε x ε λ ε ε
. (4.45)
Theorem [Fiacco, 1983]
If the linear independence condition and the second order sufficient condition of the
problem are satisfied, and if the problem functions are twice continuously
differentiable in ( near then, in a neighborhood of
( ) P ε
, ) x ε
*
( , ) x 0 = ε 0 , we have
62
* *
( ) ( ) f L = ε ε (4.46a)
* * * *
( ) ( ( ), ) ( ) ( ( ), )
i i
i I
d
f f h
d
λ
∈
∂ ∂
= −
∂ ∂
∑
ε x ε ε ε x ε ε
ε ε ε
(4.46b)
2 2 * *
* *
2
* * 2 * *
2
( ( ), ( ), )
( ) ( )
( ( ), ) ( ) ( ( ), ( ), )
T
i i
i I
d L d
f
d d
h L λ
∈
∂
=
∂ ∂
⎛ ⎞ ∂ ∂ ∂
− +
⎜ ⎟
∂ ∂ ∂
⎝ ⎠
∑
x ε λ ε ε
ε x ε
ε x ε ε
x ε ε ε x ε λ ε ε
ε ε ε
(4.46c)
and
2 2 T
L L ∇ ⎧ ⎫ ⎡ ⎤ −∇ ∇ −∇ ⎧ ⎫
=
⎨ ⎬ ⎨ ⎬ ⎢ ⎥
∇ −∇ ∇
⎩ ⎭ ⎣ ⎦ ⎩ ⎭
ε εx x x
ε x ε
x h
λ h 0 h
(4.46d)
where [ ]
1 2
, ,...
T
h h = h
(
. The sensitivity analyses (4.46) provide a quantitative measure of the
first and secondorder change in the optimal value function of . It should be noted
that the value of the optimal value function and its gradient can be calculated once the
optimal solutions
( ) P ε
)
* *
( ), ( ) ε λ ε
lim
x have been determined. However, for the general problem,
the value of Hessian matrix of the optimal value function requires the determination of
both the optimal solutions and its first derivatives.
4.4.2 Definition of the limit state function
As mentioned above, the limit state function contains the parameters of structural
resistance R and loading S. If we defined the limit load factor α as follows
lim
lim
P
P
α = (4.47)
where are limit load and actual load of the structure. For the shake of simplicity, the
limit state function can be normalized with the actual load and then becomes
lim
, P P
g P
lim
1 g α = − (4.48)
63
Deterministic
model and data
Random variables x
and its starting value x
FEMbased deterministic LISA
Limit state function
g(x) = α  1
lim
Sensitivity analysis
∂g/∂x, ∂ g/∂x
2 2
∂x/∂u
SQP algorithm
to find the design point
0
Failure probabilities
(FORM/SORM approximations)
descent direction
line search
∂ x/∂u
2 2
Figure 4.3 Flowchart of the probabilistic limit and shakedown analysis
From previous chapter, the shakedown load factor can be calculated by the non
linear programming (3.15)
( )
lim
,
1 1
min
. : 1 0
ik
P
NG m
T
ik ik
i k
F
s t
α
= =
=
⎛ ⎞
− =
⎜ ⎟
⎝ ⎠
∑∑
e u
e t
(4.49)
with
64
( )
( )
2 2
0 0 1 1 2 2 0
1
1
1 1
1 1 1 1
ˆ
3
ˆ ˆ
2
ˆ ˆ
2
m
NG
k
P
T
i m m
ik i ik i
k k
T
NG m m m
ik ik i ik i
i k k k
P
YN K
F
c
c
Y
κ
ε β β γ β β γ γ
η
=
=
= =
= = = =
⎧ ⎫
⎛ ⎞
⎪ ⎪ ⎜ ⎟ + + + +
⎜ ⎟
⎪ ⎪
⎝ ⎠
=
⎨ ⎬
⎪ ⎪
⎛ ⎞ ⎛ ⎞
+ − −
⎪ ⎪
⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠ ⎩ ⎭
⎧ ⎫
⎛ ⎞ ⎛ ⎞ ⎪ ⎪
= + − −
⎨ ⎬
⎜ ⎟ ⎜ ⎟
⎝ ⎠ ⎝ ⎠
⎪ ⎪
⎩ ⎭
∑
∑
∑ ∑
∑ ∑ ∑ ∑
e B u e B u
e B u e B u
(4.50)
with
( )
2 2
0 0 1 1 2 2 0
ˆ
3
ik
P
N K
κ
η ε β β γ β β γ γ = + + +
+ . (4.51)
It can be seen that the limit state function is the function of yield stress variable Y ,
thickness variable Z and load variables. The shakedown range obtained from (4.49) is
linear function of the yield stress variable Y (see 4.4.4). In the case of heterogeneous
material, we will obtain at different Gaussian points eventually different yield stress
variables . Then the limit state function is no more a linear function of these variables.
The actual load is defined in components by using the concept of a constant referent
load as follows
i
i
Y
P n
0
P
0
0 0
P P
0 0
1 2
0 0
1 1 2 2
...
...
n
n n
P P
P x P x P x P
= + +
= + + +
+
(4.52)
where
j
x is the realization of the
th
j basic load variable
j
X ( 1, j n = ). The corresponding
actual stress field can also be described in the same way
0 0
1 1 2 2
ˆ ˆ ˆ ˆ ...
n n
0
x x x = + + + σ σ σ σ . (4.53)
From (3.10) and (4.53), the corresponding normalized stress fields are obtained
0 0 0
0 0 0 0 1 1 2 2 1 1 2 2
ˆ ˆ ˆ ˆ ... ...
n n n n
N N x x x x x x ε ε ⎡ ⎤ = = + + + = + + +
⎣ ⎦
t σ σ σ σ t t
t (4.54)
or
1
n
j j
j
x
=
=
∑
t t . (4.55)
4.4.3 First derivatives of the limit state function
With the nonlinear programming (4.51), the Lagrange function becomes
65
( )
1 1
1 1 1 1 1 1
1
ˆ ˆ
1
2
NG m
T
PL P ik ik
i k
T
NG m m m NG m
T
ik ik i ik i ik ik
i k k k i k
F F
c
Y
λ
η λ
= =
= = = = = =
⎛ ⎞
= − −
⎜ ⎟
⎝ ⎠
⎧ ⎫
⎛ ⎞ ⎛ ⎞ ⎛ ⎪ ⎪
= + − − −
⎨ ⎬
⎜ ⎟ ⎜ ⎟ ⎜
⎝ ⎠ ⎝ ⎠ ⎝
⎪ ⎪
⎩ ⎭
∑∑
∑ ∑ ∑ ∑ ∑∑
e t
e B u e B u e t
⎞
−
⎟
⎠
lim
(4.56)
In order to get the sensitivity of the limit state function, one require is that the
derivatives of the limit load factor
* * *
( , , )
ik
α must be available. Let λ e u
lim
are the solutions
of the optimization problem. At optimal point, the first derivative of the limit load factor
th
j load variable α versus
j
X can be calculated as follows
( ) ( )
( )
* *
* * * * *
lim
,
1 1 1 1
,
, , , ik
ik ik
NG m NG m
T T ik PL
ik ik ik j
i k i k
j j j
F
X X X
λ
λ λ
λ λ
= = = =
∂ ∂ ∂
= = − = −
∂ ∂ ∂
∑∑ ∑∑
e
e u e
t
e e t
th
α
(4.57)
where is the stress vector at Gaussian point i and load vertex due to the
, ik j
t k j load
case. The derivatives of the limit load factor
lim
α versus yield stress variable Y and
thickness variable Z can be determined in the same way and have the form
( )
( )
* * *
*
* lim lim
1 1 1 1 , ,
ik
ik
NG m NG m
PL
ik ik
i k i k
F
Y Y Y Y
λ
Y
α α
η η
= = = =
∂ ∂ ∂
= = = =
∂ ∂ ∂
⎝ ⎠
∑∑ ∑∑
e u
e
⎛ ⎞
⎜ ⎟
(4.58)
( )
( )
( )
* * *
*
*
2 2 lim
0 0 1 1 2 2 0
1 1 , ,
2 2
0 0 1 2
1 1
ˆ
3
ˆ
3
ik
ik
ik
NG m
k PL
i k
NG m
k
i k
P F
YN K
Z Z Z
P
YN
λ
α
ε β β γ β β γ γ
ε β γ β γ
= =
= =
⎛ ⎞ ⎛ ⎞
∂ ∂ ∂
⎜ ⎟ ⎜ ⎟ = = + + + +
⎜ ⎟ ⎜ ⎟ ∂ ∂ ∂
⎝ ⎠ ⎝ ⎠
⎛ ⎞
⎜ ⎟ = + + +
⎜ ⎟
⎝ ⎠
∑∑
∑∑
e u
e
e
(4.59)
4.4.4 Second derivatives of the limit state function
In SORM algorithm, as discussed above, the Hessian matrix which summarizes
second partial derivatives of the limit state function is needed. They can be obtained from a
directly analytical derivation of the first derivatives. Consider first the derivatives of
lim
α
versus load variables. Taking the derivatives of (4.57) versus
l
X , one gets
( )
( ) ( )
* *
* * * *
2
lim
,
1 1
,
, ,
1 1 1 1
, ,
ik
ik ik
NG m
T
ik ik j
i k l j l
NG m NG m
T T ik
ik ik j ik j
i k i k l l
X X X
X X
λ
λ λ
α
λ
λ
λ
= =
= = = =
∂ ∂ ⎛ ⎞
= −
⎜ ⎟
∂ ∂
⎝ ⎠
∂ ∂
= − −
∂ ∂
∑∑
∑∑ ∑∑
e
e e
e t
e
e t t
(4.60)
66
It can be seen that the derivative
2
lim
/
j
2
X α ∂ ∂ is a special case of (4.60) when l j ≡ .
To complete the calculation in (4.60) the derivatives of λ and versus
ik
e
l
X are also
needed. Since the KarushKuhnTucker conditions and the strict complementary slackness
hold at the optimal point, it follows that
1
1
ˆ
m
⎛ ⎞
= − + i, k
ˆ
PL
ik ik ik i ik
k
ik
F
c
P
κ
λ
=
∂
− ∀
⎜ ⎟
∂
⎝ ⎠
∑
H e e B u t 0
e
= (4.61a)
1 1
ˆ ˆ
NG m
T PL
i ik i
i k
F
c
= =
∂ ⎛
= − − =
⎜
∂
⎝ ⎠
∑ ∑
B e B u
u
⎞
⎟
0
⎞
(4.61b)
1 1
1 0
NG m
T
ik ik
i k = =
− =
∑∑
e t . (4.61c)
Since this relation
1
ˆ
0
m
ik i
k =
⎛
− =
⎜
⎝ ⎠
∑
e B u
⎟
l
holds at optimal point, then we can delete it in
equation (4.61a) and equation (4.61b) is automatically fulfilled. According to implicit
function theorem, by taking derivative of (4.61a, c) versus X and using the chain rule for
two variables ,
ik
λ e one gets
,
i, k
ik
ik ik ik l
l l
X X
λ
λ
∂ ∂
− −
e
G t
= ∀
∂ ∂
t 0 (4.62a)
1 1 1 1
NG m NG m
T ik ik
ik ik
i k i k l l
T
X X
= = = =
∂
= −
∂ ∂
∑∑ ∑∑
e
t
∂t
e (4.62b)
with
(
2
3 2
3
12 1
ˆ ˆ
T PL
ik ik ik ik ik ik
ik
k k
F
P P
∂
= = − + +
∂
G H e e P H
e
) D (4.63)
with are obtained from (3.20), (3.23) and (1.49).
3
, ,
ik ik
H D P
At the optimal point, must be positive definite due to the second order
sufficiency condition, thus it is invertible. Referring to (4.57) the equation (4.62b) can be
written as follows
ik
G
lim
1 1
1
NG m
T ik
ik
i k l l
X X
α
λ
= =
∂ ∂
=
∂ ∂
∑∑
e
t
. (4.64)
From (4.62a) we obtain the derivative of versus
ik
e
l
X
1
,
ik
ik ik ik ik l
l l
X X
1
λ
λ
−
∂ ∂
= +
∂ ∂
e
G t G t
−
. (4.65)
67
Substituting (4.65) into (4.64) and after some manipulations we obtain as well the
derivative of λ versus
l
X
1
lim
,
1 1
1
1 1
1
NG m
T
ik ik ik l
i k l
NG m
T
l
ik ik ik
i k
X
X
α
λ
λ λ
−
= =
−
= =
∂
−
∂ ∂
=
∂
∑∑
∑∑
t G t
t G t
. (4.66)
Analogously, by taking derivatives of (4.58) versus ,
j
X Y and referring to (4.57)
one has
( ) ( )
( )
* *
* * * *
2
lim lim lim
,
1 1
,
, ,
1
ik
ik ik
NG m
T
ik ik j
i k j j j
X Y X Y Y X Y
λ
λ λ
α α α λ
= =
∂ ∂ ∂
⎛ ⎞
= = = −
⎜ ⎟
∂ ∂ ∂ ∂
⎝ ⎠
∑∑
e
e e
e t
(4.67)
* *
2
lim lim lim lim lim lim
2 2
1 1
0
ik ik ik
Y Y Y Y Y Y Y Y Y
α α α α α α ∂ ∂ ∂
⎛ ⎞ ⎛ ⎞ ⎛ ⎞
*
2
= = − = −
⎜ ⎟ ⎜ ⎟ ⎜ ⎟
∂ ∂ ∂
⎝ ⎠ ⎝ ⎠ ⎝ ⎠
e e
=
e
,
ik
e u
. (4.68)
It is obvious from (4.68) that for the case of homogeneous material the shakedown
load factor is actually a linear function of yield stress, i.e. the optimisation variables
are independent of Y . Moreover, the Lagrange multiplier λ is as well a linear function of
. These can be proved by taking derivatives of (4.61a, c) versus , it follows that Y Y
i, k
ˆ
ik ik ik
ik ik
Y Y
Y P
κ
λ ∂ ∂
− + = ∀
∂ ∂
e H e
G t 0
(4.69a)
1 1
0
NG m
T ik
ik
i k
Y
= =
∂
=
∂
∑∑
e
t
. (4.69b)
Substituting (4.69a) into (4.69b) with referring to the relation (4.61a) and after
some manipulations we obtain
, 0
ik
Y Y Y
λ λ ∂ ∂
=
∂ ∂
e
= . (4.70)
The second derivatives of
lim
α versus thickness variable Z are also necessary.
Since and
ik
e Z , from (4.59) and (4.57) one has λ are functions of
( )
( )
*
*
2
2 2 lim
0 0 1 2 2
1 1
2 2
0 0 1 2
1 1
ˆ
3
ˆ
3
ik
ik
T
NG m
ik k
i k ik
NG m
k
i k
P
YN
Z Z
P
YN
Z
α
ε β γ β γ
ε β γ β γ
= =
= =
⎧ ⎫
⎛ ⎞
∂ ∂ ∂ ⎪ ⎪ ⎛ ⎞
⎜ ⎟ = + + +
⎨ ⎬
⎜ ⎟
⎜ ⎟
∂ ∂ ∂
⎝ ⎠
⎪ ⎪
⎝ ⎠ ⎩ ⎭
⎧ ⎫
⎛ ⎞
∂ ⎪ ⎪
⎜ ⎟ + + + +
⎨ ⎬
⎜ ⎟
∂
⎪ ⎪
⎝ ⎠ ⎩ ⎭
∑∑
∑∑
e
e
e
e
(4.71)
68
( )
( ) ( )
* *
* * * *
2
lim
,
1 1
,
, ,
1 1 1 1
, ,
ik
ik ik
NG m
T
ik ik j
i k
j
NG m NG m
T T ik
ik ik j ik j
i k i k
X Z Z
Z Z
λ
λ λ
α
λ
λ
λ
= =
= = = =
∂ ∂ ⎛ ⎞
= −
⎜ ⎟
∂ ∂ ∂
⎝ ⎠
∂ ∂
= − −
∂ ∂
∑∑
∑∑ ∑∑
e
e e
e t
e
e t t
(4.72)
( )
( )
*
*
2
2 2 lim
0 0 1 2
1 1
2 2
0 0 1 2
1 1
ˆ
3
ˆ
3
ik
ik
NG m
k
i k
NG m
k
i k
P
YN
Y Z Y
P
N
α
ε β γ β γ
ε β γ β γ
= =
= =
⎛ ⎞ ⎛ ⎞
∂ ∂
⎜ ⎟ ⎜ ⎟ = + +
⎜ ⎟ ⎜ ⎟ ∂ ∂ ∂
⎝ ⎠ ⎝ ⎠
⎛ ⎞
⎜ ⎟ = + + +
⎜ ⎟
⎝ ⎠
∑∑
∑∑
e
e
+
(4.73)
The first term in (4.71) can be simplified by
( )
( )
2 2
0 0 1 2
2 2
1 2
1 2 3
2 2
1 2
0 0
2 2
1 2
ˆ
3
16 8
3
ˆ
4
1 1
k
ik
ik
k
P
YN
YN
P
ε β γ β γ
β β
β γ β γ
β γ β γ
ε
β γ β γ
⎛ ⎞
∂
⎜ ⎟ + + +
⎜ ⎟
∂
⎝ ⎠
⎧ ⎫
⎛ ⎞
⎪ ⎪ ⎜ ⎟ + + + + −
⎜ ⎟
⎪ ⎪ + +
⎪ ⎝
=
⎨ ⎬
⎛ ⎞
⎪ ⎪
⎜ ⎟ + +
⎪ ⎪
⎜ ⎟
+ +
⎪ ⎪ ⎝ ⎠ ⎩ ⎭
e
P A
e
B
⎠ ⎪
(4.74)
with are given by (1.49) and (3.23). The second term in (4.71) can be rewritten
as follows
3
, , P A B
( )
2 2
0 0 1 2
1 1
1 2
0 0
2 2
1 1
1 2
ˆ
3
ˆ
3
2 2
NG m
k
i k
NG m
k
i k
P
YN
Z
P
YN
ε β γ β γ
β β
ε
β γ β γ
= =
= =
⎧ ⎫
⎛ ⎞
∂ ⎪ ⎪
⎜ ⎟ + + +
⎨ ⎬
⎜ ⎟
∂
⎪ ⎪
⎝ ⎠ ⎩ ⎭
⎛ ⎞ ⎛ ⎞
⎜ ⎟ ⎜ ⎟ = +
⎜ ⎟ ⎜ ⎟
+ +
⎝ ⎠ ⎝ ⎠
∑∑
∑∑
(4.75)
In order to complete (4.71), the derivatives of λ and versus
ik
e Z are needed. By
taking derivatives of (4.61a, c) versus Z and using the chain rule for two variables ,
ik
λ e ,
one has
1
i, k
ˆ
ik ik
ik ik ik
Z Z Z
P
κ
λ ∂ ∂ ∂
− + = ∀
∂ ∂ ∂
e H
G t e
0 (4.76a)
1 1
0
NG m
T ik
ik
i k
Z
= =
∂
=
∂
∑∑
e
t
(4.76b)
Referring to (3.20) and after some manipulations we obtain
69
( )
2 2
1 2 3
1 2
0 0
2 2
1 2
2 2
1 2
4 3
2 3
3 1 1
4
ik
YN
Z
β γ β γ
β β
ε
β γ β γ
β γ β γ
⎛ ⎞
⎜ ⎟
+ + +
⎜ ⎟
⎜ ⎟
⎛ ⎞ ⎜ ⎟
∂
⎜ = + − ⎜ ⎟
⎜
∂
+ + ⎜ ⎟
⎝ ⎠
⎜ ⎟
⎛ ⎞
⎜ ⎟
⎜ ⎟ + +
⎜ ⎟
⎜ ⎟ ⎜ ⎟
+ +
⎝ ⎠ ⎝ ⎠
P
H
A
B
⎟
⎟
(4.77)
From (4.76a) we obtain the derivative of versus
ik
e Z
1 1
1
ˆ
ik ik
ik ik ik ik
Z Z Z
P
κ
λ
− −
∂ ∂ ∂
= −
∂ ∂ ∂
e
G t G e
H
(4.78)
Substituting (4.78) into (4.76b) one gets
1
1 1
1
1 1
1
ˆ
NG m
T
ik
ik ik ik
i k
NG m
T
ik ik ik
i k
Z
P
Z
κ
λ
−
= =
−
= =
∂
∂
∂
=
∂
∑∑
∑∑
H
t G e
t G t
. (4.79)
4.4.5 Special case of probabilistic shakedown analysis
Consider a special case of probabilistic shakedown analysis, where is a
stochastic oneparameter load, i.e. with a reference load we have and
P
0
P
0
P P x =
0
ik ik
x = t t
0
ik
t
0
P where is the stress vector due to referent load . Otherwise we have as well
ik ik
X X
∂
=
∂
t t
. (4.80)
In this case, the nonlinear optimisation (4.49) becomes
( )
lim
,
0
1 1
min
. : 1 0
ik
P
NG m
T
ik ik
i k
F
s t x
α
= =
=
⎛ ⎞
− =
⎜ ⎟
⎝ ⎠
∑∑
e u
e t
(4.81)
If we now define a new strain vector ˆ
ik ik
x = e e
and new displacement vector
ˆ x = u u , then (4.81) can be rewritten as follows
( )
lim
ˆ ˆ ,
0
1 1
min
ˆ . : 1 0
ik
P
NG m
T
ik ik
i k
F
s t
α
= =
=
⎛ ⎞
− =
⎜ ⎟
⎝ ⎠
∑∑
e u
e t
(4.82)
with
70
1 1 1 1
ˆ
ˆ ˆ ˆ
2
T
NG m m m
P ik ik i ik
i k k k
Y c
F
X
η
= = = =
⎧ ⎫
⎛ ⎞ ⎛ ⎪ ⎪ ⎛ ⎞
= + − −
⎨ ⎬
⎜ ⎟ ⎜ ⎟ ⎜
⎝ ⎠ ⎝ ⎠ ⎝ ⎠
⎪ ⎪
⎩ ⎭
∑ ∑ ∑ ∑
e B u e B u
ˆ
ˆ
i
⎞
⎟
. (4.83)
Since the constraint is now no longer depended on the load variable X , it follows
that
( )
* * * *
lim lim
2
1 1 ˆ ˆ , , ˆ
ik ik
NG m
PL
ik
i k
F Y
X X X
λ
X
α α
η
= =
∂ ∂
⎛ ⎞
= = − =
⎜ ⎟
∂ ∂
⎝ ⎠
∑∑
e u e
− (4.84)
2
lim lim lim lim lim lim lim
2 2
1 1
2 2
2
X X X X X X X X X X
α α α α α α ∂ ∂ ∂
⎛ ⎞
= − = − + = + =
⎜ ⎟
∂ ∂ ∂
⎝ ⎠
α
(4.85)
2
lim lim lim lim
1
Y X Y X X Y XY
α α α ∂ ∂ ∂
⎛ ⎞
= − = − = −
⎜ ⎟
∂ ∂ ∂ ∂
⎝ ⎠
α
(4.86)
2
lim lim lim
1
Z X Z X X Z
α α ∂ ∂ ∂
⎛ ⎞
= − = −
⎜ ⎟
∂ ∂ ∂ ∂
⎝ ⎠
α
. (4.87)
These above derivatives can also be easily obtained by using the same procedure as
applied for the material variable Y . By this way, it is easy to prove that
71
ˆ
, 0
ik
X X X
λ λ ∂ ∂
= − =
∂ ∂
e
, and then we find again the above expressions finally.
The flowchart in figure 4.3 contains the logical connections of the main analysis
steps as they have been implemented in our finite element code based probabilistic limit
and shakedown analysis. In each probabilistic iteration, i.e. the iteration for finding the
design point, two deterministic loops are required, the first one provides information for
sensitivity analysis and the second one for the simple line search algorithm, as mentioned
in (4.35).
5 MULTIMODE FAILURE AND THE IMPROVEMENT OF
FORM/SORM RESULTS
In the previous chapter, the concept of reliability has mainly been concerned with
individual structural elements and failure mechanisms. The reliability of a real structure is
usually much more difficult to evaluate since more than one element (member) can fail and
because there is the possibility of more than one failure mode for the system. In this case,
analysis of the structural system is required to evaluate the safety of the structure as a
whole [Ditlevsen and Bjerager, 1986]. To handle problems of this kind, the real structure is
sometimes modelled by an equivalent system in such a way that all relevant failure modes
can be treated [ThoftChristensen and Baker, 1982]. Structural systems can generally be
characterized as series or parallel systems or some combination of the two [Madsen et al.,
1986]. In series system, the formation of any individual failure mode or mechanism is
defined as system failure. For example, in statically determinate or rigidplastic structures,
formation of a collapse mechanism will result in failure of the total system and therefore
they can be modelled as series system with each element of the series being a failure
mechanism. In parallel system, failure in a single element will not result in failure of the
system, because the remaining elements may be able to sustain the external loads by
redistributing of the loads. A typical example of a parallel system is a statically
indeterminate structures. Failure of such structures will always require that more than one
element fails before the structure loses integrity and fails.
The nonlinear optimization algorithm which was developed in section 4.3 is
globally convergent [Tran et al., 2007b, d], i.e., the sequence is guaranteed to converge to a
minimumdistance point on the limit state surface, provided that the limit state function is
continuous and differentiable. However, as with any nonconvex optimisation problem, it
is not guaranteed that the solution point will be the global minimumdistance point when
the system has more than one failure mode (multimode failure or multiple design points).
This chapter aims at presenting a method to successively find the multiple design points of
a component reliability problem, when they exist on the limit state surface. Each design
point corresponds with an individual failure mode or mechanism. FORM and SORM
approximations are applied at each design point followed by a series system reliability
analysis to lead to improved estimates of the system failure probability.
5.1 Multimode failure
5.1.1 Bounds for the system probability of failure
In the previous chapter, a probabilistic algorithm was developed to identify the
failure mode and to evaluate the failure probability of an elastoplastic shell structure. In
structural analysis, it is rather an exception to deal with a problem where the failure of a
structure can be satisfactorily defined by one limit state equation only. Thus, the identified
mode is only one of many modes that can occur, and the probability of the mode occurring
may or may not provide a good estimate of the system probability of failure. If there are
failure mechanisms and the limit state surface is respectively described by equations
m
m
m
1
( ) ( ,..., ) 0, 1,
i i n
g g X X i = = = X (5.1)
and if we denote the failure due to the ith mode as the random event
i
E
{ }
( ) 0
i i
E g = ≤ x X (5.2)
then the probability that the system fails is the probability that any failure mechanism
occurs. It means that
( )
1 2
1
...
m
f m i
i
P P E E E P E
=
⎛ ⎞
= =
⎜ ⎟
⎝ ⎠
∪ ∪ ∪
∪
. (5.3)
If the joint probability density function of the failure events is known, then the
system probability of failure can be calculated by the ndimensional integral
( )
E
f e
0 0
1 1
1
... ( ,..., ) ...
m
f i E m
i
P P E f e e de de
=
−∞ −∞
⎛ ⎞
= =
⎜ ⎟
⎝ ⎠
∫ ∫
∪ m
i
P
. (5.4)
Several techniques such as simulation methods or reducing the dimension of the
original integral (e.g. [Dunnet et al., 1955], [Rackwitz, 1978], [Hohenbichler, 1981]) exist
for evaluation of equation (5.4). However, in general, evaluation of the system probability
of failure through direct integration may not be feasible, even if an expression exists for the
joint density function of failure modes and all failure modes have been identified. In this
case, bounds relieve the necessity of evaluating the mdimensional integral either
analytically, numerically or through Monte Carlo simulation [Ramachandran, 1984],
[Melchers, 1987]. Several firstorder bounds exist (e.g., [Cornell, 1967]) which only
require knowledge of the individual probabilities of failure resulting directly from the
axioms of the probability theory. Unfortunately, these bounds may be quite wide for
structural reliability application [Grimmelt and Schueller, 1982]. Closer or secondorder
bounds can be given in term of the individual failure probabilities and the joint failure
probabilities between any two modes. If we denote the individual failure probabilities as
[ ] ( ) 0 , 1,
i i
P P g i m = ≤ = X (5.5)
73
then, the bounds of the system probability of failure for a series system are [Ditlevsen,
1979]
1
1
2 1 1 2
max , 0 max
m i m m
i ij f i
j i
i j i i
P P P P P
−
<
= = = =
⎧ ⎫
+ − ≤ ≤ −
⎨ ⎬
⎩ ⎭
∑ ∑ ∑ ∑ ij
P (5.6)
with the notation has been used for the joint failure probability
ij
P
( ) 0, ( ) 0
ij i j
P P g g ⎡ ⎤ = ≤ ≤
⎣ ⎦
X X . (5.7)
Since not all couples of the random events are taken into account in equation
(5.6) the ordering of the modes will have an effect on the bounds. Practical experience
suggested that ordering the failure modes according to decreasing values may
correspond to the better bounds. In structural reliability, these bounds are frequently used
and are considered sufficiently accurate for most structural systems [Grimmelt and
Schueller, 1982]. Next section will present the calculation of these secondorder bounds
based on the firstorder system reliability analysis. Another bounds can, of course, be
derived which include the joint probabilities of failure of any three failure modes. It
appears, however, that for practical problems this improvement is small compared to the
improvement in using secondorder bounds instead of firstorder bounds [Hohenbichler
and Rackwitz, 1983].
i
E
i
P
*
, 1, 2...
i
u i m =
5.1.2 Firstorder system reliability analysis
In a firstorder system reliability analysis, the failure set is approximated by the
polyhedral set bounded by the tangent hyperplanes at the design points. Each design point
corresponds to a failure mode and they are the points on the limit state surface that have
smallest distances to the origin in a standard normal space. We denote the design points in
standard normal space as and associated with each design point, we define
the distance
*
HLi i
u β =
i
P
Li
, which is the corresponding reliability index. The individual
probabilities of failure are determined as
( )
i H
P = β Φ − . (5.8)
The firstorder approximation to is obtained by approximating the joint failure
set by the set bounded by the tangent hyperplanes at the design points for the two failure
modes. Figure 5.1 presents the projection of the limit state surface for the two failure
modes on the plane containing the origin and the two design points and u . The joint
failure probability is thus calculated as
ij
P
*
i
u
*
j
ij
P
)
y dxdy
( ) (
, ; , ;
HLj HLi
ij HLi HLj ij ij
P x
β
β
β β ρ ϕ ρ
−
−
−∞ −∞
= Φ − − =
∫ ∫
(5.9)
74
with the correlation coefficient between two failure modes
ij
ρ are
( )
* *
cos
T
i
ij ij
j
HLi HLj
ρ ν
β β
= =
u u
(5.10)
and ( , ; x y ) ϕ ρ is the probability density function for a bivariate normal vector with zero
mean values, unit variances and correlation coefficient ρ
( )
2 2
2
2
1 1 2
, ; exp
2 1
2 1
x y xy
x y
ρ
ϕ ρ
ρ
π ρ
⎡ ⎤ + −
= − . (5.11)
⎢ ⎥
−
− ⎣ ⎦
β
HLi
β
HLj
Origin
ν
ij
i
u
*
u
*
j
Safe region
Failure region
Figure 5.1 Geometrical illustration of firstorder system failure set
Substituting the density function in (5.9) by the corresponding cumulative
distribution function ( ) , ; x y ρ Φ , which gives
( )
( )
( ) ( ) ( )
0
0
, ;
, ; 0
, ;
ij
ij
HLi HLj
ij HLi HLj
HLi HLi HLi HLj
z
P d
z dz
ρ
ρ
β β
β β
ρ
β β ϕ β β
∂Φ − −
= Φ − − +
∂
= Φ − Φ − + − −
∫
∫
z
(5.12)
Numerical techniques are available for evaluating the joint failure probability in
equation (5.12), e.g. NewtonCodes method. Simple bounds on the joint failure probability,
which is based on geometrical illustration of multimode failure system can also be given,
thus avoiding any numerical integration [Madsen et al., 1986]. It should be noted that the
bounds (5.6) still estimate a solution of the generally unknown region which respect to the
exact value of the probability of failure. If we do not know where the values of the
75
probabilities are placed with respect to the exact values, we can not confirm that the
bounds given above estimate the probability of failure. They bound some approximation
and we can only moreorless reasonably expect that the approximation is close to the
exact result and the bounds remain meaningful.
5.2 Solution technique
As discussed above, a real structure may have several failure modes or failure
mechanisms and the existence of multiple failure modes (or multiple design points) may
cause the following problems in FORM and SORM. That is, the optimization algorithm
which was developed in section 4.3 may converge to a local design point. In that case, the
FORM/SORM solution will miss the region of dominant contribution to the failure
probability integral and, thus, has large error. Even if the global design point is found, the
neighborhoods of the local design points may also have significant contribution to the
failure probability integral. Approximating the limit state surface only at the global design
point will lose these contributions.
In this section, a simple method is presented for finding the multiple design points
of a reliability analysis problem, when they exist on the limit state surface. The method
was developed by [Der Kiureghian and Dakessian, 1998]. Once all the design points are
known, the failure probability of series system is calculated by using firstorder system
reliability method and secondorder bounds as presented above.
5.2.1 Basic idea of the method
In optimization theory, one method often suggested to find the global and local
solutions is to repeat the analysis with different starting points and hope that all optimal
points will be found. Unfortunately, it is possible for all trials to converge to the same point
even when the other optimal points exist. A more effective method is to construct
“barriers” around previously found solutions, thus forcing the algorithm to seek a new
solution. In our problem to find the points which have the minimum distances from origin,
the “barriers” around found solutions can be construct by moving the limit state surface in
the neighborhood of them away from the origin. Supposing that the first design point u is
already found by the nonlinear optimization algorithm in section 4.3. We construct a
“barrier” for this point by adding a ‘bulge’ to the limit state surface. Thus, the limit state
function for the deformed surface is
*
1
( ) ( ) ( )
1
g g B = + u u
1
u
( )
1
g u
*
2
u
*
3
u ( )
2
B u
*
2
(5.13)
where defines the bulge fitted at . Solving the optimisation problem with the new
limit state function leads to a second design point . In order to seek the third
solution point , a bulge is now added at u resulting in the limit state function
( )
1
B u
*
1
u
76
( ) ( ) ( ) ( ) ( ) ( )
2 1 2 1 2
g g B g B B = + = + + u u u u u u
 m th
. Repeating the process until all design
points are found. The limit state function for finding the design point thus, is
( ) ( )
1
g g
−
u u ( ) B u
1
1
m
m i
i
−
=
= +
∑
(5.14)
β
HLi i
u
*
γ
β
H
L
i
u
1
2
u
Figure 5.2 Definition of a bulge at design point
*
i
u
As is shown in the figure 5.2, it is possible for the optimisation algorithm to
converge to the points located at the feet of the bulge, which are actually the spurious
minimumdistance points. However, practical experience showed that this occurs only
when there is no other genuine design point. Thus, convergence to a spurious point usually
means that no other genuine design point exists [Der Kiureghian and Dakessian, 1998].
This nature can be used as the stopping criterion of the algorithm.
5.2.2 Definition of the bulge
The derivation of a bulge ( )
i
B u at a design point is based on the following
requirements
*
i
u
*
i
u • must be positive in the neighborhood of (in order to move the limit state
surface away from the origin) and zero elsewhere.
( )
i
B u
• The bulge should have a strong outward curvature at its feet in order to make the
search algorithm converging to an existing genuine design point rather than to its
feet.
• must be continuous and differentiable since the optimization algorithm
requires a continuous and differentiable limit state function.
( )
i
B u
77
Basing on the above considerations, Der Kiureghian and Dakessian proposed a
spherical shape of the bulge as follows
( )
( )
2
2
2 * *
if
0 elsewhere
i i i i i
i
s r r
B
⎧
− − − ⎪
=
⎨
⎪
⎩
u u u u
u
≤
Li
(5.15)
where
i H
r γβ = is the radius of the bulge. The parameter γ should be chosen carefully
since a small γ is conservative but may produce insufficient curvature at the foot of the
bulge, whereas a large one may result in a bulge that conceals other significant design
points. The scale factor governs the height of the bulge and, thus, the distance by which
the limit state surface in the neighborhood of is moved away from the origin. It must be
chosen large enough to assure a strong outward curvature at the food of the bulge. Suppose
we want to move the design point u away from the origin by the amount
i
s
*
i
u
*
i
δ
HLi
β
0
, where
( )
78
δ γ < <
*
i
u . Since the slope of the limit state surface at is
*
i
g ∇ u , it is seen that the
necessary height of the bulge at that point is
( )
*
HLi i
g δβ ∇ u . From equation (5.15), it
follows that
( )
( ) ( )
*
i
2
2 2
HLi
i
HLi HLi
s
δβ
γβ δβ
=
g ∇
⎡ ⎤
−
⎦
u
. (5.16)
⎣
( )
*
i
g ∇ u The definition of the bulge at each design point in term of
HLi
β , and two
parameters γ , δ is complete through the two equations (5.15) and (5.16). To complete the
method one needs to know whether a newly found design point is genuine or spurious.
This is easily solved by calculating its distances from previously found design points. If
any one of these distances is smaller than or equal to the radius of the corresponding bulge,
then the newly found design point is at the foot of this bulge and, therefore, is spurious.
Convergence to a spurious design point usually means that no more genuine design points
exist. However, in order to make sure that this is the case, one may wish to add a new
bulge at this spurious design point and continue the search. Of course, the spurious design
points are not included in the reliability calculations.
6 LIMIT AND SHAKEDOWN ANALYSIS OF DETERMINISTIC
PROBLEMS
The deterministic and probabilistic shakedown optimizations described in the above
chapters is programmed and implemented in the finite element package Code_Aster, a free
open source computing software [Code_Aster 7.3, 2003]. The 3node triangular and 4
node quadrangular isoparametric flat shell elements, which are based on the Kirchhoff’s
hypothesis, were applied. Although the flat shell elements may fail to satisfy the basic
convergence properties that should be expected from sound finite element procedures in
general, they can produce reasonably accurate approximations [Chapelle and Bathe, 2003].
The presented shakedown method applies also to more effective shell elements that work
with generalized variables. But currently only flat elements are available in Code_Aster.
The finite element discretisations were realized by the personal pre and postprocessor
GiD [GiD 7.2, 2002].
Nine examples are presented below for the sake of evaluation the algorithm,
including both plate and shell structures. The structures under consideration can be
subjected to mechanical and thermal loads and are made of elasticperfectly plastic
material. In each numerical test, some existing analytical, numerical and experimental
solutions found in literature are briefly represented and compared. Perfectly plastic LISA
methods cover two mandatory design checks of the new European pressure vessel standard
[European standard, 200506], [Taylor et al., 1999].
• GP check: Limit analysis with the Tresca yield function and a typical safety factor of
1.5 can be used to exclude plastic collapse under monotonic loading by gross plastic
deformation. In practice the nonsmooth Tresca function is replaced by the von Mises
function so that an additional safety factor of
2
1.1547
3
= is effective in pipes and
cylindrical pressure vessels. The total safety factor is 3 1 = .732 , [Vu et al., 2007].
• PD check: Shakedown analysis with the von Mises yield function may be used
without a safety factor to exclude incremental collapse by ratcheting under cyclic
loading.
With respect to pressure vessel design it can be concluded that shakedown analysis is
decisive if the shakedown load is less than
1
0.57735
3
= –times the limit load. If only the
safety factor 1.5 is applied in the GP check then the PD check is decisive if the shakedown
load is less than 0. –times the limit load. The present shakedown analysis includes also
the AP check against failure by alternating plasticity. However, this check is not
mandatory.
6
6.1 Limit analysis of a cylindrical pipe under complex loading
Consider an example of a cylindrical pipe under complex loading, which was
examined by Yan [1997], Vu [2001] and [Larson et al., 1975]. The pipe is subjected to
bending and torsion moments ,
b t
M M as well as internal pressure p and axial tension F
(figure 6.1). We denote here L the length, r the mean radius and h the thickness of the
pipe.
p
r
h
L
F
M
b
t
M
Figure 6.1 Cylindrical pipe under complex loading
Firstly, consider the pipe under pure bending. The analytic solution of plastic
collapse limit may be easily obtained by Bernoulli's theory
2
2
4 (
12
bl y pl el y
h
M W h r σ η σ = = + ) (6.1)
where
y
σ denotes the material yield stress, M
bl
is the maximum bending moment we can
apply on the pipe while other loads are removed.
If only torsion moment is applied on the pipe, the plastic collapse limit is then
2
2
3
tl y
M r h π σ = . (6.2)
Furthermore if bending moment, internal pressure and axial tension apply at the
same time on a thinwalled pipe, the following analytical solution may be used
80
(
(
¸
(
¸
π
−
−
−
=
ϕ
ϕ
ϕ
2
3 4
2
cos
2
3 4
2
2
n
n n
n
m
x
(6.3)
in which , , m n n
ϕ χ
are the normalized quantities
2
= / , =4
= / , = /
= / , =2
l l
l l y
x l l y
m M M M r h
n p p p h r
n F F F rh
ϕ
y
σ
σ
πσ
(6.4)
The formulations (6.1÷6.4) are taken from Yan [1997]. If axial tension is caused by
only internal pressure (the pipe has closed ends), it is easy to find that n and the
formulation (6.3) now becomes
2 /
ϕ
n
x
=
2
3 4
2
ϕ
n
m
−
= . (6.5)
In our numerical analysis the whole pipe is discretized by 300 quadrangular flat 4
node shell elements as shown in figure 6.2. The following geometrical and physical
parameters are adopted: 2700mm, 300mm, 60mm, 160MPa σ = = = =
y
L r h .
If only a bending moment is applied with its limit value M
(formula 6.1), the limit load factor is 1.010. In this case, the numerical error is 1%
compared with the analytical solution. On the other hand, if only torsion moment is applied
with its limit value M (formula 6.2), the limit load factor is 1.0053,
and the numerical error is only 0.53% compared with the analytical solution.
kNm 52 . 3467 = =
bl b
M
kNm 24 . 3134 = =
tl t
M
Figure 6.2 FEmesh of cylindrical pipe
When the two loads (bending moment and internal pressure) are applied, the limit
load factor can be estimated by using (6.5). The interaction diagram (figure 6.3) presents
both numerical and analytical results. It is observed that the numerical solution agrees very
81
well with the analytical one: the maximum numerical error is less than 1%.
On the other hand, if only axial tension is applied, the limit load factor is 1.0015. In
this case, numerical error is only 0.15% compared with the analytical solution.
Figure 6.3 Cylindrical pipe under bending and internal pressure
6.2 Limit and shakedown analysis of a thinwalled pipe subjected to internal pressure
and axial force
In this example, a thinwalled pipe with thickness h and mean radius as shown
in figure 6.4 is considered. The pipe is subjected to internal pressure
r
p together with axial
force . This problem was investigated analytically by Cocks and Leckie [1988] using the
Tresca yield criterion and by Yan [1997] using the von Mises yield criterion.
F
When both internal pressure and axial force increase monotonically and
proportionally, the plastic collapse limit can be computed by using the condition:
1
2
2
2
2
= − +
l l l l
F
F
p
p
F
F
p
p
(6.6)
where
y
l
h
p
r
σ
β = , 2
l y
F rh π σ = with 1 = β for a long pipe without the end constraining
effect.
82
p
F F
r
h
Figure 6.4 Thinwalled pipe
If internal pressure remains constant while axial force varies within the range
, the shakedown limit can be computed by using the following condition:  F F, − 
1
2
2
2
2
= + +
l l l l
F
F
p
p
F
F
p
p
. (6.7)
Figure 6.5 FEmesh of the pipe
Note that (6.6) and (6.7) were found by using the von Mises yield criterion [Yan
1997]. On the other hand, if the Tresca yield criterion is used, the shakedown range is
limited by the condition [Cocks and Leckie 1988]:
l l
F
F
p
p
− =1 . (6.8)
In our numerical analysis the whole pipe is discretized by 400 3node triangular flat
shell elements as shown in figure 6.5 with the following geometrical properties and
material yield stress: , 300 mm r = 10 mm h = , 2700 mm L = , 116 MPa
y
σ = .
The interaction diagram for limit and shakedown analysis with normalized pressure
/
l
p p and normalized axial force F is plotted in figure 6.6. Analytical solutions are
computed by using the formulations (6.6) and (6.7). In all cases, the diagram shows good
agreements between numerical and analytical solutions. The numerical error is less than
1%.
/
l
F
83
Figure 6.6 Interaction diagram
6.3 Cylindrical shell under internal pressure and temperature change
In this continuous wellknown example, consider an axially restrained cylindrical
shell of mean radius r and thickness . The shell is subjected to a constant internal
pressure
h
p and a uniformly or linearly distributed temperature T which may vary within a
certain range. Numerical analysis is carried out by using 200 quadrangular flat 4node shell
elements as presented in figure 6.7. The following material properties and geometrical data
are adopted: 116.2 MPa
y
σ = , 151000 MPa E = , 0.3 υ = ,
, h in which is Young’s modulus,
5 o 1
1.0 10
T
C α
− −
= ×
500 = r mm 10 m = m E
T
α is the coefficient of thermal
expansion and
y
σ is the yield limit. Two following cases are examined
Uniformly distributed temperature
This problem was examined by Zhang [1995b] and Vu [2001]. The inner and outer
temperature are the same T T and vary within the range [ , . Omitting the
failure caused by axial instability, the shakedown load can be computed by using the
following condition [Zhang, 1995b]
i u
T = =
0 0
] T T T + ∆
2 2
1
3
,
2 2
p T
T
p T
y y
s s
E T pr
s s
h
α
σ σ
+ =
∆
= =
(6.9)
84
In the case of ∆ = the analytical solution leads to 0 T 2.684 MPa
sd
p = . On the
other hand, when only temperature applies 0 p = , it results in ∆ = . 102.605
o
sd
T C
T
p
u
T
i
r
h
Figure 6.7 FEmesh and geometrical dimensions
Figure 6.8 Interaction diagram (uniformly distributed temperature)
Numerical results with normalized loads ( ) ,
sd sd
p T ∆ are presented in the interaction
diagram, figure 6.8. It can be shown that these results agree well with the analytical
85
approximation. The maximum error is less than 1%. It is noted that the collapse mode at
shakedown limit is alternating plasticity in the axial direction of the pipe only when
pressure is absent and purely incremental plasticity in the other cases.
Linearly distributed temperature
This case was investigated by Heitzer [1999] when the temperature field is linearly
distributed between inner and outer wall of the shell. The inner temperature T may vary
within the range [0 while the outer one is set to zero T
i
0
, ] T 0
u
= .
In the case of shakedown analysis gives the shakedown load
. Numerical results with normalized loads
0 p =
o
136.248
sd
T = C ( ) ,
sd sd
p T are presented in the
interaction diagram, figure 6.9. It can be shown that the collapse mode at shakedown limit
is alternating plasticity when pressure is small and purely incremental plasticity when
pressure becomes bigger.
Figure 6.9 Interaction diagram (linearly distributed temperature)
6.4 Pipejunction subjected to varying internal pressure and temperature
Thermal loading is known to have a great effect on the shakedown or inadaptation
behaviour of structures. In this example a pipejunction subjected to internal pressure p
and temperature loads is examined. The supposed temperature field T is a linearly
distributed function through the thickness with outer value T and inner value T . The
essential dimensions are chosen as follows
u
0
i
=
Internal diameter of the pipe 39 mm D = .
Internal diameter of the junction d 15 mm = .
86
Thickness of the pipe and junction 3.44 mm s t = = .
In the work of Heitzer and Staat [1999], which is based on the lower bound
approach, 125 solid 27node hexahedron elements were used to model one fourth of the
structure due to its symmetry. Analogously, the problem was analyzed by Vu [2001] with a
primaldual shakedown algorithm using 720 solid 20node hexahedron elements. In our
analysis, numerical computations are carried out by using 800 flat 4noded shell elements
as presented in figure 6.10. The following material properties are adopted:
160 MPa, 210000 MPa, 0.3
y
E σ υ = =
5 o 1
1.0 10
T
C α
− −
= ×
= with the coefficient of thermal expansion
.
s
D
d t
Figure 6.10 FEmesh and geometrical dimensions
Pipejunction under internal pressure
For the sake of comparison, consider firstly the case of having only mechanical
load. Numerical limit analysis leads to the plastic collapse pressure
0.145 23.2 MPa
l y
p σ = = compared with other results (table 6.1). The convergence
property of this solution presented in figure 6.11 shows that the limit load factor becomes
stationary after only five iteration steps.
In the case of shakedown analysis, the internal pressure may vary within range
 
0
0, p . Numerical analysis gives the shakedown load p 0.10267 16.43 MPa
sd y
σ = =
compared with other known solutions (table 6.2). It is noted that in this case the structure
fails due to the alternalting plasticity. Convergence analysis of the optimization algorithm
shows that the shakedown load factor becomes stationary also after five iteration steps.
87
Table 6.1: Limit analysis of pipejunction
Limit load
normalized with
y
σ
Nature of solution Source
Staat and Heitzer 0.134 lower bound [Heitzer, 1999]
ADMerkblatt B9 0.136 German design rules [Heitzer, 1999]
Vu 0.14424 lower bound [Vu, 2001]
Vu 0.14434 upper bound [Vu, 2001]
SAMCEF 0.161 step by step method [Vu, 2001]
Vu 0.141 double elastic slope method [Vu, 2001]
Present 0.145 upper bound 
Table 6.2: Shakedown analysis of pipejunction
Shakedown load
normalized with
y
σ
Nature of solution Source
Staat and Heitzer 0.0952 lower bound [Heitzer, 1999]
Vu 0.10983 lower bound [Vu, 2001]
Vu 0.11044 upper bound [Vu, 2001]
Present 0.10267 upper bound 
Figure 6.11 Convergence of limit and shakedown load factors
Pipejunction under temperature change
Consider the case of only thermal load being applied, where T may vary within the
range 
u

0
0,T . Numerical analysis gives the shakedown load T C . It is also
noted that in this case the structure fails due to the alternating plasticity.
o
93.5086
sd
=
88
Pipejunction under internal pressure and temperature
For this case, two loading ranges are examined
1) The pressure and temperature vary simultaneously and dependently (one
parameter loads)
0
0
0
0 , 0 1
u
p p
T T
µ
µ µ
≤ ≤
≤ ≤ ≤ ≤
(6.10)
2) The pressure and temperature vary independently (twoparameter loads)
1 0 1
2 0 2
0 , 0 1
0 , 0 1
u
p p
T T
µ µ
µ µ
≤ ≤ ≤ ≤
≤ ≤ ≤ ≤
(6.11)
Figure 6.12 Interaction diagram of pipejunction^
89
In order to get the general solutions, both positive and negative temperatures
are included. Numerical results are presented in figure 6.12.
0 0
0 and 0 T T ≥ ≤
From the interaction diagram it can be shown that a thermal load has no effect on
the limit load of the pipejunction. In the case of twoparameter loads, the shakedown loads
are symmetric via pressure load axis, i.e. they are the same for both positive and negative
ranges of the outer temperature, while the elastic limits are not. Otherwise the elastic limits
and shakedown loads become bigger when the loads vary simultaneously and dependently
(oneparameter loads) with the temperature varying in the positive range, but they are
identical when the temperature is in the negative range. In all cases, the shakedown limit is
twice the elastic limit which indicates that inadaptation will occur due to alternating
plasticity. In this case the shakedown check is not mandatory although this leads to a
design which is prone to LCF. Staat et al. [2005] have discussed this example
demonstrating that twice the elastic limit is a PD check which should be used with great
case or better avoided at all because it is an AP check and it may be not safe for
independently varying loads.
6.5 Grooved rectangular plate subjected to varying tension and bending
In order to examine the limit and shakedown loads of thin plates, we consider
firstly a grooved rectangular plate subjected to inplane tension and bending
(figure 6.13). The load domain is defined by
N
p
M
p
0 ,
0 ,
M y
N y
p
p
σ
σ
( ∈
¸ ¸
( ∈
¸ ¸
(6.12)
where
y
σ denotes the material yield stress.
R
p
M
R
L
N
p
Figure 6.13 FEmesh and geometrical dimensions
90
Limit analysis of the structure was studied by Prager and Hodge [1951], Casciaro
and Cascini [1982] and Yan [1997] for pure tension load 0 , 0 = ≠
M N
p p . Heitzer [1999]
and Vu [2001] investigated the problem for the more complicated case with
. In the present analysis, the structure is modelled by 300 quadrangular flat
4node shell elements as shown in figure 6.13. The following geometrical and material
properties are chosen: R ,
0 , 0 ≠ ≠
M N
p p
mm 250 = 4 L R = , 50 h mm = , MPa E 21000 = , 3 . 0 = ν ,
116.2 MPa
y
σ = .
Consider the case of constant pure tension p , 0
N y M
p σ = = . The plastic collapse
load factor is presented in table 6.3 together with known solutions for comparison. It can
be shown that our solution is well accordant with the others, which used volume finite
elements basing on plane stress assumption.
Table 6.3: Limit analysis
( )
, 0
N y M
p p σ = =
Plane stress Plane strain
Nature of
solution
Yield criterion
Prager 0.500 0.630÷0.695 analytical Tresca
Casciaro 0.568 0.699 numerical von Mises
Yan 0.500÷0.577 0.727÷0.800 analytical von Mises
Yan 0.558 0.769 numerical von Mises
Vu 0.557 0.799÷0.802 numerical von Mises
Present 0.572  numerical von Mises
For the case of having both the inplane tension and bending, limit and shakedown
analysis are implemented. Figure 6.14 shows the evolutions of limit and shakedown load
factors. In the case of limit analysis, the upper bound converges rapidly to the solution
0.30498
l
α = . Numerical result gives the shakedown load factor 0.23624
sd
α = compared
with upper bound factor of obtained by Vu [2001]. 0.23494
The interaction diagram of the plate in figure 6.15 shows three limit lines: the
plastic collapse limit, shakedown limit and elastic limit. It is observed that, in almost cases,
inadaptation will occur due to alternating plasticity. These results agree well with the
solutions of Vu [2001] and Heitzer [1999].
91
Figure 6.14 Limit and shakedown analysis 0, , 0,
N y M
p p
y
σ σ ( ∈ ∈ (
¸ ¸ ¸ ¸
Figure 6.15 Interaction diagram of grooved rectangular plate
6.6 Square plate with a central circular hole
In this continuous wellknown example, a square plate with central circular hole
subjected to two loads
1
p
and
2
p which can vary dependently or independently is
considered. The limit load of the problem was examined analytically by Gaydon and
McCrum [1954] by using plane stress hypothesis and von Mises yield criterion. Numerical
limit and shakedown analyses were also investigated by some authors, e.g. Heitzer [1999],
Vu [2001] for different ratios / R L to evaluate the elasticplastic behaviour of the
structure.
92
L
R
2
1
p
Figure 6.16 FEmesh and geometrical dimensions
In our analysis, due to the symmetry, one fourth of the plate is discretized by 600
quadrangular flat 4node shell elements as shown in figure 6.16. The following material
properties are chosen: , MPa E 21000 = 3 . 0 = ν , 116.2 MPa
y
σ = . The two cases: one and
two applied loads are examined.
One applied load
In this case one load is set to zero and the other one can vary within a range
2 1
0, 0,
y
p p σ = ∈
¸
(
¸
. With 0 204 . 0 / ≤ < L R , the exact plastic collapse limit is found since
in this range the lower and the upper bound are coincident
( )
lim
1 /
y
p R L σ = − . (6.13)
As example, the exact collapse limit load in the case of 2 . 0 / = L R is
lim
0.8
y
p σ = .
Our numerical solution obtained in this case is 0.80149
y
σ . Based on an elastic analysis,
the alternating (plastic fatigue) limit can be estimated by using formula (2.47). The
numerical results obtained by our unified shakedown algorithm represent the minimum
between the alternating limit and the incremental limit. For 2 . 0 / = L
y
R , the collapse mode
is alternating plasticity. Our obtained alternating limit is 0.59904σ and the unified
shakedown algorithm gives 0.60332
y
σ . The small gap between them may be caused by
the use of the quadratic approximation of exact Ilyushin yield surface in the calculation of
alternating plasticity, while the exact one was used in the unified shakedown analysis. The
convergence property of these solutions presented in figure 6.17 shows that the limit and
shakedown load factor become stationary after seven iteration steps.
93
Figure 6.17 Limit and shakedown analysis
1 2
0, , 0
y
p p σ ( ∈ =
¸ ¸
In order to examine the geometrical effect of the circular hole, various values of
ratio / R L were also considered. The obtained numerical results are introduced in table
6.4, compared with the solutions of Heitzer [1999], which are obtained by lower bound
algorithm. It can be shown that our solutions agree well with those of Heitzer.
Table 6.4: Limit and shakedown analysis
1 2
0, , 0
y
p p σ ( ∈ =
¸ ¸
R/L
Shakedown load
factors (Heitzer)
Limit load
factors (Heitzer)
Shakedown load
factors (present)
Limit load
factors (present)
0.1 0.671 0.8951 0.6546 0.90172
0.2 0.6157 0.7879 0.60332 0.80149
0.3 0.5212 0.691 0.52012 0.70221
0.4 0.4361 0.572 0.43367 0.59139
0.5 0.3302 0.4409 0.31600 0.40117
0.6 0.2104 0.2556 0.21323 0.24249
0.7 0.1327 0.1378 0.12378 0.12541
0.8 0.0557 0.0565 0.05225 0.05227
0.9 0.0191 0.0193 0.01226 0.01226
Two applied loads
For this case, exact value of the plastic collapse limit load are known in the range
where analytical lower bound coincides with upper one: 1 / 483 . 0 ≤ < L R
( )
( )
3
6
lim 2
2 1 3
sin ,
6 2cos
3 /
y
p e
R L
π
α
π
α σ
α
 
−

\ .
 
= − =

\ .
(6.14)
94
Details about the upper and lower bounds within the range 0 483 . 0 / < ≤ L R can be
found in the work of Gaydon and McCrum [1954]. Numerical limit load factors for
different values of R are introduced in tables 6.5, compared with analytical results
[Gaydon and McCrum, 1954] and numerical results of Vu [2001], which are obtained by a
dual algorithm.
L /
Table 6.5: Limit analysis
1 2 y
p p σ = =
R/L
Lower bound
(analytical)
Upper bound
(analytical)
Lower bound
(Khoi [2001])
Upper bound
(Khoi [2001])
Present
0.1 0.97063 0.99215 0.97082 0.97104 0.96951
0.2 0.89425 0.92376 0.89374 0.89472 0.89404
0.3 0.79122 0.80829 0.79075 0.79125 0.79167
0.4 0.67602 0.69048 0.67585 0.67592 0.67814
0.5 0.55682 0.55682 0.55666 0.55679 0.56133
0.6 0.43801 0.43801 0.43791 0.43819 0.44109
0.7 0.32195 0.32195 0.32196 0.32221 0.32504
0.8 0.20991 0.20991 0.21010 0.21016 0.21285
0.9 0.10249 0.10249 0.10264 0.10267 0.10457
Table 6.6: Shakedown analysis
1 2
0, , 0,
y
p p
y
σ σ ( ∈ ∈ (
¸ ¸ ¸ ¸
, twoparameter loads
R/L
Alternating limit
(Khoi [2001])
Lower bound
(Khoi [2001])
Upper bound
(Khoi [2001])
Present
0.1 0.49082 0.49082 0.49086 0.48718
0.2 0.43384 0.43384 0.43390 0.43136
0.3 0.36128 0.36128 0.36131 0.35575
0.4 0.27635 0.27635 0.27638 0.27300
0.5 0.19442 0.19442 0.19445 0.19052
0.6 0.12360 0.12360 0.12364 0.12102
0.7 0.06763 0.06763 0.06765 0.06940
0.8 0.02903 0.02903 0.02905 0.02772
0.9 0.00709 0.00709 0.00710 0.00640
In the case of shakedown analysis three loading ranges are examined including both
positive and negative loads .
2 2
0 and 0 p p ≥ ≤
1) Two loads vary simultaneously and dependently (oneparameter loads)
1
2 2
0 , 0 1
0 , or 0
y
y y
p
p p
µσ µ
µσ µσ
≤ ≤ ≤ ≤
≤ ≤ ≥ ≥ −
(6.15)
2) Two loads vary independently (twoparameter loads)
95
1 1 1
2 2 2
0 , 0 1
0 , 0 1
y
y
p
p
µ σ µ
µ σ µ
≤ ≤ ≤ ≤
≤ ≤ ≤ ≤
(6.16)
3) Two loads vary independently (twoparameter loads)
1 1 1
2 2 2
0 , 0 1
0 , 0
y
y
p
p 1
µ σ µ
µ σ µ
≤ ≤ ≤ ≤
≥ ≥ − ≤ ≤
(6.17)
Figure 6.18 Interaction diagram of a square plate with a central hole, R/L = 0.2
For the sake of comparison, consider first the twoparameter load case with positive
load ranges. Numerical shakedown load factors for different values of are introduced
in tables 6.6, compared with numerical results of Vu [2001], which are obtained by a
primaldual algorithm.
L R/
96
Figure 6.18 presents both the numerical limit and shakedown analyses for the case
of . From this interaction diagram it can be shown that, in the case of two
parameter loads, the shakedown loads are symmetric via
/ 0. R L = 2
s
1
/  axi
y
p σ , i.e. they are the
same for both positive and negative ranges of
2
p , while the elastic limits are not.
Otherwise the elastic limits and shakedown loads become bigger when the loads vary
simultaneously and dependently (oneparameter loads) with one load varies in the positive
range, but they are identical when this load is in the negative range. It is also noted that
when one load varies within the negative range, the collapse mode is purely alternating
plasticity.
6.7 Elbow subjected to bending moment
The pipe bends are a problem of great interest to many designers. They have a
complex response to inplane and outofplane bending moments. When an external
moment is applied to one of its ends, the cross section tends to deform significantly both in
and out of its own plane. The momentend rotation curves show a defined limit load
behaviour for the closing mode of inplane bending but not for the opening mode
[Chattopadhyay et al., 2000], [Diem, 1994]. This difference can be explained as an effect
of large displacements. Due to their curved geometry, the stresses and strains are much
higher than those present in a straight pipe of the same size and material, under the same
loading conditions. For this reason, pipe bends are considered as critical components of a
piping system. In addition, the variation of the limit moment with the change of the
internal pressure has been of great interest and has been studied numerically, analytically
and experimentally [Kussmaul et al., 1995], [Yan, 1996, 1997], [Shalaby et al., 1998a, b],
[Mourad et al., 2002], [Chattopadhyay et al., 1999, 2000], [Fanous et al., 2005], [Tran et
al., 2007a,e]. However, the examination of shakedown limit is still rare.
Consider an 90 elbow with mean radius r , bend radius of curvature
o
R and
thickness . One of its ends is supposed clamped and the other one is subjected to an in
plane closing moment
h
I
M or outofplane bending moment
II
M as shown in figure 6.19.
The curvature factor is defined as follow
2
r
Rh
= λ (6.18)
Generally, 0.5 λ ≤ corresponds to a highlycurved pipe, while λ → ∞ corresponds
to a straight pipe. In order to evaluate the model different values of λ within the range
  0.1, 1.2 are examined. Our model for elasticplastic analysis is meshed by 700
quadrangular flat 4node shell elements as shown in figure 6.19. The following material
properties are chosen: , 208000 MPa E = 3 . 0 = ν , 250 MPa
y
σ = .
97
2r
R
h
I
M
M
II
Figure 6.19 FEmesh and geometrical dimensions
Elbow under inplane closing bending moment
We define the limit load factor as follows
/
e
I I
s
I
M M α = (6.19)
in which
e
I
M is inplane limit moment of elbow,
s
I
M is the limit moment of the straight
pipe, which has the same radius as the elbow. For a straight thinwalled pipe,
s
I
M is
determined as (see formular 6.4)
2
4
s
I y
M hr σ = . (6.20)
For this kind of pipe bend, some analytical solutions are available in the literature
1) Based on small displacement analysis and perfectly plastic material behavior
• Calladine’s and Yan’s solutions
By using an elastic solution of Clark and Reissner [1951] and the criterion of
Hodge [1954], which is a kind of sandwich approximation of Mises' criterion in the case of
a cylindrical shell, Calladine [1974] proposed a lower bound solution for an infinite highly
curved pipe ( 0.5 λ ≤ )
2/ 3
0.9346
C
I
α λ = . (6.21a)
This solution is considered in the literature to come close to the experimental limit
load factor [Bolt and Greenstreet, 1972], [Goodall, 1978a, b], [Griffiths, 1979]. According
98
to Yan [1996, 1997], it is a good approximation when 0.7 λ < . For a slightlycurved pipe
( 0.7 λ ≥ ), he proposed an approximate solution which is validated by numerical analysis
cos( )
6
Y
I
π
α
λ
= . (6.21b)
• Desquines’s solution
In the framework of his PhD thesis, Desquines et al. [1997] proposed a more
general analytical solution as a lower bound, which can be applied for any value of λ
2
1
0.3015
1
De
I
α
λ
=
+
. (6.22)
• Spence and Findlay’s solution
In order to get an analytical solution for the limit load of an elbow, Spence and
Findlay [1973] already developed two theoretical approaches. The first one utilises a linear
elastic analysis to estimate limit loads as similar to Calladine. The second one is to
manipulate the results of a creep analysis of bends to give an approximate solution of the
limit load. They result with 1.45 λ <
0.6
0.8 , 1.45
SF
I
α λ λ = <
SF
, (6.23a)
1, 1.45
I
α λ = > . (6.23b)
2) Based on large displacement analysis
• Goodall’s solution
Goodall [1978a] proposed the maximum loadcarrying capacity of the elbow
subjected to closing bending moment as
2/ 3
1.04
1
G
I
λ
α
β
=
+
(6.24)
where
( )
2/ 3
2
4 3(1 ) 3
2
3
y
r
Eh
ν σ λ
β
π
(
−
= +
(
(
¸ ¸
. (6.25)
• Touboul’s solution
Based on the experimental study at CEA DEMT, Touboul et al. [1989] proposed
the following equations of collapse moments of elbows
99
2/ 3
Closing collapse: 0.715
T
I
α λ = , (6.26a)
1/ 3
Opening collapse: 0.722
T
I
α λ = . (6.26b)
• Drubay’s solution
Drubay et al [1995] proposed another closing mode collapse moments of elbows as
2/ 3
0.769
Dr
I
α λ = . (6.27)
Our numerical results are introduced in table 6.7 and figure 6.20, compared with
these above analytical solutions and a numerical solution of Yan [1997]. It is shown that
our solutions compare well with the other analytical solutions, which are based on small
displacement theory, but are bigger than those which are based on large displacement
theory. They converge as an upper bound of Calladine’s solution and lower bound of
Desquines’s solution.
Elbow under outofplane bending moment
We define the limit load factor for this case as follows
/
e
II II II
t
M M α = (6.28)
where
e
II
M is outofplane limit moment of the elbow,
t
II
M is the torsion limit moment of
the axle which has the same radius as the elbow and determined by
2
2
3
t
II y
M hr π σ = . (6.29)
By this definition, Yan [1996, 1997] proposed an analytical solution for the elbow
subjected to outofplane bending moment
0.6
1.1 , 0.5
Y
II
α λ λ = <
≤
, (6.30)
1/ 3
0.9 , 0.5 1.4
Y
II
α λ λ = ≤ . (6.31)
Numerical results are introduced in table 6.8 and figure 6.21, compared with the
analytical solution of Yan [1997]. It is shown that our solutions compare well with the
analytical solution outside the range 0.4 0.7 λ ≤ ≤ .
100
Table 6.7: Limit load factors of an elbow under inplane closing bending moment
λ
Calladine
(6.21a)
Yan
(6.21b)
Desquines
(6.22)
Spence
Findlay
(6.23)
Goodall
(6.24)
Touboul
(6.26)
Drubay
(6.27)
Present
0.100 0.2013  0.1791 0.2001 0.1489 0.154 0.1657 0.2155
0.200 0.3196  0.3422 0.3046 0.2817 0.2445 0.263 0.3279
0.250 0.3709  0.4144 0.3482 0.3401 0.2838 0.3052 0.3900
0.300 0.4188  0.4794 0.3885 0.3947 0.3204 0.3446 0.4614
0.363 0.4756  0.5515 0.4355 0.4593 0.3638 0.3913 0.5200
0.400 0.5074  0.5888 0.4617 0.4955 0.3882 0.4175 0.5589
0.500 0.5887  0.6732 0.5278 0.5879 0.4504 0.4844 0.6260
0.600 0.6648  0.7377 0.5888 0.6741 0.5087 0.5471 0.6930
0.650 0.7013  0.7639 0.6178 0.7153 0.5365 0.577 0.7227
0.700 0.7368 0.7330 0.7868 0.6458 0.7554 0.5637 0.6063 0.7494
0.750  0.7660 0.8069 0.6732 0.7945 0.5902 0.6348 0.7773
0.800  0.7933 0.8244 0.6998 0.8327 0.6162 0.6627 0.7974
0.903  0.8365 0.8544 0.7525 0.909 0.668 0.7184 0.8341
1.000  0.8660 0.8765 0.8000 0.9782 0.715 0.769 0.8617
1.200  0.9063 0.9093 0.8925 1.1138 0.8074 0.8684 0.9006
0.2 0.4 0.6 0.8 1 1.2
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Curvature factor
L
i
m
i
t
l
o
a
d
f
a
c
t
o
r
CalladineYan
Desquines
Spence & Findlay
Yan (numerical)
Present
Figure 6.20 Limit load factors of an elbow under inplane closing bending moment
101
Table 6.8: Limit load factors of an elbow under outofplane bending moment
λ Yan (6.30) Yan (6.31) Present
0.100 0.2763  0.2476
0.200 0.4188  0.4047
0.250 0.4788  0.4709
0.300 0.5341  0.5244
0.363 0.5989  0.5675
0.400 0.6348  0.6063
0.450 0.6813  0.6337
0.500 0.7257 0.7143 0.6575
0.550  0.7374 0.6924
0.600  0.7591 0.7245
0.650  0.7796 0.7539
0.700  0.7991 0.7808
0.750  0.8177 0.8053
0.800  0.8355 0.8276
0.903  0.8699 0.8674
1.000  0.9000 0.8984
1.200  0.9564 0.9467
0.2 0.4 0.6 0.8 1 1.2
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Curvature factor
L
i
m
i
t
l
o
a
d
f
a
c
t
o
r
Yan
Present
Figure 6.21 Limit load factors of an elbow under outofplane bending moment
102
6.8 Limit and shakedown analysis of pipeelbow subjected to complex loads
In this continuous example of pipe bends a system of two straight pipes connected
to the ends of an elbow is considered. This pipeelbow can be subjected to internal pressure
p , axial force F , transversal forces F and bending moments
p
,
I II
M M as shown in
figure 6.22. One of its ends is clamped and the other one is free. In the case of transversal
forces, the two pipe ends are free. In order to evaluate the model, some subproblems are
examined [Tran et al., 2007c].
2r
R
h
I
M
M
II
F
p
p
L
F
F
p
Figure 6.22 Pipeelbow subjected to complex loads
Pipeelbow under transverse forces F
This is an experimental model presented in [Diem, 1994], [Kussmaul et al., 1995].
The geometrical data and material properties are chosen as: h , mean
radius ,
30.5 = mm
229.25 r m = m 900 R mm = , 5100 L mm = (curvature factor 0.522 λ = ),
, 208000 E MPa = 0.3 ν = , 387
y
MPa σ = .
In this case, the structure can be considered as being subjected to the inplane
closing bending moment which is determined by
( )
1
2
M F L R R
(
= − +
(
¸ ¸
. (6.32)
For this reason the available analytical solution in the literature can be applied to
calculate the limit load . In the analysis, due to the symmetry, one fourth of the structure
is modelled by 300 quadrangular flat 4node shell elements as shown in figure 6.23. Table
l
F
103
6.9 shows the comparison of the numerical result. It is seen that the present result is
compatible with the result of Goodall [1978] but higher than that of Diem [1994],
Kussmaul et al. [1995] and Calladine [1974]. This is due to two reasons: the stiffening
effect of connecting straight pipes and not taking into account the nonlinear geometrical
effect in the present method. It was noted by Griffiths [1979] that the effect of connecting
straight pipes can increase the limit load up to 33% with respect to Calladine’s solution.
Table 6.9: Limit load of the pipeelbow under transverse force (kN)
l
F
Diem
[1994]
Goodall
eq. (6.24)
Touboul
eq. (6.26)
Drubay
eq. (6.27)
Calladine
eq. (6.21)
Desquines
(6.22)
Spence &
Findlay (6.23)
Present
exp. large displacements small displacements LISA
372 403.26 297.22 319.67 387.45 441.79 347.29 404.83
Figure 6.23 FEmesh of a pipeelbow subjected to transversal forces
Pipeelbow under inplane closing bending moment and internal pressure
A pipeelbow is often subjected to combined internal pressure and bending moment
during service, which may vary independently in the range  
0
0, p and  
0 0
, M M − ,
respectively. It is now well known that the internal pressure has a significant effect on the
limit load of a pipeelbow. In all the foregoing analytical equations, the effect of internal
pressure is not taken into account. Goodall [1978b] was the first to propose the closed
form equation of the limit load of pipeelbows under combined internal pressure and in
plane bending moment by the small displacement analysis. The proposed equation was
1/ 3
2/ 3
1.04 1
2
G
I
y
pr
h
α λ
σ
 
= −
\ .


. (6.33)
From (6.33), it can be seen that the internal pressure reduces the limit moment.
Based on nonlinear finite element analyses and twice the elastic slope method,
Chattopadhyay [1999, 2000] investigated various sizes of pipeelbows subjected to
104
combined internal pressure and inplane bending moment. He proposed another analytical
solution for the pipeelbow which takes the strainhardening of the material into account
by
2/ 3 2
Closing collapse: 1.122 0.175 0.508
p C
I p
α
α λ α
λ
= + − (6.34a)
1/ 3 2
1.2
Opening collapse: 1.047 0.124 0.568
p C
I p
α
α λ α
λ
= + − (6.34b)
Applicablity: 0.24 0.6 and 0 1.0
p
λ α ≤ ≤ ≤ ≤
y
where /
p
pr h α σ =
and 45 h m
is the normalized pressure. In our numerical analysis one half of the
pipeelbow is meshed by using 640 quadrangular 4node flat shell elements as shown in
figure 6.24. The following geometrical data and material properties are adopted:
, , 40 = m 250 r m = m 750 R mm = , 1500 L mm = (curvature factor 0.480 λ =
and 0.540 λ = ), 203000 E MPa = , 0.3 ν = , 270
y
MPa σ = , 513 MPa
u
σ = . The elastic
data is only needed for the calculation of the elastic limit. Closedend condition of the
pipeelbow is simulated by applying an axial force of intensity
2
p r π at the end of
connecting straight pipe.
Figure 6.24 FEmesh of pipeelbow under internal pressure and inplane bending
Figure 6.25 shows the effect of internal pressure on the limit moment of the pipe
elbow. It can be seen that the limit moment decreases when the pressure increases, as
presented by Fanous et al. [2005]. Our solutions agree well with those of Goodall [1978b]
but not with those of [Chattopadhyay, 2000]. This is due to the omission of geometrical
nonlinearity and strainhardening of the material by the present method and in Goodall’s
analysis. Hence, the analysis of the pipeelbow without the large displacement effect gives
a conservative and acceptable result [Fanous et al., 2005].
105
0 0.2 0.4 0.6 0.8 1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
curvature factor = 0.48
Normalized pressure
L
o
a
d
f
a
c
t
o
r
Goodallcollapse
Chattopadhyaycollapse
Presentcollapse
Shakedown limit
Elastic limit
0 0.2 0.4 0.6 0.8 1
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
curvature factor = 0.54
Normalized pressure
L
o
a
d
f
a
c
t
o
r
Goodallcollapse
Chattopadhyaycollapse
Presentcollapse
Shakedown limit
Elastic limit
Figure 6.25 Effect of internal pressure on inplane limit moments
Contrary to the case of limit analysis, internal pressure has a very little effect on the
shakedown limit moment. For small pressure the shakedown limit is between 57.7% and
of the plastic collapse load. This small shakedown loads limits the benefit of the
strengthening effect of the geometrical nonlinear analysis in [Chattopadhyay et al, 2000].
At low pressure also the difference between the elastic limit and the shakedown load is
small so that fatigue and ratchetting could be of concern. The allowance of cyclic loading
up to the shakedown limit without any safety factor may only be acceptable in connection
with a planned monitoring and maintenance program.
66.7%
Pipeelbow under outofplane bending moment and internal pressure
This pipeelbow analyzed by Mourad and Younan [2002], Fanous et al. [2005] is
considered. The curvature factor of the pipe bend is 0.1615, the mean radius
, the bend radius R 203.2 mm r = 609.6 mm = , and the thickness h . The
length of two straight pipes is four times the diameter L
10.94 mm =
m 1625.6 m = in order to reduce
the effect of the pipe ends on the response of the elbow [Fanous et al., 2005]. Both internal
pressure and bending moment can vary independently in the range  
0
0, p and  
0 0
, M M − ,
respectively.
In our analysis, the model is discretised by 1500 quadrangular flat 4node shell
elements as shown in figure 6.26. An axial force is applied at the end of connecting straight
pipe to account for the closedend condition. The material model used is perfectly plastic
stainless steel 304 which has an Young’s modulus E 193743 MPa = , a yield stress
272 MPa
y
σ = and Poisson’s ratio 0.47 ν = .
The applied internal pressure effects the plastic collapse limit moment. Our
numerical results are presented in figure 6.27, compared with three plastic collapse limit
solutions of Fanous et al. [2005], which are obtained by RNode (redistribution of nodes),
Iterative RNode and plastic analysis methods. It is shown that these results compare well
with those of Fanous et al. [2005]. The plastic collapse limit moment decreases when
106
pressure increases noticeably above 4 MPa . The thesis [Rahman, 2006] demonstrates that
the existing cyclic plasticity models (Chaboche, OhnoWang, and Abdel KarimOhno) are
not capable of simulating straight and elbow pipe ratchetting responses satisfactorily when
the model parameters are determined from material level responses. The direct shakedown
analysis gives robust predictions of shakedown limits on the basis of restricted information
about the material, [Staat and Heitzer, 2002]. Figure 6.27 shows that large pressure (greater
than ) reduces the shakedown limit moment considerably. But smaller pressure has
no effect in contrast to the elastic limit. With respect to the design code the same
comments apply as for figure 6.25.
9 MPa
Figure 6.26 FEmesh of pipeelbow under internal pressure and outofplane bending
For both pipeelbows no PD check has been published which could be used to
verify the calculated shakedown limits. The shakedown analysis of pipe bends in [Abdalla
et al., 2006] is a nonmandatory AP check. In order to show the convergence property of
the algorithm, convergence analysis and the influences of optimisation parameters such as
the penalty parameter c and
2
η were included for the reference loads M
and . Figure 6.28 demonstrates that the direct plasticity methods are
extremely effective. Shakedown analysis converges in only three iterations which each
needs approximately the computing time of a linear elastic analysis. This is an important
saving compared to the analysis of a large number of plastic cycles and makes limit and
shakedown analysis a first candidate for structural optimization or reliability analysis [Tran
et al., 2007d]. It is shown in figures 6.296.30 that the parameters
225.97 kNm =
6.895 MPa p =
2
η and c have very little
107
influence on the both solutions in the ranges of
2 12
10 ,10 η
− −28
∈
and c . The
numerical results remain virtually constant over these intervals.
8 17
10 ,10 ∈
Figure 6.27 Effect of internal pressure on outofplane limit moments
0 2 4 6 8 10 12 14 16
2
4
6
8
10
12
Number of Iterations
L
o
a
d
f
a
c
t
o
r
Limit analysis
Shakedown analysis
Figure 6.28 Limit load convergence of the pipe bend
108
10 15 20 25
0.5
1
1.5
2
2.5
3
3.5
log(Eta square)
L
o
a
d
f
a
c
t
o
r
Limit analysis
Shakedown analysis
Figure 6.29 Influence of
2
η
8 10 12 14 16 18
0.985
0.99
0.995
1
1.005
1.01
1.015
1.02
log(c)
L
i
m
i
t
l
o
a
d
f
a
c
t
o
r
8 10 12 14 16 18 20
0.685
0.686
0.687
0.688
0.689
0.69
log(c)
S
h
a
k
e
d
o
w
n
l
o
a
d
f
a
c
t
o
r
Figure 6.30 Influence of penalty parameter c
6.9 Nozzle in the knuckle region of a torispherical head
Torispherical drumheads under internal pressure loading are known to have high
stresses in the ‘knuckle region’ due to the change of the curvature of the shell. In designing
such vessels it is usually necessary to have piping nozzle connections welded in the head.
To assess the plastic collapse limit and shakedown limit of such structures is of great
interest to many designers and has been included in some design code rules, e.g. the UK
pressure vessel design code BS 5500, the German pressure vessel design code AD
Merkblatt B3 and the European standard for unfired pressure vessels CEN. But they are
still limited to some specific head shapes and loadings and not applicable for nozzles
109
particularly for nozzle encroaching in the knuckle region [Hsieh et al., 2000]. Recently, a
parametric study of the knuckleencroaching nozzles was conducted by Bauer and Saal
[1997] by using finite element analysis. Using the von Mises yield criterion and 15times
the elastic slope method, that is by drawing a line at a slope 15 times that of the elastic
portion of the relevant load versus deformation curve, Hsieh et al. [2000] investigated a
series of knuckleencroaching nozzles for both internal pressure and nozzle loads. A study
of both individually and combined limit loads has also been undertaken. However, in the
above studies, the cyclic loads were not considered.
In the present example, a single nozzle on a Klöpperboden head has been selected
for investigation. Both internal pressure p and nozzle loads including axial force F , in
plane bending moment
I
M , outofplane bending moment
o
M , and twisting moment
t
M
were considered. The loads may be constant or vary independently from zero to the
maximum value. The nozzle axis is parallel to the vessel axis as presented in the figure
6.31a. The dimensions are selected from the study in [Hsieh et al., 2000],[Bauer and Saal,
1997], [Tran et al., 2007c].
o
M
M
I
D
a
c
a
d
i
R
ts
te
r
M
t
F
p
Figure 6.31 Geometrical dimensions of nozzle and FEmesh
Vessel outer diameter and thickness 2000 mm, 17 mm
a e
D t = =
Head crown and knuckle radii 2000 mm, 200 mm R r = =
110
Nozzle inner diameter and thickness 400 mm, 13.6 mm
i s
d t = =
Nozzle offset c 910 mm
a
=
The elasticperfectly plastic material model was assumed with yield stress
340
y
MPa σ =
i
d
. In order to ensure that the boundary condition effects do not influence the
nozzle junction behaviour, the vessel cylinder length and nozzle length were chosen as
and 3 , respectively. For all load cases the end of the main vessel is fully clamped. In our
numerical analysis the full model was meshed by using 3032 quadrangular 4noded flat
shell elements as shown in figure 6.31b.
a
D
Numerical results for individual loads are introduced in table 6.10. Together with
the first yield loads and limit loads, shakedown limits are also presented. It is seen that in
all cases the results obtained with the present shell model compare very well with those
gained by Hsieh et al. [2000] with 20noded volume elements. The maximum difference of
the limit analyses is 2.7% appearing in the case of internal pressure. It is noted by Hsieh et
al. [2000] in this case that the limit load of using linear finite elements, i.e. the 8noded
brick finite elements, is 29% higher than that of using 20noded elements.
Table 6.10: Elastic limit, limit and shakedown loads of the nozzle
Source,
type of
element
Limit criterion
Pressure
[MPa]
Axial
force
[MN]
Inplane
moment
[kNm]
Outofplane
moment
[kNm]
Torsion
[kNm]
Elastic limit 1.37 0.45 64.7 66.5 193.8
Hsieh et al.
[2000],
volume
elements
Plastic
collapse limit
3.54 1.63 282.6 265.8 625
Elastic limit 1.439 0.391 64.4 68.2 193.3
Plastic
collapse limit
3.636 1.669 284 270.8 635
Present,
shell
elements
Shakedown
limit
2.878 0.783 128.8 136.4 386.6
The elastic analysis of the load case of axial force shows the exceptionally large
difference of 13 . Therefore two mesh convergence analyses have been performed. First
local mesh refinements with nearly square finite elements have been made around the
intersection of the nozzle with the vessel head (figure 6.32a). Next the local refinement has
been achieved by changing the shell elements from squares to bad shaped rectangular
elements with aspect ratio 12 (figure 6.32b). The size of the elements in the refinement
region has been changed by a factor 12 from the locally coarse mesh to the fine one. Figure
6.32 shows that even bad shaped shell elements do not cause convergence problems. Only
very coarse meshes cause convergence problems for linearly elastic analyses. In the case of
.1%
111
alternating plasticity these problems carry over to shakedown analyses. The plastic
collapse limit is virtually independent on the mesh. Obviously limit analysis is very robust.
2000 4000 6000 8000 10000 12000
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Nozzle under tension
Number of finite elements
L
o
a
d
l
i
m
i
t
s
Elastic limit
Shakedown limit
Plastic collapse limit
(a)
(b)
Figure 6.32 Mesh convergence analyses for the case of axial force
The interaction diagrams for load combinations between internal pressure and
nozzle loads are presented in figures 6.336.36. In all cases the loads are normalized with
the respective plastic collapse limit. These figures illustrate clearly the need to consider
shakedown as a design criterion for knuckleencroaching nozzles since the shakedown
limits are considerably smaller than 57.7% of the plastic collapse limits in a wide range of
parameters. On the other hand, in almost all cases it is twice the elastic limit which
indicates that inadaptation will occur due to alternating plasticity. In this case the
shakedown check is not mandatory although this leads to a design which is prone to LCF.
In the case of internal pressure and axial force (figure 6.33), the behaviour is of
course signdependent. For compressive nozzle loading, a strengthening effect on plastic
collapse limit appears as presented in Hsieh et al. [2000]. Contrarily, as would not be
expected, there is a weakening effect on the shakedown limit in this case.
Figure 6.34 presents the interaction between internal pressure and inplane bending
moment. There is also a strengthening effect on plastic collapse limit in the case of positive
bending moment (nozzle rotated inwards) as mentioned by Hsieh et al. It is observed that
for this case while inplane moment plays an important role, internal pressure has a small
influence on the shakedown limit.
Figures 6.35 and 6.36 show the interactions of pressure versus outofplane bending
and torsion moments. For both cases the behaviours of plastic collapse limits can be
considered to be bilinear. A topologically different mesh has been used in Tran et al.
[2007c] which exhibited sensitivity.
112
Figure 6.33 Pressure versus axial force interaction
Figure 6.34 Pressure versus inplane bending moment interaction
113
Figure 6.35 Pressure versus outofplane bending moment interaction
Figure 6.36 Pressure versus torsion moment interaction
114
7 PROBABILISTIC LIMIT AND SHAKEDOWN ANALYSIS OF
STRUCTURES
In this chapter numerical applications of the probabilistic limit and shakedown
programming are introduced. Four examples are presented below for the sake of evaluation of
the algorithm, including both plate and shell structures. The first two examples illustrate the
FORM/SORM calculations for structures in which the loads and yield stress are considered as
random variables. In the third example the effect of the thickness imperfection on the limit
load of a shell is studied by considering the thickness as another stochastic variable. The
fourth example presents the estimation of the system failure probability of a frame formed of
three plates subjected to vertical and horizontal loads. In each numerical test, some existing
analytical and numerical solutions found in literature are briefly represented and compared.
7.1 Square plate with a central circular hole
In this example a square plate with central circular hole which was investigated in
section 6.6 and in [Tran et al., 2007b] is considered. It has the ratio and is
subjected to a uniaxial tension
/ 0.2 R L =
1
p which can be constant or can vary within the range
[ ]
max
0, p
y
. Both yield limit σ and load
1
p are supposed to be random variables. This
structure was also studied analytically and numerically by Staat and Heitzer [2003b] using
FORM and volume element. The following two cases are examined.
Limit load analysis
( )
lim
1 /
y
p R L For this case the exact plastic collapse limit is given by σ = −
/ 0.2 R L ≤
,
r s
in the
range , since the lower and upper bound are coincident. If both material and load
random variables are supposed to be normally distributed with means μ μ
,
r s
and standard
deviations σ σ respectively, then the reliability index may be given [Staat and Heitzer,
2003]
( )
( )
2 2
2 2
1 / 0.8
0.64
1 /
r s r
2
HL
r s
r s
R L
R L
μ μ μ
s
μ
β
σ σ
σ σ
− − −
= =
+
− +
. (7.1)
If both random variables are lognormally distributed, Staat and Heitzer also
introduced the analytical expression
( ) ( ) ( )
2 2
log 1 / log
r s
HL
r s
R L m m
β
δ δ
− −
=
+
( 7.2)
where and ,
r
m m
s
,
r s
δ δ are calculated as follows
2
,
2
/ 2 ,
, , , 2
2
,
,
2
,
, log 1
1
r s
r s r s
r s r s r s
r s
r s
r s
m e
δ
μ σ
μ δ
μ
σ
μ
−
⎛ ⎞
= = =
⎜ ⎟
⎜ ⎟
⎛ ⎞
⎝ ⎠
+
⎜ ⎟
⎝ ⎠
,
+ . (7.3)
In our numerical analysis, both kinds of random variables above were tested.
Numerical results of failure probability
f
P
/
for both normal and lognormal distributions are
introduced in tables 7.17.7, compared with the exact solutions and numerical solutions of
Staat and Heitzer. The numerical error involves both deterministic shakedown analysis and
reliability analysis is very small and acceptable. It is shown that SORM solutions are almost
the same of FORM due to weak nonlinearity of the limit state function. The weak non
linearity may be lost in the case of multiparameter loads or by an extreme value distribution
of the random variables. Figure 7.1 presents the failure probabilities versus
s
μ
r
μ for the
case of lognormal distributions, which are obtained by SORM.
Table 7.1: Numerical results and comparison for normal distributions
Limit analysis
, ,
0.1
r s r s
σ μ =
/
s r
μ μ
f
P (anal.)
f
P (FORM)
f
P (SORM)
f
P (num.)
[Staat et al., 2003b]
0.2 1.718E13 1.880E13 1.878E13 2.643E13
0.3 2.426E09 2.720E09 2.716E09 3.843E09
0.4 3.872E06 3.872E06 3.872E06 6.112E06
0.5 7.364E04 7.424E04 7.421E04 1.093E03
0.6 2.275E02 2.050E02 2.140E02 3.049E02
0.7 1.734E01 1.606E01 1.606E01 2.067E01
0.8 5.000E01 4.835E01 4.835E01 5.550E01
0.9 7.969E01 7.981E01 7.986E01 8.305E01
1.0 9.408E01 9.419E01 9.419E01 9.544E01
1.1 9.863E01 9.848E01 9.848E01 9.900E01
1.2 9.972E01 9.973E01 9.973E01 9.981E01
1.3 9.995E01 9.996E01 9.996E01 9.996E01
1.4 9.999E01 9.999E01 9.999E01 9.999E01
116
Table 7.2: Numerical results and comparison for normal distributions
Limit analysis 0.2 , 0.1
r r s s
σ μ σ μ = =
/
s r
μ μ
f
P (anal.)
f
P (FORM)
f
P (SORM)
f
P (num.)
[Staat et al., 2003b]
0.1 6.313E06 5.901E06 5.902E06 6.455E06
0.2 9.919E04 9.436E04 9.437E04 1.085E04
0.3 1.065E03 1.045E03 1.046E03 1.205E03
0.4 7.646E03 7.494E03 7.496E03 8.763E03
0.5 3.675E02 4.053E02 4.052E02 4.163E02
0.6 1.209E01 1.318E01 1.318E01 1.361E01
0.7 2.835E01 3.028E01 3.028E01 3.103E01
0.8 5.000E01 5.314E01 5.314E01 5.319E01
0.9 7.070E01 7.042E01 7.042E01 7.397E01
1.0 8.554E01 8.621E01 8.621E01 8.765E01
1.1 9.388E01 9.451E01 9.451E01 9.499E01
1.2 9.772E01 9.850E01 9.850E01 9.821E01
1.3 9.925E01 9.955E01 9.955E01 9.943E01
1.4 9.976E01 9.987E01 9.987E01 9.983E01
1.5 9.993E01 9.995E01 9.995E01 9.995E01
1.6 9.998E01 9.999E01 9.999E01 9.999E01
Table 7.3: Numerical results and comparison for normal distributions
Limit analysis 0.2 , 0.2
r r s s
σ μ σ μ = =
/
s r
μ μ
f
P (anal.)
f
P (FORM)
f
P (SORM)
f
P (num.)
[Staat et al., 2003b]
0.1 7.085E06 6.866E06 6.866E06 7.271E06
0.2 1.374E04 1.312E04 1.312E04 1.514E04
0.3 1.717E03 1.887E03 1.887E03 1.948E03
0.4 1.267E02 1.247E02 1.248E02 1.437E02
0.5 5.590E02 6.252E02 6.250E02 6.271E02
0.6 1.587E01 1.723E01 1.723E01 1.742E01
0.7 3.190E01 3.366E01 3.366E01 3.427E01
0.8 5.000E01 5.240E01 5.239E01 5.256E01
0.9 6.610E01 6.851E01 6.850E01 6.843E01
1.0 7.826E01 8.019E01 8.018E01 8.021E01
1.1 8.649E01 8.795E01 8.794E01 8.775E01
1.2 9.172E01 9.277E01 9.276E01 9.258E01
1.3 9.493E01 9.563E01 9.562E01 9.550E01
1.4 9.686E01 9.730E01 9.730E01 9.729E01
1.5 9.802E01 9.831E01 9.831E01 9.823E01
1.6 9.873E01 9.892E01 9.892E01 9.882E01
117
Table 7.4: Numerical results and comparison for lognormal distributions
Limit analysis 0.1 , 0.1
r r s s
σ μ σ μ = =
/
s r
μ μ
f
P (anal.)
f
P (FORM)
f
P (SORM)
f
P (num.)
[Staat et al., 2003b]
0.3 1.790E12 1.704E12 1.704E12 9.593E12
0.4 4.473E07 5.097E07 5.097E07 1.409E06
0.5 4.315E04 5.205E04 5.205E04 1.009E03
0.6 2.071E02 1.814E02 1.814E02 3.485E02
0.7 1.719E01 1.604E01 1.604E01 2.409E01
0.8 5.000E01 4.794E01 4.794E01 5.396E01
0.9 7.981E01 7.839E01 7.839E01 8.575E01
1.0 9.431E01 9.373E01 9.373E01 9.648E01
1.1 9.880E01 9.867E01 9.867E01 9.935E01
1.2 9.979E01 9.976E01 9.976E01 9.990E01
1.3 9.997E01 9.997E01 9.997E01 9.998E01
1.4 9.999E01 9.999E01 9.999E01 9.999E01
Table 7.5: Numerical results and comparison for lognormal distributions
Limit analysis 0.1 , 0.2
r r s s
σ μ σ μ = =
/
s r
μ μ
f
P (anal.)
f
P (FORM)
f
P (SORM)
f
P (num.)
[Staat et al., 2003b]
0.2 3.090E10 2.146E10 2.152E10 3.566E10
0.3 6.586E06 4.490E06 4.497E06 6.861E06
0.4 1.107E03 8.235E04 8.235E04 1.168E03
0.5 2.000E02 1.645E02 1.645E02 2.114E02
0.6 1.090E01 1.011E01 1.011E01 1.115E01
0.7 2.959E01 2.738E01 2.738E01 2.981E01
0.8 5.263E01 5.024E01 5.024E01 5.338E01
0.9 7.248E01 7.061E01 7.061E01 7.337E01
1.0 8.582E01 8.472E01 8.472E01 8.635E01
1.1 9.334E01 9.269E01 9.269E01 9.353E01
1.2 9.709E01 9.676E01 9.676E01 9.715E01
1.3 9.879E01 9.866E01 9.866E01 9.884E01
1.4 9.951E01 9.947E01 9.947E01 9.954E01
1.5 9.981E01 9.979E01 9.979E01 9.982E01
118
Table 7.6: Numerical results and comparison for lognormal distributions
Limit analysis 0.2 , 0.1
r r s s
σ μ σ μ = =
/
s r
μ μ
f
P (anal.)
f
P (FORM)
f
P (SORM)
f
P (num.)
[Staat et al., 2003b]
0.2 1.327E10 1.678E10 1.678E10 8.296E10
0.3 3.574E06 7.072E06 7.050E06 1.303E05
0.4 7.067E04 1.062E03 1.060E03 1.807E03
0.5 1.442E02 1.782E02 1.781E02 2.848E02
0.6 8.638E02 1.078E01 1.077E01 1.407E01
0.7 2.520E01 2.873E01 2.872E01 3.518E01
0.8 4.736E01 5.121E01 5.121E01 5.860E01
0.9 6.790E01 7.124E01 7.124E01 7.744E01
1.0 8.264E01 8.510E01 8.509E01 8.884E01
1.1 9.146E01 9.297E01 9.296E01 9.509E01
1.2 9.610E01 9.689E01 9.689E01 9.796E01
1.3 9.831E01 9.870E01 9.870E01 9.918E01
1.4 9.930E01 9.947E01 9.947E01 9.969E01
1.5 9.971E01 9.979E01 9.979E01 9.988E01
Table 7.7: Numerical results and comparison for lognormal distributions
Limit analysis 0.2 , 0.2
r r s s
σ μ σ μ = =
/
s r
μ μ
f
P (anal.)
f
P (FORM)
f
P (SORM)
f
P (num.)
[Staat et al., 2003b]
0.1 5.655E14 3.885E14 3.887E14 1.337E13
0.2 3.715E07 4.828E07 4.820E07 6.786E07
0.3 2.308E04 2.458E04 2.451E04 4.601E03
0.4 6.664E03 6.485E03 6.500E03 9.443E03
0.5 4.665E02 4.363E02 4.363E02 5.864E02
0.6 1.521E01 1.497E01 1.497E01 1.816E01
0.7 3.167E01 3.120E01 3.121E01 3.601E01
0.8 5.000E01 4.899E01 4.908E01 5.458E01
0.9 6.629E01 6.541E01 6.541E01 7.046E01
1.0 7.871E01 7.786E01 7.786E01 8.213E01
1.1 8.722E01 8.665E01 8.665E01 8.959E01
1.2 9.261E01 9.225E01 9.225E01 9.414E01
1.3 9.584E01 9.562E01 9.562E01 9.674E01
1.4 9.771E01 9.762E01 9.763E01 9.828E01
1.5 9.875E01 9.868E01 9.868E01 9.907E01
119
Figure 7.1 Probability of failure versus /
s r
μ μ for lognormally distributed variables
Figure 7.2 Comparison of the results for probability of failure for normally distributed
variables
Shakedown load analysis
For this case, the tension p varies within the range [ ]
max
0, p and only the amplitudes
but not the uncertain complete load history influences the solution. Consider the case where
the maximal magnitude
max
p is a random variable and the minimal magnitude is held zero.
120
From the deterministic analysis we got the shakedown load 0.60332
sh y
p σ = . The numerical
probability of failure for normally distributed variables are presented in table 7.8, compared
with the analytical solutions, which are calculated by (7.1). In figure 7.2 the failure
probabilities (SORM approximations) are shown versus /
s
μ
r
μ
, ,
0.1
r s r s
for both two cases limit and
shakedown analysis and for the case of normal distributions with σ μ = . It is worth to
note that the shakedown probabilities of failure are considerably smaller than those of limit
analysis. Thus, the loading conditions should be considered carefully when assessing the load
carrying capacity of the structure.
Table 7.8: Numerical results and comparison for normal distributions
, , r s
0.1 Shakedown analysis
r s
σ μ =
/
s r
μ μ
f
P (anal.)
f
P (FORM)
f
P (SORM)
0.2 1.109E10 9.333E11 9.333E11
0.3 3.371E06 2.833E06 2.833E06
0.4 2.486E03 2.168E03 2.168E03
0.5 9.366E02 8.644E02 8.648E02
0.6 4.844E01 4.732E01 4.732E01
0.60332 5.000E01 4.891E01 4.891E01
0.65 7.007E01 6.906E01 6.906E01
0.7 8.523E01 8.464E01 8.464E01
0.8 9.752E01 9.735E01 9.735E01
0.9 9.969E01 9.967E01 9.967E01
1.0 9.996E01 9.996E01 9.996E01
1.1 9.999E01 9.999E01 9.999E01
7.2 Pipejunction subjected to internal pressure
In this example, the pipejunction which was investigated in section 6.3 and in [Tran et
al., 2007d] is considered. It is subjected to internal pressure p which can be constant or vary
within the range [ ]
max
0, p
y
. Both yield limit σ and load p are considered as random
variables. Numerical deterministic analyses lead to a collapse pressure
lim
0.145
y
p σ =
0.10267
and
shakedown limit
sh y
p σ =
,
r s
. If both material and load random variables are supposed to
be normally distributed with means μ ,
r s
and standard deviations μ σ σ respectively, then
the reliability index may be given [Staat and Heitzer, 2003b]
2 2
0.145
f mit ana
0.021
HL
r s
or li lysis
r s
μ μ
β
σ σ
−
= (7.4a)
+
121
Table 7.9: Numerical results and comparison for normal distributions,
, ,
0.1
r s r s
μ = σ
Limit analysis Shakedown analysis
/
s r
μ μ
f
P (anal.)
f
P
(FORM)
f
P
(SORM)
/
s r
μ μ
f
P (anal.)
f
P
(FORM)
f
P
(SORM)
0.04 1.469E12 1.239E12 1.345E12 0.03 5.456E12 4.349E12 4.352E12
0.06 3.036E08 2.764E08 2.976E08 0.04 6.441E09 9.304E09 9.302E09
0.08 4.336E05 4.225E05 4.474E05 0.05 1.992E06 1.758E06 1.804E06
0.10 5.312E03 5.108E03 5.312E03 0.06 1.665E04 1.423E04 1.433E04
0.12 9.204E02 8.946E02 9.157E02 0.07 4.280E03 3.391E03 3.401E03
0.14 4.020E01 3.953E01 4.000E01 0.08 4.078E02 3.761E02 3.783E02
0.145 5.000E01 4.934E01 4.983E01 0.09 1.767E01 1.648E01 1.663E01
0.15 5.947E01 5.890E01 5.890E01 0.10 4.261E01 4.180E01 4.182E01
0.16 7.564E01 7.537E01 7.537E01 0.10267 5.000E01 4.869E01 4.873E01
0.18 9.350E01 9.346E01 9.346E01 0.11 6.869E01 6.698E01 6.698E01
0.20 9.870E01 9.869E01 9.869E01 0.12 8.637E01 8.683E01 8.677E01
0.22 9.978E01 9.978E01 9.978E01 0.14 9.842E01 9.866E01 9.865E01
0.24 9.997E01 9.997E01 9.997E01 0.16 9.987E01 9.986E01 9.986E01
0.26 9.999E01 9.999E01 9.999E01 0.18 9.999E01 9.999E01 9.999E01
Table 7.10: Numerical results for Weibull distributions,
, ,
0.1
r s r s
σ μ =
Limit analysis Shakedown analysis
/
s r
μ μ
f
P (FORM)
f
P (SORM) /
s r
μ μ
f
P (FORM)
f
P (SORM)
0.04 7.238E08 6.504E08 0.03 6.945E07 6.242E07
0.06 9.514E06 8.580E06 0.04 2.448E05 2.212E05
0.08 3.280E04 2.982E04 0.05 3.818E04 3.473E04
0.10 4.807E03 4.437E03 0.06 3.152E03 2.897E03
0.12 3.955E02 3.734E02 0.07 1.975E02 1.846E02
0.14 2.051E01 2.005E01 0.08 8.956E02 8.581E02
0.145 2.812E01 2.775E01 0.09 2.836E01 2.798E01
0.15 3.689E01 3.663E01 0.10 5.798E01 5.782E01
0.16 5.567E01 5.556E01 0.10267 6.538E01 6.508E01
0.18 8.367E01 8.326E01 0.11 8.120E01 8.074E01
0.20 9.476E01 9.449E01 0.12 9.236E01 9.200E01
0.22 9.827E01 9.819E01 0.14 9.873E01 9.870E01
0.24 9.937E01 9.945E01 0.16 9.974E01 9.990E01
0.26 9.977E01 9.993E01 0.18 9.994E01 9.999E01
122
2 2
0.10267
for shakedown analysis
0.01054
r s
HL
r s
μ μ
β
σ σ
−
=
+
(7.4b)
Numerical probabilities of failure for limit and shakedown analyses for normal
distribution are presented in table 7.9, compared with corresponding analytical solutions. Both
random variables have standard deviations
, ,
0.1
r s r s
σ μ = . The numerical error now comes
only from reliability analysis. It is shown that numerical results of FORM and SORM are very
close to the exact ones, especially for the case of limit analysis. Otherwise SORM gives
slightly better results compared with FORM.
Besides normal and lognormal distributions, the Weibull distribution is frequently
used for modelling loads and resistance. Table 7.10 shows the numerical results of
FORM/SORM for the case that both random variables have Weibull distribution. The Weibull
distribution leads to much larger failure probabilities in shakedown analysis but to much
smaller ones in limit analysis (figure 7.3). No analytical results are known. The limit state
function is nonlinear after the transformation into the u space. In this case SORM would
give improved results.
Figure 7.3 Probabilities of failure for normally and Weibull distributed variables
7.3 Limit analysis of cylindrical pipe under complex loading
Beside the loading and material strength, it is well known that the load carrying
capacity of shell structures is generally influenced by their initial imperfections which occurs
during the manufacturing and construction stages such as variability of thickness. In this
example, the effect of thickness imperfection on the limit loads of a shell structure is
examined. The cylindrical pipe subjected to complex loading which was investigated in
123
section 6.1 is considered here. The following geometrical and physical parameters are
adopted: 2700mm, 300mm, 120MPa
y
L r σ = = =
h
lim
/
y
. For this purpose, only the loading and the
thickness of the pipe are modelled as random variables. The following loading cases are
examined
Internal pressure
p h r For this case, the exact plastic collapse limit pressure is given by σ = . Thus,
the resistance R depends linearly on the realization of the thickness basic variable h Z . The
magnitude of the internal pressure is the second basic variable X . The limit state function is
defined by
( , )
y
g X Z Z
σ
X . (7.5) =
r
−
,
t s
If both thickness and load random variables are supposed to be normally distributed with
means μ ,
t s
and standard deviations μ σ σ ( ) g U
y
respectively, then the limit state function
in the standard Gaussian space is a linear function. Note that σ is the yield stress and not a
standard deviation here. The mean and standard deviation of the limit state function can be
calculated as follows
2
2
,
y y 2
g t s g t
r r
σ σ
s
μ μ μ σ σ σ
⎛ ⎞
= − = +
⎜ ⎟
⎝ ⎠
(7.6)
from which, the reliability index becomes
( )
( )
2
2 2
/
/
y t s g
HL
g
y t
r
r
s
σ μ μ μ
β
σ
σ σ σ
−
= =
+
. (7.7a)
The limit state function becomes nonlinear if both basic variables are lognormally
distributed. Analogously with (7.2), we obtain the exact reliability index for FORM
( ) ( )
( )
2 2
log / log
y t
HL
t s
r m m σ
β
δ δ
−
=
+
s
s s
(7.7b)
where and ,
t
m m ,
t
δ δ are calculated as in (7.3).
Bending moment
The exact plastic collapse limit moment is linearised by
lim 2
4
b y
M r h σ = (see (6.1)).
The limit state function is a linear function of basic variables X Z and defined by ,
2
( , ) 4
y
g X Z r Z X σ = − . (7.8)
124
The reliability indexes for both cases of normally and lognormally distributed random
variables are obtained, respectively
( )
2
2
2 2
4
4
y t s
HL
y t s
r
r
σ μ μ
β
2
σ σ σ
−
=
+
(7.9a)
( ) ( )
2
2 2
log 4 log
y t s
HL
t s
r m m σ
β
δ δ
−
=
+
. (7.9b)
Torsion moment
In this case the exact plastic collapse limit moment is given by
lim 2
2 / 3
t y
M r h π σ =
,
.
The limit state function is a linear function of basic variables X Z and defined by
2
2
( , )
3
y
g X Z r Z X π σ = − . (7.10)
The reliability indexes for both cases of normally and lognormally distributed random
variables are obtained, respectively
2
2
2 2
2
3
2
3
y t s
HL
y t
r
r
π σ μ μ
β
2
s
π σ σ σ
−
=
⎛ ⎞
+
⎜ ⎟
⎝ ⎠
(7.11a)
( )
2
2 2
2
log log
3
y t l
HL
t l
r m m π σ
β
δ δ
⎛ ⎞
−
⎜ ⎟
⎝ ⎠
=
+
. (7.11b)
Axial load
The exact plastic collapse limit load is given by
lim
2
y
F r h π σ = . The limit state
function is a linear function of the basic variables X Z and defined by ,
( , ) 2
y
g X Z r Z X π σ = − . (7.12)
The reliability indexes for both cases of normally and lognormally distributed random
variables are obtained, respectively
( )
2
2 2
2
2
y t s
HL
y t s
r
r
π σ μ μ
β
π σ σ σ
−
=
+
(7.13a)
( ) ( )
2 2
log 2 log
y t s
HL
t s
r m m π σ
β
δ δ
−
=
+
. (7.13b)
125
Table 7.11: Numerical results and comparison for pressure case
Normal distributions:
, ,
0.1
t s t s
σ μ = Lognormal distributions:
, ,
0.1
t s t s
σ μ =
/
s t
μ μ
3
6
10
m
kN
−
×
f
P (anal.)
f
P
(FORM)
f
P
(SORM)
/
s t
μ μ
3
6
m
−
10
kN
×
f
P (anal.)
f
P
(FORM)
f
P
(SORM)
0.10 1.718E13 1.720E13 1.724E13 0.14 4.964E14 3.160E14 3.179E14
0.15 2.428E09 1.935E09 1.935E09 0.15 1.791E12 1.175E12 1.175E12
0.20 3.872E06 3.651E06 3.700E06 0.20 4.473E07 3.277E07 3.274E07
0.25 7.364E04 6.117E04 6.117E04 0.25 4.316E04 3.522E04 3.524E04
0.30 2.275E02 1.976E02 2.017E02 0.30 2.071E02 1.844E02 1.849E02
0.35 1.734E01 1.580E01 1.582E01 0.35 1.719E01 1.596E01 1.596E01
0.40 5.000E01 4.740E01 4.741E01 0.40 5.000E01 4.752E01 4.753E01
0.45 7.969E01 7.787E01 7.792E01 0.45 7.981E01 7.798E01 7.799E01
0.50 9.408E01 9.332E01 9.333E01 0.50 9.432E01 9.356E01 9.356E01
0.55 9.863E01 9.840E01 9.840E01 0.55 9.880E01 9.859E01 9.860E01
0.60 9.972E01 9.967E01 9.967E01 0.60 9.980E01 9.975E01 9.975E01
0.65 9.995E01 9.994E01 9.994E01 0.65 9.997E01 9.996E01 9.996E01
0.70 9.999E01 9.999E01 9.999E01 0.70 9.999E01 9.999E01 9.999E01
Table 7.12: Numerical results and comparison for bending moment case
Normal distributions:
, ,
0.1
t s t s
σ μ = Lognormal distributions:
, ,
0.1
t s t s
σ μ =
/
s t
μ μ
5
1
10
kN
−
×
f
P (anal.)
f
P
(FORM)
f
P
(SORM)
/
s t
μ μ
5
1
10
kN
−
×
f
P (anal.)
f
P
(FORM)
f
P
(SORM)
0.15 3.488E10 3.360E10 3.378E10 0.15 3.231E14 3.030E14 3.075E14
0.20 5.484E07 5.284E07 5.324E07 0.20 2.393E08 2.257E08 2.279E08
0.25 1.330E04 1.284E04 1.292E04 0.25 5.282E05 5.039E05 5.076E05
0.30 6.041E03 5.859E03 5.886E03 0.30 4.871E03 4.704E03 4.729E03
0.35 7.013E02 6.849E02 6.869E02 0.35 6.783E02 6.629E02 6.649E02
0.40 2.934E01 2.888E01 2.893E01 0.40 2.927E01 2.891E01 2.895E01
0.432 5.000E01 4.945E01 4.949E01 0.432 5.000E01 4.946E01 4.950E01
0.45 6.135E01 6.083E01 6.079E01 0.45 6.139E01 6.085E01 6.081E01
0.50 8.483E01 8.449E01 8.447E01 0.50 8.500E01 8.466E01 8.463E01
0.55 9.542E01 9.528E01 9.528E01 0.55 9.565E01 9.552E01 9.551E01
0.60 9.885E01 9.880E01 9.880E01 0.60 9.901E01 9.897E01 9.896E01
0.65 9.974E01 9.973E01 9.973E01 0.65 9.981E01 9.980E01 9.980E01
0.70 9.994E01 9.994E01 9.994E01 0.70 9.997E01 9.997E01 9.997E01
0.75 9.999E01 9.999E01 9.999E01 0.75 9.999E01 9.999E01 9.999E01
126
Table 7.13: Numerical results and comparison for torsion moment case
Normal distributions:
, ,
0.1
t s t s
σ μ = Lognormal distributions:
, ,
0.1
t s t s
σ μ =
/
s t
μ μ
5
1
10
kN
−
×
f
P (anal.)
f
P
(FORM)
f
P
(SORM)
/
s t
μ μ
5
1
10
−
×
kN
f
P (anal.)
f
P
(FORM)
f
P
(SORM)
0.10 2.673E13 2.811E13 2.803E13 0.14 1.497E13 1.127E13 1.127E13
0.15 4.123E09 3.585E09 3.585E09 0.15 5.031E12 3.832E12 3.832E12
0.20 6.507E06 5.670E06 5.670E06 0.20 9.381E07 7.756E07 7.756E07
0.25 1.142E03 1.017E03 1.017E03 0.25 7.250E04 6.584E04 6.584E04
0.30 3.145E02 2.889E02 2.889E02 0.30 2.924E02 2.821E02 2.821E02
0.35 2.132E01 2.021E01 2.021E01 0.35 2.120E01 2.009E01 2.009E01
0.39178 5.000E01 4.845E01 4.845E01 0.39178 5.000E01 4.846E01 4.846E01
0.40 5.584E01 5.429E01 5.429E01 0.40 5.585E01 5.431E01 5.431E01
0.45 8.354E01 8.257E01 8.257E01 0.45 8.370E01 8.272E01 8.272E01
0.50 9.558E01 9.521E01 9.522E01 0.50 9.581E01 9.545E01 9.545E01
0.55 9.904E01 9.895E01 9.895E01 0.55 9.919E01 9.910E01 9.910E01
0.60 9.982E01 9.980E01 9.980E01 0.60 9.987E01 9.986E01 9.986E01
0.65 9.997E01 9.996E01 9.996E01 0.65 9.998E01 9.998E01 9.998E01
0.70 9.999E01 9.999E01 9.999E01 0.70 9.999E01 9.999E01 9.999E01
Table 7.14: Numerical results and comparison for tension case
Normal distributions:
, ,
0.1
t s t s
σ μ = Lognormal distributions:
, ,
0.1
t s t s
σ μ =
/
s t
μ μ
6
10
m
kN
−
×
f
P (anal.)
f
P
(FORM)
f
P
(SORM)
/
s t
μ μ
6
10
m
−
×
kN
f
P (anal.)
f
P
(FORM)
f
P
(SORM)
0.05 1.414E14 1.385E14 1.383E14 0.10 3.601E09 4.715E09 4.790E09
0.06 6.153E13 5.718E13 5.752E13 0.125 1.309E05 1.427E05 1.427E05
0.07 2.102E11 1.909E11 1.913E11 0.15 1.796E03 1.729E03 1.730E03
0.08 5.525E10 4.999E10 5.000E10 0.175 3.444E02 3.421E02 3.427E02
0.09 1.105E08 9.983E09 9.991E09 0.20 1.914E01 1.866E01 1.884E01
0.10 1.674E07 1.516E07 1.530E07 0.2262 5.000E01 4.890E01 4.869E01
0.15 2.496E03 2.312E03 2.313E03 0.25 7.609E01 7.522E01 7.522E01
0.20 1.928E01 1.854E01 1.850E01 0.275 9.169E01 9.126E01 9.126E01
0.2262 5.000E01 4.890E01 4.882E01 0.30 9.773E01 9.758E01 9.758E01
0.25 7.599E01 7.512E01 7.512E01 0.325 9.949E01 9.945E01 9.945E01
0.30 9.753E01 9.737E01 9.737E01 0.35 9.990E01 9.989E01 9.989E01
0.35 9.985E01 9.984E01 9.984E01 0.375 9.998E01 9.998E01 9.998E01
0.40 9.999E01 9.999E01 9.999E01 0.40 9.999E01 9.999E01 9.999E01
127
The probabilities of failure
f
P are presented in tables 7.117.14 for both distributions
and in figures 7.47.5 for lognormal distributed variables. Numerical solutions of the limit
analyses are compared with the analytical ones resulting from exact solutions. For each
loading case, both random variables are normally or lognormally distributed with standard
deviations
, ,
0.1
t s t s
σ μ = . It is shown that our results compare well with the analytical ones for
all cases.
Figure 7.4 Comparison of the results for lognormally distributed variables
Figure 7.5 Comparison of the results for lognormally distributed variables
128
7.4 Folding shell subjected to horizontal and vertical loads
In this example, a wellknown problem of having several failure modes is investigated.
Consider the frame formed of three plates of figure 7.6a which is generated by prolonging a
plane frame in the third direction. It is made of elasticperfectly plastic material and subjected
to two uniform horizontal and vertical loads . The loading and geometrical data were
selected to match those of the plane frame included in the book of Madsen et al. [1986].
Loads and limit plastic bending moment (material strength) are random variables which are
assumed mutually independent and lognormally distributed. There are three basic variables
and their mean values and standard deviations are given by
, H V
Horizontal load H 50 kNm, =15 kNm
H H
μ σ =
Vertical load V 40 kNm, =12 kNm
V V
μ σ =
Material strength
y
σ
2 2
3372.5 kN/m , =337.25 kN/m
y y
σ σ
μ σ =
p
The thickness of the shell is supposed to be constant . The plastic moment capacity =0.4 m h
M at each section is then a random variable with mean value and standard deviation as
follows
2
2
134.9 kNm
4
13.49 kNm
4
p y
p y
M
M
bh
bh
σ
σ
μ μ
σ σ
= =
= =
Hinge lines are thought to form at the end of elements (beam and columns) or at lines
of load application. As in the original plane frame, three failure modes caused by plastic hinge
mechanisms are expected to occur, those are sway mode, frame mode and beam mode (figure
7.7).
V
H
b
=
1
m
l=5 m l=5 m
l
=
5
m
Figure 7.6 Geometrical dimensions and FEmesh of the frame
129
H V H
V
H
V
Frame mode
Sway mode
Beam mode
Figure 7.7 Three failure modes of the frame
Numerical computation is carried out by using 300 quadrangular flat 4node shell
elements as shown in figure 7.6b. The ‘barriers’ technique developed by Der Kiureghian and
Dakessian is performed with 0.4 = 0.3 , γ δ = in order to find all the three design points as
expected. Our numerical results of design points are presented in table 7.15. The global design
point [ ]
*
1
3.083 0.944 0.024 0.329
T
= u
1
3.083
HL
with β = was found firstly. This design
point corresponds to the Sway mode since the effect of the horizontal load is dominant.
After adding a bulge at , the algorithm converges to the second design point
H
1
( ) B u
*
1
u
[ ]
*
2
3.24 0.776 0.47 0.422
T
= u
2
3.240
HL
with β = . This design point corresponds clearly to
the Frame mode because both of loads have big contributions to the failure of the structure.
We continuously added a bulge at and found the third design point
2
( ) B u
*
2
u
[ ]
*
3
3.461 0.457 0.783 0.421
T
= u
3
3.461
HL
with β = which corresponds to the Beam mode.
Now we suppose to proceed further and place a bulge at . Our search algorithm now
converges to
3
( ) B u
*
3
u
[ ]
*
4
3.307 0.925 0.325 0.194
T
=
4
3.307
HL
u with = . The distance β
* *
4 1
1.218 − = u u
1
0.4 3.083 1.233 r = × =
*
4
4
( ) B u
*
4
u
[
between the two design points is less than but close to the radius
of the bulge, thus confirming that u is a spurious design point. If we
further place a bulge at and continue the algorithm, the point
]
*
5
4.07 0.667 0.59 0.454
T
= u
5
4.070
HL
with β = is found. Obviously it is also a spurious
design point since the distances
* *
5 3
1.24 − = u u
*
3
u
*
5
u between the two design points , and
* *
5 2
1.028 − = u u
*
2
u
*
5
u
3
0.4 3.461 1.384 r = × =
2
0.4 3.24 1.296 r = × =
between the two design points , are less than the radius
and of the bulges. Thus, at this stage, we stop
130
to search and assume that there are only three design points for this problem. It should be
noted here that, the SQP algorithm worked well in this case to seek all the optimal points.
Conversely, the simple gradient search algorithm failed after adding the first bulge. It
oscillates around a foot of the bulge unstoppably.
Table 7.15: Multiple design points and search steps
Step
1
a
2
a
3
a
HL
β
Nature Mode
1 0.944
0.024
0.329 3.083 global design point Sway mode
2
0.776 0.470
0.422 3.240 local design point Frame mode
3
0.457 0.783
0.421 3.461 local design point Beam mode
4
0.925 0.325
0.194 3.307 spurious design point 
5
0.667 0.590
0.454 4.070 spurious design point 
The linear approximation of the failure set is now constructed by the tangent hyper
planes at the three design points. The corresponding approximations of failure mode
correlations and joint failure mode probabilities are listed in table 7.16. The single and system
failure probability and reliability indices are presented in table 7.17 and in table 7.18,
respectively, compared with those of the original plane frame obtained by Madsen et al. It is
shown that our two first failure probabilities of sway and frame modes are smaller than those
of plane frame while the last one compares well with the solution of Madsen et al. leading to a
smaller failure probability of the system. It is understandable since the stress state is now
threedimensional, not only the bending moment but also the compression force has an
influence on the loadcarrying capacity of the structure. By comparing the differences
between present results and those of Madsen et al. in the tables 7.17 and 7.18, one can easily
observe the sensitivity of the failure probability
f
P as discussed previously. It is shown that
about 10% difference in the reliability indices
HL
β leads to about 330% difference in the
failure probabilities
f
P . This sensitivity even increases very fast in the case of small
probabilities of failure.
Table 7.16: Failure mode correlations and joint failure mode probabilities (first order)
Failure mode correlations
ij
ρ
Joint failure probabilities
2
10
ij
P ×
1 2 3 1 2 3
1 1.0 0.883 0.589 0.1023 0.0301 0.0037
2 0.883 1.0 0.899 0.0301 0.0598 0.0155
3 0.589 0.899 1.0 0.0037 0.0155 0.0269
131
Table 7.17: Failure probability of the frame ( )
2
10
f
P ×
*
1
u alone (sway)
*
2
u alone (frame)
*
3
u alone (beam)
*
1
u , and
*
2
u
*
3
u
Method
FORM SORM FORM SORM FORM SORM FORM
Present 0.1023 0.1026 0.0598 0.110 0.0269 0.0166 0.140.155
Madsen et al. 0.336 0.322 0.199 0.267 0.0291 0.0283 0.467
Table 7.18: Reliability indices of the frame
HL
β
*
1
u alone (sway)
*
2
u alone (frame)
*
3
u alone (beam)
*
1
u , and
*
2
u
*
3
u
Method
FORM SORM FORM SORM FORM SORM FORM
Present 3.083 3.082 3.240 3.062 3.461 3.589 2.9582.989
Madsen et al. 2.710 2.725 2.880 2.786 3.440 3.447 2.600
132
8 SUMMARY
Integrity assessment of pressure vessels and piping by means of direct plasticity
methods has been a problem of great interest to many designers, especially in the design of
industrial and nuclear power plants. The new European pressure vessel standard EN
134453 is based on perfectly plastic limit and shakedown analysis (LISA) [European
standard, 200506] thus indicating the industrial need for LISA software for safe and
reliable design of such structures. Additionally, practical design codes often prescribe what
kind of values to choose for the safety factor of the resistance and of the loads for a given
problem since all resistance and loading variables are generally random. To this purpose,
the present thesis provides an effectively plastic analysis method for the integrity
assessment of general shells. Besides a deterministic LISA algorithm, structural reliability
analysis based on probabilistic LISA is also performed to establish a rational basis for the
choice of safety factors.
In the deterministic LISA procedure three failure modes of structures such as
plastic collapse, low cycle fatigue and ratchetting are analyzed based upon an upper bound
approach. The direct method involves the use of the exact Ilyushin yield surface and the
solution of the nonlinear constraint optimization problems is solved by Newton’s method
in conjunction with a penalty method and the Lagrange dual method. Line search is
performed to improve the current kinematic solution. A special “smooth regularization
method” was also used for overcoming the nondifferentiability of the objective function.
The actual Newton directions are updated at each iteration by solving a purelyelasticlike
system of linear equations which ensures automatically the kinematical condition of the
displacements.
The proposed method appears to be capable of identifying reasonable estimates of
the limit and shakedown load factors for a wide range of thin shell problems. It has been
tested against experiments and several load limits which have been calculated in literature
with different numerical methods using shell or volume elements. The load limits are
obtained here together with tests of mesh convergence and robustness. Some of them could
be checked against twice the elastic limit because they appear to represent the alternating
plasticity limit under the considered loading conditions. A numerically very effective
method is achieved from the lesser computational cost by using shell elements compared
with volume elements and by direct plasticity methods which achieve plastic solutions in
the computing time of only several linear elastic steps. Parametric studies carried out in
some examples show that different choices of optimization parameters have quite different
effects on the behaviour of the algorithm.
In the probabilistic LISA, the loading and strength of the material as well as the
thickness of the shell are considered as random variables. The procedure involves a
deterministic limit and shakedown analysis for each probabilistic iteration. Different kinds
of distribution of basic variables are taken into consideration and performed with
FORM/SORM for the calculation of the failure probability of the structure. A nonlinear
optimization was implemented, which is based on the Sequential Quadratic Programming
for finding the design point. Nonlinear sensitivity analyses are also performed for
computing the Jacobian and the Hessian of the limit state function.
The advantage of the method in structural reliability analysis is that the failure
under cyclic loading is treated as a timeinvariant problem and it is applicable with
incomplete data. On the other hand, sensitivity analyses are obtained directly from a
mathematical optimization with no extra computational cost. Numerical examples were
tested against literature with analytical methods and with a numerical method using
volume elements. It is shown that the proposed method is capable of identifying good
estimates of the failure probability, even in the case of very small probabilities. The results
are achieved just after several deterministic steps even if the starting points are chosen far
from the actual design point. If the limit state function is linear with the limit and
shakedown analysis, then the use of FORM is sufficient and SORM is only necessary if the
linearity is lost after the transformation into the u space.
Practical experience showed that the existence of multimode failure (multiple
design points) in component reliability analysis could give rise to large errors in
FORM/SORM approximations of the failure probability. In this thesis, a technique which
was developed by [Der Kiureghian and Dakessian, 1998] has been performed to
successively find the multiple design points of a reliability problem, when they exist on the
limit state surface. This technique is based upon a ‘barrier’ method by constructing a bulge
around previous found design points, thus forcing the algorithm to seek a new design point.
Secondorder bounds of the reliability of series system are then calculated based on the
firstorder system reliability analysis. Numerical example of a frame made of plates
showed that it improves considerably the estimates of the system failure probability.
To the author’s knowledge the above mentioned results contain the following new
contributions which have never been published elsewhere:
• The development of a kinematic algorithm of deterministic limit and shakedown
analysis for general plate and shell structures using the exact Ilyushin yield surface.
The exact Ilyushin yield surface has not been used so far in LISA.
• The development of a kinematic algorithm of probabilistic limit and shakedown
analysis for general plate and shell structures using the First and Second Order
Reliability Methods. The present thesis is the first one dealing with reliabilitybased
LISA for shells and using SORM.
134
135
• The introduction and formulation of the probabilistic LISA in which the thickness
is considered as a random variable. Analytical solutions for a simple problem were
derived in order to compare with the numerical solutions.
• The sensitivity analysis of the limit state function versus load, material strength and
shell thickness variables.
• The introduction and formulation of the reliability problem for the structural
systems (series systems). The implementation of a special technique to successively
find all the multiple design points on the limit state surface allows to calculate the
secondorder bounds of the reliability for series systems.
It should be noted that in the present work we accepted some simplified hypotheses,
such as perfectly plastic material, small deformation and displacement models. The
perfectly plastic material is sufficient with respect to the design codes. However, the
present results may be safe to be applied in some practical engineering since some practical
materials often exhibit strain hardening. In this meaning limit and shakedown analysis
based on the kinematic hardening may improve the present design procedures. On the other
hand, the geometrically nonlinear effects may become important to very thinwalled
structures. On this basis, we would propose here some possibilities for future work:
• Extend the present work for both deterministic and probabilistic LISA to the real
materials in which we take into account the bounded kinematic hardening of the
material.
• Extend the present work for both deterministic and probabilistic LISA to the real
behaviour of thin shell structures in which we consider the large deformation and
large displacement hypotheses.
• Perform the deterministic and probabilistic LISA by the application of the primal
dual theory, in which both lower and upper bounds are obtained at each iteration.
• Extend the present work for probabilistic LISA to the real shell structures in which
the thickness of the shell is considered a random field, i.e. the random thickness is a
function in space. In a similar way temperature fields and temperature dependent
material strength can be considered.
REFERENCES
Abdalla H. F., Megahed M. M. and Younan M. Y. A. (2006). Determination of Shakedown Limit Load for a
90Degree Pipe Bend Using a Simplified Technique. ASME J. Pressure Vessel Technol., 128 (4), pp. 618
624.
Andersen K. D. and Christiansen E. (1995). A Newton barrier method for minimizing a sum of Euclidean
norms subject to linear equality constraints. IMADA preprint PP199507. Department of Mathematics and
Computer Science at SDU, Odense University, Denmark.
Andersen K. D. and Christiansen E. (1998a). Minimizing a sum of norms subject to linear equality
constraints. Comput. Optim. Appl., 11, pp. 6579.
Andersen K. D., Christiansen E. M. and Overton, L. (1998b). Computing limit loads by minimizing a sum of
norms. SIAM J. Sci. Comput. 19, pp. 10461062.
Andersen K. D., Christiansen E., Conn A. R., Overton M. L. (2000). An efficient primaldual interiorpoint
method for minimizing a sum of Euclidean norms. SIAM J. Sci. Comput., 22, pp. 243262.
Augusti G., Baratta A., Casciati F. (1984). Probabilistic Method in Structural Engineering. Chapman and
Hall, London.
Bauer H., Saal H. (1997). Nozzles in the knuckle region of pressure vessels heads – strength and stiffness
under various loading conditions. In Proc. Int. Conf. Carrying capacity of steel shell structures, Brno, pp.
138144.
Bazaraa M. S., Sherali H. D., Shetty C. M. (1993). Nonlinear programming Theory and Algorithms. John
Wiley & Sons, Inc., New York.
Bisbos C. D. and Papaioannou G. (2006). Shakedown analysis of FEMdiscretized steel shell structures
under the Ilyushin yield criterion. In Baniotopoulos C. C. (ed.): II. Int. Conference on Nonsmooth
Nonconvex Mechanics with Applications in Engineering (II. NNMAE), Thessaloniki, Juli 78, pp. 197204.
Bisbos C. D., Pardalos P. M. (2007). Secondorder cone and semidefinite representations of material failure
criteria. Journal of Optimization Theory and Applications, 134 (2), pp. 275301.
http://dx.doi.org/10.1007/s1095700792438
Bjerager P. (1989). Plastic systems reliability by LP and FORM. Computers and Structures, 31, pp. 187196.
Bjerager P. (1991). Methods for structural reliability computations. In Course on Reliability Problems:
General Principles and Applications of Solids and Structures, International Center for Mechanical Sciences,
Udine, Italy, 1990. CISM lecture notes, Springer, Wien.
Bleich H. (1932). Über die Bemessung statisch unbestimmter Stahlwerke unter der Berücksichtigung des
elastischplastischen Verhaltens des Baustoffes. Bauingenieur, 13, pp. 261267.
Bolt S. E. and Greenstreet W. L. (1971). Experimental determination of plastic collapse loads for pipe
elbows. ASME paper 71PVP37.
Borkowski A. and Kleiber M. (1980). On a numerical approach to shakedown analysis of structures.
Computer methods in applied mechanics and engineering, 22, pp. 110119.
Bree J (1967). Elasticplastic behaviour of thin tubes subjected to internal pressure and intermittent highheat
fluxes with applications to fastnuclearreactor fuel elements. Journal of Strain Analysis, 2, pp. 226238.
Bridgman P. W. (1923), The compressibility of thirty metals as a function of pressure and temperature. Proc.
Am. Acad. Arts Sci., 58, pp. 163242.
Bridgman P. W. (1952). Studies in Large Plastic Flow and Fracture. McGrawHill, New York.
Burgoyne C. J. and Brennan M. G. (1993a). Calculation of elastoplastic rigidities using the exact Ilyushin
yield surface. Int. J. Sol. Struct., 30 (8), pp. 11331145.
Burgoyne C. J. and Brennan M. G. (1993b). Exact Ilyushin yield surface. Int. J. Sol. Struct., 30 (8), pp. 1113
1131.
Calladine C. R. (1974). Limit analysis of curved tubes. J. Mech. Eng. Science, 16, pp. 8587.
Capsoni A. and Corradi L. (1997). A finite element formulation of the rigidplastic limit analysis problem.
Int. J. Num. Meth. Eng., 40, pp. 20632086.
Carvelli V., Cen Z. Z., Liu Y., Maier G. (1999). Shakedown analysis of defective pressure vessels by a
kinematic approach. Archive of Applied Mechanics, 69, pp. 751764.
Casciaro R. and Cascini L. (1982). A mixed formulation and mixed finite elements for limit analysis. Int. J.
Num. Meth. in Eng., 18, pp. 211243.
Chapelle D., Bathe K. J. (2003). The finite element analysis of shells – fundamentals. Berlin: Springer.
Chattopadhyay J., Venkatramana W., Dutta B. K. and Kushwaha H.S. (1999). Limit Load of Elbows Under
Combined Internal Pressure and Bending Moment. Proceedings of the 15th International Conference on
Structural Mechanics in Reactor Technology, SMiRT, Vol. V, Korea, pp. 281288.
Chattopadhyay J., Nathani D. K., Dutta B. K. and Kushwaha H. S. (2000). ClosedForm collapse moment
equations of Elbows under combined internal pressure and inplane bending moment. ASME J. Pressure
Vessel Technol., 122, pp. 431436.
Clark R. A. and Reissner E. (1951). Bending of curved tubes. Advances in Applied Mechanics, 2, pp. 93122.
Cocks A. C. F. and Leckie F. A. (1988). Deformation bounds for cyclically loaded shell structures operating
under creep condition. Journal of Applied Mechanics, Transactions of the ASME, 55, pp. 509516.
Code_Aster 7.3 (2003). User’s manual. A product of EDF R&D, Clamart, France. http://www.codeaster.org
Cornell C. A. (1967). Bounds on the Reliability of structural systems. Journal of the Structural Division,
ASCE, 93, ST1, Feb., pp. 171200.
Crisfield M. A. (1973). Large deflection elastoplastic buckling analysis of plates using finite elements.
Transport and Road Research Lab. Report 51UC.
Der Kiureghian A., Dakessian T. (1998). Multiple design points in first and secondorder reliability.
Structural Safety, 20, pp. 3749.
Desquines J., Plancq D. and Wielgosz C. (1997). Inplane limit moment for an elbow lower bound analytical
solution and finite element processing by elastic compensation method. Int. J. Pres. Ves. & Piping, 71,
pp. 2934.
137
Diem H. (1994). Untersuchungen zum Geometrieeinfluß auf das Verformungs und Versagensverhalten von
Rohrbogen. Dissertation, Universität Stuttgart.
Ditlevsen O. (1979). Narrow reliability bounds for structural systems. Journal of Structural Mechanics, 7 (4),
pp. 453472.
Ditlevsen O. and Bjerager P. (1986). Methods of structural system reliability. Structural Safety, 3, pp.
195229.
Dronia S. (2005). Sequentielle Quadratische Programmierung für effektive Zuverlässigkeitsanalyse mit der
Finite Element Methode. Diplomarbeit, Labor für Biomechanik, Fachhochschule Aachen Campus Jülich,
Germany.
Drubay B., et al. (1995). A16: Guide for Defect Assessment and LeakBeforeBreak Analysis. Third Draft,
Commissariat A L’energie Atomique, Rapport DMT 96.096, France.
Dunny C. W., Sobel M. (1955). Approximations to the probability integral and certain percentage points of
multivariate analogue of Student’s tdistribution. Biometrica, 42, pp. 258260.
European Standard (200506). Unfired pressure vessels  Part 3: Design, Annex 2: Annex B Direct route for
design by analysis and Annex C Stress categorisation route for design by analysis. CEN European
Committee for Standardization, EN 134453:2002, Issue 14.
Fanous I. F. Z., AdibiAsl R. and Seshadri R. (2005). Limit load analysis of pipe bends using the Rnode
method. ASME J. Pressure Vessel Technol., 127, pp. 443448.
Fiacco A. (1983). Introduction to Sensitivity and Stability Analysis in Nonlinear Programming. Academic
Press, New York.
Fiessler B., Neumann H. J., Rackwitz R. (1979). Quadratic Limit States in Structural Reliability Theory.
ASCE, Journal of Engineering Mechanics, 105, pp. 661676.
Franco J. R. Q., Ponter A. R. S. (1997.a). A general approximate technique for the finite element shakedown
and limit analysis of axisymmetrical shells. Part 1: Theory and fundamental relations. Int. J. for Num. Meth.
in Engng., 40 (19) pp. 34953513.
Franco J. R. Q., Ponter A. R. S. (1997.b). A general approximate technique for the finite element shakedown
and limit analysis of axisymmetrical shells. Part 2: Numerical applications. Int. J. for Num. Meth. in Engng.,
40 (19) pp. 35153536.
Frieze P. A. (1975). Ultimate load behaviour of steel box girders and their components. PhD Thesis,
University of London.
Gaydon F. A., McCrum A. W. (1954). A theoretical investigation of the yield point loading of a square plate
with a central circular hole. Journal of the Mechanics and Physics of Solids, 2, pp. 156169.
GiD 7.2 (2002). User’s manual. A product of Center for Numerical Methods in Engineering (CIMNE),
Barcelona. http://www.gid.cimne.upc.es
Gollwitzer S., Abdo T., Rackwitz R. (1988). FORM Manual. RCP GmbH, München.
Gokhfeld D. A. and Cherniavsky O. F. (1980). Limit analysis of structures at thermal cycling. Sijthoff &
Noordhoff, The Netherlands.
Goodall I. M. (1978a). Large deformations in plastically deformed curved tubes subjected to inplane
bending. Research Division Report CEGBRD/B/N4312, Central electricity generating board, UK.
138
Goodall I. W. (1978b). Lower Bound Limit Analysis of Curved Tubes Loaded by Combined Internal Pressure
and Inplane Bending Moment. Research Division Report RD/B/N4360, Central electricity generating board,
UK.
Gorman M. R. (1981). Automatic generation of collapse mode equations. J. Struct. Div., ASCE, Vol. 107,
No. ST7, July, pp. 13501354.
Griffiths J. E. (1979). The effect of cracks on the limit load of pipe bends under inplane bending
experimental study. Int. J. Sci., 21, pp. 119130.
Grimmelt M. J., Schuëller G. I. (1982). Benchmark Study on methods to determine collapse failure
probabilities of redundant structures. Structural Safety, 1, pp. 93106.
Hasofer A. M., Lind N. C. (1974). An exact and Invariant First Order Reliability Format. ASCE, Journal of
Engineering Mechanics Division, Vol. 100, No. EM1, pp. 111121.
Heitzer M. (1999). Traglast und Einspielanalyse zur Bewertung der Sicherheit passiver Komponenten.
Berichte des Forschungszentrums Jülich, Jül3704, Dr.Ing. Thesis, RWTH Aachen.
Heitzer M. and Staat M. (1999). FEMcomputation of load carrying capacity of highly loaded passive
components by direct methods. Nuclear Engineering and Design, 193 (3), pp. 349358.
http://dx.doi.org/10.1016/S00295493(99)001909
Heitzer M., Staat M. (2000). Reliability Analysis of ElastoPlastic Structures under Variable Loads. In:
Inelastic Analysis of Structures under Variable Loads: Theory and Engineering Applications (edited by
Weichert D., Maier G.) Kluwer, Academic Press, Dordrecht, pp. 269288.
Heitzer M., Staat M. (2002). Limit and Shakedown Analysis with Uncertain Data. In: Stochastic
Optimization Techniques, Numerical Methods and Technical Applications (edited by Marti K.), Lecture
Notes in Economics and Mathematical Systems, Vol. 513, Springer Heidelberg, pp. 253267.
Hilsenkopf P. et al. (1988). Experimental Study of Behavior and Functional Capability of Ferritic Steel
Elbows and Austenitic Stainless Steel ThinWalled Elbows. Int. J. Pres. Vessels and Piping, 33 (2),
pp. 111128.
Hodge P. G. Jr. (1954). The Rigidplastic analysis of symmetrically loaded cylindrical shells. J. App. Mech.,
21, pp. 336442.
Hodge P. G. Jr. (1959). Plastic analysis of structures. McGrawHill, New York.
Hodge P. G. Jr. (1961). The Mises yield condition for rotationally symmetric shells. Quarterly of Applied
Mathematics, Vol. XVIII, No. 4, 305311.
Hodge P. G. Jr. (1963). Limit analysis of rotationally symmetric plates and shells. Prentice Hall, Englewood,
New Jersey.
Hohenbichler M. (1981). Approximate evaluation of the multinormal distribution function. Berichte zur
Zuverlässigkeitstheorie der Bauwerke, Tech. Univ. München, Heft 58.
Hohenbichler M., Rackwitz R. (1983). Firstorder concepts in system reliability. Structural Safety, 1, pp.
177188.
Hohenbichler M., Gollwitzer S., Kruse W., Rackwitz R. (1987). New light on first and secondorder
reliability methods. Structural Safety, 4, pp. 267284.
139
Hsieh M. F., Moffat D. G., Mistry J. (2000). Nozzles in the knuckle region of a torispherical head: limit load
interaction under combined pressure and piping loads. Int. J. Pres. Vessels and Piping, 77, pp. 807815.
Ilyushin A. A. (1948). Plasticity. Gostekhizdat, Moscow. In Russian.
Ivanov G. V. (1967). Approximating the final relationship between the forces and moments of shell under the
Mises plasticity condition. Inzhenernyi Zhurnal Mekhanica Tverdogo Tela, 6, pp. 7475. In Russian.
Jospin R. J. (1992). Etats limites des tuyauteries par la méthode des éléments finis et la programmation
mathématique. Thèse de doctorat, Université de Liège.
Kalisky S. (1985). Plasticity, Theory and Engineering Applications. Elsevier, Amsterdam.
Klingmüller O. (1979). Anwendung der Traglastberechnung für die Beurteilung der Sicherheit von
Konstruktionen. Thesis Forschungsberichte aus dem Fachbereich Bauwesen 9, Universität Essen
Gesamthochschule.
Koiter W. T. (1960). General theorems for elastic plastic solids. In: Progress in Solid Mechanics (edited by
Sneddon I. N. and Hill R.), pp. 165221, NordHolland, Amsterdam.
König J. A. (1966). Theory of shakedown of elasticplastic structures. Arch. Mech. Stos., 18, pp. 227238.
König J. A. (1969). A shakedown theorem for temperature dependent elastic moduli. Bull. Acad. Pol. Sci.,
Sér. Sci. Techn., Vol. 17, pp. 161165.
König J. A. (1972). On shakedown criterion. Proc. Int. Symp. Foundations of Plasticity, Vol. 2, pp. 5660,
Warsaw.
König J. A., Kleiber M. (1978). On a new method of shakedown analysis. Bull. Acad. Pol. Sci., Sér. Sci.
Techn., 4, pp. 165171.
König J. A. (1987). Shakedown of elasticplastic structures. Elsevier, New York.
Kussmaul K., Diem H. K., Uhlmann D. and Kobes H. (1995). Pipe Bend Behaviour at Load Levels Beyond
Design. Proceedings of 13th International Conference on Structural Mechanics in Reactor Technology,
SMiRT. Brazil. Vol. G., pp. 187198.
Larson L.D., Stokey W.F. and Franzen W.E. (1975). An approximate Model for an elasticplastic pipe
element under combined loading. Journal of Pressure Vessel Technology, Transaction of the ASME, 97,
pp. 2228.
Little G. H. (1977). Rapid analysis of plate collapse by live energy minimization. Int. J. Mech. Sci., 19,
pp. 725744.
Liu Y. H., Carvelli V., Maier G. (1997). Integrity assessment of defective pressurized pipelines by direct
simplified methods. Int. J. Pres. Ves. & Piping, 74, pp. 4957.
Locci J. M. (1995). Automatization of limit analysis calculations, application to structural reliability
problems. In: Proceedings of the ICASP7 conference, Paris (edited by Lemaire M., Favre J. L., Mébarki A.),
Vol. 2: Applications of Statistics and Probability. A. A. Balkema Rotterdam, Broookfield, pp. 10951110.
Lubliner J. (2005). Plasticity Theory. Revised edition, University of California at Berkeley.
Madsen H.O., Krenk S., Lind N.C. (1986). Methods of Structural Safety. PrenticeHall, Englewood Cliffs.
Maier G. (1969). Shakedown theory in perfect elastoplasticity with associated flowlaws. Meccanica, 4,
pp. 250260.
140
Maier G. (1970). A matrix structural theory of picewiselinear plasticity with interacting yield planes.
Meccanica, 7, pp. 5166.
Maier G. (1973). A shakedown matrix theory allowing for work hardening and secondorder geometric
effects. In: Foundations of plasticity (edited by Sawczuk A.), NorthHolland, Amsterdam, pp. 417433.
Martin J. B. (1975). Plasticity. MIT Press, Cambridge.
Melan E. (1936). Theorie statisch unbestimmter Systeme aus ideal plastischem Baustoff. Sitzber. Akad. Wiss.
Wien IIa 145, pp. 195218.
Melchers R. E. (1987). Structural Reliability – Analysis and Prediction. Ellis Horwood limited, Chichester.
Morelle P. (1984). Structural shakedown analysis by dual finiteelement formulations. Eng. Struct., 6 (1),
pp. 7079.
Morelle P. (1989). Analyse duale de l'adaptation plastique des structures par la méthode des éléments finis et
la programmation mathématique. Thèse de Doctorat, Université de Liège, Belgique.
Mourad H. M. and Younan M. Y. A. (2002). Limit load analysis of pipe bends under outofplane moment
loading and internal pressure. ASME J. Pressure Vessel Technol., 124, pp. 32–37.
Moxham K. E. (1971). Theoretical prediction of the strength of welded steel plates in compression.
Cambridge University Report No CUED/CStruct/TR2.
Nguyen Dang H., König J. A. (1976). A finite element formulation for shakedown problems using a yield
criterion of the mean. Comp. Meth. Appl. Mech. Eng., 8, 179192.
Nguyen Dang H., Morelle P. (1990). Plastic shakedown analysis. In: Mathematical programming methods in
structural plasticity. (edited by Smith, D. L.), pp. 183205.
Nguyen Dang H., Tran T. N. (2004). Analysis of cracked plates and shells using “metis” finite element
model. Finite Element in Analysis and Design, Vol. 40 (8), pp. 855878. http://dx.doi.org/10.1016/S0168
874X(03)001173.
Panagiotopoulos P. D. (1985). Inequality problems in Mechanics and Applications, Convex and Nonconvex
Energy Functions. Birkhäuser, Boston.
Polizzotto C., Borino G., Caddemi S., Fuschi P. (1991). Shakedown problems for material with internal
variables. Eur. J. Mech. A/Solids, 10, (6), pp. 621639.
Polizzotto C. (1993.a). On the condition to prevent plastic shakedown of structures: Part I – Theory. J. Appl.
Mech. ASME, 60, pp. 1519.
Polizzotto C. (1993.b). On the condition to prevent plastic shakedown of structures: Part II – The plastic
shakedown limit load. J. Appl. Mech. ASME, 60, pp. 2025.
Ponter A. R. S., Carter K. F. (1997.a). Shakedown state simulation techniques based on linear elastic
solution. Comput. Methods Appl. Mech. Engrg., 140, pp. 259279.
Ponter A. R. S., Carter K. F. (1997.b). The calculation of limit load and shakedown boundaries by effective
modulus method. Transactions of the 14
th
International Conference on Structural Mechanics in Reactor
Technology (SmiRT 14), Lyon, France, August 1722.
Ponter A. R. S., Fuschi P., Engelhardt M. (2000). Shakedown limits for a general yield condition:
implementation and application for a Von Mises yield condition. European Journal of Mechanics  A/Solids,
19 (3), pp. 423445.
141
Powell M. J. D. (1978). Algorithms for Nonlinear Constraints that Use Lagrangian Functions. Mathematical
Programming, 14, pp. 224248.
Prager W., Hodge P.G.Jr. (1951). Theory of perfectly plastic solids. Wiley, New York.
Prager W. (1959). Probleme der Plastizitätstheorie. Birkhäuser, Basel 1955; English edition: Introduction to
plasticity, AddisonWesley, Boston.
Prager W. (1972). Limit analysis: The development of a concept. In: Foundation of plasticity 2 (edited by
Sawczuk A.), Noordhoff, Groningen.
Preiss R. (2000). Ratcheting and shaekdown analysis of pressure equipment using elastoplastic Finite
ElementAnalysis. PhD Thesis, TU Wien.
Rackwitz R. (1978). Close bounds for the reliability of structural systems. Berichte zur
Zuverlässigkeitstheorie der Bauwerke, Tech. Univ. München, Heft 29.
Rackwitz R. (19932004). Zuverlässigkeit und Lasten im konstruktiven Ingenieurbau. Lecture Notes,
Technische Universität München.
Rahman S.M. (2006). Finite element analysis and related numerical schemes for ratchetting simulation.
Dissertation, North Carolina State University, Raleigh.
Ramachandran K. (1984). Systems Bounds: A critical study. Civil Engineering Systems, 1, March,
pp. 123128.
Rashedi M. R., Moses F. (1983). Studies on reliability of structural systems. Report No. R833, Department
of Civil Engineering, Case Western Reserve University, Cleveland, Ohio.
Robinson M. (1971). A comparison of yield surfaces for thin shells. Int. J. Mech. Sci., 13, pp. 345354.
Sawczuk A. (1969a). On incremental collapse of shells under cyclic loading. In: Theory of the Thin Shells
(edited by Niordson F. I.), SpringerVerlag, Berlin, pp. 328340.
Sawczuk A. (1969b). Evaluation of upper bounds to shakedown loads for shells. J. Mech. Phys. Solids, 17,
pp. 291301.
Schittkowski K. (1983a). Theory, implementation and test of a nonlinear programming algorithm. In:
Optimization Methods in Structural Design (edited by Eschenauer H., Olho N.), BI Wissenschaftsverlag,
Mannheim.
Schittkowski K. (1983b). On the convergence of a sequential quadratic programming method with an
augmented Lagrangian search direction. Optimization, 14, pp. 197216.
Schittkowski K. (1988). Solving nonlinear least squares problems by a general purpose SQPmethod. In:
Trends in Mathematical Optimization (edited by Hoffmann K.H., HiriartUrruty J.B., Lemarechal C., Zowe
J.), International Series of Numerical Mathematics, Vol. 84, Birkhäuser, pp. 295309.
Schittkowski K. (2004). NLPQLP20: A Fortran implementation of a sequential quadratic programming
algorithm with distributed and nonmonotone line search  user’s guide. Report, Department of Computer
Science, University of Bayreuth.
Seitzberger M. (2000). Contributions to an Efficient Numerical Analysis of the Plastic Behaviour of Thin
Walled Structures. FortschrittBerichte VDI, Reihe 18, Nr. 247, VDIVerlag, Düsseldorf.
Shalaby M. A. and Younan M. Y. A. (1998a). Limit loads for pipe elbows with internal pressure under in
plane closing bending moment. ASME J. Pressure Vessel Technol., 120, pp. 3542.
142
Shalaby M. A. and Younan M. Y. A. (1998b). Limit Loads for Pipe Elbows Subjected to Inplane Opening
Moment and Internal Pressure. Paper Presented at the 1998 ASME/JSME Joint Pressure Vessels and Piping
Conference held at San Diego, California, July 2630, PVP Vol. 368, 1998, pp. 163170.
Siemaszko A., Bielawski G., Knabel J. (2001). Shakedown and limit reliabilitybased design. Transactions
SMiRT16 Structural Mechanics in Reactor Technology, Washington, DC USA, August 1217.
Spence J. and Findlay G. E. (1973). Limit loads for pipe bends under inplane bending. Proc. 2nd Int. Conf.
on Pressure Vessel Technology, ASME Vol. 1, pp. 393399.
Staat M., Heitzer M. (1997). Limit and shakedown analysis for plastic safety of complex structures.
Transactions of the 14
th
International Conference on Structural Mechanics in Reactor Technology (SMiRT
14), Lyon, France, August 1722. http://www.iasmirt.org/SMiRT14/B022
Staat M., Heitzer M. (2001). LISA – a European project for FEMbased limit and shakedown analysis.
Nuclear Engineering and Design, 206 (23), pp. 151166. http://dx.doi.org/10.1016/S00295493(00)004155
Staat M. (2002). Some Achievements of the European Project LISA for FEM Based Limit and Shakedown
Analysis. In N. Badie (ed.) Computational Mechanics: Developments and Applications  2002. ASME PVP
Vol. 441, Paper PVP20021300, pp.177185.
Staat M., Heitzer M. (2002). The restricted influence of kinematic hardening on shakedown loads.
Proceedings of WCCM V, 5th World Congress on Computational Mechanics, Vienna, Austria, July 712.
http://opus.bibliothek.fhaachen.de/opus/volltexte/2005/79/
Staat M. and Heitzer M. (ed.) (2003a). Numerical Methods for Limit and Shakedown analysis – Deterministic
and Probabilistic Problems. NIC Series Vol. 15, John von Neumann Institute for Computing, Jülich.
http://www.fzjuelich.de/nicseries/volume15/nicseriesband15.pdf
Staat M. and Heitzer M. (2003b). Probabilistic limit and shakedown problems. In [Staat and Heitzer (2003a)].
Staat M., Heitzer M., Lang H., Wirtz K. (2005). Direct Finite Element Route for DesignbyAnalysis of
Pressure Components. Int. J. Pressure Vessels Piping, 82 (1), pp. 6167.
http://dx.doi.org/10.1016/j.ijpvp.2004.04.006
Taylor N. et al. (1999). The designbyanalysis manual. Report EUR 19020 EN, European Commission, Joint
Research Centre, Petten, The Netherlands. For error corrections and remarks see:
http://info.tuwien.ac.at/IAA/news/dba1_engl.htm
ThoftChristensen P. and Baker M. J. (1982). Structural reliability theory and its applications. Springer
Verlag, Berlin.
Touboul F., et al. (1989). Design Criteria for Piping Components Against Plastic Collapse: Application to
Pipe Bend Experiments. Proceedings of 6
th
International Conference of Pressure Vessel Technology, Beijing,
China, September 1115, eds., Cengdian, Liu, and Nichols, pp. 7384.
Tran T. N., Staat M., Kreißig R. (2007a). Calculation of load carrying capacity of shell structures with
elastoplastic material by direct methods. Proceedings of the International Conference on Material Theory
and Nonlinear Dynamics, Hanoi, Vietnam, September 2426. http://opus.bibliothek.fh
aachen.de/opus/volltexte/2007/227/ .
Tran T. N., Staat M., Kreißig R. (2007b). Finite element shakedown and limit reliability analysis of thin
shells. Proceedings of 9
th
International conference on computational plasticity, COMPLAS IX, Barcelona,
Spain, September 0507, eds., Oñate E. and Owen D. R. J.
http://opus.bibliothek.fhaachen.de/opus/volltexte/2007/233/
143
Tran T. N., Kreißig R., Vu D. K., Staat M. (2007c). Upper bound limit and shakedown analysis of thin shells
using the exact Ilyushin yield surface. Computer & Structures, in press.
http://dx.doi.org/10.1016/j.compstruc.2008.02.005
Tran T. N., Kreißig R., Staat M. (2007d). Probabilistic limit and shakedown analysis of thin plates and shells.
Structural Safety, in press. http://dx.doi.org/10.1016/j.strusafe.2007.10.003.
Tran T. N., Kreißig R., Staat M. (2007e). Load bearing capacity of thin shell structures with elastoplastic
material by direct methods. Technische Mechanik, in press.
Tran T. N., Pham P. T., Vu D. K., Staat M. (2008). Reliability analysis of inelastic shell structures under
variable loads. In: A.R.S. Ponter, D. Weichert, (eds.) Limit States of Materials and Structures: Direct
Methods. Springer, Berlin, to appear.
Tvedt L. (1983). Two Secondorder Approximations to the Failure Probability. Veritas Report RDIV/20
00483, Det Norske Veritas, Oslo, Norway.
Vu D. K. (2001). Dual Limit and Shakedown analysis of structures. PhD Thesis, Université de Liège,
Belgium.
Vu D. K,. Yan A. M., Nguyen Dang H. (2004). A primaldual algorithm for shakedown analysis of structure.
Comp. Meth. Appl. Mech. Engng., 193, pp. 46634674.
Vu D. K., Staat M., Tran I. T. (2007). Analysis of pressure equipment by application of the primaldual
theory of shakedown. Communications in Num. Meth. in Engng., 23 (3), pp. 213225.
http://dx.doi.org/10.1002/cnm.891
Wang, Tsau Yueh, Corotis R. B., Ellingwood B. R. (1990). Limit state sensitivity of structural frames
subjected to cyclic forces. J. Struct. Engng., ASCE, 116 (10), pp. 28242841.
Weichert D., GrossWeege J. (1988). The numerical assessment of elasticplastic sheets under variable
mechanical and thermal loads using a simplified towsurface yield condition. Int. J. Mech. Sci., 30, pp. 757
767.
Weichert D., Hachemi A. (2002). Shakedown of thinwalled structures with geometrical nonlinear effects.
In: Inelastic behaviour of structures under variable repeated loads: Direct analysis methods (edited by
Weichert D., Maier G.), Springer, Wien NewYork, pp. 229238.
Wierskowski C. (2005). Probabilistische Einspielanalyse mit der Finite Elemente Methode und
FORM/SORM. Diplomarbeit, Labor für Biomechanik, Fachhochschule Aachen Campus Jülich, Germany.
Yan A. M. (1996). Limit analysis of general pipe elbow subjected to complex loading. LTAS
FRAMATOME, Report CL26, University of Liege, June.
Yan A. M. (1997). Contribution to the direct limit state analysis of plastified and cracked structures. PhD
Thesis, Université de Liège, Belgium.
Yan A. M., Nguyen Dang H. (2000). Direct finite element kinematical approaches in limit and shakedown
analysis of shells and elbows. In: Inelastic Analysis of Structures under Variable Loads: Theory and
Engineering Applications (edited by Weichert D., Maier G.) Kluwer, Academic Press, Dordrecht,
pp. 233254.
Yan A. M., Vu D. K., Nguyen D. H. (2003). Kinematical formulation of limit and shakedown analysis. In
[Staat and Heitzer (2003a)].
144
145
Yang W. H. (1982). A variational principle and an algorithm for limit analysis of beams and plates. Comp.
Meth. Appl. Mech. Eng., 33, pp. 575582.
Zhang Y. G., Lu M. W. (1995a). An algorithm for plastic limit analysis. Comp. Meth. Appl. Mech. Eng., 126,
pp. 333341.
Zhang Y. G. (1995b). An iteration algorithm for kinematic shakedown analysis. Comp. Meth. Appl. Mech.
Eng., 127, pp. 217226.
Zimmermann J. J. (1991). Analysis of structural system reliability with stochastic programming. PhD Thesis,
John Hopkins University, Baltimore, Maryland.
Appendix
PROBABILITY DISTRIBUTIONS AND TRANSFORMATION TO
THE STANDARD GAUSSIAN SPACE
Let X be a random variable with mean μ , standard deviation σ , variance
2
σ ,
coefficient of variation / v σ μ = and shift parameter a . This parameter gives the smallest
possible value of X . The Probability Density Function (PDF) and Cumulative Distribution
Function (CDF) of f and , respectively. F X are denoted by
The mean and variance of X can be calculated from the PDF f or measurements
of X . Suppose first, that f is known. Then,
( )
b
a
x
x
xf x dx μ =
∫
(A.1)
and
( )
2
2
( )
b
a
x
x
x f x dx σ μ = −
∫
(A.2)
where the limits of integration
a
x and
b
x extend over the range of possible values of X .
For example, this range is ( when ) , +∞ X follows a Gaussian probability. When −∞ f is
not available,
2
μ and σ can be estimated from a set of m measurements
( ) ( ) ( )
{ }
1 2
, ,...,
m 2
x x x of random parameter X . The estimates of μ and σ are, respectively,
( )
1
1
ˆ
m
k
k
x
m
μ
=
=
∑
(A.3)
and
( )
2
2 ( )
1
1
ˆ ˆ
1
m
k
k
x
m
σ μ
=
= −
−
∑
2
(A.4)
and approach the actual values of the mean and variance of X as increases indefinitely. m
Probabilities and distributions that are available in our code are given in this
appendix when available in closed form. Closed form results are also presented for the
transformation from the standard Gaussian U with mean zero and unit variance to variable
X .
1. Normal distribution,
( )
2
, N μ σ
PDF:
2
1
( ) exp 0.5 ,
2
x
f x x
μ
σ πσ
⎧ ⎫
− ⎪ ⎪ ⎡ ⎤
= − − ∞ ≤
⎨ ⎬
⎢ ⎥
⎣ ⎦
⎪ ⎪
⎩ ⎭
≤ +∞. (A.5)
Parameters:
2
( ) , ( ) E X Var X μ σ = =
CDF:
( ) ( )
x
x
F x f t dt
μ
σ
−∞
−
⎡ ⎤
= = Φ
⎢ ⎥
⎣ ⎦
∫
(A.6)
where
( ) ( )
x
x f u du
−∞
Φ =
∫
. (A.7)
Transformation:
X U μ σ = + . (A.8)
2. LogNormal distribution
PDF:
2
1 ln( )
( ) exp 0.5 ,
2
x
f x x
x
μ
σ π σ
⎧ ⎫
− ⎪ ⎪ ⎡ ⎤
= −
⎨ ⎬
⎢ ⎥
⎣ ⎦
⎪ ⎪
⎩ ⎭
a ≥ (A.9)
Parameters:
2
( ) , ( ) E X Var X μ σ = = , shift a and coefficient of variation /( ) v a σ μ = −
2 2
v
2
ln 1
ln( ) 0.5
σ
μ μ σ
⎡ ⎤ = +
⎣ ⎦
= −
. (A.10)
Transformation:
( ) exp X a U μ σ = + + . (A.11)
3. Exponential distribution, ( ) , E a λ
PDF:
( ) { }
( ) exp , f x x a λ λ = − − x a ≥ . (A.12)
Parameters:
2
( ) 1/ , ( ) 1/ E X a Var X λ λ = + = .
CDF:
147
( ) { }
( ) 1 exp , F x x a x a λ = − − − ≥ . (A.13)
Transformation:
(
1
) X a
λ
U = − Φ − . (A.14)
4. Uniform distribution, [ ] , U a b
PDF:
1
( ) , a f x x
b a
b = ≤ ≤
−
. (A.15)
Parameters: ( ) ( )
2
( ) / 2, ( ) /12 E X a b Var X b a = + = −
CDF:
( ) , a
x a
F x x b
b a
−
== ≤
−
≤ . (A.16)
Transformation:
( ) ( ) X a b a U = + − Φ . (A.17)
5. Gamma distribution, ( ) , G k vx
PDF:
( ) ( )
( )
1
exp
( ) ,
k
v vx vx
f x
k
−
−
=
Γ
x a ≥
du
. (A.18)
where the gamma function
( ) ( )
1
0
exp
k
k u u
∞
−
Γ = −
∫
. (A.19)
Parameters:
2
( ) / , ( ) / E X a k v Var X k v = + =
CDF:
( )
( )
,
( ) , 0
k vx
F x x
k
Γ
= ≥
Γ
. (A.20)
where
( ) ( )
1
0
, exp
y
k
k y u u du
−
Γ = −
∫
. (A.21)
148
Transformation: not available in closed form. Therefore a NewtonRaphson iteration
scheme can be employed in order to obtain the inverse.
6. Weibull Distribution
PDF:
1
( ) exp ,
k k
k x a x a
f x x
v a v a v a
−
⎧ ⎫
− − ⎪ ⎪ ⎛ ⎞ ⎛ ⎞
= −
⎨ ⎬ ⎜ ⎟ ⎜ ⎟
− − −
⎝ ⎠ ⎝ ⎠
⎪ ⎪
⎩ ⎭
a ≥ . (A.22)
Parameters: ( ) ( ) 0, ( ) 1 1/ k E X a v a k > = + − Γ + .
( ) ( ) (
2
2
( ) 1 2 / 1 1/ Var X v a k k)
⎡ ⎤ = − Γ + − Γ +
⎣ ⎦
CDF:
( ) 1 exp ,
k
x a
F x x a
v a
⎧ ⎫
− ⎪ ⎪ ⎛ ⎞
= − − ≥
⎨ ⎬
⎜ ⎟
−
⎝ ⎠
⎪ ⎪
⎩ ⎭
. (A.23)
Transformation:
( ) ( ) { }
1/
ln
k
X a v a U = + − − Φ − ⎡ ⎤
⎣ ⎦
. (A.24)
7. Extreme Type I Distribution
PDF:
( ) ( ) { }
( ) exp exp , f x x v x v x α α α = − − − − − − ∞ ≤ ≤ ⎡ ⎤
⎣ ⎦
+∞. (A.25)
Parameters: ( ) ( )
2 2
( ) 0.577/ , ( ) / 6 E X v Var X α π α = + = .
CDF:
( ) { }
( ) exp exp , F x x v x α = − − − − ∞ ≤ ≤ + ⎡ ⎤
⎣ ⎦
∞. (A.26)
Transformation:
( ) { }
1
ln ln X v U
α
= − − Φ ⎡ ⎤
⎣ ⎦
. (A.27)
149
Trần, Thanh Ngọc Limit and shakedown analysis of plates and shells including uncertainties Dissertation an der Fakultät für Maschinenbau der Technischen Universität Chemnitz, Institut für Mechanik und Thermodynamik, Chemnitz 2008 149 + vii Seiten 55 Abbildungen 28 Tabellen 162 Literaturzitate Referat The reliability analysis of plates and shells with respect to plastic collapse or to inadaptation is formulated on the basis of limit and shakedown theorems. The loading, the material strength and the shell thickness are considered as random variables. Based on a direct definition of the limit state function, the nonlinear problems may be efficiently solved by using the First and Second Order Reliability Methods (FORM/SORM). The sensitivity analyses in FORM/SORM can be based on the sensitivities of the deterministic shakedown problem. The problem of the reliability of structural systems is also handled by the application of a special barrier technique which permits to find all the design points corresponding to all the failure modes. The direct plasticity approach reduces considerably the necessary knowledge of uncertain input data, computing costs and the numerical error. Die Zuverlässigkeitsanalyse von Platten und Schalen in Bezug auf plastischen Kollaps oder NichtAnpassung wird mit den Traglast und Einspielsätzen formuliert. Die Lasten, die Werkstofffestigkeit und die Schalendicke werden als Zufallsvariablen betrachtet. Auf der Grundlage einer direkten Definition der Grenzzustandsfunktion kann die Berechnung der Versagenswahrscheinlichkeit effektiv mit den Zuverlässigkeitsmethoden erster und zweiter Ordnung (FROM/SORM) gelöst werden. Die Sensitivitätsanalysen in FORM/SORM lassen sich auf der Basis der Sensitivitäten des deterministischen Einspielproblems berechnen. Die Schwierigkeiten bei der Ermittlung der Zuverlässigkeit von strukturellen Systemen werden durch Anwendung einer speziellen Barrieremethode behoben, die es erlaubt, alle Auslegungspunkte zu allen Versagensmoden zu finden. Die Anwendung direkter Plastizitätsmethoden führt zu einer beträchtlichen Verringerung der notwendigen Kenntnis der unsicheren Eingangsdaten, des Berechnungsaufwandes und der numerischen Fehler. Schlagworte: Limit analysis, shakedown analysis, exact Ilyushin yield surface, nonlinear programming, first order reliability method, second order reliability method, design point Archivierungsort: http://archiv.tuchemnitz.de/pub/2008/0025
ACKNOWLEDGEMENTS
This work has been carried out at the Biomechanics Laboratory, Aachen University of Applied Sciences, Campus Jülich. The author gratefully acknowledges the Deutscher Akademischer Austausch Dienst (DAAD) for a research fellowship award under the grant reference A/04/20207. The author is indebted to Prof. Dr.Ing. M. Staat who has been the constant source of caring and inspiration for his helpful guidance and encouragement. His commitment and assistance were limitless and this is greatly appreciated. The author would like to express his deep gratitude to Prof. Dr.Ing. R. Kreißig for giving him the permission to complete Doctorate of Engineering at the Chemnitz University of Technology and for kindly assistance and supervision. The author would like to thank Prof. Dr.Ing. C. Bisbos, Aristotle University of Thessalonoki, Greece for having kindly accepted to review this thesis. The author is thankful to Dr.Ing. Vũ Đức Khôi for help and advice, to Ms Wierskowski and Ms Dronia for their programming as part of their diploma theses in some parts of FEM source code. The author’s thanks are also extended to Prof. Dr. rer. nat. Dr.Ing. S. Sponagel and to the other colleagues at the Biomechanics Laboratory for their helpful assistance. The author is immensely indebted to his father Trần Thanh Xuân and his mother Nguyễn Thị Hòa who have been the source of love and discipline for their inspiration and encouragement throughout the course of his education including this Doctorate. Last but not least, the author is extremely grateful to his wife Mrs. Nguyễn Thị Thu Hà who has been the source of love, companionship and encouragement, to his daughters My and Ly who have been the source of joy and love.
iii
iv .
..........................................................1....................................... 8 1.................2..............................................2 Description of the exact Ilyushin yield surface ...........2 General theorems of limit analysis ..................... 27 2....... 9 1............1 Derivation ...................................................... 14 1.......... MATHEMATICAL FORMULATIONS OF LIMIT AND SHAKEDOWN ANALYSIS IN GENERALIZED VARIABLES ....................................................................................................3 Fundamental of shakedown theorems..............3........2.........................3 Exact Ilyushin yield surface.............................................................................1 Introduction............... 4 1......... 24 2.....4 Plastic dissipation function ..... 25 2..................5 Unified shakedown limit......2 Definition of load domain......................................................................... 3 1............... 9 1........ 23 2.....................................2 Fundamental principles in plasticity .............................3 Strain quantities .............................................2 Normalized shell quantities ...........................1.....................2.....................................2 Theory of shakedown analysis..............................................................................................2.....................4 Yield criteria ...........................2....1 Elastic and rigid perfectly plastic materials..... 30 2............... 21 2........1 Theory of limit analysis ............ 22 2..............3........................................................................ 10 1...... 22 2...........................4 Separated shakedown limit ........1................. 16 1......................2...................5 Reformulation ..1....................................... 3 1.........3.............................................. 24 2...................................... 11 1.1 Reference quantities.................................................. 1 1.............3........ 33 v ......................................................3 Reparameterization .2....1 Basic concepts of plasticity....2 Stress quantities ........................................................................... 12 1..3 Drucker’s postulate .................... 19 2.........TABLE OF CONTENTS INTRODUCTION ............................. 9 1.......................................................................................................................................2.................................................................................................................................................5 Plastic dissipation function in local variables....................................... 12 1..................................... 3 1......................... 18 1........................4 StressStrain relation.................. 6 1...............................................2......1 Introduction....................1................3..1............................................ 7 1....................................................1.................................... FUNDAMENTALS...................................
.................................. 50 4... 92 6..... 41 4....... 76 5... 73 5.......... 77 6...............2 Random variables ............................................................... 72 5.............. 70 5......................3 Moments ... 51 4...............2.......................................................................1............7 Elbow subjected to bending moment.......................................................................... 76 5.................. 65 4.........................5 Grooved rectangular plate subjected to varying tension and bending ......... 55 4......3 Calculation of design point ....................1.....................Table of Contents 3.............................6 Square plate with a central circular hole.................................. 53 4.............................................. 61 4.................................................................... LIMIT AND SHAKEDOWN ANALYSIS OF DETERMINISTIC PROBLEMS 79 6........................................... 50 4........................... 109 vi .................................1 Limit analysis of a cylindrical pipe under complex loading............................... 80 6...........4........................ 49 4.............................2 Solution technique ............. 66 4.1 Failure function and probability ....4.......9 Nozzle in the knuckle region of a torispherical head............2.......................1...........................................1 Finite element discretization..................4.................................................................... 97 6...1 Mathematical sensitivity .......................................................................................................................................1 Sample space......................... 53 4...... 39 3.and SecondOrder Reliability Method .................... 58 4.......2.... 63 4. 74 5.5 Special case of probabilistic shakedown analysis.................. DETERMINISTIC LIMIT AND SHAKEDOWN PROGRAMMING.......................................2 First................................................................4 Pipejunction subjected to varying internal pressure and temperature .2 Firstorder system reliability analysis................................4 Second derivatives of the limit state function................1 Bounds for the system probability of failure ..........................................1 Multimode failure ...........................................................4....4 Sensitivity of the limit state function............4. 50 4...........1....................................... 73 5.............2 Limit and shakedown analysis of a thinwalled pipe subjected to internal pressure and axial force ..........1 Basic idea of the method...........................2 Definition of the bulge..................... MULTIMODE FAILURE AND THE IMPROVEMENT OF FORM/SORM RESULTS .......................................... 82 6.......2 Reliability analysis..... 90 6........................................................3 Cylindrical shell under internal pressure and temperature change ... 84 6.............................. 86 6.......................................................................................3 First derivatives of the limit state function ......... PROBABILISTIC LIMIT AND SHAKEDOWN PROGRAMMING..........................1..............2................ 62 4....2 Definition of the limit state function...............................8 Limit and shakedown analysis of pipeelbow subjected to complex loads ........................................... 103 6..........................................................1 Basic concepts of probability theory ...2 Kinematic algorithm ...................... 38 3........
........................... 146 1............................................................ Exponential distribution....... Gamma distribution .................. 147 2................................................................................................. 147 4...........................................1 Square plate with a central circular hole. Uniform distribution ................. 149 vii .................................................................................. Weibull Distribution .............................. PROBABILISTIC LIMIT AND SHAKEDOWN ANALYSIS OF STRUCTURES ..................... 148 6.................................................2 Pipejunction subjected to internal pressure ................... 148 5................................. 121 7............................... Extreme Type I Distribution ...................................... LogNormal distribution .. 115 7.............................................................................................................................................. 149 7......................................................................................................... SUMMARY....................7............................................................... 115 7.................... 123 7.......................................................................................................... 147 3........................ 133 REFERENCES .......... 136 APPENDIX: PROBABILITY DISTRIBUTIONS AND TRANSFORMATION TO THE STANDARD GAUSSIAN SPACE ................................................ 129 8..............4 Folding shell subjected to horizontal and vertical loads............... Normal distribution....3 Limit analysis of cylindrical pipe under complex loading ...............................................
Based on a direct definition of the limit state function. Since the deterministic problem is a subroutine of the probabilistic one. an extension for lower and upper bound theorems in terms of generalized variables are proposed and formulated. The thesis consists of two parts: the theory part (chapters 15) and numerical part (chapters 67). Studies show. we present the two fundamental theorems of limit and shakedown analysis. Probabilistic limit and shakedown analysis deals with uncertainties originating from the loads. For deterministic problem. several relations between physical and normalized values for plates and shells are presented. will be applied instead of simplified ones (linear or quadratic approximations). Based on the LoveKirchhoff theory. The problem of reliability of structural systems (series systems) will also be handled by the application of a special technique which permits to find all the design points corresponding to all the failure modes. In chapter 3. a kinematic approach of limit and shakedown analysis. including the fundamental principles and yield criteria. Based on the original ones which were proposed by Melan and Koiter. that it improves considerably the FORM/SORM results. Chapter 1 introduces some basic concepts of plasticity theory.INTRODUCTION The present work aims at providing an effective numerical method for the limit and shakedown analysis (LISA) of general shell structures with the help of the finite element method. which is adopted for shell structures is developed (the deterministic LISA). In chapter 2. in this case. namely the exact Ilyushin yield surface. Starting from a finite element discretization. The derivation and description of the exact Ilyushin yield criterion is briefly summarized. material strength and thickness of the shell. a yield criterion which is exact for rigidperfectly plastic material behaviour and is expressed in terms of stress resultants. To this reason. A simple approach for the direct calculation of the shakedown limit as the minimum of incremental plasticity limit and alternating plasticity limit is also presented. low cycle fatigue and ratchetting are analysed based upon an upper bound approach. thus. Both deterministic and probabilistic limit and shakedown analyses are presented. the static and kinematic theorems. even a small error in the deterministic model can lead to a big error in the reliability analysis because of the sensitivity of the failure probability. three failure modes of structure such as plastic collapse. a detailed kinematic algorithm in terms of generalized variables will . the calculation of the failure probability may be efficiently solved by using the First and Second Order Reliability Methods (FORM/SORM).
The loads and material strength as well as the thickness of the shell are to be considered as random variables. A simple technique for overcoming numerical obstacles. when they exist on the limit state surface. which is based on the kinematical approach. such as the nonsmooth and singular objective function. For each test case. which is based on the Sequential Quadratic Programming (SQP). The FORM approximation is. 2 . some existing analytical and numerical solutions found in literature are briefly represented and compared. Finally chapter 8 contains some main conclusions and future perspectives. Uncertainties which originate from the loads. Numerical results are tested against analytical solutions. In order to get the design point. Each design point corresponds with an individual failure mode or mechanism. experiments and several limit loads which have been calculated in literature with different numerical methods using shell or volume elements. then applied at each design point followed by a series system reliability analysis leading to improved estimates of the system failure probability. In chapter 6. Nonlinear sensitivity analyses are also performed for computing the Jacobian and the Hessian of the limit state function. we aim at presenting various typical examples of deterministic limit and shakedown analyses to illustrate and validate the theoretical methods. Numerical studies of limit and shakedown analysis for probabilistic problems are introduced in chapter 7. is also proposed.be formulated and introduced. Many different kinds of distribution of basic variables are taken into consideration and performed with First and Second Order Reliability Methods (FORM/SORM) for calculation of the failure probability of the structure. Chapter 5 presents a method to successively find the multiple design points of a component reliability problem. Chapter 4 focuses on presenting a new algorithm of probabilistic limit and shakedown analysis for thin plates and shells. the strength of material and the thickness of the shell are all analyzed. a nonlinear optimization was implemented.
Based on the LoveKirchhoff theory. Stresses are not allowed to become higher than this threshold. If some thermal θ effects occur. It can be proved that elastic characteristics do not affect the plastic collapse limit state and thus the application of the elastic perfectly plastic material model becomes same to that of the rigid perfectly plastic model for limit analysis. which is closely related to the works of Burgoyne and Brennan [1993b]. For convenience. In the geometrically linear theory the total strain ε ij is assumed to be decomposed e additively into an elastic or reversible part ε ij and an irreversible part ε ijp . Furthermore. instead of plates and shells.1 Basic concepts of plasticity 1. This is equivalent to the rigid plastic material model.1) The elastic part of the strain obeys Hooke’s law. 1. which are necessary for the developments in the subsequent chapters. a thermal strain term ε ij should be added and thus e θ ε ij = ε ij + ε ijp + ε ij .1 FUNDAMENTALS In the following. The derivation and description of the exact Ilyushin yield criterion is summarized.1.2) . its relationship with stress is linear e −1 ε ij = Cijklσ kl (1. we will use only the concept of shells. the elastic deformation can usually be disregarded when compared with the plastic deformation. nonhardening solid body is idealized by the elastic perfectly plastic model. including the fundamental principles and yield criteria. We start with a brief introduction of plasticity theory. Seitzberger [2000]. In this model. the material behaves elastically below the yield stress and will begin to yield if the stress intensity reaches the yield stress. (1. several relations between physical and normalized values for plates and shells are presented.1 Elastic and rigid perfectly plastic materials Mechanical behaviour of rate intensities elasticplastic. some theoretical foundations are stated.
1. Any stress field respecting conditions (1. β = i. For an isotropic material.6b) (1.1. l ) the Kronecker delta. δ αβ (α . f (σ ij ) = 0 corresponds to appearance of plastic deformation. The inverse relationship of (1.where Cijkl .8) is called a statically admissible field. this tensor is expressed in the form below Cijkl = (1 + ν )(1 − 2ν ) νE δ ijδ kl + E (δ ikδ jl + δ ilδ jk ) 2 (1 + ν ) (1.7) and (1. 4 .3) where E denotes the Young’s modulus and ν the Poisson ratio. called the elastic constants.4) where G = E is the shear modulus of elasticity.5) (1. are components of a tensor of rank 4. Plastic flow can occur only when the stress point is on the yield surface and for elasticperfectly plastic material the additional loading σ ij can only move along the tangential direction. The stresses σ ij are said to be in equilibrium if they satisfy the equations of internal equilibrium ∂ jσ ij + f i = 0 in V and the conditions of equilibrium at the surface of the body ti = n jσ ij = ti (1.6a) ∂f ∂σ ij (1. (1. j.8) Furthermore. it is called a plastically admissible or licit stress field. f (σ ij ) > 0 corresponds to a region inaccessible for the material. 2(1 + ν ) The plastic strain rate obeys an associated flow law ε ijp = λ where λ is a nonnegative plastic multiplier and f = f (σ ij ) represents a timeindependent yield surface such as f (σ ij ) < 0 corresponds to elastic behaviour.7) on ∂Vσ .2) can be written as e σ ij = 2Gε ij + 2ν e Gδ ijε kk 1 − 2ν ) ( (1.2 Fundamental principles in plasticity Consider a structure subjected to volume loads f i and surface loads ti . f (σ ij ) ≤ 0 . (1. k .6c) The definition of yield function means that the stress point cannot move outside the yield surface. if this stress field nowhere violates the yield criterion.
Furthermore.9) (1. (1.12) ∂Vσ Principle of complimentary virtual power For an arbitrary set of infinitesimal virtual variations of the stress tensor δσ ij that are statically admissible.11) then it is called a licit mechanism. the necessary and sufficient condition to make the strain rate tensor ε ij and velocity vector ui compatible is to satisfy the following equation ∫ ε δσ ij V ij dV = ∂Vu ∫ n δσ j ij i u dS . and for all stresses σ ij that are statically admissible we have the following virtual work equation ∫ σ ε dV = ∫ f u dV + ∫ ij ij i i V V ti ui dS + ∂Vσ ∂Vu ∫ nσ j ij i u dS . A kinematically admissible strain and displacement field can be defined in a similar manner. Any mechanism ( ui . one can easily deduce that for all strain rate tensors ε ij and velocity vectors ui that are kinematically admissible. the necessary and sufficient condition to make the stress field σ ij equilibrium is to satisfy the following equation ∫ σ δε dV = ∫ f δ u dV + ∫ ij ij i i V V tiδ ui dS . ε ij ) respecting conditions (1.The actual flow mechanism composed of the velocities ui and strain rate ε ij in the body which satisfy the compatibility and kinematical boundary conditions ε ij = 1 (∂ jui + ∂ iu j ) 2 in V (1.10) ui = ui on ∂Vu .10) is called kinematically admissible.13) Equation of virtual power From the two above principles.14) 5 .9) and (1. (1. (1. if this mechanism furnishes a non negative external power WE = ∫ f i ui dV + V ∂Vσ ∫ ti ui dA ≥ 0 (1. which states that for an arbitrary set of infinitesimal virtual velocity variation δui that are kinematically admissible. Principle of virtual power One of the main tools in the mechanics of continua is the principle of virtual power.
It is valid for both workhardening and perfectly plastic materials.1. (1. 6 .3 Drucker’s postulate "Over the cycle of application and removal of the set of forces. see figure 1. As a result.1.16a) is violated. Starting from this ∫ is the integral taken over a cycle of applying and removing the added stress set. 1. (1. 2005]: Principle of maximum energy dissipation 0 (σ ij − σ ij )ε ijp ≥ 0 .15) σ ij is the stress tensor on the yield surface satisfying the yield condition f (σ ij ) = 0 . postulate. In the case of having n intersected differentiable yield surfaces at a singular point.It is well known that the above variational principles are independent of the constitutive equation of the material.16) one can deduce that the plastic strain rates tensor ε ij must be normal to the yield surface at a smooth point or lie between adjacent normals at a corner (nonsmooth point). the new work performed by the external agency on the changes in the displacements it produces is nonnegative" [Martin.17) Convexity of yield surface 0 It is clear form figure 1. This is expressed mathematically by the following inequality ∫ (σ where ij 0 − σ ij )dε ij ≥ 0 (1. In other words.1 that if there is any σ ij lying on the outside of the tangent.16b) sometimes known simply as Drucker’s inequality.5) should be replaced by: ε ijp = ∑ λk k =1 n ∂f k . Normality rule p From (1. the inequality (1.16a) or σ ijε ijp ≥ 0 . the yield surface is convex. ∂σ ij (1. and 0 0 σ ij is the plastically admissible stress tensor such that f (σ ij ) ≤ 0 . (1. 1975]. the entire elastic region must lie to one side of the tangent. we have the following important consequences [Lubliner.
The convexity of the yield surface has a very important role in plasticity. It permits the use of convex programming tools in limit and shakedown analysis. It should be noted that Drucker’s postulate is quite independent of the basic laws of thermodynamics. It does not hold if internal structural changes occur or for temperature dependent behaviour [Kalisky, 1985]. Furthermore, the yield surface fails to be convex if there is an interaction between elastic and plastic deformations, i.e. if the elastic properties depend on the plastic deformation [Panagiotopoulos, 1985].
εp elastic domain σ−σ 0 σ
εp
σ0
yield surface
inaccessible domain
Figure 1.1 Normality rule
1.1.4 Yield criteria
The yield criterion defines the elastic limits of a material under a combined state of stress. The yield function f in stress space may be written with no loss of generality in terms of the stress deviator and the first invariant of stress, that is
f ( σ, ξ ) = f ( s, I1 , ξ )
(1.18)
where I1 = σ kk = δ ijσ ij is the trace of σ ij , ξ are internal variables which are determined experimentally and sij is the deviator defined as
1 1 ⎛ ⎞ skl = σ kl − I1δ kl = ⎜ δ ikδ jl − δ ijδ kl ⎟ σ ij . 3 3 ⎝ ⎠ (1.19)
Since the concept of plasticity was first applied to metals, in which the influence of mean stress on yielding is generally negligible [Bridgman, 1923 and 1952], the oldest and most commonly used yield criteria are those that are independent of I1 . They are therefore formulated with sij using J 1 = skk = 0 .We present here two most wellknown yield criteria in plasticity.
7
Tresca criterion
The Tresca yield criterion is historically the oldest one; it embodies the assumption that plastic deformation occurs when the maximum shear stress over all planes attains a critical value, namely, the value of the current yield stress in shear, denoted k T . This criterion may be represented by the yield function f ( σ ) = Max ( σ 1 − σ 2 , σ 2 − σ 3 , σ 3 − σ 1 ) − 2kT = 0 where kT = (1.20)
, σ y is yield stress. It can also be expressed explicitly in terms of the 2 invariants J 2 and J 3 of the stress deviator tensor as
3 2 2 4 6 f ( J 2 , J 3 ) = 4 J 2 − 27 J 32 − 36kT J 2 + 96kT J 2 − 64kT = 0 .
σy
(1.21)
Von Mises criterion
The von Mises criterion is known as the maximumoctahedralshearstress or maximumdistortionalenergy criterion, which states, that yielding begins when the octahedral shearing stress reaches a critical value kv such as f ( J 2 ) = J 2 − kv2 = 0 where kv = (1.22)
. We may also formulate the von Mises yield criterion in the form of 3 principal stresses (σ 1 − σ 2 ) 2 + (σ 2 − σ 3 ) 2 + (σ 3 − σ 1 ) 2 = 6kv2 . (1.23)
σy
The projection of the Tresca yield surface in the (σ 1 − σ 3 )(σ 2 − σ 3 ) plane takes the form of an irregular hexagon, where the von Mises criterion is an ellipse. In the next section the von Mises criterion will be used to derive the socalled exact Ilyushin yield surface for the elasticplastic analysis of shelllike structures.
1.1.5 Plastic dissipation function in local variables
The plastic dissipation function is defined by
* D p = max(σ ijε ijp ) = σ ijε ijp
(1.24)
* * where σ ij is a plastically admissible stress tensor, i.e. satisfying f (σ ij ) ≤ 0 , σ ij is the
* definition, while the second one is due to the Drucker stability postulate: (σ ij − σ ij ) ε ijp ≥ 0 .
stress tensor satisfying yield condition f (σ ij ) = 0 . The first equality of (1.24) is the
The plastic dissipation for the Mises criterion and associated flow rule is given by [Lubliner, 2005]
8
D p (ε ijp ) = kv 2ε ijpε ijp .
(1.25)
1.2 Normalized shell quantities 1.2.1 Reference quantities
The representation of the exact Ilyushin yield surface and the description of its applications for the limit and shakedown analysis of shell structures are given in terms of normalized generalized stresses and strains. The membrane forces and bending moments are normalized with respect to the plastic limit loads in uniaxial tension and bending, respectively N 0 = σ y h, M0 = σ y h2 4 (1.26)
where σ y and h are yield stress and shell thickness, respectively. Physical strain values are referred to the reference strain ε 0 . A convenient measure for ε 0 is given by (1 − ν 2 ) ε0 = σ y . E For the sake of simplicity, the reference “curvature” κ 0 is defined as follows (1.27)
ε 0 corresponds to the elastic strain of an uniaxially stretched plate at the yield stress σ y .
4ε 0 . h
κ0 =
(1.28)
so that N 0ε 0 = M 0κ 0 , [Burgoyne and Brennan 1993b], [Seitzberger, 2000]. This definition does not represent a kinematic relation. A dimensionless thickness coordinate z is used for throughthickness integrations. The relation between z and the thickness coordinate s3 is as follows z= s3 . h (1.29)
1.2.2 Stress quantities
Physical and normalized stress vectors for a state of plane stress are given by, respectively
⎛ σ 11 ⎞ σ = ⎜ σ 22 ⎟ and ⎜ ⎟ ⎜σ ⎟ ⎝ 12 ⎠ ⎛ σ 11 ⎞ 1 ⎜ ⎟ σ= ⎜ σ 22 ⎟ . σy ⎜ ⎟ ⎝ σ 12 ⎠
(1.30)
9
From these. β ∈ {1. respectively.33) The LoveKirchhoff assumptions state that the normals (i. 2} ) are the physical membrane force and bending moment components of the shell.31) where the terms Nαβ and M αβ . ⎟ κ0 ⎜ 2κ 12 ⎟ ⎠ ⎝ (1.e.e. s3 q M 22 s2 M11 s1 N11 N12 M12 N12 M12 N 22 Figure 1. unstretched and normal (i. the normalized membrane force and bending moment vectors n and m can be obtained as follows ⎛ N11 ⎞ h/2 1/ 2 1/ 2 1 ⎜ 1 ⎟= 1 n= N 22 ⎟ ∫ σds3 = σ y h −1/ 2 σhdz = −1/ 2 σdz ∫ ∫ N0 ⎜ ⎜ N ⎟ N0 −h / 2 ⎝ 12 ⎠ ⎛ M 11 ⎞ 1/ 2 1/ 2 h/2 1 ⎜ 4 ⎟= 1 2 m= M 22 ⎟ ∫ ∫ σs3ds3 = σ y h 2 −1/ 2 σh zdz = 4−1/ 2 σzdz ∫ M0 ⎜ ⎜ M ⎟ M 0 −h / 2 ⎝ 12 ⎠ (1.32) are normalized with respect to the ε 0 and κ 0 .2. ⎜ ⎟ ⎜ 2ε ⎟ ⎝ 12 ⎠ ⎛ κ 11 ⎞ κ = ⎜ κ 22 ⎟ ⎟ ⎜ ⎜ 2κ ⎟ ⎝ 12 ⎠ (1.2 Static shell quantities 1. which gives ⎛ ε11 ⎞ 1⎜ e = ⎜ ε 22 ⎟ . (α . ⎟ ε0 ⎜ 2ε12 ⎟ ⎝ ⎠ ⎛ κ 11 ⎞ 1 ⎜ k = ⎜ κ 22 ⎟ . the lines perpendicular to the shell’s midplane) remain straight. they always make a 10 .3 Strain quantities The physical strain and curvature vectors ⎛ ε11 ⎞ ε = ⎜ ε 22 ⎟ .
39) where B. it is convenient to group both the strain vectors and the curvature vectors as well as the section force and moment vectors into “engineering” vectors. (1. ⎝ m⎠ (1. (1. 2κ 12 = 2 − 1 ∂ s1 ∂ s2 ∂ s2 ∂ s1 ε11 = (1.34) where ε and κ are physical midplane strain and curvature vectors ( s3 = 0 ).right angle to the midplane) after loading. By the introduction of z and ε 0 .35) with uα . 2ε12 = 2 + 1 ∂ s1 ∂ s2 ∂ s1 ∂ s2 ∂θ ∂θ ∂θ ∂θ κ 11 = 2 . as follows (1. C.2.36) By using of (1. For purely elastic material behaviour we have 11 . κ 22 = − 1 .37) ⎛e⎞ ε = ⎜ ⎟. The relation between them and the displacements delivers ∂ u1 ∂u ∂u ∂u . ⎝k⎠ ⎛n ⎞ σ= ⎜ ⎟ . respectively.38) For an elastoplastic material behaviour.28). ε 22 = 2 .4 StressStrain relation As will be shown in the subsequent chapters.34) becomes 1 ε0 ε= 1 ε0 ε+z 4 κ0 κ. 1. the tangential stressstrain relations for a Kirchhoff shell may be written as ⎛ dn ⎞ ⎛ C B ⎞ ⎛ d e ⎞ ⎛ C B ⎞ dσ = ⎜ ⎟=⎜ ⎟⎜ ⎟ = ⎜ ⎟ dε ⎝ dm ⎠ ⎝ B D ⎠ ⎝ dk ⎠ ⎝ B D ⎠ (1.θα (α ∈ {1. the Kirchhoff hypothesis may be written in normalized form as e ( z ) = e + 4 zk . D are submatrices of the physical tangential stiffness matrix. Based on the LoveKirchhoff hypothesis the physical strain ε( s3 ) = ε + s3κ vector can be written in the standard form as follows (1. These mean that the transversal shear strains in the thickness coordinate s3 are negligible which is only valid for thin shells with small displacement u3 ( u3 h ).2} ) being translations and rotations of the midplane.
A reparameterization of the exact Ilyushin yield surface for thin shells which produces a simpler (though still exact) form was presented by Burgoyne and Brennan [1993b].3 Exact Ilyushin yield surface In order to calculate the maximum collapse load of a shell. At the limit state. it is of course important to identify a yield surface. e. each material point through the thickness has a plastic material behaviour. which marks the limiting values of the stress resultants. When dealing with stress resultants.42) 12 . a linear approximation proposed by Ilyushin himself.1 Derivation of the exact Ilyushin yield surface The derivation of the exact Ilyushin yield surface is based on the following assumptions • • • • perfectly plastic isotropic material behaviour obeying the von Mises yield criterion. a criterion is needed to assess when the shell reaches a situation where the behaviour is governed by plasticity. are valid f ( σ ) = σ Pσ − 1 = 0. plane stress conditions in each material point. Little [1977]) which vary through the thickness. the von Mises yield condition and the normality rule. the Ivanov yield surface. Some approximations have been used.41) de p = dξ ∂f = 2dξ Pσ . when a more complex fully plasticity yield surface is needed. Thus both. ⎛ ⎜1 ν 0 ⎜ 0 C = ⎜ν 1 ⎜ 1 −ν ⎜0 0 2 ⎝ ⎞ ⎟ ⎟ ⎟. In 1948 Ilyushin published the derivation of a stress resultant yield surface describing the case where a crosssection of a shell is fully plastified and thus reaches its load capacity.g.3. in which case a yield criterion such as the von Mises criterion is used. This yield surface.40) 1. or in terms of stress resultants (Crisfield [1973]. validity of the Kirchhoff hypothesis for both total and plastic strains. Frieze [1975]). Their work opens the way for the practical use of the exact Ilyushin yield surface in structural calculations. beyond which the shell may not be loaded. Two approaches have been identified. has not been used because the parametric form in which it was described by Ilyushin was not amenable to calculation. T −1/ 2 0 ⎞ ⎛ 1 ⎜ −1/ 2 ⎟ 1 0⎟ P=⎜ ⎜ 0 0 3⎟ ⎝ ⎠ (1. 3 (1.B = 0. It is possible to work either in terms of stresses (Moxham [1971]. ⎟ ⎟ ⎠ 4 D = C. 1. validity of the normality rule for the plastic deformations. however. ∂σ (1.
(1. the stress resultants may finally be written as p ⎛ n ⎞ 3 ⎛ J 0P −1 4 J 1P −1 ⎞ ⎛ d e ⎞ 3 ⎛ J 0P −1 4 J 1P −1 ⎞ p σ=⎜ ⎟= ⎜ ⎟= ⎜ ⎟⎜ ⎟ dε −1 16 J 2P −1 ⎠ ⎝ dk p ⎠ 2 ⎝ 4 J 1P −1 16 J 2P −1 ⎠ ⎝ m ⎠ 2 ⎝ 4 J 1P (1. P3 = ⎜ −1 ⎟ 0 ⎠ ⎝0 P ⎠ ⎝ 0 0⎠ ⎝P /2 (1.50).50) From Eqs.From Eqs.44) It is assumed that the plastic strain increment resultants obey the Kirchhoff hypothesis. From (1. 4 (1. (1. P2 = ⎜ −1 ⎟ .48) These incremental plastic strain resultant intensities are subject to the condition (by the Schwarz inequality) Pε Pκ ≥ Pεκ 2 Substitution of Eqs.45) and (1.31) and (1.43) d ξ ( de p ) = (1.44).37) we have de p ( z ) = d e p + 4 z d k p .46) in Eq.43) gives (1. which gives the consistency parameter dξ = 1 Pε + 2 Pεκ z + Pκ z 2 3 (1. (1. (1.45) into (1. P1 = ⎜ ⎟ .51) where the integrals J i (not to confuse with the invariants of the deviator) can be calculated as follows 13 . (1.41) and (1.45) Substituting (1. (1.42) σ and dξ can be expressed as functions of de p σ ( de p ) = 1 P −1de p p 2dξ ( de ) T 1 ( de p ) P−1de p .47) ( ≥ 0) Pεκ = 3 ( d e p ) P −1dk p = 3 ( dε p ) P2 dε p Pκ = 12 ( dk p ) P −1dk p = 12 ( dε p ) P3dε p T T where ⎛ P −1 0 ⎞ ⎛ 0 P −1 / 2 ⎞ ⎛0 0 ⎞ .46) with the incremental plastic strain resultant intensities Pε = T T 3 3 ( d e p ) P−1d e p = 4 ( dε p ) P1dε p 4 T T ( ≥ 0) (1.49) σ= 3 2 1 Pε + 2 Pεκ z + Pκ z 2 P −1 ( d e p + 4 z dk p ) .
1. which can be represented in parameter form as a function of two independent parameters.54). ⎟ ⎟ ⎟ 2 ⎟⎜ P J1 J 2 J2 ⎠⎝ κ ⎠ J 0 J1 (1. Ilyushin [1948] represented the surface in parameterized form. Qm ) = 0 . If the direction of the plastic strain increment resultants is given.54) describes a surface. ( ≥ 0) (1. by introducing the two following parameters ⎛ P − P + Pκ / 4 ⎞ ζ = ⎜ ε εκ ⎟ ⎝ Pε + Pεκ + Pκ / 4 ⎠ 1/ 2 .51) can be regarded as a sixdimensional stress resultant yield surface for the limit that the shell is wholly plastic and thus in each point over the thickness the von Mises yield criterion and the normality rule are satisfied. (1. (1.51) can be reduced to a surface in the threedimensional Qspace as follows ⎛ J 02 ⎛ Qt ⎞ ⎜ ⎜ ⎟ ⎜ Qtm / 4 ⎟ = 3 ⎜ J 0 J 1 ⎜ ⎜ ⎟ ⎜ 2 ⎝ Qm /16 ⎠ ⎝ J1 J 12 ⎞ ⎛ Pε ⎞ ⎟⎜ ( J 0 J 2 + J12 ) / 2 J1 J 2 ⎟ ⎜ 2 Pεκ ⎟ . Ji = ∫ 3 −1/ 2 Pε + 2 Pεκ z + Pκ z 2 1/ 2 (1. provided the integrals J i can be evaluated numerically. An implicit form of Eq. Qtm . The resulting equations of the exact Ilyushin yield surface are 14 . can not be obtained. Equation (1.3. 1/ 2 ⎛ ⎞ Pε Pκ − Pεκ 2 μ =⎜ ⎟ ⎝ Pκ ( Pε + Pεκ + Pκ / 4 ) ⎠ (1. the surface (1.2 Description of the exact Ilyushin yield surface Corresponding to the quadratic strain resultant intensities. i.53) From these.52) Equation (1. F = F (Qt . however.55) which corresponds to Pε Pκ ≥ Pεκ 2 .1 zi dz .e.56) . quadratic stress resultant intensities can also be defined Qt = nT Pn Qtm = nT Pm Qm = mT Pm ( ≥ 0) .54) The surface is bounded by the condition that QtQm ≥ Qtm 2 (1.51). the stress resultants can be obtained from Eq.
see figure 1. It introduces a discontinuity at the line Qtm = 0 . 1993b].3 shows a graphical representation of the exact and the linear approximation of the yield surface in the Qspace. Figure 1. In his original paper Ilyushin proposed a linear approximation to his exact surface.60) This crude approximation consists of two planes in the Qspace. Thus. which is usually referred to as Ilyushin yield surface F1 = Qt + 1 Qtm + Qm = 1 . 1993b]. Since the lines of constant ζ and μ are virtually parallel in many cases and.57) ( ) where (1 + ψ = ln ϕ = ζ −1 1− μ2 μ ) ± ln (ζ + ζ 2 − μ2 μ ) (1. which are governed by different equations (due to the alternative signs). 3 (1. known as the “inplane dominant” and “bending dominant” regions. Ilyushin’s original parameterization makes it necessary to divide the surface into four regions. it can also be plotted in twodimensional form as Qtm over Qt − Qm without loss of clarity [Burgoyne and Brennan. 15 . (1.3.59) The boundary is given by μ = 0 . As can be seen. a numerical algorithm based on these parameters will be illconditioned and numerically unstable [Burgoyne and Brennan.Qt = 1 μ 2ψ 2 + ϕ 2 ) 2 ( Δ1 2 μ 2 Δψ 2 + Δϕ 2 + μ 2ϕψ + ϕχ ) 3 ( Δ1 4 μ 2ψ 2 ( μ 2 + Δ 2 ) + ϕ 2 ( 4 μ 2 + Δ 2 ) + 2 μ 2 Δϕψ − 2 μ 2ψχ + 2 Δϕχ + χ 2 4 Δ1 Qtm = Qm = (1. μ ≤ ζ ≤ 1.58) χ = 1− μ2 ±ζ ζ 2 − μ2 Δ1 = 1 − μ 2 ± ζ 2 − μ 2 Δ= subject to the conditions 1−ζ 2 Δ1 0 ≤ μ ≤ 1. the surface is symmetric with respect to the Qt − Qm plane.
In structural reliability analyses such errors in the deterministic model are not acceptable because of the sensitivity of the failure probability. can be found e. the error can increase up to approximately 10% according to [Preiss. 2000]. 1.5% on the unsafe side. 1993b) Ivanov [1967] proposed a quadratic approximation of the exact Ilyushin yield surface 2 Qm 1 ⎛ QtQm − Qtm ⎞ 2 2 F2 = Qt + + Qm / 4 + Qtm − ⎜ ⎟ =1. 1971]. it has no discontinuities in slope except one at Qt = 1 . However. According to Robinson.3.48Qm ⎠ (1. and always lies within 1% of the exact Ilyushin yield surface. in [Robinson.3 Reparameterization In order to avoid the difficulties arising with the parameterization of Ilyushin and open a possibility using the exact yield surface in practical computations. 2 4 ⎝ Qt + 0.g. Further suggestions of approximate full plasticity yield surfaces.61) The Ivanov yield surface overcomes many of the difficulties associated with the approximate Ilyushin yield surface. where the exact surface also has a slope discontinuity.Figure 1. partly including the effect of transverse shear as well as hardening effects. the maximum error of the linear approximation is 6% on the safe side and 3.3 Exact and linear approximation of Ilyushin yield surfaces (from Burgoyne and Brennan. Burgoyne and Brennan [1993b] introduced the parameters 16 .
64) Figure 1.62) where β and γ are proposed as two independent parameters for the description of the yield surface. β = − εκ Pκ Pκ and γ =υ − β 2 (1.υ= Pε P . where the consistency parameter dξ in Eq. With these parameters. β has the physical meaning of being the position within the thickness of the shell.65) 17 . (1. 1993b) This yield surface is subject to the conditions 0 ≤ β 2 ≤ υ ≤ ∞.4 One half of the exact Ilyushin yield surface (Qtm ≤ 0 ) constructed in terms of β and γ (from Burgoyne and Brennan.46) is a minimum. the yield surface assumes the form Qt = ( β K 0 − K1 ) + γ K 02 2 Qtm = 4 ( β K 0 − K1 )( β K1 − K 2 ) + 4γ K 0 K1 Qm = 16 ( β K1 − K 2 ) + 16γ K12 2 (1.63) where the integrals Ki are given by 1/ 2 Ki = 3Pκ J i = −1/ 2 ∫ zi Pε + 2 Pεκ z + Pκ z 2 dz (1. 0 ≤ γ ≤ ∞. The integrals Ki can be evaluated analytically giving (1.
5 − β ) +γ − ( 0.4 Plastic dissipation function The derivation and description of the exact Ilyushin yield surface presented above was performed with incremental strain quantities. if the strain rate quantities are used throughout in stead of the incremental strain quantities.5 + β ) ⎟ ⎠ 2 2 ( 0. According to Eq. This representation is the most convenient for calculations. 2 ( 0.51) (written in rate form).4 shows a twodimensional representation of one half of the yield surface (Qtm ≤ 0) in terms of the two parameters β and γ . Figure 1.5 − β ) ( 0.67) replacing the generalized strain increment vector dε . For the evaluation of the power of internal forces. provided the reference values ε 0 and κ 0 are replaced by reference strain and curvature rates ε 0 and κ 0 .3.68) where D p and d p are the physical and normalized plastic dissipations per unit area of the midplane of the shell. however.5 + β ) + γ − ( 0. −1 2 16 J 2 P −1 ⎠ ⎝ 4 J 1P p p (1.69) Analytical evaluation of this relation gives d p (ε p ) = 2 2 ∫ 2 Pε + 2 Pεκ z + Pκ z dz 3 −1/ 1/ 2 (1. since nowhere on the yield surface do lines of constant values of the two parameters become parallel.5 + β ) ( 0.70) 18 . (1. The plastic dissipation function for a shell structure may be written in the form p ⎛n ⎞ ⎛ e ⎞ Dp = ⎜ ⎟ ⎜ p ⎟ = σT ε p d = N 0ε 0 ⎝ m ⎠ ⎝ k ⎠ T p (1.5 − β ) ⎞ ⎟. It is to be noted that the state relations are not affected.5 + β ) 2 2 + γ + β K0 . (1. (1.34) a dimensionless generalized strain rate vector ε is defined εT = ( e ⎛1 k ) = ⎜ ε11 ⎝ ε0 1 ε0 ε 22 1 ε0 2ε12 1 κ0 κ 11 1 κ0 κ 22 ⎞ 2κ 12 ⎟ κ0 ⎠ 1 (1. With the sixdimensional representation of the exact Ilyushin yield surface Eq.5 − β ) + γ + ( 0. 1.5 − β ) + γ + ( 0.⎛ K 0 = ln ⎜ ⎜ ⎝ K1 = ( 0. a description in terms of strain rate quantities is more convenient.66) 2 2 K 2 = ( 0. respectively. d p may be expressed as a function of the strain rate resultant quantities ε p −1 4 J 1P −1 ⎞ p 3 p T ⎛ J 0P d (ε ) = (ε ) ⎜ ⎟ε .5 + β ) + γ + 2 β K1 − γ K 0 .
in reliability analysis. Thus some expressions of quantities which are necessary for analysis algorithm should be reformulated. So we always can write σ y = Y σ 0 . in general. 2000] ⎧ Pε ⎪2 3 ⎪ p d =⎨ ⎪ Pκ β β 2 + γ + β β 2 + γ + γ K 2 2 0 ⎪ 3 1 1 ⎩ for Pκ = 0 (a ) (1. without the loss of generality. (1.71) is amenable to a numerical evaluation for all values of ε p . in this case. as will be discussed later. Thus.5 and γ = 0 are fulfilled. nonsmooth. (1. By that way the normalized quantities in Eqs. it is supposed that Pκ is always positive in order to have the expression of the plastic dissipation function as described in (1. However. material strength and thickness random variables.73) where σ 0 .27) and (1.3. To this end it is necessary to add to γ and to Pκ a small positive number. as long as γ is not exactly equal to zero.5 Reformulation For our general purpose to deal with probabilistic problems.5 − β and β 2 = 0. Additionally. as will be discussed later. In order to allow a direct nonlinear. d p is convex but not everywhere differentiable [Capsoni and Corradi.72) It is to be noted that the value of K 0 will become indefinite if both conditions β ≤ 0. the sensitivities are required which contain first and second derivatives versus loading. a “smooth regularization method” will be used for overcoming the nondifferentiability of the objective function. but to some small positive numbers.74a) 19 . 1.5 + β . Let us restrict ourselves to the case of homogeneous material and shells of constant thickness in which the yield limit σ y and thickness h are the same at every Gaussian point of the structure.71) for Pκ > 0 ( b) ( ) where β1 and β 2 are β1 = 0. Eq. the yield limit σ y and thickness h might be no longer constant but random variables. On the other hand. h0 are constant reference values and Y .which finally may be written as [Seitzberger. 2000]. a “regularized” evaluation of K 0 may be obtained [Seitzberger.71b). constrained optimization problem. Otherwise. 1997]. (1.28) assume the new form as follows N 0 = σ 0h0 = σ yh YZ = N0 YZ (1. Z are random variables. and then the reference quantities are also changed together with the different ‘realizations’ of random variables. (1.26). h = Zh0 (1.
h0 Yh Y (1. γ which were introduced in (1.78) 20 . β 2 = + β .M0 = σ0 h02 σ y h 2 M = = 02 2 4 Y 4Z YZ = (1.77) ˆ for Pκ > 0 (b) (β 1 β12 + γ + β 2 β 22 + γ + γ K 0 ) with the new β1 . σ= ⎜ 2 ⎟ ⎜ k⎟ ⎝ YZ m ⎠ ⎝Z ⎠ (1. β1 = − β . β 2 and K 0 are ⎛ β2 +γ + β ⎞ Z Z 1 ⎟. K 0 = ln ⎜ 1 2 ⎜ β +γ − β ⎟ 2 2 2 2 ⎠ ⎝ (1.62) with these new strain resultant intensities and then substitute them into (1.74d) With the new normalized quantities. the new “engineering” strain and stress vectors are obtained ⎛Ye ⎞ ⎛ YZn ⎞ ˆ ˆ ε = ⎜ Y ⎟ . β .75) and the incremental plastic strain resultant intensities have the form T 3 (Yε p ) P1 (Yε p ) = Y 2 Pε ( ≥ 0) 4 2 T ˆ = 3 ε p P ε p = 3 (Y ε p )T P ⎛ Y ε p ⎞ = Y P ˆ ˆ Pεκ εκ 2 2⎜ ⎟ ⎝Z ⎠ Z ˆ 3 ˆ Pε = ε p 4 ( ) T ˆ P1ε p = ( ) (1.74b) ε0 = and σ 0 (1 − ν 2 ) E σ y (1 − ν 2 ) YE = ε0 Y (1.74c) κ0 = 4ε 0 4 Z ε 0 Z = = κ0 . we obtain the new expression of the plastic dissipation function ⎧ ⎪ 2YZN 0ε 0 ⎪ p D =⎨ ˆ ⎪ P YN 0ε 0 κ ⎪ 3 ⎩ ˆ Pε 3 ˆ for Pκ = 0 (a) (1.76) ˆ ˆ Pκ = 12 ε p ( ) T 2 ⎛Y ⎞ ⎛Y ⎞ Y ˆ P3ε = 12 ⎜ ε p ⎟ P3 ⎜ ε p ⎟ = 2 Pκ ⎝Z ⎠ ⎝Z ⎠ Z p T ( ≥ 0) Adapting the parameters υ .71).
Such materials can deform considerably without breaking. many materials. All design codes are based on perfectly plastic models. exhibit distinct.. plastic properties. 1999] thus indicating the industrial need for LISA software. The extension of LISA to hardening materials is no problem [Staat and Heitzer. limit analysis evaluates the plastic collapse load or the largest load to which the structure would be subjected during its lifetime. not only can lowcycle fatigue cause structural failure below the plastic collapse load . the structure will fail due to global plastic flow. In Europe LISA has been developed as a direct plasticity method for the design and the safety analysis of severely loaded engineering structures. Prager [1972]. Among the plasticity methods. elasticplastic structural analysis becomes more general than the classical elastic one. This implies that structural failure occurs before yielding. such as nuclear power plants and chemical plants. the structure does not necessarily fail or deform extensively. Limit and Shakedown Analysis (LISA) seems to be the most powerful one. In the early 20th century. This implies that if the stress intensity reaches the critical (yield) value. the loads are generally timedependent or may vary independently. Limit analysis was pioneered approximately from the works of Kazincky in 1914 and Kist in 1917. offshore structures etc. Based on the elasticperfectly plastic or rigidperfectly plastic models of material and considering the loads as monotonic and proportional. Since then. 2003a]. 1961. such as the works of Hodge [1959. However. [Staat. Martin [1975]. However. [Staat and Heitzer. Lubliner [2005] and Capsoni and Corradi [1997]. it has been relatively easily defined by forcing the stress intensity at a certain point of the structure to attain the yield stress of the material.2 MATHEMATICAL FORMULATIONS OF LIMIT AND SHAKEDOWN ANALYSIS IN GENERALIZED VARIABLES It is the objective of structural analysis to determine the load carrying capacity. 2002]. To this case. Beyond this limit. Annex B of the new European pressure vessel standard EN 134453 is based on LISA [European standard. the applications of limit analysis theory in engineering have been widely reported. for example the majority of metals. 1963]. 2002]. in order to permit higher loads. 200506]. [Taylor et al. The first complete formulations of both upper bound and lower bound theorems were established later by Drucker. Greenberg and Prager and the alternative formulation using rigidplastic material was given by Hill in 1951. even after the stress intensity attains the yield stress. in practice. Practical experience showed that in the case of variable repeated loads. Maier [1970].
calculated with limit analysis but also an accumulation of plastic deformations may occur. [1997a. 2000]. a direct application of the limit and shakedown theorems is cumbersome if possible at all. [1993a. [1976. Weichert et al. The mathematical difficulties arising from the need to use a threedimensional analysis are difficult to be overcome. The theory of limit and shakedown analysis were established long time ago. Koiter [1960] developed a general kinematic shakedown theorem based on an analogy to limit analysis. [1997. Such an approach allows one to reduce complicated threedimensional problems to simple plane or onedimensional ones. two fundamental theorems of limit and shakedown analysis are introduced. The external loading P consists of general body force f in V and surface traction t on ∂Vσ . integrated variables. e. In this way. resulting in excessive deflections of the structure [König.e. As α increases and reaches a special value. Based on the original ones which were proposed by Melan and Koiter.1) where P0 = ( f0 . is also presented. 1972]. an extension for lower and upper bound theorems in terms of generalized variables are proposed and formulated.1 Introduction Let us consider a structure of volume V made of elasticperfectly plastic or rigidperfectly plastic material and subjected to external loading P . Polizzotto et al. This state of stress is called elastic limit. In the case of more complex structures. i. Further 22 . This static theorem was then extended by Melan [1936] to more the general case of a continuum. the static and kinematic theorems. Maier [1969. A simple approach for direct calculation of the shakedown limit as the minimum of incremental plasticity limit and alternating plasticity limit. namely the unified shakedown limit method.1. Ponter et al. We assume that all loads are applied in a monotonic and proportional way P = α P0 (2. 2002]. He stated and proved the plastic analysis theorems. 1973]. Since then. t0 ) denotes the nominal or initial load. the limit analysis and the shakedown ones in the form used nowadays. The first static shakedown theorem was formulated by Bleich [1932] for a system of beam of ideal Icrosssections.1. which are used to describe the static and kinematic quantities. 2003b]. a new branch of plasticity. 1990]. the theory of shakedown came to existence. Staat et al. large amount of work were reported in the literature. Theory of limit analysis 2. cf. In this case it is convenient to introduce some generalized. the body behaves elastically. [1988. However their numerical applications encountered some difficulties and were limited to some relatively simple structures. the first point in the body reaches the plastic state. 1969. Nguyen Dang et al. shelllike structures. In this chapter. If the value of α is sufficiently small. b]. It may also happen that the structure comes back to its elastic behaviour after a certain time period. 2001. 2. 1997b. In an alternative way. 1987]. König [1966.g.
1959]. bypassing the spreading process of the plastic flow.3) where Eq. L denotes a linear operator (usually differential one) and f ( σ ) is the yield function. (2. [Lubliner. the task of computing the limit load factor becomes a nonlinear optimization problem α l = max α l− ⎧ L [ σ ] + α l− P0 = 0 ⎪ s. the structure fails to support the applied load and collapses. The structure gradually forms a collapse mechanism.: ⎨ ⎪ f (σ) ≤ 0 ⎩ (a) (b) (2. 2. that is αl− ≤ αl .2 General theorems of limit analysis Lower bound theorem Based on the variational principle of Hill for perfectly plastic material. the value α l corresponding to the plastic collapse state is called the safety factor of the structure or the limit load multiplier. (a) is equilibrium condition.t.2) To prove this theorem the principle of virtual work and the property of convexity of the yield surface are used. 2005]. At limit state. The theory of limit analysis offers a way to solve directly the problem of estimating the plastic collapse load.1. also known as the principle of maximum plastic work. 1959]. Then. If P0 represents the applied load.increase of α will lead to the expansion of plastic region in the body. For the continuum we have L[σ ] = divσ . [Prager. The limit load and stresses so obtained are of great interest in practical engineering whenever the perfectly plastic model and small deformation assumption constitute a good approximation of the material.4) . the lower bound theorem of limit analysis can be stated as follows The exact limit load factor α l is the largest one among all possible static solutions α l− corresponding to the set of all licit stress fields σ . The limit value of the load is estimated and at the same time the limit state of stress in the whole structure can be evaluated. Upper bound theorem The upper bound theorem can be demonstrated as The actual limit load multiplier α l is the smallest one of the set of all multipliers α l+ corresponding to the set of all licit velocity fields u αl ≤ αl+ where 23 (2. [Hodge.
monotonic and proportional mechanical load (selfequilibrium thermal load has no effect on classical limit analysis). The upper bound theorem permits us to estimate the limit load factor by solving the following optimization problem (written in a normalized form) αl = min Win s. imposed on the actual stresses shift them towards purely elastic behaviour. These loads may be repeated (cyclic) or varying arbitrarily in certain domain. which are presented in the Breediagram (figure 2. Since for the elastic regime there are no plastic effects at all. They also avoid any plastic effects in the future provided that the loads are smaller than the initial overload. shakedown (adaptation). 1967]) together with the evolution of the structural response: elastic.5) (c) (d) u=0 on ∂Vu with Win and Wex the total power of the internal deformation and the power of the external loads of the structure.αl+ = Win Wex Win = ∫ D p (ε p )dV V (a) (b) Wex = ∫ f0T udV + V ∂Vσ ∫ t0T udS > 0 (2. both regimes are considered as safe working ones and they constitute a foundation for the structural design. [Bree. loads which are less than plastic collapse limit may cause the failure of the structure due to an excessive deformation or to a local break after a finite number of loading cycles. namely. however. inadaptation (nonshakedown). Inelastic structures such as for example pressure vessels and pipelines subjected to variable repeated or cyclic loading may work in four different regimes. (2. They clear out effects of all preceding smaller loads. whereas for the adaptation regime the plastic effects are restricted to the initial loading cycles and then they are followed by asymptotically elastic behaviour. structures are often subjected to the action of varying mechanical and thermal loading.1. The structural shakedown takes place due to development of permanent residual stresses which. In practice. Residual stresses are a result of kinematically inadmissible plastic strains introduced to the structure by overloads. At the limit load the structure looses instantaneously its load bearing capacity. Limit and shakedown analyses deal directly with the calculation of the load capacity or the maximum load intensities that the structure is able to support.6) 2.2. The inadaptation phenomena such as low cycle fatigue and or ratchetting should be avoided since they lead to a rapid structural failure.t.1 Introduction It has been understood in limit analysis considered above that the loading is simple. 24 . and limit (ultimate) state. Theory of shakedown analysis 2.2. In this case.: Wex = 1 . We do not consider elastic failure such as buckling or high cycle fatigue here.
The maximum safe load is defined as the shakedown load avoiding low cycle fatigue and ratchetting. the main problem of shakedown theory is to investigate whether or not a structure made of certain material will shake down under the prescribed loads. one can see that the first and second situations may not become dangerous but maximum exploitation of materials can only be attained in the adaptation or shakedown case. Pk+ ⎤ = ⎡ μk− .1 Breediagram of a pressurized thin wall tube under thermal and mechanical loads Viewing the situations above.7) 25 . σ0 σ σ ε ε ratchetting low cycle fatigue σ 2 σ ε collapse ε 1 elastic shakedown pure elastic behaviour 0 σmech. in limit and shakedown analyses the knowledge of the exact load history is not necessary. σther. σ0 1 Figure 2. Thus.Therefore. μk+ ⎤ Pk0 . k = 1.2 Definition of load domain We study here the shakedown problem of a structure subjected to n timedependent (thermal and mechanical) loads Pk0 (t ) with time is denoted by t . n . ⎣ ⎦ ⎣ ⎦ (2. 2.2. Only the maximum loads (limits) count and the envelopes should be taken into consideration. each of them can vary independently within a given range Pk0 ( t ) ∈ I k0 = ⎡ Pk− .
8) where μk− ≤ μk ( t ) ≤ μk+ . P2 αL P3 μ+ 2 L (2.2 for two variable loads.8) ˆ ˆ σ E ( x. t ) = ∑ μk ( t ) σ Ek ( x ) k =1 n (2. (1. To this end. t ) + ρ ( x ) (2. we use here the notion of a fictitious infinitely elastic structure which has the same geometry and elastic properties as the actual one.9) P2 μ1 μ2 P4 + μ1 P1 P1 Figure 2. it is useful to describe this load domain in the generalized stress space. This fictitious elastic generalized stress vector may be written in a form similar to (2.2 Two dimensional load domain L and L '=α L In the case of shell structures. Let the crosssections of the shell be identified by a vectorvariable x .10) ˆ with the fictitious elastic generalized stress vector σ E ( x. The actual ‘engineering’ stress field in Eq. t ) is defined as that would appear in the fictitious infinitely elastic structure if this structure was subjected to the same loads as the actual one.75) in the elasticplastic shell can be expressed in the following way ˆ ˆ σ ( x.11) 26 . This load domain can be represented in the following linear form P(t ) = ∑ μ k (t )Pk0 k =1 n (2. k = 1.These loads form a convex polyhedral domain L of n dimensions with m = 2 n vertices in the load space as shown in figure 2. n . t ) = σ E ( x.
most of the authors have formulated shakedown criteria in this way.3. t ) of the structure under ˆ consideration is equal to σ E ( x. 2.ˆ where σ Ek ( x ) denotes the generalized stress vector in the infinitely elastic (fictitious) structure when subjected to the unit load mode Pk0 . This residual generalized stress field satisfies the homogeneous static equilibrium and boundary conditions (2.3 Fundamental of shakedown theorems Static shakedown theorem The decomposition (2. boundedness of the total plastic work. t →∞ t →∞ lim ρ ( x ) = 0. Following (2. It is shown in this case.13) where S denotes the middle surface of the shell. ρ ( x ) denotes a timeindependent residual generalized stress field.a) for P = 0 . without its maximum value being specified. perfectly plastic material to shake down elastically is that the plastic generalized strains and therefore the residual generalized stresses become stationary for given loads P ( t ) ˆ lim ε p ( x.2. which expresses 27 . let us define a load domain L '=α L such that when subjected to ′ ˆ L ' the fictitious elastic generalized stress vector σ E ( x. plastic strain ceases to develop and the structure returns to elastic behaviour. t ) multiplied by a load factor α n ′ ˆ ˆ ˆ σ E ( x . t ) = α σ E ( x . t ) = 0. k =1 (2. ∀ x∈S (2.11). This situation means that after some time t or some cycles of loading. the problem is to find the largest value of α which still guarantees elastic shakedown. Indeed.14) may be considered as an intuitive examination which verifies if a given structure is going to shake down. the total amount of plastic energy dissipated over any possible load path within the domain L ' must be finite ˆ ˆ Win = N 0ε 0 ∫ σT ε p dt < ∞ . However.10) is an extension of the classical formulae.12) In shakedown analysis. for example when low cycle fatigue is considered. Furthermore. seems to be sometimes too weak a requirement. 0 t (2. it should be noted that such a concept leads to an approximate description: the total energy dissipated may be finite even if plastic strain increments appear at every load cycle but comprise a convergent series. One criterion for an elastic.14) The inequality (2. t ) = α ∑ μ k ( t ) σ Ek ( x ) .
statically admissible. By comparing with the classical static shakedown criterion. t ) + ρ ( x ) ) ≤ 0 .: ⎧ L [ρ] = 0 ⎪ ⎨ ˆE ⎪ f (α σ ( x. 2. The same as proposed by Koiter for plastic strain field. t ) ) = f ( σ E ( x. one can easily see that the shakedown of a shell structure is equivalent to the existence of a timeindependent residual generalized stress field ρ ( x ) in structure such that it does not anywhere violate the yield criterion ˆ ˆ f ( σ ( x. t ) + ρ ( x ) ) ≤ 0 ∀t ⎩ (a) (b) (2. From the above static theorem. The following theorem shows that this is the necessary and sufficient condition for a structure to shake down Theorem II. The obtained shakedown load multiplier α − is generally a lower bound. Shakedown occurs if there exists a timeindependent residual generalized stress field ρ ( x ) . Sawczuk [1969a. t ) ) = f ( σ E ( x.1: 1. The plastic ˆ generalized strain rate ε p may not necessarily be compatible at each instant during the time cycle T but the plastic generalized strain accumulation over the cycle 28 . b]. we ˆ introduce here an admissible cycle of plastic generalized strain field Δε p . t ) ) = f ( σ E ( x.18) Kinematic shakedown theorem Using plastic strain field to formulate shakedown criterion. t ) + ρ ( x ) ) ≤ 0 (2.16) ˆ ˆ f ( σ ( x.t.17) Based on the above static theorem. The theorem was given by Koiter [1960] and some of its applications in analysis of incremental collapse were derived by Gokhfeld [1980]. kinematic shakedown theorem is the counterpart of the static one.15) where f denotes the yield function in term of generalized variables.17). (2. t ) + ρ ( x ) ) < 0 . which states for stress variables. such that ˆ ˆ f ( σ ( x. we can find a permanent statically admissible residual generalized stress field in order to obtain a maximum load domain α L that guarantees (2. the shakedown problem can be seen as a mathematical maximization problem in nonlinear programming α − = max α s. Shakedown will not occur if no ρ ( x ) exists such that (2.in terms of stresses.
(2. differential or algebraic operator.t: ⎧ p ˆ ˆp ⎪ Δε = ∫ ε dt ⎨ 0 ⎪ Δε p = R u ( x ) [ ] ⎩ ˆ T 29 .24) 0 S 0 S Based on the kinematic theorem. the following extension of Koiter theorem holds Theorem II. t )ε ijp ) dV + ∫ ti ui dSσ ⎥dt = ∫ ∫ N 0ε 0 ( σ E ( x.22) θ where σ ij ( x. T T T (2.23) 0 S 0 S 2. an upper bound of the shakedown limit load multiplier α + can be computed. t )) ε dSdt > ∫ ∫ D (ε )dSdt . u(x ) denotes the vector of displacements on the middle surface. The shakedown problem can be seen as a mathematical minimization problem in nonlinear programming α + = min T ∫∫D 0 S T p ˆ (ε p )dSdt T ˆE ˆp ∫ ∫ N 0ε 0 ( σ ( x.19) 0 is required to be kinematically compatible such that ˆ Δε p = R [ u( x )] and ˆE ˆp ∫ ∫ N 0ε 0 ( σ ) ε dtdS > 0 T S 0 T (2. Shakedown can not happen when the following inequality holds p ˆE ˆp ˆp ∫ ∫ N 0ε 0 ( σ ( x. t )) ε dSdt 0 S (2. From Eq.22).ˆ ˆ Δε = ∫ ε p dt p T (2. Shakedown may happen if the following inequality is satisfied p ˆE ˆp ˆp ∫ ∫ N 0ε 0 ( σ ( x. t ) ) ε p dSdt ∫ ⎢V ( ∫ ⎥ ∂Vσ 0⎢ 0 S ⎣ ⎦ T (2. The virtual power principle (1.2: 1. t ) is a selfequilibrium thermal stress due to the temperature field θ (t ) .21) where R is a linear.20) (2.14) permits us to write the external power in a more general form which contains the generalized variables T ⎡ ⎤ T θ ˆ ˆ f i ui + σ ij ( x. t )) ε dSdt ≤ ∫ ∫ D (ε )dSdt . T T T (2.25) s.
t ) + ε ( x. T ) . These inadaptation modes can be defined precisely in the following way Theorem II.3: 1.t ) in which a kinematically admissible plastic generalized strain increment Δ ε ( x. 1987]. While the former analyses separately two different failure modes: incremental plasticity (ratchetting) and alternating plasticity.28) 30 . t ) ⋅ Δ ε ( x ) Λ ( x. which leads to a kinematically admissible plastic generalized strain increment within a periodic interval (0. T ) such that the total increment of the plastic generalized strain Δε ( x ) over this period is zero. the latter analyses them simultaneously. an extension for generalized variables is presented. Both methods deserve special attention due to their role in structural computation. 0 T (2. T ) ) is a process of plastic deformation ε ( x. the two following methods can be applied: separated and unified methods. can be decomposed into two components of perfectly incremental collapse and alternating plasticity ˆ ε p ( x.27) into the shakedown condition (2. it is easy to see that any plastic ˆ generalized strain history ε p ( x. From the above definitions (2. (2.26) or (2.26) 2.2. T ) − ε ( x.0 ) is attained in a proportional and monotonic way.t ) within a certain time interval (0.27).23).0 ) = 0 Λ ( x. In the following sections. An alternating plasticity process is any process of plastic deformation ε ( x. T ) = 1 (2.26) and (2. incremental collapse and alternating plasticity may happen to combine. t ) = ε ( x. namely Δ ε = R [ u( x ) ] ε = Λ ( x. 2. t ) = ε ( x.27) The criteria of safety with respect to alternating plasticity or incremental collapse may be obtained by substituting the plastic strain history (2.In order to calculate the shakedown limit load multiplier.4 Separated shakedown limit As was mentioned above. t ) ≥ 0 Λ ( x.t ) dt = 0 . Δε ( x ) = ∫ ε ( x. from the original ones which state for local variables [König. t ) . A perfect incremental collapse process (over a certain time interval (0. t ) .
2.11).29).26) into (2. S S (2. p 0 S 0 S S T T (2. t ) ≠ 0 only when the product σ E ( x. Substituting (2. one obtains ˆ Wex = ∫ ∫ N 0ε 0 σ E ( x. t ) Λ ( x. In this case. the smallest upper bound of incremental limit could be attained when the external power Wex assumes its maximum and the internal dissipation Win takes its minimum. the shakedown load multiplier against incremental collapse α + can be formulated as a nonlinear minimization problem 31 .29) By taking into account the properties of the dissipation function and the plastic strain history (2.34) in which ⎧ μk+ μk = ⎨ − ⎩ μk if if ˆ σ Ek ( x ) Δ ε ( x ) ≥ 0 ˆ σ Ek (x ) Δ ε ( x ) < 0 (2.26). t ) Δ ε ( x ) takes its maximum possible value for a given load domain L .31) in which ˆ ˆ σ E ( x ) Δ ε ( x ) = max ( σ E ( x. the plastic strain field is assumed by (2.32) By this way. namely n independently varying loads.34).33) becomes ˆ α ∫ ∑ N 0ε 0 μk σ Ek ( x ) Δ ε ( x )dS ≤ ∫ D p ( Δ ε )dS S k =1 S n (2. t ) Δ ε ( x )dSdt = ∫ N 0ε 0σ E ( x ) Δ ε ( x )dS 0 S S T (2. the safety condition against any form of perfectly incremental collapse thus has the form ˆ α ∫ N 0ε 0σ E ( x ) Δ ε (x )dS ≤ ∫ D p ( Δ ε )dS . we can write Win = ∫ ∫ D ( ΛΔ ε )dSdt = ∫ Λ ∫ D p ( Δ ε )dSdt = ∫ D p ( Δ ε )dS . To this end. the formulation (2. (2.30) From the shakedown condition (2.8).23). t ) Δ ε ( x ) ) . (2.35) From condition (2. t ) is selected in such a ˆ way that Λ (x.1 Incremental collapse criterion If the safety condition against any form of perfectly incremental collapse is considered. t ) Δ ε ( x )dSdt ≤ Win = ∫ ∫ D p ( ΛΔ ε )dSdt . the external power Wex can be written as ˆ ˆ Wex = ∫ ∫ N 0ε 0σ E ( x.26). t ) Λ ( x.9) and (2.33) If the load variation domain is prescribed by (2.4. 0 S 0 S T T (2.2. the function Λ(x.
Alternating plasticity criterion If the safety condition against alternating plasticity is considered. does not violate the yield condition ˆ f ( σ E ( x. t ) dt ε ij α (x) 0 ⎧T p ⎪ ∫ D ( ε )dt = 1 ⎪0 ⎨T ⎪ ε ( x. the static shakedown condition against any form of alternating plasticity can be obtained A given structure is safe against alternating plasticity if there exists a timeindependent generalized stress field ρ which. T ∫ ε ( x.37) 2. t ) + ρ) ≤ 0 . if superimposed on the envelope of elastic generalized stresses. The shakedown condition (2. (2.41) It should be noted that the stress field ρ in (2.2. If we define a general stress response 32 . t ) dSdt ≤ ∫ ∫ D p ( ε )dSdt 0 S T (2.41) is an arbitrary timeindependent generalized stress field and not necessarily selfequilibrated as that in Melan’s theorem (2.40) s. the optimization problem leading to the most stringent limit condition can be established at each point x separately.4.36) α + = min Win s. t ) dt = 0 0 for all x ∈ S .24) in this case has the form ˆ α ∫ ∫ N 0ε 0 σ 0 S T E ( x.2. (2. t ) ε ( x.t : Wex = 1 .38) with α > 1.39) Starting from the kinematic theorem and the last constraint in (2. t )dt = 0 ⎪∫ ⎩0 T (2. t ) ε ( x.39).α + = min⎜ ⎜ or in normalized form ⎛ Win ⎝ Wex ⎞ ⎟ ⎟ ⎠ (2.16) and (2. The global safety factor against alternating plasticity limit will be the minimum of local α ( x ) defined as 1 ˆ = max ∫ N 0ε 0σ E ( x. the plastic strain field must be satisfied (2.t : By solving this problem. (2.17).27).
μk = μk+ − μk− 2 . in most cases it is impossible to apply lower and upper theorems to find directly the shakedown limit defined by the minimum of incremental 33 . k =1 n (2. (2. 2003]. see Polizzotto [1993a. the optimal timeindependent generalized stress field ρ can be defined by ˆ ρ = − ∑ μk σ Ek .5 Unified shakedown limit In practical computation.43) In view of (2.44) where f = F −1. the plastic shakedown load multiplier (lower bound) may be calculated as α = min x 1 ˆ F ( σ ( x ) + ρ( x ) ) * (2.ˆ ˆ σ* ( x ) = ∑ ( μk + μk )σ Ek ( x ) k =1 n (2. 2.40).45) The sign of μk must be chosen so that the value of function F is maximal.46) Then the plastic shakedown limit load multiplier can be finally represented as α = min x 1 ⎛ ⎞ ˆ F ⎜ ∑ μk σ Ek ( x ) ⎟ ⎝ k =1 ⎠ n (2.40)..2.42) ˆ where σ Ek ( x ) is the elastic generalized stress field in the reference structure when subjected to the kth load and μk = μk+ + μk− 2 . As counterparts of the static condition (2. the alternating plasticity limit is proportional to the current material strength for an isotropic strain hardening material [Yan et al.47) it can be shown that the constant loads have no influence on the plastic fatigue limit if these constant loads do not change the geometry and material properties. On the other hand. a kinematic condition as well as an upper bound of the plastic shakedown load multiplier can be found. Kinematical strain hardening has also no influence on the alternating plasticity limit because it does not change the allowed stress variation.47) where all the combinations of the signs ± must be accounted for. (2. b] for further details. From equation (2. By considering the alternating characteristic of the stress corresponding to an alternating strain rate.
These obstacles can be overcome with the help of the following two convexcycle theorems. Polizzotto [1991].3 Critical cycles of load for shakedown analysis These theorems. Another scheme (figure 2. 34 . Based on these theorems. Theorem II. introduced by König and Kleiber [1978].4: “Shakedown will happen over a given load domain L if and only if it happens over the convex envelope of L ”. P2 P3 P2 P3 P2 P2 P1 P1 P4 P1 P4 P1 a) A cycle connecting all vertices of L with the origin b) A cycle covering the vertices of L on the entire boundary Figure 2. König and Kleiber suggested a load scheme as shown in figure 2. This scheme was applied in a simple stepbystep shakedown analysis by Borkowski and Kleiber [1980]. The difficulty here is the presence of the ˆ timedependent generalized stress field σ E ( x.25). They allow us to examine only the stress and strain rate fields at every vertex of the given load domain instead of computing an integration over the time cycle.b) was adopted later by Morelle [1984]. which hold for convex load domains and convex yield surfaces.3. Extensions and implementations of these theorems can also be found in the works of Morelle [1989].plasticity limit and alternating plasticity limit.5: “Shakedown will happen over any load path within a given load domain L if it happens over a cyclic load path containing all vertices of L ”. t ) in (2. permit us to consider one cyclic load path instead of all loading history. Nguyen and Morelle [1990]. Jospin [1992].3.16)÷(2.17) or the time integration in formulation (2. Yan [1997].a for two independently varying loads. Theorem II.
n is the total number of varying loads. (2. Obviously.48) where m = 2 n is the total number of vertices of L .Let us restrict ourselves to the case of a convex polyhedral load domain L . the generalized strain at any instant t is represented by ˆ( ˆ ε t ) = ∑ μ ( t − tk ) ε k .49) μ (t − tk ) = ∫ δ [ x − (t − tk )]dx . We have the following static shakedown theorem 35 .48) ∞ −∞ ∫ δ ( x )dx = 1 if if t = tk t ≠ tk (2. This remark permits us to ˆ replace the timedependent generalized stress field σ E ( x. the Melan condition required in the whole load domain will be satisfied if and only if it is satisfied at all vertices (or the above special loading cycle) of the domain due to the convex property of load domain and yield function.51) must be kinematically compatible. t ) and time integrations in the lower and upper shakedown theorems. The question is how to apply the above theorems to eliminate timedependent elastic ˆ generalized stress field σ E ( x. In order to do so. t ) by its values calculated only at load vertices.50) At each instant (or at each load vertex). k (2. −∞ ∞ Over this load path. δ (t k ) is the Dirac distribution with the property that f ( tk ) = −∞ ∫ ∞ f ( x )δ (t − tk )dx . T ) passing through all vertices of the load domain L such as ˆ P( x. however the accumulated generalized strain over a load cycle ˆ ˆ Δε = ∑ ε k k =1 m (2. let us consider a special load cycle (0. the kinematical condition may not be satisfied. Intuitively it is understood by ⎧ ⎪ ∞ and δ ( t − tk ) = ⎨ ⎪ ⎩0 We define for eq. t ) = ∑ μ (t − tk )Pk ( x ) k =1 m (2.
54) ∀k = 1. such that ˆ ˆ f σ E ( x.53) 1. Pk ˆ ∫ ⎪ k =1 S ⎩ m (a) (b) (c) (2.t.52) and the kinematic theorem (2.: ( ( )) T ˆ ε k dS = 1 (d) 36 .52) The application of loading cycle (2.6: The necessary and sufficient condition for shakedown to occur is that there exists a permanent residual generalized stress field ρ ( x ) .48) also leads to the elimination of time integration in kinematic shakedown condition as stated in the theorem hereafter Theorem II.18) and (2. m (c) 2. m . Pk ) + ρ ( x ) ≤ 0 ( ) ∀k = 1.7: The necessary and sufficient condition for shakedown to occur is that there exists a plastic ˆ accumulation mechanism ε k such that ⎧m ˆ ˆE ⎪ ∑ ∫ N 0ε 0 σ x.t.25) now can be reformulated in simpler forms corresponding to the static theorem (2.: (a) (b) ⎧ L [ρ] = 0 ⎪ ⎨ ˆ ˆE ⎪ f α σ x.55) s.Theorem II. The lower bound α − = max α s. statically admissible. Pk k =1 S ⎪ ⎪ m ⎨ ˆ ˆ Δε = ∑ ε k ⎪ k =1 ⎪ ⎪ Δε = R [ u( x )] ⎩ ˆ ( ( )) T ˆ ˆ ε k dS ≤ ∑ ∫ D p (ε k )dS k =1 S m (2. The upper bound (in normalized form) ˆ α = min ∑ ∫ D p (ε k )dS + k =1 S m ⎧ ˆ ˆ Δε = ∑ ε k ⎪ k =1 ⎪ ⎪ ˆ ⎨ Δε = R [ u ( x ) ] ⎪m ˆ ⎪ ∑ N 0ε 0 σ E x. Pk + ρ ( x ) ≤ 0 ⎩ ( ( ) ) (2.53) The bounds of shakedown limit load multiplier (2. (2.
9) one has μ1− = μ1+ . 37 . This fact means that limit analysis can be considered as a special case of shakedown analysis. (2.3) and (2.Let us note that if there is only one load and this load does not vary then according to load domain definition (2.6).56) In this case it is easy to see that the above upper bound and lower bound reduce to the formulations of upper and lower bounds of limit load factor (2.
2004. Being considered as a nonlinear programming technique. Displacement finite elements work best with the upper bound approach in which kinematic conditions are satisfied. we determine the maximum load factor for shakedown. i.e. In lower bound shakedown analysis. the lower bound and upper bound theorems. 1999].e. This can be achieved by performing a multilayer analysis which requires the knowledge of the stress distribution through the thickness and thus requires extensive computing resources both of time and storage.. the minimum and maximum problems resulting from the lower bound and upper bound theorems are dual. Studies showed that they converge to the same load factor [Vu et al.. The . Direct procedures are thus necessary. 2007]. Recently many largescale optimization methods have been developed for LISA with continuum finite elements in an European research project [Staat et al. Actually. The research showed that today LISA can achieve exact values of the carrying capacity under monotone and under cyclic loading for bounded linearly kinematic hardening material with the potential inclusion of continuum damage and moderately large deformations. the problem of finding the limit and shakedown load factors can be discretized and transformed into a problem of mathematical programming. 2003a].. the elastic compensation method modifies the Young's modulus of each element during an iterative linearelastic finite element calculation in order to obtain an optimized statically admissible stress field. An upper bound and a pseudolower bound can be obtained after each iteration. The elastoplastic analysis of failure loads and failure modes for shell structures has been a problem of great interest to many designers. With the help of the finite element method. In [Taylor et al. elastic compensation and the deviatoric map have been used for LISA of perfectly plastic shell structures by the finite element method. Users of the upper bound method sometimes state that the lower bound approach should be used with stressbased finite elements because otherwise the bounding character is lost. i.3 DETERMINISTIC LIMIT AND SHAKEDOWN PROGRAMMING Based on two theorems. the exterior approximation of the load domain will be sought in upper bound analysis. we calculate the interior approximation of the load domain. different numerical methods for LISA which deal with both linear and nonlinear programming were built to analyse complicated structures which analytical tools fail to deal with. In fact it is very cumbersome to use “stepbystep” procedures in solving the problem of limit and shakedown analysis. Alternatively. we determine the minimum load factor for nonshakedown.
Dealing with singular dissipation function in limit analysis. [1995.1) where m is the number of vertices of the load domain L and ne is the total number of elements. e]. However.33. 2000] introduced a 39 .a) becomes m ne ∑∑ ∫ D k =1 e =1 Se p ˆ (ε k )dS (3. Franco and Ponter [1997a. Bisbos and Pardalos [2007] formulated a lower bound shakedown analysis of steel shells as a specific mathematical optimization problem known as secondorder cone programming by using the linear approximation of the exact Ilyushin yield surface (usually referred to as Ilyushin’s generalized yield model) In this chapter. c. Carvelli et al. The approach was initially proposed by Zhang [1995b] and further developed by Liu et al. Starting from a finite element discretization. the objective function (3.more effective method is to deal with a single layer analysis which relates to stress resultants. • if the rigid or elastic perfectly plastic material model is adopted with the Ilyushin yield criterion. Yan and Nguyen Dang [2000] proposed an approximate stress distribution through the thickness of the shell and simplified the von Mises criterion for generalized variables to get a yield condition in stress resultants. such as nonsmooth and singular objective functions. the integration (2. which is adopted for shell structures [Tran et al. 1998a.1 Finite element discretization By subdividing the whole middle surface S into elements Se . Bisbos and Papaioannou [2006]. 3. A simple technique for overcoming numerical obstacles. 1998b. a detailed kinematic algorithm in terms of generalized variables will be formulated and introduced. 2007a. limit analysis based on kinematic formulation suffers from the same difficulty. Being considered as a special case of shakedown analysis.1) is only differentiable in the plastisfied region of structures while powerful optimization methods require its gradient to be available everywhere. The two major numerical obstacles that appear here in dissipation function are ˆ • the value of K 0 in the expression of the dissipation function D p (ε k ) may become infinite in the case of γ = 0 (as mentioned in the first chapter). is also proposed.. in this case wellformulated finite element methods for the direct evaluation of shakedown limits are currently still rare due to the complexity of geometrical shape and the mathematical difficulties in the stress resultant yield criterion. Vu [2001] applied this in conjunction with Newton’s method to establish a numerical algorithm for dual shakedown limits. [1999]. b] developed an approximate technique which is based on the reformulation of the kinematic shakedown theorem for an axisymmetric shell with piecewise linear yield conditions. we present a kinematic approach of limit and shakedown analysis. [1997]. Andersen et al.
“smooth regularization method” by replacing the original dissipation function D p ( ε k ) by
2 its disturbed one D p ( ε k ,η0 ) . In this function η0 is a very small number and by that way
all elements are seen as plastified or on the plastified verge. The technique is actually the wholeregion regularization method mentioned in [Yan, 1997] and will be adopted here in our analysis. In order to overcome the two obstacles above, we replace γ and the bending strain
ˆ resultant intensity Pκ in the expression of the dissipation function by the new ones ˆ γ + η02 and Pκ + η02 . By that way, the dissipation function is finite and differentiable
everywhere, even in the nonplastisfied region. Furthermore, since the bending strain ˆ resultant intensity P is nonnull, thus, the dissipation function now is
κ
ˆ 2 D p ε k ,η0 = YN 0ε 0
(
)
ˆ Pκ β1 β12 + γ + β 2 β 22 + γ + γ K 0 3
(
)
(3.2)
with β1 , β 2 , K 0 , γ calculated by Eqs. (1.78) and (1.62). It is to be noted that, the quadratic strain resultant intensities now become
ˆ 3 ˆ Pε = ε k 4 ˆ ˆ P =3 ε
εκ
ˆ Pκ
ˆ ( ) Pε ˆ ( ) Pε ˆ ˆ = 12 ( ε ) P ε +η
T 1 k T k 2 k T k 3 k
( ≥ 0) (3.3)
2 0
(>0)
ˆ ˆ where ε k is the generalized strain rate vector corresponding to load vertex Pk ,
ˆ εk
⎡1 k = ⎢ ε11 ⎣ε0
1
ε0
ε
k 22
1
ε0
2ε
k 12
1
κ0
κ
k 11
1
κ0
κ
k 22
⎤ 2κ ⎥ . κ0 ⎦ 1
k 12
T
(3.4)
It should be noted that the incompressibility condition, although it is true for plastic deformation of metal, introduces some numerical difficulties. Fortunately, by using plane stress or shelltype finite elements, this condition can be naturally achieved by adopting the Kirchhoff’s hypothesis. This can reduce the computational cost in the optimization process. The integrations
∫D
Se
p
ˆ (ε k )dS in objective function (3.1) are calculated by Gauss
Legendre integration technique. From equations (3.1) and (3.2) one has
ˆ ∑∑ ∫ D p (ε k )dS = ∑∑ ∫ YN 0ε 0
k =1 e =1 Se k =1 e =1 Se m NG m ne m ne
ˆ Pκ β1 β12 + γ + β 2 β 22 + γ + γ K 0 dS 3 ˆ Pκ β1 β12 + γ + β 2 β 22 + γ + γ K 0 3
(
)
= ∑∑ wiYN 0ε 0
k =1 i =1
(
)
(3.5)
40
where wi is the weighting factor of the Gauss point i . The same integration technique is applied to evaluate external power (2.33.d) ˆ ˆ ∑ ∫ N 0ε 0 σ E x, Pk
k =1 S m
( (
))
T
ˆ ε k dS
ˆ ˆ = ∑∑ ∫ N 0ε 0 σ E x, Pk
k =1 e =1 Se m NG
m
ne
( (
))
T
ˆ ε k dS (3.6)
ˆ ˆ = ∑∑ wi N ε ε σ
k =1 i =1 T 0 0 ik
E ik
ˆ where εik denotes the plastic generalized strain rate vector corresponding to vertex k of
the load domain and calculated at the Gauss point i in the structure. The compatible condition (2.33.c) is verified at each Gauss point i
ˆ ∑ε
k =1 m ik
= Bi u
(3.7)
with Bi denotes the deformation matrix B(x ) at Gauss point i . From (3.5÷3.7) the shakedown limit load multiplier may be formulated as
α + = min ∑∑ wiYN 0ε 0
k =1 i =1
m NG
ˆ Pκ β1 β12 + γ + β 2 β 22 + γ + γ K 0 3
(
)
(3.8)
s.t:
⎧m ˆ ∀i = 1, NG ⎪ ∑ εik = Bi u ⎪ k =1 ⎨ m NG ⎪ ˆ ˆ w N ε εT σ E = 1 ⎪ ∑∑ i 0 0 ik ik ⎩ k =1 i =1
In the above formulation, for simplicity, the sums over all elements and over all Gauss points of one element (in the expressions of internal and external powers) are unified in the sum over all Gauss points of discretized structure. The total number of Gauss points in structure is NG = ne × ng , where ng is the number of Gauss point on an element.
3.2 Kinematic algorithm
By restricting ourselves to polyhedral form of load domain, in this section we introduce a kinematic algorithm for limit and shakedown analysis of shell structures. To begin with, for the sake of simplicity, let us rewrite the upper bound limit (3.8) in a simpler form by introducing some new notations • The new strain rate vector eik
ˆ e ik = wi εik .
(3.9)
•
The new fictitious elastic generalized stress field t ik
ˆE t ik = N 0ε 0σik .
(3.10)
41
•
ˆ The new deformation matrix B i ˆ Bi = wi Bi .
(3.11)
Following these definitions, the objective function in (3.8) becomes
α + = min ∑∑ YN 0ε 0
k =1 i =1
m NG
ˆ Pκ β1 β12 + γ + β 2 β 22 + γ + γ K 0 3
(
)
(3.12)
ˆ where β1 , β 2 , γ , K 0 , Pκ are calculated by Eqs. (1.72), (1.62), (1.66) and (3.3) in which the ˆ generalized strain vector εik is replaced by the rate eik . It is to be noted that this replacing ˆ does not affect the value of β1 , β 2 , γ , K0 , but Pκ .
By substituting (3.9)÷(3.12) into (3.8) one obtains a simplified version for upper bound of shakedown limit load
α + = min ∑∑ YN 0ε 0
k =1 i =1
m NG
ˆ Pκ β1 β12 + γ + β 2 β 22 + γ + γ K 0 3
(
)
(3.13)
s.t:
⎧m ˆ ⎪ ∑ e ik = Bi u ⎪ k =1 ⎨ m NG ⎪ ∑∑ eT tik = 1 ⎪ k =1 i =1 ik ⎩
∀i = 1, NG
Dealing with the nonlinear constrained optimization problem (3.13), the penalty method is used for satisfying the compatibility condition. To this purpose, let us write the penalty function as
⎧ ˆ P ⎪m ⎛ FP = ∑ ⎨ ∑ ⎜ YN 0ε 0 κ β1 β12 + γ + β 2 β 22 + γ + γ K 0 3 i =1 ⎪ k =1 ⎜ ⎩ ⎝
NG
(
)
T ⎫ m ⎞ c⎛ m ˆ u ⎞ ⎛ ∑ e − B u ⎞⎪ ˆ ⎟ + ⎜ ∑ eik − Bi ⎟ ⎜ ik i ⎟⎬ ⎟ 2 ⎝ k =1 ⎠ ⎝ k =1 ⎠⎪ ⎠ ⎭ (3.14)
where c is penalty parameter such that c >> 1 . This parameter c may be dependent on integration points or load vertices and c should be adjusted to fit different compatibility criteria. However, at this stage, for the sake of simplicity, c is let to be constant everywhere. Theoretically, when c goes to infinity we will recover related conditions. Following (3.14) the modified kinematic formulation (3.13) becomes
α + = min FP
s.t:
(a)
(3.15)
∑∑ e
i =1 k =1
NG m
T ik ik
t =1
(b)
The corresponding Lagrange function of (3.15) is
42
⎛ NG m ⎞ FPL = FP − λ ⎜ ∑∑ eT t ik − 1⎟ . ik ⎝ i =1 k =1 ⎠
(3.16)
Obviously, the normalization can be also treated by a penalty method, which may help to remove the Langrangian multiplier λ from our problem. However, the application of a penalty method to (3.15.b) leads to a system of equations which requires more effort and computer memory to solve. For this reason that λ is retained as variable and will be handled later on. The stationarity condition for the Lagrange function FPL states that
⎧ ⎪ ⎪ ⎪ ⎪ ∂FPL ⎪ ∂e ⎪ ik ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎪ ⎪ ∂FPL ⎪ ∂u ⎪ ⎪ ∂FPL ⎪ ⎩ ∂λ ⎛4 3 ⎞ ⎜ β1 β12 + γ + β 2 β 22 + γ + γ K 0 P3 ⎟ ˆ ⎜ Pκ ⎟ ⎜ ⎟ ⎜ 4 3 ⎟ 2 2 = YN 0ε 0 ⎜ + β1 + γ − β 2 + γ ( P2 + 4 β P3 ) ⎟ eik ˆ Pκ ⎜ ⎟ ⎜ ⎟ ⎜ + 3 K 0 P1 + 8β P2 − 16 (γ − β 2 ) P3 ⎟ ⎜ 2 P ⎟ ˆ κ ⎝ ⎠
(
)
(
)
(
)
(3.17)
(a) (b) (c)
⎛ m ˆ ⎞ +c ⎜ ∑ e ik − Bi u ⎟ − λ t ik = 0 ⎝ k =1 ⎠
NG m ˆ ⎛ ˆ ⎞ = − c ∑ BT ⎜ ∑ eik − Bi u ⎟ = 0 i i =1 ⎝ k =1 ⎠
∀ i, k
⎛ NG m ⎞ = ⎜ ∑∑ eT t ik − 1⎟ = 0 ik ⎝ i =1 k =1 ⎠
The first equation of (3.17) can be rewritten in the following form
⎛ ⎞ 2 2 ⎜ 4 3 β1 β1 + γ + β 2 β 2 + γ + γ K 0 P3 ⎟ ⎜ ⎟ m ⎜ +4 3 β 2 + γ − β 2 + γ ( P + 4 β P ) ⎟ e +c P ⎛ e − B u ⎞ − λ P t = 0 . ˆ ∑ ˆ ˆ YN 0ε 0 κ ⎜ κ ik ik i ⎟ 1 2 2 3 ⎜ ⎟ ik ⎝ k =1 ⎠ ⎜ ⎟ ⎜ + 3 K 0 P1 + 8β P2 − 16 (γ − β 2 ) P3 ⎟ ⎜ ⎟ ⎝ 2 ⎠ (3.18)
(
(
)
)
(
)
This form of stationary condition will help us reduce numerical difficulty in singular situation when ˆ Pκ → 0 . This means that (3.18) may be satisfied even when there is no plastic deformation and η0 is set equal to zero while in the same conditions the equation (3.17.a) is indefinite. For convenience, let us rewrite Eqs. (3.18), (3.17.b) and (3.17.c) as follows
43
19) where ⎛ ⎞ 2 2 ⎜ 4 3 β1 β1 + γ + β 2 β 2 + γ + γ K 0 P3 ⎟ ⎜ ⎟ ⎜ +4 3 β 2 + γ − β 2 + γ ( P + 4 β P ) ⎟ Hik = YN 0ε 0 1 2 2 3 ⎜ ⎟ ⎜ ⎟ ⎜ + 3 K 0 P1 + 8β P2 − 16 (γ − β 2 ) P3 ⎟ ⎜ ⎟ ⎝ 2 ⎠ ( ( ) ) (3.m ⎧ ˆ ⎛ ˆ ⎞ ˆ Hik e ik + c Pκ ⎜ ∑ e ik − Bi u ⎟ − λ Pκ t ik = 0 = fik ⎪ ⎝ k =1 ⎠ ⎪ NG ⎪ NG T ⎛ m ⎪ ˆ ˆ ⎞ ⎨ ∑ Bi ⎜ ∑ e ik − Bi u ⎟ = 0 = ∑ hi i =1 ⎝ k =1 ⎠ ⎪ i =1 ⎪ NG m T ⎪ ∑∑ e ik t ik − 1 = 0 ⎪ i =1 k =1 ⎩ (a) (b) (c) (3.22) with 44 .21) where Mik = Hik + Dik + ⎞ T 12 ⎛ ⎛ m ˆ ⎞ ⎜ c ⎜ ∑ eik − Bi u ⎟ − λ t ik ⎟ eik P3 ˆ ⎠ Pκ ⎝ ⎝ k =1 ⎠ (3.19) one gets m ⎧ ⎞ ˆ ⎛ ˆ ˆ M ik de ik + c Pκ ⎜ ∑ de ik − Bi du ⎟ − d λ Pκ t ik = −fik ⎪ ⎝ k =1 ⎠ ⎪ NG m NG ⎪ ⎞ ⎪ ˆT ⎛ ˆ ⎨ ∑ Bi ⎜ ∑ deik − Bi du ⎟ = − ∑ hi i =1 ⎝ k =1 ⎠ ⎪ i =1 NG m NG m ⎪ T T ⎪ ∑∑ de ik t ik = − ∑∑ e ik t ik + 1 i =1 k =1 ⎪ i =1 k =1 ⎩ (a) (b) (c) (3.20) ( ) By applying NewtonRaphson’s method to solve the modified system (3.
The substitution of (3.25) where ˆ − Ei = Ii + c ∑ Pκ Mik1 k m (3. m and summing them up.23) In the system (3. after some manipulations one has ∑ de k m ik − ˆ − ˆ ˆ − = −Ei−1 ∑ Mik1fik + cEi−1 ∑ Pκ Mik1Bi du + d λ Ei−1 ∑ Pκ Mik1t ik k k k m m m (3.26) and I i is the identity matrix.24) for k = 1.a) to lefthand side and then. multiplying both sides − with Mik1 we get m ⎞ − ˆ − ⎛ ˆ ˆ − de ik = − Mik1fik − c Pκ M ik1 ⎜ ∑ de ik − Bi du ⎟ + d λ Pκ Mik1t ik .25) in (3. By moving the last two terms in (3. ⎝ k =1 ⎠ (3. du denote the incremental vectors of generalized strain rate and displacement respectively while d λ denotes the incremental value of λ .21.21) de ik .24) Writing (3.B = P1 + 8β P2 − 16 (γ − β 2 ) P3 ⎛ ⎛ 48 3 ⎜⎜− β12 + γ − ˆ ⎜⎜ Pκ ⎜⎜ ⎜⎜ 3 3 ⎛ 1 ⎜⎜+ ⎜ − ˆ ⎜ ⎜ Pκ ⎜ β12 + γ ⎝ ⎜⎝ ⎜ ⎛ ⎛ 1 ⎜ ⎜3 3⎜ − ˆ Dik = YN 0ε 0 ⎜ ⎜ Pκ ⎜ β12 + γ ⎝ ⎜ ⎜ +⎜ ⎜ ⎛ ⎜ ⎜ 3 3⎜ ⎜ ⎜ + 8P ⎜ ˆ 2 κ ⎜ β1 + γ ⎜ ⎜ ⎝ ⎝ ⎜ ⎜ 6 3 T ⎜ − ˆ K 0 P3eik e ik C ⎜ Pκ ⎜ ⎝ A = P2 + 4 β P3 ( β 22 + γ P3 + 1 ) β2 24 3 ⎛ β1 ⎜ + ˆ Pκ ⎜ β12 + γ β 22 + γ ⎝ ⎞ 24 3 ⎟B − K 0 P2 ˆ Pκ β 22 + γ ⎟ ⎠ ⎞ ⎟A β 22 + γ ⎟ ⎠ 1 1 ( β12 + γ + β1 ) − β 22 + γ ( ⎞ ⎞ ⎞ ⎟ ⎟A⎟ ⎟ ⎟ ⎟ ⎠ ⎟ T eik eik A ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ ⎠ ⎟ ⎟ ⎞ ⎟ ⎟ ⎟ ⎟ ⎟ ⎟ T ⎟ eik e ik B ⎟ ⎞ 1 ⎟ ⎟B⎟ ⎟ ⎟ ⎟ β 22 + γ − β 2 ⎟ ⎟ ⎟ ⎠ ⎠ ⎟ ⎟ ⎟ ⎟ ⎟ ⎠ ) C = P1 + 24 β P2 − 16 (γ − 5β 2 ) P3 (3.b) leads to 45 .21.
30) due to a nonsymmetry of the matrix S . A difficulty arises here in the system (3.24) and after some manipulations we get the incremental vectors of displacement and strain rate as ⎧ du = du1 + du 2 ( λ + d λ ) ⎪ ⎨ ⎪ de ik = ( de ik )1 + ( de ik ) 2 ( λ + d λ ) ⎩ where (a) (b) (3.28) Starting from the definition of Ei .30) where NG ⎧ ˆ ˆ S = ∑ BT Ei−1Bi i ⎪ i =1 ⎪ NG m ⎪ − ˆ f1 = ∑ BT Ei−1 ∑ M ik1 ( M ik − Hik )e ik ⎨ i i =1 k =1 ⎪ NG m ⎪ ˆ ˆ − f2 = ∑ BT Ei−1 ∑ Pκ Mik1t ik ⎪ i i =1 k ⎩ (a) (b) (c) (3. The matrix Ei−1 plays the role of the elastic ⎣ ⎦ matrix at the Gauss point i while the vector f = f1 + f2 ( λ + d λ ) (3. (3. In order to reduce the computational costs.26).21.b).29) By introducing (3. and (3.34) 46 .30) with the last two terms on the righthand side may be interpreted as the linear system arising in purely elastic computations with the global stiffness matrix S and applied force vector ⎡ f1 + f2 ( λ + d λ ) ⎤ .30). (3. 2 ( ) (3.32) can be considered as the vector of nodal forces applied on the structure. m k (3.27) where K = I − cE −1 i ∑ k m ˆ − Pκ Mik1 .27) and take into account Eqs.21. we can further simplify (3. Eq. (3.29) in (3.31) The system (3.− ˆ − ˆ ˆ ˆ ∑ BT KBi du = ∑ hi − ∑ BT Ei−1 ∑ Mik1fik + d λ ∑ BT Ei−1 ∑ Pκ Mik1tik i i i i =1 i =1 i =1 k i =1 k NG NG NG m NG m (3. (3.28) as ˆ − K = I − cEi−1 ∑ Pκ Mik1 = I − Ei−1 ( Ei − I ) = Ei−1 .33) From (3. one can make the matrix Ei−1 symmetric by Ei−1 = T 1 −1 Ei + ( Ei−1 ) . (3.25).a). one has the following system Sdu = −Su + f1 + f2 ( λ + d λ ) (3.
.34. maximum number of iterations.⎧ du1 = −u + S −1f1 ⎪ −1 ⎪ du 2 = S f 2 ⎪ m ⎪ ⎛ m −1 −1 ˆ −1 −1 ⎛ ˆ ˆ ⎞⎞ ⎨( deik )1 = c Pκ M ik Ei ⎜ ∑ M ik Hik e ik + Bi du1 − ⎜ ∑ eik − Bi u ⎟ ⎟ − M ik Hik e ik ⎝ k =1 ⎠⎠ ⎝ k ⎪ ⎪ m ˆ − ˆ ˆ − ˆ − ⎪( deik ) = −c Pκ M ik1Ei−1 ⎛ −Bi du 2 + ∑ Pκ M ik1t ik ⎞ + Pκ M ik1t ik ⎜ ⎟ 2 ⎪ k ⎝ ⎠ ⎩ (3. In order to compute ( d λ + λ ) we substitute (3. and (3.37) The new vectors of u and e tend to satisfy (3.c) and have NG m ⎡ ⎤ 1 − ∑∑ t T ( e ik + ( deik )1 ) ⎥ ik ⎢ k ⎥.17.34) and (3. de ik are actually Newton’s directions which assure that a suitable step along them will lead to a decrease of the objective function. deik from Eqs.36).31) at current values of u and e • Step 3: Calculate ( d λ + λ ) . e ik ⎣ .b). du. • Step 2: Calculate S .38) Normally the fictitious solution must be computed first in order to define the load ˆE domain L in terms of the fictitious elastic generalized stress σ ik .b) and (3.m ⎤ . (3.17. f1 and f2 from (3.21. Details of the iterative algorithm are presented hereafter.18) simultaneously.36) Based on (3..18).17. e .. ⎦ T (3. Hence u0 and e 0 may assume fictitious values (after being normalized) for their initialization. Iterating these steps may drive us to a stable set of u ..c) is satisfied ∑∑ t i =1 k =1 NG m T 0 ik ik e = 1. • Step 1: Initialize displacement and strain rate vectors: u0 and e 0 such that the normalized condition (3.c) and (3. By forcing them to fulfil (3. Set up initial values for the penalty parameter c and for η0 .36) 47 .c) we get Lagrange multiplier λ updated as ( d λ + λ ) in (3. λ satisfying all conditions (3. Set up convergence criteria.34) we can update the displacement vector u and the global generalized strain rate vector e e = ⎡e11 .17.35) The vectors du. (3. ( λ + d λ ) = ⎢ i =1NG=1m T ⎢ ∑∑ tik ( deik )2 ⎥ ⎢ ⎥ i =1 k =1 ⎣ ⎦ (3.b) into (3. e NG .17.
(3. • Step 6: Stop The algorithm can fail due to some reasons: failure in computing the inverse matrix S or unsuccessful initialization step.• Step 4: Perform a line search to find Λ s such that FP ( u + Λ s du. If those obstacles do not exist.39) Update displacement. which after some iterations results in an unexpected form −1 ∑∑ t i =1 k =1 NG m T 0 ik ik e =0 or the maximum number of iterations is insufficient to get a convergent solution. • Step 5: Check convergence criteria: if they are all satisfied go to step 6 otherwise repeat steps 2. e . strain rate and λ as u = u + Λ s du = λ + dλ (a ) ( b) (c) (3. 48 .40) e ik = e ik + Λ s de ik λ As was shown above. e. λ ) can be found as proved in [Vu. 3 and 4. a line search without using the derivatives is recommended here. e + Λ s de ) → min .g. the Golden section method. Thus. 2001]. it is costly to get the analytical form of the derivatives of the objective function. a solution set ( u.
d. Use of large safety factors can lead to the view that “absolute” safety can be achieved. which is based on the direct computation of the loadcarrying capacity or the safety margin. Absolute safety is. The more effective method of structural reliability analysis is probabilistic limit and shakedown analyses. This may be defined by first yield or by some member failure if the structure can be designed on an element basis. 2007b. this gives quite pessimistic reliability estimates. Many different kinds of distribution of basic variables are taken into consideration and performed with First and Second Order . 2008]. However. full coverage of the present random variables is also necessary for a meaningful reliability assessment. undesirable if not unobtainable. Structural reliability analysis deals with all these uncertainties in a rational way. Reliability assessment of structures requires on the one hand mechanical models and analysis procedures that are capable of modeling limit states accurately. Progressive member failures of such systems reduce redundancy until finally the statically determined system fails. which is based on the kinematic approach [Tran et al. The realistic evaluation of structural performance can be conducted only if the uncertainty of the actual loadcarrying capacity of the structure is taken into consideration. The loading. Present structural reliability analysis is typically based on the limit state of initial or local failure.4 PROBABILISTIC LIMIT AND SHAKEDOWN PROGRAMMING The traditional approach to safety assessment and design is based on a deterministic model which invariably involves a large safety factor usually assigned from heuristic and somewhat arbitrary decisions. from limited information and from model idealizations of unknown credibility. This system approach is not defined in an obvious way for a finite element representation of a structure. Moreover. This approach has almost certainly been reinforced by the very large extent to which structural engineering design is codified and the lack of feedback about the actual performance of the structure. since it could only be approached by deploying infinite resources. these safety factors do not provide any information regarding the probability that would lead to the loss of structural integrity. because virtually all structures are redundant or statically undetermined. material as well as the thickness of the shell are to be considered as random variables.. On the other hand. of course. The aim of this chapter is to present a new algorithm of probabilistic limit and shakedown analysis for thin plates and shells. Uncertainty may originate from random fluctuations of significant physical properties.
4. A subset of sample space is called an event. A family { X i }i of random variables with X i : Ω → R is called stochastically independent or independent for short.2 Random variables An event can be identified through the value of a function called a random variable. nonnegative and σ additive measure 1) 2) 3) 0 ≤ P( A) ≤ 1. A random variable is a function which maps events ω in the sample space Ω into the real line R . The values of X = x (ω ) are called realizations of X and denoted by x . In order to get the design point. The set of all possible outcomes of such tests is called the sample space. The cumulative probability distribution function (CDF) of the random X is defined as follows F ( x ) = P( X < x ) .Reliability Methods (FORM/SORM) for calculation of the failure probability of the structure. By each test a value for the yield stress is determined but this value will probably be different from test to test. P(Ω) = 1.1. 4. which is based on the Sequential Quadratic Programming (SQP).1 Basic concepts of probability theory 4. The space of event Ω is characterized by a probability measure P . (4.1 Sample space A standard way of determining the yield stress of a material such as steel is to perform a number of simple tensile tests with specimens made from the material in question. An event is the set of sample points. The sample space for the yield stress is the set of all positive real numbers and is continuous. If it contains no sample points. a nonlinear optimization was implemented. Therefore. P( A ∪ B ) = P( A) + P( B ) if A ∩ B = ∅ .2) . in this connection. which satisfies the properties of a normed. the yield stress must be taken as an uncertain quantity and it is in accordance with this point of view said to be random quantity. it is called an impossible event. P(∅) = 0.1.1) A random variable can be continuous or discrete. Usually a random variable is denoted by a capital letter X X : Ω→ R. 50 (4. if for every choice of a subset Ai ⊂ R the event { X i ∈ Ai }i are independent of each other. Nonlinear sensitivity analyses are also performed for computing the Jacobian and the Hessian of the limit state function.
Normal distribution and standard normal distribution (σ = 1. A random variable is characterized by its CDF. x →∞ x →−∞ lim F ( x ) = 0.1. (4.This means. The appendix summarizes the most important distribution functions.3) The probability density function has the properties 1) 2) 3) P ( a ≤ X ≤ b) = ∫ f (t )dt . that the random variable X has a realization lower than x . that the value of F at x is the probability of the event.F ( a ) . ∞ (4.5 x 2 e . We summarize the most important characteristics of distribution functions 1) 2) F ( x ) is nondecreasing and continuous lim F ( x ) = 1. Then its probabilistic characteristics are described by the cumulative distribution function F . We have F ( x) = −∞ ∫ x f (t )dt .5) 51 . The expected value E [ X ] of X with the density f is defined as E[X ] = −∞ ∫ xf ( x )dx . the form of F is not known in all details. However. μ = 0) with the respective densities f ( x) = 1 2πσ 2 e−( x−μ ) 2 / 2σ 2 and f ( x ) = 1 −0. densities and further details. In the following it is assumed that all the random variables are continuous.3 Moments Let X be a continuous random variable. it is often useful to use the derivative probability function. For the continuous random variable. It is therefore useful to have an approximate discription of a random variable by its most important features. This function is called the probability density function (PDF) f : R → R .4) 4. P( a < X < b) = F (b) . The following distributions play a special role in structural reliability analysis. 2π (4. in many applications. a b f ( x) ≥ 0 . −∞ ∫ ∞ f (t )dt = 1 .
. E [Y ] if and only if X and Y are independent . σ j are the standard deviations of the random variables X i and X j .. The ratio ρij = Cov ⎡ X i . E [ λ X ] = λ E [ X ] . 52 .Y ]) = E [ X ] .. E ([ X .7) The mixed central moment of random variables is defined as ⎡ ⎤ ⎡ ⎤ Σij = Cov ⎣ X i . The following simple rules for the expectation hold 1) 2) 3) E [ a ] = a. independent random variables are uncorrelated but not vice versa. is called standard deviation of X . The positive square root of the variance. It can be shown that −1 ≤ ρij ≤ 1 . E [ X + Y ] = E [ X ] + E [Y ] . ⎜ i =1 ⎟ ⎝ ⎠ −∞ −∞ x1 xn (4. is called the correlation coefficient.. λ ∈ R .. X j ⎤ ⎣ ⎦ σ iσ j (4. ∫ f (t1 . This definition can easily be extended to a vector valued random variable X : Ω → R n called a random vector. Therefore. X j ⎦ = 0 (or ρij = 0) holds. The covariance matrix Σij is symmetric semidefinite. ⎣ ⎦ (4. a ∈ R . the mean or the first moment of X and the symbol μ is often used for it. Two random variables X i and ⎡ ⎤ X j are said to be uncorrelated if Cov ⎣ X i .dtn . It can be used as a measure of mutual linear dependence between a pair of random variables.6) The variance reflects the expected deviation of a realization x from the expected value E [ X ] . The joint cumulative distribution function for the random vector X is defined as n F ( X ) = P ⎛ ∩ ( X i ≤ xi ) ⎞ = ∫ . The n th central moment of X is defined by E ⎡( X − μ ) n ⎤ .9) where σ i .The expected value is also called the ensemble average.8) is called covariance of random variables (written with random vector components X i and X j ). The concept of random variable has been used only in onedimensional sense. σ [ X ] . X j ⎦ = E ⎣( X i − μi ) ( X j − μ j )⎦ = E ⎡ X i X j ⎤ − μi μ j ⎣ ⎦ (4. In the appendix the expectations and variances of the most important distribution functions for structural reliability analysis are listed. The second central ⎣ ⎦ moment of X is called the variance of X and is denoted by σ 2 or Var [ X ] 2 ⎡ ⎤ Var [ X ] = σ 2 = E ⎡ ( X − μ ) ⎤ = E ⎣ X 2 ⎦ − μ 2 .. tn )dt1...
.e. Typical examples of engineering interest are earthquake ground motion.1 Failure function and probability In reliability based structural analysis all probabilistic characteristics are modelled as random variables. Reliability analysis should not.4. i. The probabilistic characteristics of the processes are known from various available measurements and investigation in the past. systems engineering. wind turbulence.).2 Reliability analysis The behavior of a structure is influenced by various typically uncertain parameters (loading type. or material data. the available probabilistic characteristics of random quantities affecting the loading of the mechanical system often cannot be utilized directly to account for the randomness of the structural response due to its complexity.. Data with random fluctuations in time and space is adequately described by stochastic processes and fields. ⎧< 0 ⎪ g ( X ) = R(X ) − S ( X ) ⎨ = 0 ⎪> 0 ⎩ for failure for limit state for safe structure (4.. The principles of reliability analysis have been applied to a very large class of problems. S which are generally unknown or difficult to establish. With R. Figure 4.1 shows the densities of two random variables R. The deterministic safety margin M is based on the comparison of a structural resistance (threshold) R and loading S (which is usually an invariant measure of local stress at a hot spot or in a representative crosssection).. The failure probability Pf is the probability that g ( X ) is nonpositive. X 1 . These socalled physical basic random variables are collected in a vector X = ( X 1 . quality control engineering and statistical acceptance testing. be thought of as an isolated discipline as it is closely related to the theory of statistics and probability and to such fields as operations research. 4.. imperfections. In engineering mechanics.e. loading magnitude. dimensions. sea waves. S functions of X the structure fails for any realization with nonpositive failure function or limit state function g ( X ) . In structural response calculations a distinction is made between the involved structural model properties which are either considered as being deterministic or stochastic.. X n ) where n is the number of stochastic variables.. ranging from the design of control systems for complex nuclear and chemical plants to the design of specific mechanical and structural components. however.2.11) 53 . Pf = P ( g ( X ) ≤ 0 ) = ∫ f X ( x )dx F (4. as well as more generally in the field of aerospace industries. i.10) The limit state g ( X ) = 0 defines the limit state hypersurface ∂F which separates the failure region F = {x g ( X ) < 0} from safe region.
Φ (. Usually. the sample size required for direct MCS must be at least Pc /10 leading to a minimum sample size in the range of 10−5 to 10−7 . it is not possible to calculate Pf analytically. They are even much lower in nuclear reactor technology.1 Density functions of R. The numerical effort can be reduced considerably by variance reduction methods like Importance Sampling and by Response Surface Methods (RSM).s Figure 4. the most effective analysis is based on First. β HL = min uT u . In special cases. However.13) For the general cases. i. It is defined as the shortest distance from origin to the failure surface ∂F in the standardised and uncorrelated u space (figure 4.e. f R(r) f S (s) Loading S Resistance R r. there are several approximate methods to compute the failure probability Pf . Direct Monte Carlo Simulation (MCS) becomes increasingly expensive with the increase of the structural reliability but it is independent on the dimension n of the space of basis variables. The concept of a reliability index which is invariant with respect to the formulation of the failure function was proposed by Hasofer and Lind [1974]. Acceptable failure probabilities might be in the range of 10−4 to 10−6 . it can be shown that Pf = Φ ( − β c ) (4. For a validation that the failure probability Pf is less than an accepted limit Pc . Such a large number exceeds particularly for complex FEmodels.and SecondOrder Reliability Methods (FORM/SORM) if gradient information is available [Gollwitzer et al. g ( u ) =0 (4.where f X (x ) is the ndimensional joint probability density function. S presenting on one axis 54 . 1988].2). available resources by far.12) where β c is the second moment reliability index defined as the mean of M divided by the standard deviation of M. However the numerical effort of FORM/SORM increases with the dimension n ..) is the standard Gaussian distribution function. where the basic variables are jointly Gaussian distributed and the failure surface is a hyperplane.
σ = 1) .. on the other hand.and SecondOrder Reliability Methods (FORM/SORM) are efficient means and standard methods of numerical probability calculations for multiple random variables. The failure probability is computed in three steps • Transformation: The physical space x of uncertain parameters. 55 . FORM/SORM apply to problems.2. The appendix provides a description of some independent random variables. such as the Central Processing Unit (CPU). any effective analysis is based on FORM/SORM [Hohenbichler et al.g. where the set of basic variables are continuous). Therefore. can deal with random variables having arbitrary probability distributions and can estimate reliability results for large numbers of random variables. Especially in the case of small failure probability (large reliability). is transformed into a new ndimensional space. u space constant probability safe region g(u) > 0 βHL g(u) = 0 failure region g(u) < 0 gL(u) = 0 gQ(u) = 0 Figure 4.2 Domain based on linear and quadratic approximations in u space FORM/SORM are analytical probability integration methods. X . 1987]. They are often confused with other socalled first order methods that may rely on assumptions of normality and/or small numbers of random variables to achieve closedform solutions. FORM/SORM are extremely efficient compared with the MCS method regarding the requirement of computer time. Because of the large computational effort of MCS due to small failure probabilities.2 First. Such a transformation is always possible for continuous random variables. u . in which the numerical effort depends on the number of stochastic variables but not on Pf (contrary to MCS). the defined problem has to fulfill the necessary analytical requirements (e. Practical experience with FORM/SORM algorithm indicates that their estimates usually provide satisfactory reliability measure. FORM/SORM.and SecondOrder Reliability Method First. consisting of independent standard Gaussian variables U ( μ = 0. These methods have only become popular since the advent of inexpensive computing.4.
e. the original limit state function g ( x ) = 0 is mapped into the new limit state function g ( u) = 0 in the u space.18) 56 .16) where β HL is calculated from (4.17) Consider a small vector ε which can be interpreted as an error appearing in the calculation of the design point (i. the probability of failure Pf can be determined with good accuracy by Pf . (4. More details of this algorithm will be discussed in next section. This corresponds to an approximating hyperplane g L (u) = 0 (linear or firstorder) or hyperparaboloid gQ ( u) = 0 (quadratic or secondorder).15) Here g L ( u) = 0 is the Hesse normal form of the tangential hypersurface to the limit state function at the design point. is determined by an appropriate nonlinear optimization algorithm. The failure probability Pf P ( gQ ( u ) < 0 ) in SORM.13).14) (4.equations to map them into the standard Gaussian space. ∇g ( u0 ) (4. • Computation of design point: The design point or β HL point .2. This transformation is exact and not an approximation [Bjerager. respectively. I 1 = Φ ( − β HL ) = 2π − β HL −∞ ∫ e −0. If the failure function is not strictly nonlinear. By this transformation. 1991]. which is the point on the limit state surface having the shortest distance to the origin in the u space. I ) . which is given by g L ( u) = aT ( u − u0 ) = aT u + β HL where a is the vector of direction cosines and can be obtained from a= ∇g ( u0 ) . 2004] β HL ( ε ) = −Φ −1 ( Pf . The resulting disturbance of the reliability index β HL (ε ) can be calculated as follows [Rackwitz. • Approximating the limit state surface and calculation of the failure probability: The limit state surface g ( u) = 0 is approximated by a tangential hypersurface at the design point. u0 is replaced by u0 + ε ).5 z dz 2 (4.1 FirstOrder Reliability Method (FORM) is thus approximated by P ( g L ( u ) < 0 ) in FORM or Consider a tangential linearization at the design point u0 on the limit state surface g ( u) = 0 .2.I ( u0 + ε ) ) ≈ aT ε + β HL . (4. The firstorder reliability index can be defined as following β I = β HL = −Φ −1 ( Pf . 4.
The general idea is to approximate the limit state surface by a quadratic hypersurface rather than by a hyperplane. II = S1 + S2 + S3 S1 = Φ ( − β HL )∏ (1 − β HLκ j ) j =1 n −1 −1/ 2 (4. The main curvatures of the quadratic hypersurface at the design point are equal to those of the limit state surface. I ( ε ) ∂ε i ε i →0 (4.19) = −ϕ ( − β HL ) ai .20a) (4. 4.20c) ⎧ n −1 ⎡ n −1 −1/ 2 −1/ 2 ⎤ ⎫ ⎪ ⎪ S3 = ( β HL + 1) [ β HLΦ ( − β HL ) − φ ( β HL )] ⎨∏ (1 − β HLκ j ) − Re ⎢∏ (1 − ( β HL + i )κ j ) ⎥ ⎬ ⎪ j =1 ⎣ j =1 ⎦⎪ ⎩ ⎭ (4. S1 is an asymptotic approximation of Pf being exact for β HL → ∞ . ai represents the sensitivities of β HL and Pf ..i .2 SecondOrder Reliability Method (SORM) Various methods have been suggested to improve the accuracy of FORM calculations and to give a rough estimate of the approximation quality [Madsen et al. ∂ε i ε →0 i ∂Pf . 1979]. [Fiessler et al. In other words. 1986]. First. Obviously. This leads to the socalled SecondOrder Reliability Methods (SORM). I versus the changes of the mean value of u0.2. The curvatures κ j are obtained in the following way.. so that the design point becomes part of the last axis of the rotated normal y space.20d) where i = −1 . the coordinate system of the standard normalized u space is rotated.20b) n −1 ⎧ n −1 −1/ 2 −1/ 2 ⎫ S2 = [ β HLΦ ( − β HL ) − φ ( β HL ) ] ⎨∏ (1 − β HLκ j ) − ∏ (1 − ( β HL + 1)κ j ) ⎬ j =1 ⎩ j =1 ⎭ (4. Using all n − 1 main curvatures κ j . This is achieved through the following transformation 57 . S2 and S3 are correction terms.] represents the real part of the complex argument and φ (. the last basic vector of the y space coincides with the design point position vector in the normalized u space. the failure probability is calculated as a three term approximation suggested by Tvedt [1983] as Pf . Re [.and therefore ∂β HL ( ε ) = ai .2.) is the standard Gaussian PDF.
This method has proved to be suitable for tasks in the area of the reliability theory [Rackwitz. However. 2004]. this algorithm is only guaranteed to converge towards a locally most likely failure point in each sequence of points on the failure surface if the safe region is quasiconvex or concave (see section 7. [2003] and Bjerager [1989] got good results in probabilistic limit analyses with Rackwitz’s simple gradient search algorithm. The open source code of FORM/SORM is available in [Wierskowski. employs Newton’s method (or quasiNewton methods) to solve the KKT conditions for the 58 . 1999] minimize: f ( u) = s. 1988. 4. Bazaraa et al. which is based on a linearization of the limit state function at each step. e.Y = DT U and. 2005]. SQP becomes an efficient tool for solving nonlinear optimization problems and shall be introduced for this reason here in short statements. (4.22) Defining H y0 as the matrix of second derivatives of the limit state function in y space at the design point y 0 .t. the design point must be identified. More details can be found in the literature. where a coefficient 1 T u u 2 g ( u) ≤ 0 (4.3 Calculation of design point In order to apply FORM/SORM.23) where I is the identity matrix. A more general algorithm is the Sequential Quadratic Programming (SQP). also known as successive or recursive quadratic programming. 1983b.25) 1 is added for technical reasons. Example of applying SQP in reliability analysis can be found in [Dronia.g Schittkowski [1983a. 2004]. Many algorithms have been 2 suggested to deal with this problem. provided that D −1 = DT (4.24) It should be noted that β I has a geometric meaning but β II has no such interpretation.21) U = DY . 2005]. the curvatures are obtained as solutions of a characteristic equation ⎛ ⎛ ∂g ( y ) ⎞ −1 ⎞ 0 det ⎜ H y ⎜ ⎟ −κI⎟ = 0 ⎟ ⎜ ⎝ ∂yn ⎠ ⎝ ⎠ (4.II ) (4.4). [1993]. The SQP method. This leads to a nonlinear constrained optimization problem as follows [Heitzer. Staat et al. Due to the complex shape of the limit state surface. The secondorder reliability index can be defined as following β II = −Φ −1 ( Pf .
28) where α k ∈ ( 0. λ k ) . λ k +1 the corresponding multiplier of this subproblem. Since ∇2 L( x k ) might not be positive definite. it can be replaced by a positive definite approximation B k .26) where x is an ndimensional parameter vector containing the design variables.27) is formulated and must be solved in each iteration.t. Then a quadratic programming subproblem of the form 1 minimize: f ( x k ) + ∇f (x k )T d + dT ∇2 L(x k )d 2 T s. λ k +1 ) − ∇L( x k . As a result. gi ( x k ) + ∇gi ( x k ) d ≤ 0 i∈I (4.2pT B k p k k k otherwise (4. Define ∇2 L( x k ) = ∇2 f ( x k ) + ∑ λki ∇2 gi ( x k ) to be the usual Hessian i∈I of the Lagrangian at x k . given an iterate ( x k . gi ( x ) ≤ 0 i∈I (4. a certain irregularity leading to infinite cycles. 1978] B k +1 = B k − with sk = θ q k + (1 − θ ) B k p k q k = ∇L( x k +1 .original problem directly. then the new iterate is obtained by ⎛ x k +1 ⎞ ⎛ x k ⎞ ⎛ dk ⎞ ⎜ λ ⎟ = ⎜ λ ⎟ + αk ⎜ λ − λ ⎟ k ⎠ ⎝ k +1 ⎠ ⎝ k ⎠ ⎝ k +1 (4. the accompanying subproblem turns out to be the minimization of a quadratic approximation to the Lagrangian function optimized over a linear approximation to the constraints. where λ k ≥ 0 is the Lagrange multiplier estimating for the inequality constraints. 1983a] 59 . which is known as BFGS update [Powell.30) p k = x k +1 − x k B k p k p T B k s k sT k + T k pT B k p k sk p k k (4.8 T k k k T ⎪ p B p −p q ⎩ k k k k k when pT q k ≥ 0. λ k +1 ) ⎧1 ⎪ θ =⎨ pT B p 0.t. Consider the inequality constrained nonlinear optimization problem minimize: f ( x ) s.29) To ensure global convergence of the SQP method and avoid the Maratos effect. Let d k be the optimal solution.1] is a suitable step length parameter. we have to select a suitable steplength ξ k subject to the augmented Lagrangian merit function ψ (ξ ) [Schittkowski. For this case.
ψ (ξ ) = f (x + ξ d) + ∑ ⎜ λi gi ( x + ξ d) + ri gi (x + ξ d )2 ⎟ +
⎛ i∈I1 ⎝
1 2
⎞ ⎠
1 λi2 ∑ 2 i∈I 2 ri
(4.31)
where I1 = {i ∈ I : gi ( x ) ≤ λi / ri } and I 2 = I \ I1 , ri ≥ 1 is the penalty parameter. In order to get the value of penalty parameter, two positive constants r > 1, ε < 1 should be chosen. Then at each iteration, together with the search direction d k , two additional variables ⎧ dk 2 ⎧ dT B d ⎫ , when λk +1 ≠ λk ⎪ ⎪ ⎪ δ k = min ⎨ k k 2 k , δ k −1 ⎬ and ε k = ⎨ λ k +1 − λ k ⎪ dk ⎪ ⎪ ⎩ ⎭ , otherwise ⎩ε with δ 0 = 1 and a small positive number jk , which satisfies the inequality
16 ⎧ ⎫ ln ⎪ ε kδ k ( 4 − δ k ) ⎪ 1 1 ⎪ ⎪ ⎛ 1 ⎞ < ε kδ k ⎜ 1 − δ k ⎟ , i.e. jk = min ⎨ j ∈ N , j > ⎬ j 4 ln r r ⎝ 4 ⎠ ⎪ ⎪ ⎪ ⎪ ⎩ ⎭
(4.32)
(4.33)
must be identified. The penalty parameter ri k at iteration k can be obtained from ri k = max {ri k −1 , r jk } (4.34)
with r0 = r . The step length ξ k can be chosen by a separate algorithm, which should take the curvature of the merit function into account. If ξ k is too small, the resulting step sizes might become too small leading to a higher number of outer iteration. On the other hand, a large value close to one requires too many function calls during the line search. Thus we need some kind of compromise which is obtained by applying first a polynomial interpolation. Rackwitz [2004] proposed a simple line search, which is based on a quadratic interpolation minimization algorithm
ξk =
where
∇ψ (0) ≥γ >0 2 (ψ (0) + ∇ψ (0) − ψ (1) )
(4.35)
∇ψ (0) = ∇f ( x )T d + ∑ ( λi ∇gi ( x )T d + ri gi ( x )∇gi ( x )T d )
i∈I1
(4.36)
and γ is a small positive number, which ensures that the value of ξ k is not too small. This algorithm is very useful in reliability analysis, because it allows to reduce the number of deterministic loop in an iteration to two, for ξ = 0 and ξ = 1 , and thus considerably reduce the computational cost. Applying SQP method for solving the optimization problem (4.25), the quadratic programming subproblem (4.27) at iteration k becomes
60
1 minimize: f (u k ) + ∇f (u k )T d k + dT ∇2 f ( u k )d k k 2 s.t. g ( u k ) + ∇g ( u k ) T d k ≤ 0
(4.37)
with ∇f ( u k ) = u k and ∇2 f ( u k ) = I . The KarushKuhnTucker conditions for this problem are
⎧∇L( u k , λk ) = ∇f (u k ) + d k + λk ∇g (u k ) = 0 . ⎨ g ( u k ) + ∇g ( u k ) T d k = 0 ⎩
(4.38)
Applying Newton’s method for solving these equations leads to the system
⎡ Bk ⎢ ∇g ( u )T k ⎣ ∇g ( u k ) ⎤ ⎛ d k ⎞ ⎛ ∇f ( u k ) ⎞ ⎥ ⎜ λ ⎟ = − ⎜ g (u ) ⎟ 0 ⎦ ⎝ k +1 ⎠ k ⎝ ⎠
(4.39)
with the solutions
λk +1 = −
− ∇g ( u k )T B k 1∇f ( u k ) − g (u k ) − ∇g ( u k )T B k 1∇g ( u k )
(4.40)
− − d k = −B k 1∇g ( u k )λk +1 − B k 1∇f ( u k )
as the Lagrange multiplier and search direction for the next iteration. If d k = 0 then u k together with the Lagrange multiplier λk +1 yields the optimal solution for the problem (4.25), i.e. the design point is actually found. The calculation of the necessary derivatives are considered as the sensitivity analysis and will be discussed in the next section. 4.4 Sensitivity of the limit state function The reliability analysis described above can be carried out now with the help of a probabilistic limit and shakedown analysis. From the results of the finite element analysis, the necessary derivatives of the limit state function based shakedown analysis can be determined analytically. This represents a considerable reduction of computing time comparing with the other methods, e.g. the difference approximation, and makes such an efficient and costsaving calculation of the reliability of the structure. Contrary to the numerical calculation, the analytical calculation is faster and more exact. The necessary data for the calculation of the derivatives are available after the execution of the deterministic shakedown analysis since they are based on the limit load factor α lim . The derivatives must be calculated firstly at each iteration in the physical x space. Then it is transferred into the standard Gaussian u space by using the chain rule ∇ u g ( u ) = ∇ u g ( x ) = ∇ x g ( x )∇ u x
2 2 ∇ u g ( u ) = ∇ u ( ∇ x g ( x )∇ u x ) = ( ∇ u x ) ∇ 2 g ( x )∇ u x + ∇ x g ( x )∇ u x x
T
(4.41)
The calculation of the derivatives in the physical x space is based on a sensitivity analysis. Sensitivity arises prominently in microeconomic theory, where optimality 61
condition provide the basic for analysis. It is interested primarily not in finding an optimal solution, but rather in how the solution is affected by changes in the problem data. By this way, sensitivity analysis becomes widely in engineering, especially in reliability analisis of the structures, where the data defining problem are random variables.
4.4.1 Mathematical sensitivity
Consider a constraint optimization problem P is defined as follows minimize: s.t. f (x ) hi (x ) = 0 i∈I . (4.42)
Suppose that f , hi : R n → R are twice continuously differentiable and I is some index set. In many applications the objective function f as well as the constraint functions hi may depend on other parameters. Consider the following pertubation problem P(ε ) of the original problem P(0) minimize: s.t. f ( x, ε ) hi (x, ε ) = 0 i ∈ I , ε ∈ Rm . (4.43)
The perturbations ε can be interpreted in two ways: as a random error or as aspecific change in the parameters defining the problem functions. The optimal solution
x* (ε ) of the problem P(ε ) with the Lagrange multipliers λ * fulfills the following first
order KarushKuhnTucker conditions
∇L( x* , λ * , ε ) = ∇f ( x* , ε ) − ∑ λi*∇hi ( x* , ε ) = 0
i∈I
.
(4.44)
hi ( x , ε ) = 0
*
i∈I
If f has a continuous and nonsingular gradient matrix ∇ε f ( x, ε ) in an open set containing ( x* , ε ) , the implicit function theorem implies the existence of a unique differentiable local solution
( x (ε), λ (ε) )
* *
of the problem P(ε ) . Let ε = 0 then the
conditions are fulfilled in a local solution x* (0) of the problem P(0) . We defined the optimal value function f * (ε ) and the optimal value Lagrangian L* (ε ) as follows f * (ε ) ≡ f ( x* (ε ), ε ) L* (ε ) ≡ L ( x* (ε ), λ * (ε ), ε ) Theorem [Fiacco, 1983] If the linear independence condition and the second order sufficient condition of the problem P(ε ) are satisfied, and if the problem functions are twice continuously differentiable in (x, ε ) near ( x* , 0) then, in a neighborhood of ε = 0 , we have . (4.45)
62
4. the value of Hessian matrix of the optimal value function requires the determination of both the optimal solutions and its first derivatives. λ * (ε ). ε ) dε ∂ε ∂ε i∈I d2 * ∂ 2 L(x* (ε ). If we defined the limit load factor α lim as follows α lim = Plim P (4. For the shake of simplicity.f * (ε ) = L* (ε ) d * ∂ ∂ f (ε ) = f (x* (ε ).2 Definition of the limit state function As mentioned above.. λ * (ε ) ) have been determined.. ε ) ⎞ ∂λi* (ε ) ∂ 2 L( x* (ε ). ε ) − ∑ λi* (ε ) hi ( x* (ε ). h2 . It should be noted that the value of the optimal value function and its gradient can be calculated once the optimal solutions ( x* (ε ). P are limit load and actual load of the structure. for the general problem.and secondorder change in the optimal value function of P(ε) .46b) (4. ε ) − ∑⎜ i + ⎟ ∂ε ∂ε ∂ε 2 i∈I ⎝ ⎠ T (4.. the limit state function g can be normalized with the actual load P and then becomes g = α lim − 1 (4.46a) (4.4. the limit state function contains the parameters of structural resistance R and loading S.] . ε ) d * f (ε ) = x (ε ) dε 2 ∂x∂ε dε ⎛ ∂h ( x* (ε ). However.46) provide a quantitative measure of the first. λ * (ε ).47) where Plim . The sensitivity analyses (4.46d) where h = [ h1 .46c) and 2 ⎡ ∇2 L −∇x hT ⎤ ⎧∇ε x ⎫ ⎧ −∇εx L ⎫ x =⎨ ⎬ ⎬ ⎢ ⎥⎨ 0 ⎦ ⎩∇ ε λ ⎭ ⎩∇ε h ⎭ ⎣ −∇x h T (4.48) 63 .
1 ∂x/∂u 2 2 ∂ x/∂u Sensitivity analysis ∂g/∂x.3 Flowchart of the probabilistic limit and shakedown analysis From previous chapter.Deterministic model and data Random variables x and its starting value x0 FEMbased deterministic LISA Limit state function g(x) = α lim . ∂ 2 g/∂x 2 SQP algorithm to find the design point descent direction line search Failure probabilities (FORM/SORM approximations) Figure 4.t : ⎛ NG m T ⎞ ⎜ ∑∑ e ik t ik − 1⎟ = 0 ⎝ i =1 k =1 ⎠ (4.49) with 64 .u s.15) α lim = min ( FP ) eik . the shakedown load factor can be calculated by the nonlinear programming (3.
+ xn σ 0 ⎤ = x1t1 + x2 t 2 + .51). thickness variable Z and load variables.. Then the limit state function is no more a linear function of these variables. the corresponding normalized stress fields are obtained 0 ˆ ˆ2 ˆn⎦ t = N 0ε 0σ = N 0ε 0 ⎡ x1σ1 + x2σ 0 + .51) It can be seen that the limit state function is the function of yield stress variable Y .. The corresponding actual stress field can also be described in the same way ˆ ˆ0 ˆ2 ˆn σ = x1σ1 + x2σ 0 + .53) (4. + xn σ 0 . n ).4. The shakedown range obtained from (4.54) or t = ∑ x jt j .49) is linear function of the yield stress variable Y (see 4.. 3 ( ) (4. j =1 n (4.4.. + xn Pn0 1 (4..⎧m ⎛ ˆ P ⎪ ∑ ⎜ YN 0ε 0 κ β1 β12 + γ + β 2 β 22 + γ + γ K 0 NG 3 ⎪ k =1 ⎜ FP = ∑ ⎨ ⎝ T i =1 ⎪ m c⎛ m ˆ u ⎞ ⎛ ∑e − B u ⎞ ˆ ⎪ + ⎜ ∑ eik − Bi ⎟ ⎜ ik i ⎟ 2 ⎝ k =1 ⎠ ⎝ k =1 ⎠ ⎩ ( ) ⎞⎫ ⎟⎪ ⎟⎪ ⎠ ⎬ ⎪ ⎪ ⎭ (4.10) and (4.3 First derivatives of the limit state function With the nonlinear programming (4..52) where x j is the realization of the j th basic load variable X j ( j = 1... the Lagrange function becomes 65 .. + Pn0 1 P = x1P 0 + x2 P20 + . From (3.55) 4. In the case of heterogeneous material.50) T m ⎧m ⎫ c⎛ m ⎪ ˆ u ⎞ ⎛ ∑e − B u ⎞⎪ ˆ = ∑ ⎨ ∑ (Yηik ) + ⎜ ∑ eik − Bi ⎟ ⎜ ik i ⎟⎬ 2 ⎝ k =1 i =1 ⎪ k =1 ⎠ ⎝ k =1 ⎠⎪ ⎩ ⎭ NG with ηik = N 0ε 0 ˆ Pκ β1 β12 + γ + β 2 β 22 + γ + γ K 0 .4).53).. we will obtain at different Gaussian points i eventually different yield stress variables Yi . + xn t n ⎣ ˆ (4. The actual load P is defined in n components by using the concept of a constant referent load P 0 as follows P 0 = P 0 + P20 + .
one gets ∂ 2α lim ∂ = ∂X l X j ∂X l ⎛ NG m T ⎞ ⎜ −λ ∑∑ eik t ik .56) In order to get the sensitivity of the limit state function. λ * ) are the solutions ik of the optimization problem.λ* ) ∂Y ⎝ i =1 k =1 ⎠ (eik ) i =1 k =1 (4. At optimal point.57) versus X l . The derivatives of the limit load factor αlim versus yield stress variable Y and thickness variable Z can be determined in the same way and have the form NG m ∂α lim ∂FPL α lim ∂ ⎛ NG m ⎞ * = = ∑∑Yηik ⎟ * = ∑∑ηik = Y ⎜ ∂Y ∂Y (e*ik .59) 4.u . j ⎟ ⎝ i =1 k =1 ⎠ (e*ik . the Hessian matrix which summarizes second partial derivatives of the limit state function is needed.60) 66 . j ∂Xik * * ik ik i =1 k =1 i =1 k =1 l ( eik .⎛ NG m T ⎞ FPL = FP − λ ⎜ ∑∑ eik t ik − 1⎟ ⎝ i =1 k =1 ⎠ T m NG m ⎧m ⎫ c⎛ m ⎪ ˆ u ⎞ ⎛ ∑ e − B u ⎞ ⎪ − λ ⎛ ∑∑ eT t − 1⎞ ˆ = ∑ ⎨ ∑ (Y ηik ) + ⎜ ∑ e ik − Bi ⎟ ⎜ ik i ⎟⎬ ik ik ⎜ ⎟ 2 ⎝ k =1 i =1 ⎪ k =1 ⎠ ⎝ k =1 ⎠⎪ ⎝ i =1 k =1 ⎠ ⎩ ⎭ NG (4. as discussed above. the first derivative of the limit load factor αlim versus j th load variable X j can be calculated as follows ∂α lim ∂FPL = ∂X j ∂X j = −λ ∑∑ eT ik i =1 k =1 NG m ( e* ik .λ* ) ∂Z ⎜ i =1 k =1 ⎜ ⎝ ⎝ ( ) ⎟⎟ ⎟⎟ ⎞⎞ ⎠ ⎠ e*ik ⎛ ˆ P = ∑∑ ⎜ YN 0ε 0 k 3 i =1 k =1 ⎜ ⎝ NG m ( ⎞ β12 + γ + β 22 + γ ⎟ ⎟ ⎠ e*ik ) (4. j is the stress vector at Gaussian point i and load vertex k due to the j th load case.4 Second derivatives of the limit state function In SORM algorithm. one require is that the derivatives of the limit load factor αlim must be available.u* . u* .λ* ) ∂e ∑∑ eT tik . They can be obtained from a directly analytical derivation of the first derivatives.λ * * ) ∂t ik ∂X j ( e* ik .57) where t ik . λ * ik ) (4.λ * ) = − λ ∑∑ eT t ik .λ ) ( eik .λ ) NG m NG m ∂λ =− ∂X l (4. j ik i =1 k =1 NG m ( e* . Let (e* . j * * − λ ∑∑ tT .4. Taking the derivatives of (4.u* . Consider first the derivatives of αlim versus load variables.58) ˆ P ∂α lim ∂FPL ∂ ⎛ NG m ⎛ ⎜ ∑∑ ⎜ YN 0ε 0 k β1 β12 + γ + β 2 β 22 + γ + γ K 0 = = 3 ∂Z ∂Z (e*ik .
G ik must be positive definite due to the second order sufficiency condition. k (4.65) 67 .60) when l ≡ j . it follows that ∂FPL 1 ⎛ m ˆ ⎞ Hik eik +c ⎜ ∑ e ik − Bi u ⎟ − λ t ik = 0 = ∂eik ˆ ⎝ k =1 ⎠ P κ NG m ∂FPL ˆ ⎛ ˆ ⎞ = − c ∑ BT ⎜ ∑ eik − Bi u ⎟ = 0 i ∂u i =1 ⎝ k =1 ⎠ ∀ i. (3. At the optimal point.62b) can be written as follows ∑∑ t i =1 k =1 NG m T ik ∂e ik 1 ∂α lim . k (4. λ one gets Gik ∂eik ∂λ − t ik − λ t ik .57) the equation (4. c) versus X l and using the chain rule for two variables e ik .60) the derivatives of λ and e ik versus X l are also needed. P3 are obtained from (3.49).61b) ∑∑ e i =1 k =1 NG m T ik ik t −1 = 0 .61c) ⎛ m ˆ ⎞ Since this relation ⎜ ∑ eik − Bi u ⎟ = 0 holds at optimal point.It can be seen that the derivative ∂ 2α lim / ∂X 2 is a special case of (4. According to implicit function theorem. ∂X l ∂X l (4. Since the KarushKuhnTucker conditions and the strict complementary slackness hold at the optimal point.61a) (4. then we can delete it in ⎝ k =1 ⎠ equation (4.l . j To complete the calculation in (4.62b) with Gik = ∂ 2 FPL 12 1 =− Hik eik eT P3 + ( Hik + Dik ) ik 2 ∂eik ˆ ˆ Pk3 Pk (4. Referring to (4.l = 0 ∂X l ∂X l NG m ∂eik ∂t = − ∑∑ eT ik ik ∂X l ∂X l i =1 k =1 ∀ i.64) From (4.23) and (1. (4. Dik .63) with Hik . thus it is invertible. = ∂X l λ ∂X l (4.61a.62a) we obtain the derivative of e ik versus X l ∂eik ∂λ − − = Gik1t ik + λGik1t ik . by taking derivative of (4.61a) and equation (4.61b) is automatically fulfilled.20).62a) ∑∑ tT ik i =1 k =1 NG m (4.
the Lagrange multiplier λ is as well a linear function of Y . the optimisation variables eik .59) and (4. Y and referring to (4.67) (4.69a) into (4.69b) with referring to the relation (4.61a) and after some manipulations we obtain ∂λ λ ∂eik = .58) versus X j .Substituting (4. Moreover. = 0. c) versus Y .l ik λ ∂X l i =1 k =1 ∑∑ t i =1 k =1 NG m . i. ⎜ ⎟ =⎜ 2 ∂Y ∂Y ⎝ Y ⎠ e*ik ⎝ Y ∂Y Y ⎠ e*ik ⎝ Y Y Y ⎠ e*ik (4. u are independent of Y .68) It is obvious from (4.57) one has ∂ 2α lim ∂ ⎛ α lim ⎞ 1 ∂α lim λ NG m = = = − ∑∑ eT t ik .λ* Y i =1 k =1 (e*ik . k (4. ∂Y Y ∂Y Since eik and λ are functions of Z .e. ∂Y (4. (4. from (4.λ* Y ∂X j e* . These can be proved by taking derivatives of (4.66) T ik G t −1 ik ik Analogously.64) and after some manipulations we obtain as well the derivative of λ versus X l ∂λ = ∂X l NG m 1 ∂α lim − − λ ∑∑ t T Gik1t ik . it follows that Gik ∂eik ∂λ Hik eik − t ik + =0 ∂Y ∂Y Y P ˆ κ T ik ∀ i.71) ˆ P ∂ ⎧ NG m ⎛ ⎪ ⎜ YN 0ε 0 k + ⎨ ∑∑ 3 ∂Z ⎪ i =1 k =1 ⎜ ⎝ ⎩ ( ⎞⎫ ⎪ β12 + γ + β 22 + γ ⎟ ⎬ ⎟⎪ ⎠ ⎭ e* ) ik 68 .61a. by taking derivatives of (4.68) that for the case of homogeneous material the shakedown load factor is actually a linear function of yield stress.69b) Substituting (4.69a) ∑∑ t i =1 k =1 NG m ∂eik =0. ⎛ ˆ P ⎜ YN 0ε 0 k ⎜ 3 ⎝ ( β12 + γ + β 22 + γ ⎟ ⎬ ⎟ )⎪ ⎞⎫ ⎪ ⎠ ⎭ e*ik (4.λ* ) ( ik ) ( ik ) ∂ 2α lim ∂ ⎛ α lim ⎞ ⎛ 1 ∂α lim α lim ⎞ ⎛ 1 α lim α lim ⎞ = − 2 ⎟ =⎜ − 2 ⎟ = 0.57) one has T ∂ 2α lim ⎧ NG m ⎛ ∂eik ⎞ ∂ ⎪ = ⎨ ∑∑ ⎜ ⎟ ∂Z 2 ∂Z ⎠ ∂eik ⎪ ⎩ i =1 k =1 ⎝ (4.65) into (4.70) The second derivatives of αlim versus thickness variable Z are also necessary. j ik ⎜ ⎟ ∂X j ∂Y ∂X j ⎝ Y ⎠ e* .
23).λ* ) i =1 k =1 (e*ik . j ik eik t ik .71) can be simplified by ˆ ∂ ⎛ P ⎜ YN 0ε 0 k 3 ∂eik ⎜ ⎝ ( ( β12 + γ + β 22 + γ ⎟ ⎠ )⎟ ) ⎛ β β2 1 P3 + 8 ⎜ − 2 ⎜ β 22 + γ ⎝ β1 + γ ⎞ ⎟B ⎟ ⎠ ⎞ ⎫ ⎟ A⎪ (4. The second term in (4.∂ 2α lim ∂ ⎛ NG m T ⎞ = ⎜ −λ ∑∑ eik t ik .76b) Referring to (3. λ .72) ˆ P ∂ 2α lim ∂ ⎛ NG m ⎛ ⎜ ∑∑ ⎜ YN 0ε 0 k = 3 ∂Y ∂Z ∂Y ⎜ i =1 k =1 ⎜ ⎝ ⎝ ⎛ ˆ P = ∑∑ ⎜ N 0ε 0 k 3 i =1 k =1 ⎜ ⎝ NG m ( β12 + γ + β 22 + γ ⎟ ⎟ )⎟⎟ ⎞⎞ ⎠ ⎠ e*ik (4.71) can be rewritten as follows ∂ ∂Z ⎧ NG m ⎛ ˆ Pk ⎪ ⎜ ⎨ ∑∑ ⎜ YN 0ε 0 3 ⎪ i =1 k =1 ⎝ ⎩ ( β12 + γ + β 22 + γ ⎟ ⎬ ⎠⎭ ) ⎟⎪ ⎞⎫ ⎪ ⎞⎞ ⎟⎟ ⎟⎟ ⎠⎠ (4. c) versus Z and using the chain rule for two variables eik . the derivatives of λ and eik versus Z are needed. j ∑∑ ik ∂Z i =1 k =1 ∂Z (e*ik .73) ( ⎞ β12 + γ + β 22 + γ ⎟ ⎟ ⎠ e*ik ) The first term in (4.74) ⎟ ⎪ ⎠ ⎪ ⎬ eik ⎪ ⎪ ⎪ ⎭ ⎞ ⎧ ⎪16 β12 + γ + β 22 + γ ⎪ 3 ⎪ = YN 0ε 0 ⎨ ˆ 4 Pk ⎪ ⎛ 1 1 + ⎪+ ⎜ ⎜ β2 +γ β 22 + γ ⎪ ⎝ 1 ⎩ with P3 .61a.λ* ) (4.20) and after some manipulations we obtain 69 .49) and (3.71). B are given by (1.75) ⎛ ˆ P ⎛ β1 β2 = ∑∑ ⎜ YN 0ε 0 k ⎜ + 2 3 ⎜ 2 β1 + γ 2 β 2 2 + γ i =1 k =1 ⎜ ⎝ ⎝ NG m In order to complete (4.76a) ∑∑ t i =1 k =1 NG m T ik ∂eik =0 ∂Z (4. By taking derivatives of (4. A. j ⎟ ∂X j ∂Z ∂Z ⎝ i =1 k =1 ⎠ (e*ik . k (4.λ* ) NG m ∂e ∂λ NG m T =− − λ ∑∑ t T . one has Gik ∂eik ∂λ 1 ∂Hik eik = 0 − t ik + ∂Z ∂Z ˆ Pκ ∂Z ∀ i.
where P is a stochastic oneparameter load.4.49) becomes (4.81) ˆ If we now define a new strain vector e ik = xe ik and new displacement vector ˆ u = xu .77) From (4.t : ⎛ NG m T 0 ⎞ ˆ ⎜ ∑∑ e ik t ik − 1⎟ = 0 ⎝ i =1 k =1 ⎠ (4. i. (4.78) into (4.81) can be rewritten as follows α lim = min ( FP ) ˆ ˆ eik .80) α lim = min ( FP ) eik .t : ⎛ NG m T 0 ⎞ ⎜ ∑∑ eik xt ik − 1⎟ = 0 ⎝ i =1 k =1 ⎠ (4.u s.76a) we obtain the derivative of eik versus Z ∂eik ∂λ 1 − − ∂Hik Gik1 eik = Gik1t ik − ∂Z ∂Z ∂Z ˆ Pκ Substituting (4.78) ∑∑ i =1 k =1 NG m 1 T −1 ∂Hik t ik Gik eik ˆ ∂Z P κ NG m i =1 k =1 T ik ∑∑ t . with a reference load P 0 we have P = xP 0 and 0 0 t ik = xt ik where t ik is the stress vector due to referent load P 0 . then (4.u s. Otherwise we have as well ∂t ik t ik .⎛ ⎞ ⎜ ⎟ 2 2 ⎜ 4 3 β1 + γ + β 2 + γ P3 ⎟ ⎜ ⎟ ⎛ β ⎞ ⎟ ⎜ β2 ∂Hik 1 ⎟A⎟ = YN 0ε 0 ⎜ +2 3 ⎜ − ⎜ β2 +γ ⎟ ∂Z β 22 + γ ⎠ ⎟ ⎜ ⎝ 1 ⎜ ⎟ ⎛ ⎞ ⎟ ⎜ 3 1 1 ⎜ ⎟B ⎟ + ⎜+ ⎜ β2 +γ ⎜ 4 β 22 + γ ⎟ ⎟ ⎝ 1 ⎠ ⎠ ⎝ ( ) (4.e.5 Special case of probabilistic shakedown analysis Consider a special case of probabilistic shakedown analysis.76b) one gets ∂λ = ∂Z (4.79) G t −1 ik ik 4. the nonlinear optimisation (4.82) with 70 . = X ∂X In this case.
83) Since the constraint is now no longer depended on the load variable X .T m ⎧ m ⎛Y ⎫ ⎪ ⎞ c⎛ m ˆ ˆ u ⎞ ⎛ ∑ e − B u ⎞⎪ . the first one provides information for sensitivity analysis and the second one for the simple line search algorithm. two deterministic loops are required. the iteration for finding the design point.3 contains the logical connections of the main analysis steps as they have been implemented in our finite element code based probabilistic limit and shakedown analysis. FP = ∑ ⎨ ∑ ⎜ ηik ⎟ + ⎜ ∑ e ik − Bi ˆ ⎟ ⎜ ˆ ik ˆ i ˆ ⎟ ⎬ ⎠ 2 ⎝ k =1 i =1 ⎪ k =1 ⎝ X ⎠ ⎝ k =1 ⎠⎪ ⎩ ⎭ NG (4. = 0 .u* . i. it follows that NG m ∂α lim ∂FPL α ⎛ Y ⎞ = = ∑∑ ⎜ − 2 ηik ⎟ = − lim ∂X ∂X (e*ik . By this way.85) ∂ 2α lim ∂ ⎛ α lim ⎞ 1 ∂α lim α =− = − lim ⎜ ⎟=− ∂Y ∂X ∂Y ⎝ X ⎠ X ∂Y XY ∂ 2α lim ∂ ⎛ α lim ⎞ 1 ∂α lim . In each probabilistic iteration. and then we find again the above expressions finally. as mentioned in (4.86) (4. it is easy to prove that ˆ ∂λ λ ∂eik =− . X ∂X ∂X The flowchart in figure 4. 71 .84) ∂ 2α lim ∂ =− 2 ∂X ∂X 1 ∂α lim α lim 1 α lim α lim 2α lim ⎛ α lim ⎞ + 2 = + 2 = ⎜ ⎟=− X ∂X X X X X X2 ⎝ X ⎠ (4.35).e. =− ⎜ ⎟=− X ∂Z ∂Z ∂X ∂Z ⎝ X ⎠ (4.87) These above derivatives can also be easily obtained by using the same procedure as applied for the material variable Y .λ* ) i =1 k =1 ⎝ X X ⎠ e*ik ˆ ˆ ˆ (4.
Each design point corresponds with an individual failure mode or mechanism. FORM and SORM approximations are applied at each design point followed by a series system reliability analysis to lead to improved estimates of the system failure probability. 2007b. In series system.. because the remaining elements may be able to sustain the external loads by redistributing of the loads. i. failure in a single element will not result in failure of the system.e.. In this case. In parallel system. the concept of reliability has mainly been concerned with individual structural elements and failure mechanisms.3 is globally convergent [Tran et al. 1986]. when they exist on the limit state surface. as with any nonconvex optimisation problem. the formation of any individual failure mode or mechanism is defined as system failure. For example.. . To handle problems of this kind. d]. This chapter aims at presenting a method to successively find the multiple design points of a component reliability problem. Failure of such structures will always require that more than one element fails before the structure loses integrity and fails. the real structure is sometimes modelled by an equivalent system in such a way that all relevant failure modes can be treated [ThoftChristensen and Baker. A typical example of a parallel system is a statically indeterminate structures. The nonlinear optimization algorithm which was developed in section 4. However. provided that the limit state function is continuous and differentiable.5 MULTIMODE FAILURE AND THE IMPROVEMENT OF FORM/SORM RESULTS In the previous chapter. Structural systems can generally be characterized as series or parallel systems or some combination of the two [Madsen et al. in statically determinate or rigidplastic structures. the sequence is guaranteed to converge to a minimumdistance point on the limit state surface. formation of a collapse mechanism will result in failure of the total system and therefore they can be modelled as series system with each element of the series being a failure mechanism. 1982]. it is not guaranteed that the solution point will be the global minimumdistance point when the system has more than one failure mode (multimode failure or multiple design points). analysis of the structural system is required to evaluate the safety of the structure as a whole [Ditlevsen and Bjerager. The reliability of a real structure is usually much more difficult to evaluate since more than one element (member) can fail and because there is the possibility of more than one failure mode for the system. 1986].
em )de1..1. X n ) = 0. i = 1. 1955]. and the probability of the mode occurring may or may not provide a good estimate of the system probability of failure. If there are m failure mechanisms and the limit state surface is respectively described by m equations gi ( X ) = gi ( X 1 ..g... ∪ Em ) = P ⎜ ∪ Ei ⎟ . bounds relieve the necessity of evaluating the mdimensional integral either analytically. evaluation of the system probability of failure through direct integration may not be feasible. 1978].. even if an expression exists for the joint density function of failure modes and all failure modes have been identified. Unfortunately.3) If the joint probability density function of the failure events f E (e ) is known. a probabilistic algorithm was developed to identify the failure mode and to evaluate the failure probability of an elastoplastic shell structure. then the system probability of failure can be calculated by the ndimensional integral ⎛m ⎞ Pf = P ⎜ ∪ Ei ⎟ = ∫ . ⎝ i =1 ⎠ −∞ −∞ 0 0 (5.. ⎝ i =1 ⎠ (5.1 Multimode failure 5.5. m and if we denote the failure due to the ith mode as the random event Ei (5. It means that ⎛m ⎞ Pf = P ( E1 ∪ E2 ∪ .. it is rather an exception to deal with a problem where the failure of a structure can be satisfactorily defined by one limit state equation only. in general. [Hohenbichler. In this case... 1982].2) then the probability that the system fails is the probability that any failure mechanism occurs.. In structural analysis.... [Rackwitz. [Cornell. 1984].dem . Closer or secondorder bounds can be given in term of the individual failure probabilities and the joint failure probabilities between any two modes. the identified mode is only one of many modes that can occur. 1987]. 1981]) exist for evaluation of equation (5. Several firstorder bounds exist (e. 1967]) which only require knowledge of the individual probabilities of failure resulting directly from the axioms of the probability theory.4) Several techniques such as simulation methods or reducing the dimension of the original integral (e.1 Bounds for the system probability of failure In the previous chapter.g. [Dunnet et al. [Melchers. i = 1. these bounds may be quite wide for structural reliability application [Grimmelt and Schueller. numerically or through Monte Carlo simulation [Ramachandran.4). However. ∫ f E ( e1 . Thus...1) Ei = {x gi ( X ) ≤ 0} (5.5) 73 . m (5.. If we denote the individual failure probabilities Pi as Pi = P [ gi ( X ) ≤ 0] .
− β HLj . Each design point corresponds to a failure mode and they are the points on the limit state surface that have smallest distances to the origin in a standard normal space. be derived which include the joint probabilities of failure of any three failure modes. (5. Another bounds can.. ρ ) dxdy ij −∞ (5.. 5.9) 74 . that for practical problems this improvement is small compared to the improvement in using secondorder bounds instead of firstorder bounds [Hohenbichler and Rackwitz. The individual probabilities of failure Pi are determined as Pi = Φ ( − β HLi ) . 1979] m i −1 ⎧ P + ∑ max ⎨ Pi − ∑ Pij . however. Next section will present the calculation of these secondorder bounds based on the firstorder system reliability analysis.6) the ordering of the modes will have an effect on the bounds. of course.6) with the notation Pij has been used for the joint failure probability Pij = P ⎡ gi ( X ) ≤ 0. 1982]. In structural reliability. Figure 5. these bounds are frequently used and are considered sufficiently accurate for most structural systems [Grimmelt and Schueller.2 Firstorder system reliability analysis In a firstorder system reliability analysis. It appears. which is the corresponding reliability index. g j ( X ) ≤ 0⎤ . the failure set is approximated by the polyhedral set bounded by the tangent hyperplanes at the design points.7) Since not all couples of the random events Ei are taken into account in equation (5. we define the distance β HLi = ui* . the bounds of the system probability of failure for a series system are [Ditlevsen. ⎣ ⎦ (5. 2.m and associated with each design point. We denote the design points in standard normal space as ui* . The joint i failure probability Pij is thus calculated as Pij = Φ ( − β HLi . 1 i =2 j =1 ⎩ m m ⎫ 0 ⎬ ≤ Pf ≤ ∑ Pi − ∑ max Pij j <i i =1 i =2 ⎭ (5. ρij ) = − β HLi − β HLj −∞ ∫ ∫ ϕ ( x. Practical experience suggested that ordering the failure modes according to decreasing values Pi may correspond to the better bounds. 1983].1. i = 1.then.1 presents the projection of the limit state surface for the two failure modes on the plane containing the origin and the two design points u* and u*j . y.8) The firstorder approximation to Pij is obtained by approximating the joint failure set by the set bounded by the tangent hyperplanes at the design points for the two failure modes.
If we do not know where the values of the 75 . − β HLj ..12) Numerical techniques are available for evaluating the joint failure probability in equation (5.1 Geometrical illustration of firstorder system failure set Substituting the density function in (5.11) Failure region νij u* i Safe region βHLi u*j βHLj Origin Figure 5. − β HLj . NewtonCodes method.g. z ) ∂ρ dz = Φ ( − β HLi ) Φ ( − β HLi ) + ∫ ϕ ( − β HLi . which is based on geometrical illustration of multimode failure system can also be given. ρ ) . 1− ρ2 2π 1 − ρ 2 ⎣ 2 ⎦ 1 (5. thus avoiding any numerical integration [Madsen et al. ρ ) = ⎡ 1 x 2 + y 2 − 2 ρ xy ⎤ exp ⎢ − ⎥. e. y. z ) dz 0 ρij (5.6) still estimate a solution of the generally unknown region which respect to the exact value of the probability of failure. ρ ) is the probability density function for a bivariate normal vector with zero mean values. It should be noted that the bounds (5. 1986].10) and ϕ ( x.12). unit variances and correlation coefficient ρ ϕ ( x.with the correlation coefficient between two failure modes ρij are ρij = cosν ij (u ) = * T i u*j β HLi β HLj (5.0 ) + ρij ∫ 0 ∂Φ ( − β HLi . y. − β HLj . y. which gives Pij = Φ ( − β HLi .9) by the corresponding cumulative distribution function Φ ( x. Simple bounds on the joint failure probability.
13) * where B1 ( u ) defines the bulge fitted at u1 . a bulge B2 ( u ) is now added at u* resulting in the limit state function 3 2 76 . thus.1 Basic idea of the method In optimization theory. Approximating the limit state surface only at the global design point will lose these contributions. the neighborhoods of the local design points may also have significant contribution to the failure probability integral. has large error. the optimization algorithm which was developed in section 4. when they exist on the limit state surface. Solving the optimisation problem with the new limit state function g1 ( u ) leads to a second design point u* . Supposing that the first design point u1 is already found by the nonlinear optimization algorithm in section 4. the “barriers” around found solutions can be construct by moving the limit state surface in * the neighborhood of them away from the origin. Unfortunately. That is. They bound some approximation and we can only moreorless reasonably expect that the approximation is close to the exact result and the bounds remain meaningful. a simple method is presented for finding the multiple design points of a reliability analysis problem.3. In this section. 1998]. In that case. 5. one method often suggested to find the global and local solutions is to repeat the analysis with different starting points and hope that all optimal points will be found. A more effective method is to construct “barriers” around previously found solutions. the failure probability of series system is calculated by using firstorder system reliability method and secondorder bounds as presented above. the FORM/SORM solution will miss the region of dominant contribution to the failure probability integral and. thus forcing the algorithm to seek a new solution. In our problem to find the points which have the minimum distances from origin. We construct a “barrier” for this point by adding a ‘bulge’ to the limit state surface.2 Solution technique As discussed above.probabilities are placed with respect to the exact values.3 may converge to a local design point. the limit state function for the deformed surface is g1 ( u ) = g ( u ) + B1 ( u ) (5. a real structure may have several failure modes or failure mechanisms and the existence of multiple failure modes (or multiple design points) may cause the following problems in FORM and SORM. Thus. 5. it is possible for all trials to converge to the same point even when the other optimal points exist. Even if the global design point is found. Once all the design points are known. In order to seek the third 2 solution point u* .2. we can not confirm that the bounds given above estimate the probability of failure. The method was developed by [Der Kiureghian and Dakessian.
g 2 ( u ) = g1 ( u ) + B2 ( u ) = g ( u ) + B1 ( u ) + B2 ( u ) . • The bulge should have a strong outward curvature at its feet in order to make the search algorithm converging to an existing genuine design point rather than to its feet. convergence to a spurious point usually means that no other genuine design point exists [Der Kiureghian and Dakessian. 5. This nature can be used as the stopping criterion of the algorithm. • Bi ( u ) must be continuous and differentiable since the optimization algorithm requires a continuous and differentiable limit state function. Repeating the process until all design points are found. is g m −1 ( u ) = g ( u ) + ∑ Bi ( u ) i =1 m −1 (5.2. The limit state function for finding the mth design point thus. Thus. practical experience showed that this occurs only when there is no other genuine design point.2 Definition of the bulge The derivation of a bulge Bi ( u ) at a design point u* is based on the following i requirements • Bi ( u ) must be positive in the neighborhood of u* (in order to move the limit state i surface away from the origin) and zero elsewhere.14) u2 γβ HLi u* i βHLi u1 Figure 5. 1998]. 77 . which are actually the spurious minimumdistance points.2. However. it is possible for the optimisation algorithm to converge to the points located at the feet of the bulge.2 Definition of a bulge at design point u* i As is shown in the figure 5.
Of course. 78 . one may wish to add a new bulge at this spurious design point and continue the search. The scale factor si governs the height of the bulge and.15) and (5. To complete the method one needs to know whether a newly found design point is genuine or spurious. (5. it i follows that si = δβ HLi ∇g ( u* ) i ⎡(γβ HLi ) 2 − (δβ HLi ) 2 ⎤ ⎦ ⎣ 2 . Suppose we want to move the design point u* away from the origin by the amount δβ HLi . therefore. However. Convergence to a spurious design point usually means that no more genuine design points exist. is spurious.Basing on the above considerations. thus. It must be i chosen large enough to assure a strong outward curvature at the food of the bulge. This is easily solved by calculating its distances from previously found design points. where i 0 < δ < γ .16) The definition of the bulge at each design point in term of β HLi . Since the slope of the limit state surface at u* is ∇g ( u* ) . the spurious design points are not included in the reliability calculations. then the newly found design point is at the foot of this bulge and. whereas a large one may result in a bulge that conceals other significant design points. it is seen that the i i necessary height of the bulge at that point is δβ HLi ∇g ( u* ) . the distance by which the limit state surface in the neighborhood of u* is moved away from the origin. δ is complete through the two equations (5. ∇g ( u* ) and two i parameters γ .16). Der Kiureghian and Dakessian proposed a spherical shape of the bulge as follows ⎧ 2 * ⎪ s r − u − ui Bi ( u ) = ⎨ i i ⎪ 0 ⎩ ( 2 2 ) if u − u* ≤ ri i (5. From equation (5.15) elsewhere where ri = γβ HLi is the radius of the bulge. The parameter γ should be chosen carefully since a small γ is conservative but may produce insufficient curvature at the foot of the bulge. If any one of these distances is smaller than or equal to the radius of the corresponding bulge.15). in order to make sure that this is the case.
5 can be used to exclude plastic collapse under monotonic loading by gross plastic deformation.. numerical and experimental solutions found in literature are briefly represented and compared.3. [Vu et al. The structures under consideration can be subjected to mechanical and thermal loads and are made of elasticperfectly plastic material. 2003].and postprocessor GiD [GiD 7. 1999].732 . [Taylor et al. a free open source computing software [Code_Aster 7. which are based on the Kirchhoff’s hypothesis.1547 is effective in pipes and 3 cylindrical pressure vessels. Although the flat shell elements may fail to satisfy the basic convergence properties that should be expected from sound finite element procedures in general. 2003]. 2007]. . The total safety factor is 3 = 1. In each numerical test. including both plate and shell structures. were applied. • PD check: Shakedown analysis with the von Mises yield function may be used without a safety factor to exclude incremental collapse by ratcheting under cyclic loading. The presented shakedown method applies also to more effective shell elements that work with generalized variables. they can produce reasonably accurate approximations [Chapelle and Bathe. 2002].2. In practice the nonsmooth Tresca function is replaced by the von Mises 2 function so that an additional safety factor of = 1.6 LIMIT AND SHAKEDOWN ANALYSIS OF DETERMINISTIC PROBLEMS The deterministic and probabilistic shakedown optimizations described in the above chapters is programmed and implemented in the finite element package Code_Aster. Nine examples are presented below for the sake of evaluation the algorithm. some existing analytical. But currently only flat elements are available in Code_Aster. • GP check: Limit analysis with the Tresca yield function and a typical safety factor of 1. 200506]. Perfectly plastic LISA methods cover two mandatory design checks of the new European pressure vessel standard [European standard. The 3node triangular and 4node quadrangular isoparametric flat shell elements.. The finite element discretisations were realized by the personal pre.
The pipe is subjected to bending and torsion moments M b . the following analytical solution may be used 80 . M t as well as internal pressure p and axial tension F (figure 6. the plastic collapse limit is then M tl = 2 π r 2hσ y . 3 (6.With respect to pressure vessel design it can be concluded that shakedown analysis is 1 decisive if the shakedown load is less than = 0. The present shakedown analysis includes also the AP check against failure by alternating plasticity.1) where σ y denotes the material yield stress. r the mean radius and h the thickness of the pipe. 1975].2) Furthermore if bending moment. If only torsion moment is applied on the pipe. 6.5 is applied in the GP check then the PD check is decisive if the shakedown load is less than 0. consider the pipe under pure bending. internal pressure and axial tension apply at the same time on a thinwalled pipe. The analytic solution of plastic collapse limit may be easily obtained by Bernoulli's theory M bl = σ yη plWel = 4σ y h( r 2 + h2 ) 12 (6. However. this check is not mandatory. Mbl is the maximum bending moment we can apply on the pipe while other loads are removed. r p Mt Mb F h L Figure 6.. Vu [2001] and [Larson et al. We denote here L the length.1 Limit analysis of a cylindrical pipe under complex loading Consider an example of a cylindrical pipe under complex loading.1).1 Cylindrical pipe under complex loading Firstly. which was examined by Yan [1997]. 6 –times the limit load.57735 –times the limit load. If only the 3 safety factor 1.
and the numerical error is only 0.3) in which m. the limit load factor is 1. the limit load factor can be estimated by using (6. nϕ . h = 60mm.4) are taken from Yan [1997]. it is easy to find that n x = nϕ / 2 and the formulation (6. σ y = 160MPa . Figure 6.53% compared with the analytical solution. On the other hand. If axial tension is caused by only internal pressure (the pipe has closed ends).2 FEmesh of cylindrical pipe When the two loads (bending moment and internal pressure) are applied.1÷6. (6.5) In our numerical analysis the whole pipe is discretized by 300 quadrangular flat 4node shell elements as shown in figure 6.0053. It is observed that the numerical solution agrees very 81 .3) now becomes m= 2 4 − 3nϕ 2 .2.3) presents both numerical and analytical results.24 kNm (formula 6. if only torsion moment is applied with its limit value M t = M tl = 3134.52 kNm (formula 6.5). The interaction diagram (figure 6. In this case. pl = σ y h/r Fl =2πσ y rh (6. the numerical error is 1% compared with the analytical solution. nχ are the normalized quantities m =M/M l .010. M l =4r 2 hσ y nϕ = p/pl . the limit load factor is 1.4) The formulations (6.1). r = 300mm. If only a bending moment is applied with its limit value M b = M bl = 3467 .2). The following geometrical and physical parameters are adopted: L = 2700mm. nx =F/Fl .m= 2 4 − 3nϕ 2 n − 2n π x cos ϕ 2 4 − 3nϕ 2 (6.
well with the analytical one: the maximum numerical error is less than 1%. numerical error is only 0.3 Cylindrical pipe under bending and internal pressure 6. the limit load factor is 1. In this case. (6.2 Limit and shakedown analysis of a thinwalled pipe subjected to internal pressure and axial force In this example. the plastic collapse limit can be computed by using the condition: p2 F 2 p F + 2 − =1 2 pl Fl pl Fl where pl = β effect.4 is considered.15% compared with the analytical solution.6) σ yh r . Fl = 2π rhσ y with β = 1 for a long pipe without the end constraining 82 .0015. if only axial tension is applied. On the other hand. a thinwalled pipe with thickness h and mean radius r as shown in figure 6. The pipe is subjected to internal pressure p together with axial force F . Figure 6. When both internal pressure and axial force increase monotonically and proportionally. This problem was investigated analytically by Cocks and Leckie [1988] using the Tresca yield criterion and by Yan [1997] using the von Mises yield criterion.
p F F Figure 6. In all cases. Analytical solutions are computed by using the formulations (6.5 with the following geometrical properties and material yield stress: r = 300 mm . F ] . L = 2700 mm . h r 83 . pl2 Fl 2 pl Fl (6.5 FEmesh of the pipe Note that (6.8) In our numerical analysis the whole pipe is discretized by 400 3node triangular flat shell elements as shown in figure 6.7). the shakedown range is limited by the condition [Cocks and Leckie 1988]: p F =1− . if the Tresca yield criterion is used. the diagram shows good agreements between numerical and analytical solutions. h = 10 mm . σ y = 116 MPa .6) and (6. The interaction diagram for limit and shakedown analysis with normalized pressure p / pl and normalized axial force F / Fl is plotted in figure 6. pl Fl (6. the shakedown limit can be computed by using the following condition: p2 F 2 p F + + = 1. On the other hand. The numerical error is less than 1%.7) were found by using the von Mises yield criterion [Yan 1997].6) and (6.6.4 Thinwalled pipe If internal pressure remains constant while axial force varies within the range [− F.7) Figure 6.
Figure 6. consider an axially restrained cylindrical shell of mean radius r and thickness h . υ = 0.3 .9) 84 . the shakedown load can be computed by using the following condition [Zhang. E = 151000 MPa .6 Interaction diagram 6. αT = 1.3 Cylindrical shell under internal pressure and temperature change In this continuous wellknown example. h = 10 mm in which E is Young’s modulus. Two following cases are examined Uniformly distributed temperature This problem was examined by Zhang [1995b] and Vu [2001]. T0 + ∆T ] .7.2 MPa . Omitting the failure caused by axial instability. sT = 2σ y h 2σ y (6. The inner and outer temperature are the same Ti = Tu = T and vary within the range [T0 . The shell is subjected to a constant internal pressure p and a uniformly or linearly distributed temperature T which may vary within a certain range. Numerical analysis is carried out by using 200 quadrangular flat 4node shell elements as presented in figure 6.0 × 10−5 o C −1 r = 500 mm . 1995b] 2 s 2 + sT = 1 p sp = EαT ∆T 3 pr . The following material properties and geometrical data are adopted: σ y = 116. αT is the coefficient of thermal expansion and σ y is the yield limit.
8 Interaction diagram (uniformly distributed temperature) Numerical results with normalized loads ( psd .8.7 FEmesh and geometrical dimensions Figure 6. On the other hand. when only temperature applies p = 0 .684 MPa .605 oC . It can be shown that these results agree well with the analytical r 85 . Tu Ti p h Figure 6. ∆Tsd ) are presented in the interaction diagram. it results in ∆Tsd = 102.In the case of ∆T = 0 the analytical solution leads to psd = 2. figure 6.
Numerical results with normalized loads ( psd . figure 6. T0 ] while the outer one is set to zero Tu = 0 . The essential dimensions are chosen as follows Internal diameter of the pipe Internal diameter of the junction 86 D = 39 mm . Linearly distributed temperature This case was investigated by Heitzer [1999] when the temperature field is linearly distributed between inner and outer wall of the shell. Tsd ) are presented in the interaction diagram.4 Pipejunction subjected to varying internal pressure and temperature Thermal loading is known to have a great effect on the shakedown or inadaptation behaviour of structures.248 oC . It can be shown that the collapse mode at shakedown limit is alternating plasticity when pressure is small and purely incremental plasticity when pressure becomes bigger. The inner temperature Ti may vary within the range [0. The supposed temperature field T is a linearly distributed function through the thickness with outer value Tu and inner value Ti = 0 . . In the case of p=0 shakedown analysis gives the shakedown load Tsd = 136.9 Interaction diagram (linearly distributed temperature) 6. Figure 6. It is noted that the collapse mode at shakedown limit is alternating plasticity in the axial direction of the pipe only when pressure is absent and purely incremental plasticity in the other cases. d = 15 mm . The maximum error is less than 1%. In this example a pipejunction subjected to internal pressure p and temperature loads is examined.approximation.9.
which is based on the lower bound approach. numerical computations are carried out by using 800 flat 4noded shell elements as presented in figure 6.145σ y = 23. the problem was analyzed by Vu [2001] with a primaldual shakedown algorithm using 720 solid 20node hexahedron elements. D 87 s . 125 solid 27node hexahedron elements were used to model one fourth of the structure due to its symmetry.Thickness of the pipe and junction s = t = 3. υ = 0. consider firstly the case of having only mechanical load.10. d t Figure 6. The following material properties are adopted: σ y = 160 MPa.10267σ y = 16.10 FEmesh and geometrical dimensions Pipejunction under internal pressure For the sake of comparison.3 with the coefficient of thermal expansion αT = 1. the internal pressure may vary within range [0. E = 210000 MPa.2).43 MPa compared with other known solutions (table 6. In the work of Heitzer and Staat [1999]. p0 ] . Convergence analysis of the optimization algorithm shows that the shakedown load factor becomes stationary also after five iteration steps. In the case of shakedown analysis.0 × 10−5 oC −1 . In our analysis. Numerical analysis gives the shakedown load psd = 0.11 shows that the limit load factor becomes stationary after only five iteration steps. Analogously. The convergence property of this solution presented in figure 6.44 mm . It is noted that in this case the structure fails due to the alternalting plasticity.1). Numerical limit analysis leads to the plastic collapse pressure pl = 0.2 MPa compared with other results (table 6.
134 0. where Tu may vary within the range [ 0.136 0.10267 Nature of solution lower bound lower bound upper bound upper bound Source [Heitzer. 2001] double elastic slope method [Vu. Numerical analysis gives the shakedown load Tsd = 93. 2001]  Figure 6.145 Nature of solution Source lower bound [Heitzer. 1999] [Vu.14424 0. 1999] German design rules [Heitzer.141 0. 1999] lower bound [Vu.1: Limit analysis of pipejunction Limit load normalized with σ y Staat and Heitzer ADMerkblatt B9 Vu Vu SAMCEF Vu Present 0.161 0. T0 ] . It is also noted that in this case the structure fails due to the alternating plasticity.11 Convergence of limit and shakedown load factors Pipejunction under temperature change Consider the case of only thermal load being applied.5086 oC . 2001] upper bound  Table 6. 88 . 2001] [Vu. 2001] upper bound [Vu.11044 0. 2001] step by step method [Vu.Table 6.0952 0.14434 0.10983 0.2: Shakedown analysis of pipejunction Shakedown load normalized with σ y Staat and Heitzer Vu Vu Present 0.
two loading ranges are examined 1) The pressure and temperature vary simultaneously and dependently (oneparameter loads) 0 ≤ p ≤ µ p0 0 ≤ Tu ≤ µT0 . 0 ≤ µ ≤1 2) The pressure and temperature vary independently (twoparameter loads) 0 ≤ p ≤ µ1 p0 .Pipejunction under internal pressure and temperature For this case. 0 ≤ µ1 ≤ 1 0 ≤ µ2 ≤ 1 (6.10) Figure 6.12 Interaction diagram of pipejunction^ 89 . 0 ≤ Tu ≤ µ2T0 .11) (6.
e.5 Grooved rectangular plate subjected to varying tension and bending In order to examine the limit and shakedown loads of thin plates. both positive and negative temperatures T0 ≥ 0 and T0 ≤ 0 are included.12.13 FEmesh and geometrical dimensions 90 . we consider firstly a grooved rectangular plate subjected to inplane tension p N and bending p M (figure 6. In all cases. Otherwise the elastic limits and shakedown loads become bigger when the loads vary simultaneously and dependently (oneparameter loads) with the temperature varying in the positive range.13). σ y (6. while the elastic limits are not. the shakedown limit is twice the elastic limit which indicates that inadaptation will occur due to alternating plasticity. i. From the interaction diagram it can be shown that a thermal load has no effect on the limit load of the pipejunction. Numerical results are presented in figure 6. pM pN R R L Figure 6.12) where σ y denotes the material yield stress. the shakedown loads are symmetric via pressure load axis. In this case the shakedown check is not mandatory although this leads to a design which is prone to LCF. The load domain is defined by pM ∈ 0 . σ y pN ∈ 0 . In the case of twoparameter loads. but they are identical when the temperature is in the negative range. 6. Staat et al. [2005] have discussed this example demonstrating that twice the elastic limit is a PD check which should be used with great case or better avoided at all because it is an AP check and it may be not safe for independently varying loads.In order to get the general solutions. they are the same for both positive and negative ranges of the outer temperature.
the structure is modelled by 300 quadrangular flat 4node shell elements as shown in figure 6. which used volume finite elements basing on plane stress assumption.727÷0. These results agree well with the solutions of Vu [2001] and Heitzer [1999]. It is observed that. Heitzer [1999] and Vu [2001] investigated the problem for the more complicated case with p N ≠ 0.572 Plane strain 0. In the present analysis. In the case of limit analysis. Figure 6. The interaction diagram of the plate in figure 6. in almost cases.500 0. It can be shown that our solution is well accordant with the others.769 0.500÷0. shakedown limit and elastic limit. pM = 0 .2 MPa . σ y = 116.577 0. p M ≠ 0 .800 0.558 0.23624 compared with upper bound factor of 0.699 0. 91 . Numerical result gives the shakedown load factor α sd = 0.15 shows three limit lines: the plastic collapse limit.3 . Casciaro and Cascini [1982] and Yan [1997] for pure tension load p N ≠ 0.568 0.630÷0.14 shows the evolutions of limit and shakedown load factors. E = 21000 MPa . The plastic collapse load factor is presented in table 6.802  Nature of solution analytical numerical analytical numerical numerical numerical Yield criterion Tresca von Mises von Mises von Mises von Mises von Mises For the case of having both the inplane tension and bending. The following geometrical and material properties are chosen: R = 250 mm .Limit analysis of the structure was studied by Prager and Hodge [1951]. p M = 0 . Consider the case of constant pure tension pN = σ y . limit and shakedown analysis are implemented. the upper bound converges rapidly to the solution α l = 0.13.3 together with known solutions for comparison. h = 50 mm .695 0.30498 . Table 6. ν = 0. inadaptation will occur due to alternating plasticity.799÷0.23494 obtained by Vu [2001]. pM = 0 ) Plane stress Prager Casciaro Yan Yan Vu Present 0.3: Limit analysis ( pN = σ y . L = 4 R .557 0.
e. σ y . σ y Figure 6.g. Numerical limit and shakedown analyses were also investigated by some authors. Heitzer [1999]. a square plate with central circular hole subjected to two loads p1 and p2 which can vary dependently or independently is considered.15 Interaction diagram of grooved rectangular plate 6. The limit load of the problem was examined analytically by Gaydon and McCrum [1954] by using plane stress hypothesis and von Mises yield criterion.6 Square plate with a central circular hole In this continuous wellknown example.Figure 6.14 Limit and shakedown analysis pN ∈ 0. 92 . Vu [2001] for different ratios R / L to evaluate the elasticplastic behaviour of the structure. pM ∈ 0.
The following material properties are chosen: E = 21000 MPa .3 . the alternating (plastic fatigue) limit can be estimated by using formula (2. while the exact one was used in the unified shakedown analysis.2 is plim = 0.60332σ y . the collapse mode is alternating plasticity. σ y . Based on an elastic analysis. ν = 0.2 . the exact plastic collapse limit is found since in this range the lower and the upper bound are coincident plim = (1 − R / L ) σ y . (6. The small gap between them may be caused by the use of the quadratic approximation of exact Ilyushin yield surface in the calculation of alternating plasticity.80149σ y . The convergence property of these solutions presented in figure 6. due to the symmetry.204 . The numerical results obtained by our unified shakedown algorithm represent the minimum between the alternating limit and the incremental limit. Our obtained alternating limit is 0.59904σ y and the unified shakedown algorithm gives 0.47).8σ y . σ y = 116. p1 ∈ 0. For R / L = 0.2 p1 L R Figure 6.17 shows that the limit and shakedown load factor become stationary after seven iteration steps.13) As example.16. the exact collapse limit load in the case of R / L = 0. With 0 < R / L ≤ 0. One applied load In this case one load is set to zero and the other one can vary within a range p2 = 0. The two cases: one and two applied loads are examined.2 MPa . one fourth of the plate is discretized by 600 quadrangular flat 4node shell elements as shown in figure 6.16 FEmesh and geometrical dimensions In our analysis. 93 . Our numerical solution obtained in this case is 0.
1327 0.8951 0. σ y .2556 0.6546 0.17 Limit and shakedown analysis p1 ∈ 0. The obtained numerical results are introduced in table 6.572 0.90172 0. p2 = 0 R/L 0.1378 0.2104 0. It can be shown that our solutions agree well with those of Heitzer. σ y .4 0.7879 0. Table 6.0565 0.5 0. compared with the solutions of Heitzer [1999].05227 0.01226 Two applied loads For this case.70221 0.01226 0.40117 0.52012 0.4409 0.483 < R / L ≤ 1 where analytical lower bound coincides with upper one: plim π sin α − σ y .5212 0.0191 0.31600 0.80149 0.9 Shakedown load Limit load Shakedown load Limit load factors (Heitzer) factors (Heitzer) factors (present) factors (present) 0.12541 0.43367 0. exact value of the plastic collapse limit load are known in the range 0.24249 0. which are obtained by lower bound algorithm.7 0.59139 0.6157 0. various values of ratio R / L were also considered.4361 0. = 6 3 2 1 ( R / L) 2 3 e = 2cos (α ) π 3 α − 6 (6.05225 0. Figure 6.4.3 0.2 0.0557 0.6 0.691 0.14) 94 .12378 0.21323 0.8 0.0193 0.60332 0.671 0.3302 0.1 0. p2 = 0 In order to examine the geometrical effect of the circular hole.4: Limit and shakedown analysis p1 ∈ 0.
89374 0.19445 0.02903 0.67585 0.06765 0. twoparameter loads R/L 0.96951 0.483 can be found in the work of Gaydon and McCrum [1954].43791 0.55666 0.5: Limit analysis p1 = p2 = σ y R/L 0.6 0. Numerical limit load factors for different values of R / L are introduced in tables 6. σ y .9 Lower bound Upper bound Lower bound (analytical) (analytical) (Khoi [2001]) 0.43390 0.5 0.89472 0.10457 Table 6.32504 0.10249 0.27638 0.12360 0. 1) Two loads vary simultaneously and dependently (oneparameter loads) 0 ≤ p1 ≤ µσ y .79125 0.36128 0.67602 0.00709 0.3 0.99215 0.21010 0.21285 0.89425 0.27635 0.02772 0.10249 0.49082 0.12102 0.00710 Present 0.35575 0.48718 0.43801 0.20991 0.43819 0.49086 0.5.4 0.10267 Present 0. 1954] and numerical results of Vu [2001].97104 0.67592 0.00709 Upper bound (Khoi [2001]) 0.56133 0.1 0. σ y .8 0.55682 0.36128 0.00640 In the case of shakedown analysis three loading ranges are examined including both positive and negative loads p2 ≥ 0 and p2 ≤ 0 .02905 0.6 0.69048 0. 0 ≤ µ ≤1 0 ≤ p2 ≤ µσ y .43136 0.19442 0.92376 0.19052 0.55679 0.15) 2) Two loads vary independently (twoparameter loads) 95 .02903 0. Table 6. compared with analytical results [Gaydon and McCrum.80829 0.27635 0.7 0.43384 0.89404 0.79122 0. which are obtained by a dual algorithm.1 0. p2 ∈ 0. or 0 ≥ p2 ≥ − µσ y (6.43801 0.06763 0.32196 0.32221 0.79075 0.21016 0.32195 0.97063 0.7 0.19442 0.Details about the upper and lower bounds within the range 0 ≤ R / L < 0.36131 0.4 0.32195 0.6: Shakedown analysis p1 ∈ 0.79167 0.3 0.12360 0.27300 0.5 0.49082 0.8 0.12364 0.06763 0.2 0.2 0.55682 0.43384 0.06940 0.97082 0.10264 Upper bound (Khoi [2001]) 0.44109 0.67814 0.20991 0.9 Alternating limit Lower bound (Khoi [2001]) (Khoi [2001]) 0.
0 ≤ µ1 ≤ 1 0 ≤ µ2 ≤ 1 (6.17) Figure 6. 0 ≤ µ1 ≤ 1 0 ≥ p2 ≥ − µ2σ y .0 ≤ p1 ≤ µ1σ y .16) 3) Two loads vary independently (twoparameter loads) 0 ≤ p1 ≤ µ1σ y . R/L = 0.6. 0 ≤ p2 ≤ µ2σ y . Numerical shakedown load factors for different values of R / L are introduced in tables 6. which are obtained by a primaldual algorithm.18 Interaction diagram of a square plate with a central hole.2 For the sake of comparison. 96 . compared with numerical results of Vu [2001]. 0 ≤ µ2 ≤ 1 (6. consider first the twoparameter load case with positive load ranges.
2 . One of its ends is supposed clamped and the other one is subjected to an inplane closing moment M I or outofplane bending moment M II as shown in figure 6. The momentend rotation curves show a defined limit load behaviour for the closing mode of inplane bending but not for the opening mode [Chattopadhyay et al. For this reason.. Otherwise the elastic limits and shakedown loads become bigger when the loads vary simultaneously and dependently (oneparameter loads) with one load varies in the positive range. 1994]. under the same loading conditions. 2000]. in the case of twoparameter loads. 2005].18) Generally. 2007a. This difference can be explained as an effect of large displacements. [Tran et al. b]. while λ → ∞ corresponds to a straight pipe.. 97 . However. the stresses and strains are much higher than those present in a straight pipe of the same size and material. 1.7 Elbow subjected to bending moment The pipe bends are a problem of great interest to many designers. [Chattopadhyay et al.. When an external moment is applied to one of its ends. The following material properties are chosen: E = 208000 MPa .. pipe bends are considered as critical components of a piping system.19.1. [Diem. i. They have a complex response to inplane and outofplane bending moments. 1996. 2000]. In order to evaluate the model different values of λ within the range [0. 6. σ y = 250 MPa .19. the cross section tends to deform significantly both in and out of its own plane. [Mourad et al. 1999. Consider an 90o elbow with mean radius r . The curvature factor is defined as follow λ= Rh r2 (6. while the elastic limits are not.3 .. [Yan. Our model for elasticplastic analysis is meshed by 700 quadrangular flat 4node shell elements as shown in figure 6. analytically and experimentally [Kussmaul et al. 1998a. but they are identical when this load is in the negative range.. [Fanous et al.Figure 6. the shakedown loads are symmetric via p1 / σ y . It is also noted that when one load varies within the negative range. From this interaction diagram it can be shown that. 1997].axis . λ ≤ 0.e]. they are the same for both positive and negative ranges of p2 .18 presents both the numerical limit and shakedown analyses for the case of R / L = 0. 2002].5 corresponds to a highlycurved pipe. the examination of shakedown limit is still rare. 1995].2] are examined. the collapse mode is purely alternating plasticity. ν = 0. bend radius of curvature R and thickness h . [Shalaby et al. Due to their curved geometry. In addition.. the variation of the limit moment with the change of the internal pressure has been of great interest and has been studied numerically.e.
M Is is determined as (see formular 6. (6.9346λ 2 / 3 . 1978a. some analytical solutions are available in the literature 1) Based on small displacement analysis and perfectly plastic material behavior • Calladine’s and Yan’s solutions By using an elastic solution of Clark and Reissner [1951] and the criterion of Hodge [1954].20) For this kind of pipe bend.21a) This solution is considered in the literature to come close to the experimental limit load factor [Bolt and Greenstreet.19 FEmesh and geometrical dimensions Elbow under inplane closing bending moment We define the limit load factor as follows α I = M Ie / M Is (6. 1979]. For a straight thinwalled pipe. Calladine [1974] proposed a lower bound solution for an infinite highlycurved pipe ( λ ≤ 0.19) in which M Ie is inplane limit moment of elbow. [Goodall. which has the same radius as the elbow. b]. [Griffiths. which is a kind of sandwich approximation of Mises' criterion in the case of a cylindrical shell. (6. 1972]. According 98 . M Is is the limit moment of the straight pipe.M II h MI R 2r Figure 6.5 ) α IC = 0.4) M Is = 4hr 2σ y .
For a slightlycurved pipe ( λ ≥ 0. They result with λ < 1. it is a good approximation when λ < 0.21b) In the framework of his PhD thesis.23b) Goodall [1978a] proposed the maximum loadcarrying capacity of the elbow subjected to closing bending moment as α IG = where 1. [1989] proposed the following equations of collapse moments of elbows 99 . The second one is to manipulate the results of a creep analysis of bends to give an approximate solution of the limit load.25) • Touboul’s solution Based on the experimental study at CEA DEMT.45 .22) λ2 In order to get an analytical solution for the limit load of an elbow.to Yan [1996. Touboul et al. 1997]. (6.7 .23a) (6. he proposed an approximate solution which is validated by numerical analysis α IY = cos( • Desquines’s solution π ). β= 2 + π Eh 3 (6. [1997] proposed a more general analytical solution as a lower bound. Desquines et al.45 α ISF = 0. 2) Based on large displacement analysis • Goodall’s solution λ < 1.45 .6 . λ > 1.3015 (6.8λ 0.24) 2/3 4 3(1 − ν 2 ) rσ y ( 3λ ) . which can be applied for any value of λ α IDe = 1+ • Spence and Findlay’s solution 1 .04λ 2 / 3 1+ β (6. 6λ (6. 0. The first one utilises a linear elastic analysis to estimate limit loads as similar to Calladine.7 ). α ISF = 1. Spence and Findlay [1973] already developed two theoretical approaches.
They converge as an upper bound of Calladine’s solution and lower bound of Desquines’s solution. (6.28) e t where M II is outofplane limit moment of the elbow. It is shown that our solutions compare well with the analytical solution outside the range 0.4 ≤ λ ≤ 0. compared with the analytical solution of Yan [1997].5 .9λ 1/ 3 . compared with these above analytical solutions and a numerical solution of Yan [1997]. 1997] proposed an analytical solution for the elbow subjected to outofplane bending moment Y α II = 1.21.7 and figure 6. λ < 0. Y α II = 0.8 and figure 6.722λ 1/ 3 .6 .5 ≤ λ ≤ 1. Yan [1996.4 .29) By this definition. which are based on small displacement theory. but are bigger than those which are based on large displacement theory. (6.26a) (6.30) (6. • Drubay’s solution (6.27) Our numerical results are introduced in table 6.715λ 2 / 3 .Closing collapse: α IT = 0. 3 (6.20. M II is the torsion limit moment of the axle which has the same radius as the elbow and determined by t M II = 2 π hr 2σ y .769λ 2 / 3 .26b) Drubay et al [1995] proposed another closing mode collapse moments of elbows as α IDr = 0. 100 .1λ 0. 0.31) Numerical results are introduced in table 6. It is shown that our solutions compare well with the other analytical solutions. Elbow under outofplane bending moment We define the limit load factor for this case as follows e t α II = M II / M II (6. Opening collapse: α IT = 0.7 .
5589 0.3709 0.909 0.7525 0.2155 0.7 Limit load factor 0.3052 0.5200 0.200 SpenceCalladine Yan Desquines Goodall Findlay (6.7184 0.7868 0.3204 0.3913 0.263 0.3279 0.4144 0.8660 0.1791 0.3885 0.9063 0.Table 6.21b) (6.3422 0.8327 0.5902 0.3196 0.5087 0.668 0.4844 0.7153 0.400 0.4188 0.4175 0.8765 0.7494 0.577 0.7368 0.903 1.2 0.100 0.6 0.5888 0.3900 0.9006 0.8365 0.26) (6.769 0.8617 0.7639 0.4617 0.21a) (6.24) (6.5278 0.6998 0.5879 0.700 0.250 0.23) 0.600 0.7330 0.6732 0.3446 0.000 1.8 Curvature factor 1 1.6178 0.2 0.6260 0.154 0.4 0.22) (6.2817 0.6732 0.5637 0.9782 0.5888 0.2013 0.6741 0.4593 0.2 Figure 6.4955 0.750 0.3947 0.6063 0.6348 0.20 Limit load factors of an elbow under inplane closing bending moment 101 .3482 0.2445 0.7660 0.1489 0.500 0.363 0.1138 Touboul Drubay Present (6.7933 0.4504 0.1657 0.800 0.27) 0.5365 0.2001 0.7773 0.300 0.3 0.3046 0.2838 0.6458 0.8069 0.8341 0.5471 0.715 0.8684 0.3401 0.5 0.7554 0.8244 0.7377 0.8000 0.4355 0.3638 0.5887 0.200 0.7945 0.8074 0.8544 0.4756 0.650 0.9 0.4 0.4614 0.6648 0.7974 0.6 0.7227 0.6627 0.6162 0.7013 0.5074 0.1 CalladineYan Desquines Spence & Findlay Yan (numerical) Present 0.6930 0.9093 0.8925 1.7: Limit load factors of an elbow under inplane closing bending moment λ 0.3882 0.5515 0.8 0.4794 0.
8276 0.8699 0.100 0.3 0.4 0.9467 1 0.8177 0.7257  Yan (6.21 Limit load factors of an elbow under outofplane bending moment 102 .550 0.5675 0.9564 Present 0.000 1.8355 0.5341 0.7591 0.650 0.800 0.700 0.8 0.450 0.6924 0.600 0.8674 0.400 0.5244 0.7 0.6813 0.Table 6.6337 0.6063 0.7245 0.2476 0.8984 0.7539 0.4047 0.8: Limit load factors of an elbow under outofplane bending moment λ 0.6575 0.5 0.2763 0.7374 0.363 0.750 0.6 0.4188 0.2 Yan Present Limit load factor 0.5989 0.7143 0.200 Yan (6.250 0.4709 0.200 0.4788 0.6 0.500 0.8053 0.6348 0.7808 0.8 Curvature factor 1 1.903 1.2 0.30) 0.4 0.7796 0.9000 0.31) 0.2 Figure 6.7991 0.9 0.300 0.
One of its ends is clamped and the other one is free. M II Fp MI R h p L F p F 2r Figure 6. In the case of transversal forces.22. due to the symmetry. In this case.6. axial force Fp . one fourth of the structure is modelled by 300 quadrangular flat 4node shell elements as shown in figure 6. ν = 0. 2 (6. M II as shown in figure 6.25 mm . The geometrical data and material properties are chosen as: h = 30..23. Table 103 .22 Pipeelbow subjected to complex loads Pipeelbow under transverse forces F This is an experimental model presented in [Diem. 1994]..522 ). the structure can be considered as being subjected to the inplane closing bending moment which is determined by 1 M =F ( L − R) + R . R = 900 mm . some subproblems are examined [Tran et al.5 mm . σ y = 387 MPa . mean radius r = 229. E = 208000 MPa . L = 5100 mm (curvature factor λ = 0. In the analysis.3 . In order to evaluate the model.8 Limit and shakedown analysis of pipeelbow subjected to complex loads In this continuous example of pipe bends a system of two straight pipes connected to the ends of an elbow is considered. the two pipe ends are free.32) For this reason the available analytical solution in the literature can be applied to calculate the limit load Fl . 1995]. 2007c]. [Kussmaul et al. transversal forces F and bending moments M I . This pipeelbow can be subjected to internal pressure p .
27) eq. Goodall [1978b] was the first to propose the closedform equation of the limit load of pipeelbows under combined internal pressure and inplane bending moment by the small displacement analysis. In all the foregoing analytical equations. 372 large displacements 403. Kussmaul et al. It is now well known that the internal pressure has a significant effect on the limit load of a pipeelbow. the effect of internal pressure is not taken into account. 2000] investigated various sizes of pipeelbows subjected to 104 . (6. which may vary independently in the range [ 0.22 319. It was noted by Griffiths [1979] that the effect of connecting straight pipes can increase the limit load up to 33% with respect to Calladine’s solution.29 LISA 404.22) Findlay (6.26) eq. The proposed equation was α = 1.79 347.23) exp. This is due to two reasons: the stiffening effect of connecting straight pipes and not taking into account the nonlinear geometrical effect in the present method.04λ G I 2/3 pr . it can be seen that the internal pressure reduces the limit moment.24) eq.83 Figure 6.26 297.9 shows the comparison of the numerical result. p0 ] and [ − M 0 . [1995] and Calladine [1974].33) From (6. (6. Chattopadhyay [1999.21) (6. 1 − 2hσ y 1/ 3 (6. (6.6.33). Table 6. respectively. It is seen that the present result is compatible with the result of Goodall [1978] but higher than that of Diem [1994]. Based on nonlinear finite element analyses and twice the elastic slope method.9: Limit load Fl of the pipeelbow under transverse force (kN) Diem Goodall Touboul Drubay Calladine Desquines Spence & Present [1994] eq.45 small displacements 441. (6.23 FEmesh of a pipeelbow subjected to transversal forces Pipeelbow under inplane closing bending moment and internal pressure A pipeelbow is often subjected to combined internal pressure and bending moment during service. M 0 ] .67 387.
Figure 6. L = 1500 mm (curvature factor λ = 0. This is due to the omission of geometrical nonlinearity and strainhardening of the material by the present method and in Goodall’s analysis.6 and 0 ≤ α p ≤ 1.124 αp 2 − 0. E = 203000 MPa . Our solutions agree well with those of Goodall [1978b] but not with those of [Chattopadhyay. as presented by Fanous et al. It can be seen that the limit moment decreases when the pressure increases.combined internal pressure and inplane bending moment. In our numerical analysis one half of the pipeelbow is meshed by using 640 quadrangular 4node flat shell elements as shown in figure 6.047λ 1/ 3 + 0. ν = 0.540 ). He proposed another analytical solution for the pipeelbow which takes the strainhardening of the material into account by Closing collapse: α IC = 1.2 λ (6.34b) Applicablity: 0.122λ 2 / 3 + 0. R = 750 mm .24 FEmesh of pipeelbow under internal pressure and inplane bending Figure 6.0 where α p = pr / hσ y is the normalized pressure. the analysis of the pipeelbow without the large displacement effect gives a conservative and acceptable result [Fanous et al. 2005].3 .24.24 ≤ λ ≤ 0. [2005]. The following geometrical data and material properties are adopted: h = 40 and 45 mm .34a) (6.568α p 1. 105 .25 shows the effect of internal pressure on the limit moment of the pipeelbow. r = 250 mm .480 and λ = 0. 2000]. Closedend condition of the pipeelbow is simulated by applying an axial force of intensity pπ r 2 at the end of connecting straight pipe. σ y = 270 MPa .. Hence.175 Opening collapse: α IC = 1. The elastic data is only needed for the calculation of the elastic limit. σ u = 513 MPa .508α p λ αp 2 − 0.
27.26.7 0. The curvature factor of the pipe bend is 0. 2005].4 0.1 0 Goodallcollapse Chattopadhyaycollapse Presentcollapse Shakedown limit Elastic limit 0. the model is discretised by 1500 quadrangular flat 4node shell elements as shown in figure 6. In our analysis.8 Normalized pressure 0. An axial force is applied at the end of connecting straight pipe to account for the closedend condition.4 0.5 0.6 mm in order to reduce the effect of the pipe ends on the response of the elbow [Fanous et al.3 0. This small shakedown loads limits the benefit of the strengthening effect of the geometrical nonlinear analysis in [Chattopadhyay et al. and the thickness h = 10. The material model used is perfectly plastic stainless steel 304 which has an Young’s modulus E = 193743 MPa .6 Load factor 0.4 0.0. The allowance of cyclic loading up to the shakedown limit without any safety factor may only be acceptable in connection with a planned monitoring and maintenance program.2 0. p0 ] and [ − M 0 . which are obtained by RNode (redistribution of nodes).2 0. It is shown that these results compare well with those of Fanous et al. 2000]. [2005].6 mm .25 Effect of internal pressure on inplane limit moments Contrary to the case of limit analysis. The applied internal pressure effects the plastic collapse limit moment. Pipeelbow under outofplane bending moment and internal pressure This pipeelbow analyzed by Mourad and Younan [2002]. Fanous et al. M 0 ] . the mean radius r = 203. For small pressure the shakedown limit is between 57..3 0. a yield stress σ y = 272 MPa and Poisson’s ratio ν = 0. Both internal pressure and bending moment can vary independently in the range [ 0. the bend radius R = 609.2 0.7% and 66. The plastic collapse limit moment decreases when 106 .6 0.47 . Iterative RNode and plastic analysis methods.8 Normalized pressure 1 Figure 6.54 0. internal pressure has a very little effect on the shakedown limit moment.7 0.1615.48 curvature factor = 0. At low pressure also the difference between the elastic limit and the shakedown load is small so that fatigue and ratchetting could be of concern.6 0.2 mm . [2005] is considered.5 0.6 0. The length of two straight pipes is four times the diameter L = 1625. Our numerical results are presented in figure 6. respectively.2 1 Load factor Goodallcollapse Chattopadhyaycollapse Presentcollapse Shakedown limit Elastic limit 0.8 0.94 mm .4 0.8 0.1 0 curvature factor = 0. compared with three plastic collapse limit solutions of Fanous et al. [2005].7% of the plastic collapse load.
and Abdel KarimOhno) are not capable of simulating straight and elbow pipe ratchetting responses satisfactorily when the model parameters are determined from material level responses. The shakedown analysis of pipe bends in [Abdalla et al. OhnoWang. In order to show the convergence property of the algorithm. Figure 6. With respect to the design code the same comments apply as for figure 6. This is an important saving compared to the analysis of a large number of plastic cycles and makes limit and shakedown analysis a first candidate for structural optimization or reliability analysis [Tran et al. Figure 6.. convergence analysis and the influences of optimisation parameters such as the penalty parameter c and η 2 were included for the reference loads M = 225.25.97 kNm and p = 6.pressure increases noticeably above 4 MPa . Shakedown analysis converges in only three iterations which each needs approximately the computing time of a linear elastic analysis.28 demonstrates that the direct plasticity methods are extremely effective. Figure 6.296..26 FEmesh of pipeelbow under internal pressure and outofplane bending For both pipeelbows no PD check has been published which could be used to verify the calculated shakedown limits. 2007d].30 that the parameters η 2 and c have very little 107 .27 shows that large pressure (greater than 9 MPa ) reduces the shakedown limit moment considerably. The thesis [Rahman. [Staat and Heitzer. But smaller pressure has no effect in contrast to the elastic limit. The direct shakedown analysis gives robust predictions of shakedown limits on the basis of restricted information about the material. 2006] demonstrates that the existing cyclic plasticity models (Chaboche. 2002]. It is shown in figures 6.895 MPa . 2006] is a nonmandatory AP check.
27 Effect of internal pressure on outofplane limit moments 12 10 8 6 4 2 0 2 4 Limit analysis Shakedown analysis Load factor 6 8 10 Number of Iterations 12 14 16 Figure 6.influence on the both solutions in the ranges of η 2 ∈ 10−12 .10−28 and c ∈ 108 .1017 . Figure 6. The numerical results remain virtually constant over these intervals.28 Limit load convergence of the pipe bend 108 .
005 1 0.687 0.99 0. In designing such vessels it is usually necessary to have piping nozzle connections welded in the head.01 1.985 8 10 12 log(c) 14 16 18 0.5 1 0. To assess the plastic collapse limit and shakedown limit of such structures is of great interest to many designers and has been included in some design code rules.9 Nozzle in the knuckle region of a torispherical head Torispherical drumheads under internal pressure loading are known to have high stresses in the ‘knuckle region’ due to the change of the curvature of the shell.5 Limit analysis Shakedown analysis Load factor 10 15 20 log(Eta square) 25 Figure 6. the UK pressure vessel design code BS 5500.5 2 1. But they are still limited to some specific head shapes and loadings and not applicable for nozzles 109 .995 0. the German pressure vessel design code ADMerkblatt B3 and the European standard for unfired pressure vessels CEN.02 1.g.5 3 2.685 Shakedown load factor 8 10 12 14 log(c) 16 18 20 Figure 6. e.015 Limit load factor 1.688 0.30 Influence of penalty parameter c 6.69 0.689 0.686 0.3.29 Influence of η 2 1.
2000]. However. The loads may be constant or vary independently from zero to the maximum value. In the present example. Hsieh et al. a single nozzle on a Klöpperboden head has been selected for investigation. 2000]. A study of both individually and combined limit loads has also been undertaken. Recently.. a parametric study of the knuckleencroaching nozzles was conducted by Bauer and Saal [1997] by using finite element analysis. outofplane bending moment M o . in the above studies. that is by drawing a line at a slope 15 times that of the elastic portion of the relevant load versus deformation curve.particularly for nozzle encroaching in the knuckle region [Hsieh et al.. r = 200 mm 110 . [Tran et al. [2000] investigated a series of knuckleencroaching nozzles for both internal pressure and nozzle loads. Using the von Mises yield criterion and 15times the elastic slope method. The dimensions are selected from the study in [Hsieh et al. the cyclic loads were not considered. Both internal pressure p and nozzle loads including axial force F . 1997].31a.[Bauer and Saal. and twisting moment M t were considered.. F Mt MI ca di Mo ts r R p te Da Figure 6. The nozzle axis is parallel to the vessel axis as presented in the figure 6. 2007c]. te = 17 mm Head crown and knuckle radii R = 2000 mm. inplane bending moment M I .31 Geometrical dimensions of nozzle and FEmesh Vessel outer diameter and thickness Da = 2000 mm.
45 1. [2000] with 20noded volume elements.439 3.7 282.32b). In order to ensure that the boundary condition effects do not influence the nozzle junction behaviour. the 8noded brick finite elements.6 64. [2000] in this case that the limit load of using linear finite elements. In the case of 111 . is 29% higher than that of using 20noded elements.636 2. The size of the elements in the refinement region has been changed by a factor 12 from the locally coarse mesh to the fine one. respectively. Figure 6.37 [2000]. i.Nozzle inner diameter and thickness d i = 400 mm. In our numerical analysis the full model was meshed by using 3032 quadrangular 4noded flat shell elements as shown in figure 6. Only very coarse meshes cause convergence problems for linearly elastic analyses.4 284 128.8 625 193.6 mm Nozzle offset ca = 910 mm The elasticperfectly plastic material model was assumed with yield stress σ y = 340 MPa .32a).8 136. volume Plastic 3. Pressure type of Limit criterion [MPa] element Hsieh et al. the vessel cylinder length and nozzle length were chosen as Da and 3d i .63 0.31b.32 shows that even bad shaped shell elements do not cause convergence problems. Table 6. shell elements Plastic collapse limit Shakedown limit 1. Numerical results for individual loads are introduced in table 6. First local mesh refinements with nearly square finite elements have been made around the intersection of the nozzle with the vessel head (figure 6. It is seen that in all cases the results obtained with the present shell model compare very well with those gained by Hsieh et al. ts = 13.3 635 386. Together with the first yield loads and limit loads.1% .2 270. The maximum difference of the limit analyses is 2. Elastic limit 1.10: Elastic limit.e. Next the local refinement has been achieved by changing the shell elements from squares to bad shaped rectangular elements with aspect ratio 12 (figure 6.878 Axial force [MN] 0.783 Inplane moment [kNm] 64.10.4 193. shakedown limits are also presented.391 1.8 68. limit and shakedown loads of the nozzle Source.6 The elastic analysis of the load case of axial force shows the exceptionally large difference of 13.5 265.7% appearing in the case of internal pressure.54 elements collapse limit Elastic limit Present. It is noted by Hsieh et al. Therefore two mesh convergence analyses have been performed. For all load cases the end of the main vessel is fully clamped.669 0.8 Outofplane Torsion moment [kNm] [kNm] 66.
In the case of internal pressure and axial force (figure 6.35 and 6.32 Mesh convergence analyses for the case of axial force The interaction diagrams for load combinations between internal pressure and nozzle loads are presented in figures 6. In all cases the loads are normalized with the respective plastic collapse limit.33). The plastic collapse limit is virtually independent on the mesh. in almost all cases it is twice the elastic limit which indicates that inadaptation will occur due to alternating plasticity. as would not be expected. the behaviour is of course signdependent. For compressive nozzle loading.8 1. There is also a strengthening effect on plastic collapse limit in the case of positive bending moment (nozzle rotated inwards) as mentioned by Hsieh et al.4 2000 4000 Elastic limit Shakedown limit Plastic collapse limit 6000 8000 10000 Number of finite elements 12000 (a) (b) Figure 6. Obviously limit analysis is very robust.34 presents the interaction between internal pressure and inplane bending moment.36 show the interactions of pressure versus outofplane bending and torsion moments. 112 . there is a weakening effect on the shakedown limit in this case.36. Figure 6. A topologically different mesh has been used in Tran et al. Figures 6.6 1.6 0.2 1 0. In this case the shakedown check is not mandatory although this leads to a design which is prone to LCF. These figures illustrate clearly the need to consider shakedown as a design criterion for knuckleencroaching nozzles since the shakedown limits are considerably smaller than 57.4 Load limits 1. [2007c] which exhibited sensitivity. For both cases the behaviours of plastic collapse limits can be considered to be bilinear.7% of the plastic collapse limits in a wide range of parameters.8 0. internal pressure has a small influence on the shakedown limit. [2000]. On the other hand.336. It is observed that for this case while inplane moment plays an important role. Nozzle under tension 1. Contrarily. a strengthening effect on plastic collapse limit appears as presented in Hsieh et al.alternating plasticity these problems carry over to shakedown analyses.
34 Pressure versus inplane bending moment interaction 113 .33 Pressure versus axial force interaction Figure 6.Figure 6.
36 Pressure versus torsion moment interaction 114 .Figure 6.35 Pressure versus outofplane bending moment interaction Figure 6.
including both plate and shell structures.1) . then the reliability index may be given [Staat and Heitzer. The following two cases are examined. In the third example the effect of the thickness imperfection on the limit load of a shell is studied by considering the thickness as another stochastic variable. 2007b] is considered.6 and in [Tran et al. μ s and standard deviations σ r .σ s respectively.2 and is subjected to a uniaxial tension p1 which can be constant or can vary within the range [0.64σ r2 + σ s2 . (7. Four examples are presented below for the sake of evaluation of the algorithm.8μ r − μ s 0. pmax ] .. If both material and load random variables are supposed to be normally distributed with means μ r . In each numerical test. 2003] β HL = (1 − R / L ) μr − μ s 2 (1 − R / L ) σ r2 + σ s2 = 0.2 . It has the ratio R / L = 0. since the lower and upper bound are coincident. This structure was also studied analytically and numerically by Staat and Heitzer [2003b] using FORM and volume element.7 PROBABILISTIC LIMIT AND SHAKEDOWN ANALYSIS OF STRUCTURES In this chapter numerical applications of the probabilistic limit and shakedown programming are introduced.1 Square plate with a central circular hole In this example a square plate with central circular hole which was investigated in section 6. Both yield limit σ y and load p1 are supposed to be random variables. Limit load analysis For this case the exact plastic collapse limit is given by plim = (1 − R / L ) σ y in the range R / L ≤ 0. 7. The fourth example presents the estimation of the system failure probability of a frame formed of three plates subjected to vertical and horizontal loads. some existing analytical and numerical solutions found in literature are briefly represented and compared. The first two examples illustrate the FORM/SORM calculations for structures in which the loads and yield stress are considered as random variables.
969E01 9.835E01 7.848E01 9. s ⎞ = log ⎜ 2 + 1⎟ .880E13 2.408E01 9.973E01 9.049E02 2.981E01 9.4 1.734E01 5.996E01 9.8 0.848E01 9.996E01 9.) [Staat et al.7.872E06 7.419E01 9. 2003b] 2.900E01 9.872E06 7.s / 2 = μr .981E01 9.364E04 2.606E01 4.424E04 2.716E09 3.2) where mr . Figure 7.s ⎛ σ r2.643E13 3.720E09 3. δ s are calculated as follows mr . compared with the exact solutions and numerical solutions of Staat and Heitzer.972E01 9.112E06 1.3 0.421E04 2. s e −δ r2.) Pf (FORM) Pf (SORM) 0.3 1. Table 7.999E01 116 .s ⎛ σ r2.996E01 9.863E01 9.986E01 9.973E01 9.1 1.2 0.999E01 1.550E01 8. Staat and Heitzer also introduced the analytical expression β HL = log ( (1 − R / L ) mr ) − log ( ms ) δ r2 + δ s2 ( 7.1μ r .s ⎠ .2 1.140E02 1.000E01 7.067E01 5. Numerical results of failure probability Pf for both normal and lognormal distributions are introduced in tables 7..5 0.419E01 9.606E01 4. which are obtained by SORM.544E01 9.093E03 3.If both random variables are lognormally distributed.275E02 1.4 0.426E09 3. s ⎞ ⎜ 2 + 1⎟ ⎝ μr.050E02 1. The weak nonlinearity may be lost in the case of multiparameter loads or by an extreme value distribution of the random variables. s = 0. s μ s / μ r Pf (anal.7 0.1: Numerical results and comparison for normal distributions Limit analysis σ r . The numerical error involves both deterministic shakedown analysis and reliability analysis is very small and acceptable.999E01 1.878E13 2. ms and δ r .999E01 Pf (num.872E06 7. δ r.843E09 6.s ⎠ (7. ⎜μ ⎟ ⎝ r.995E01 9.6 0.1 presents the failure probabilities versus μ s / μ r for the case of lognormal distributions. It is shown that SORM solutions are almost the same of FORM due to weak nonlinearity of the limit state function.0 1. s = μ r .17. both kinds of random variables above were tested.835E01 7.3) In our numerical analysis.718E13 2.305E01 9.9 1.
397E01 8.436E04 1.2: Numerical results and comparison for normal distributions Limit analysis σ r = 0.851E01 8.554E01 9. σ s = 0.887E03 1.) Pf (FORM) Pf (SORM) 0.271E02 1.901E06 9.995E01 9.205E03 8.252E02 1.258E01 9.563E01 9.455E06 1.172E01 9.085E04 1.028E01 5.312E04 1.5 1.919E04 1.2 μ r .1 0.987E01 9.723E01 3.1 1.021E01 8.0 1.250E02 1.2 0.277E01 9..496E03 4.9 1. 2003b] 6.3 1.995E01 9.052E02 1.2 1.103E01 5.621E01 9.045E03 7.5 0.763E03 4.976E01 9.318E01 3.070E01 8.3: Numerical results and comparison for normal distributions Limit analysis σ r = 0.795E01 9.318E01 3.Table 7.3 0.892E01 6.821E01 9.388E01 9.723E01 3.6 7.955E01 9.028E01 5.717E03 1.850E01 9.6 0.887E03 1.514E04 1.831E01 9.562E01 9.8 0.794E01 9.042E01 8.802E01 9.993E01 9.610E01 7.163E02 1.) Pf (FORM) Pf (SORM) 0.983E01 9.3 1.1 0.765E01 9.267E02 5.675E02 1.) [Staat et al..046E03 7.2 0.850E01 9.892E01 Pf (num.999E01 5.775E01 9.366E01 5.209E01 2.276E01 9.4 1.866E06 1.948E03 1.823E01 9.313E06 9.319E01 7.314E01 7.248E02 6.247E02 6.042E01 8.1μ s μ s / μ r Pf (anal. 2003b] 7.312E04 1.866E06 1.621E01 9.5 0.999E01 Pf (num.019E01 8.366E01 5.6 0.902E06 9.314E01 7.499E01 9.9 1.256E01 6.995E01 9.730E01 9.550E01 9.873E01 6.2 μ s μ s / μ r Pf (anal.190E01 5.451E01 9.361E01 3.729E01 9.000E01 6.882E01 117 .843E01 8.4 0.000E01 7.999E01 Table 7.7 0.2 1.494E03 4.987E01 9.1 1.831E01 9.955E01 9.925E01 9.018E01 8.053E02 1.826E01 8.8 0.730E01 9. σ s = 0.943E01 9.742E01 3.065E03 7.427E01 5.772E01 9.451E01 9.7 0.590E02 1.4 0.686E01 9.998E01 5.587E01 3.0 1.835E01 5.437E04 1.5 1.850E01 8.4 1.6 6.3 0.493E01 9.271E06 1.085E06 1.649E01 9.) [Staat et al.240E01 6.646E03 3.437E02 6.374E04 1.239E01 6.2 μ r .
373E01 9.645E02 1.676E01 9.409E06 1.071E02 1.867E01 9.409E01 5.6 0.497E06 8.024E01 7.9 1.8 0.979E01 9.982E01 118 .115E01 2.2 1.997E01 9.090E01 2.000E01 7.981E01 9.205E04 1. 2003b] 9.1 1.235E04 1.954E01 9.472E01 9.4 0.090E10 6.947E01 9.981E01 5.790E12 4.485E02 2.) [Staat et al. σ s = 0.999E01 1.709E01 9.1μ r .866E01 9.2 1.3 0.738E01 5.009E03 3.5 3.976E01 9.979E01 2. σ s = 0.976E01 9.9 1.7 0.061E01 8.011E01 2.999E01 1.999E01 Table 7.000E02 1.951E01 9.814E02 1.107E03 2.575E01 9.715E01 9.947E01 9.704E12 5.998E01 9.839E01 9.) Pf (FORM) Pf (SORM) 0.205E04 1.024E01 7.997E01 9.1μ r .979E01 Pf (num.6 0..3 0.880E01 9.5: Numerical results and comparison for lognormal distributions Limit analysis σ r = 0.168E03 2.839E01 9.114E02 1.635E01 9.353E01 9.269E01 9.981E01 2.4 1.315E04 2.3 1.152E10 4.334E01 9.097E07 5. 2003b] 3.8 0.5 0.4: Numerical results and comparison for lognormal distributions Limit analysis σ r = 0.738E01 5.3 1..) [Staat et al.269E01 9.338E01 7.146E10 4.884E01 9.645E02 1.676E01 9.4 0.373E01 9.4 1.1 1.704E12 5.604E01 4.990E01 9.867E01 9.794E01 7.235E04 1.490E06 8.Table 7.7 0.586E06 1.) Pf (FORM) Pf (SORM) 0.866E01 9.997E01 9.061E01 8.2 0.935E01 9.337E01 8.604E01 4.097E07 5.248E01 8.472E01 9.648E01 9.719E01 5.431E01 9.593E12 1.794E01 7.999E01 Pf (num.566E10 6.011E01 2.396E01 8.2 μ s μ s / μ r Pf (anal.582E01 9.1μ s μ s / μ r Pf (anal.5 0.473E07 4.861E06 1.0 1.879E01 9.814E02 1.263E01 7.959E01 5.0 1.
908E01 6.884E01 9.1μ s μ s / μ r Pf (anal.629E01 7.7 0.971E01 1.407E01 3.6 0.786E07 4.979E01 1.3 0.885E14 4.870E01 9.4 0.988E01 Table 7.6 0.458E01 7.816E01 3.2 1.848E02 1.771E01 9.665E01 9.363E02 1.796E01 9.763E01 9.820E07 2.5 0.442E02 8.6: Numerical results and comparison for lognormal distributions Limit analysis σ r = 0.907E01 119 .7 0.828E01 9.2 0.744E01 8.509E01 9.574E06 7.3 1.562E01 9.451E04 6.8 0.674E01 9.873E01 5.146E01 9.518E01 5.046E01 8.584E01 9. σ s = 0.7: Numerical results and comparison for lognormal distributions Limit analysis σ r = 0.9 1.689E01 9.1 0.875E01 3.3 1.303E05 1.930E01 9.2 μ r .2 μ s μ s / μ r Pf (anal.072E06 1.601E03 9.736E01 6.807E03 2.715E07 2.121E01 7.120E01 4. 2003b] 1.5 0.) [Staat et al.828E07 2.3 0. 2003b] 8.060E03 1.308E04 6.831E01 9.8 0.786E01 8.497E01 3.4 1.5 5.541E01 7.4 1.655E14 3.918E01 9.678E10 7.762E01 9.067E04 1.610E01 9.124E01 8.0 1.337E13 6..509E01 9.000E01 6.638E02 2.500E03 4.4 0.297E01 9.213E01 8.786E01 8.062E03 1.121E01 4.225E01 9.2 0.327E10 3.665E02 1.959E01 9.899E01 6.497E01 3.790E01 8.1 1.870E01 9.) Pf (FORM) Pf (SORM) 0.) [Staat et al.860E01 7.871E01 8.864E02 1.124E01 8.969E01 9.689E01 9.2 μ r .261E01 9.414E01 9.722E01 9.868E01 Pf (num. σ s = 0.225E01 9.979E01 Pf (num.678E10 7.947E01 9.9 1.520E01 4.5 1.121E01 7.167E01 5.443E03 5.782E02 1.562E01 9.521E01 3.664E03 4.050E06 1.078E01 2.296E10 1.541E01 7.485E03 4.) Pf (FORM) Pf (SORM) 0.264E01 9.077E01 2.510E01 9.665E01 9.947E01 9.601E01 5.1 1.781E02 1.296E01 9.458E04 6.363E02 1.887E14 4.872E01 5.Table 7.868E01 3.2 1..0 1.
pmax ] and only the amplitudes but not the uncertain complete load history influences the solution.1 Probability of failure versus μ s / μ r for lognormally distributed variables Figure 7.Figure 7.2 Comparison of the results for probability of failure for normally distributed variables Shakedown load analysis For this case. 120 . the tension p varies within the range [ 0. Consider the case where the maximal magnitude pmax is a random variable and the minimal magnitude is held zero.
8: Numerical results and comparison for normal distributions Shakedown analysis σ r ..333E11 2. compared with the analytical solutions.523E01 9.021σ r2 + σ s2 β HL = for limit analysis (7.732E01 4.000E01 7.996E01 9.1).5 0.752E01 9.2 Pipejunction subjected to internal pressure In this example.486E03 9.371E06 2.1μ r .) 1. 2003b] 0.967E01 9.969E01 9. Thus. It is subjected to internal pressure p which can be constant or vary within the range [ 0.996E01 9.735E01 9.145μ r − μ s 0.3 and in [Tran et al.10267σ y .007E01 8.2 the failure probabilities (SORM approximations) are shown versus μ s / μ r for both two cases limit and shakedown analysis and for the case of normal distributions with σ r . If both material and load random variables are supposed to be normally distributed with means μ r .735E01 9. s μs / μr 0.60332σ y .464E01 9. then the reliability index may be given [Staat and Heitzer. In figure 7.4 0. s . which are calculated by (7. Both yield limit σ y and load p are considered as random variables.891E01 6.906E01 8.833E06 2. the loading conditions should be considered carefully when assessing the load carrying capacity of the structure.906E01 8.999E01 Pf (FORM) 9.109E10 3.0 1.996E01 9. s = 0.366E02 4. μ s and standard deviations σ r .6 0.9 1.1 Pf (anal.644E02 4.8.999E01 Pf (SORM) 9. The numerical probability of failure for normally distributed variables are presented in table 7.2 0.3 0.648E02 4. s = 0.145σ y and shakedown limit psh = 0.60332 0.σ s respectively.1μ r .65 0.From the deterministic analysis we got the shakedown load psh = 0.7 0. Table 7. It is worth to note that the shakedown probabilities of failure are considerably smaller than those of limit analysis.844E01 5.8 0.999E01 7.4a) 121 .732E01 4.891E01 6.967E01 9. 2007d] is considered. the pipejunction which was investigated in section 6.833E06 2.333E11 2. Numerical deterministic analyses lead to a collapse pressure plim = 0.168E03 8. pmax ] .168E03 8.464E01 9.
947E01 7.775E01 3.22 0.036E08 4.990E01 9.312E03 0.108E03 8.08 0.999E01 0.200E01 9.953E01 4.350E01 9.538E01 8.946E02 3.239E12 2. σ r .870E01 9.974E01 9.15 0.05 0.683E01 8.870E01 9.999E01 (SORM) (FORM) (SORM) 1.06 0.051E01 2.819E01 9.798E01 6.764E08 4.997E01 9.11 0.14 0.1μ r .978E01 0.10 0.074E01 9.975E02 8.) Pf 1.473E04 2.182E01 7.08 0.783E02 4. σ r .22 0.12 0.945E07 2.18 9.304E09 9.03 5.280E04 4.476E01 9.10267 5.1μ r .10 4.9: Numerical results and comparison for normal distributions.08 0.04 0.20 0.346E01 0. s Limit analysis Shakedown analysis μs / μr 0.663E01 5.698E01 9.978E01 9.10267 0.983E01 0.04 6.997E01 0.000E01 4.000E01 0.873E01 9.869E01 6.05 1.345E12 0.999E01 122 .302E09 4.18 0.16 0. s Limit analysis Shakedown analysis Pf μs / μr 0.698E01 6.14 9.504E08 8.06 1.423E04 1.07 0.869E01 9.18 0.937E01 9.Table 7.993E01 6.956E02 2.280E03 3.665E04 1.836E01 5.986E01 9.537E01 9.12 0.18 Pf (FORM) Pf (SORM) 7.663E01 5.782E01 6.078E02 3.934E01 5.537E01 0.336E05 5.992E06 1.689E01 5.352E12 2.346E01 9.349E12 4.469E12 3.456E12 4.977E01 6.580E06 2.865E01 9.978E01 9.474E05 0.812E01 3.767E01 1.212E05 3.827E01 9.581E02 2.180E01 4.11 6.000E01 5.15 0.391E03 3.) Pf Pf Table 7.020E01 5.986E01 9.26 Pf (FORM) Pf (SORM) μs / μr 0. s = 0.261E01 4.869E01 4.999E01 (FORM) 1.03 0.04 0.367E01 9.16 0.846E02 8.14 0.10 0.312E03 9.890E01 0.437E03 3. s = 0.145 0.999E01 9.758E06 1.448E05 3.04 0.12 8.401E03 4.20 0.508E01 8.152E03 1.567E01 8.09 1.994E01 6.890E01 7.24 0.556E01 8.326E01 9.06 0.145 0.120E01 9.564E01 9.09 0.648E01 1.818E04 3.807E03 3.204E02 4.433E04 9.976E08 0.999E01 μs / μr Pf (anal.07 4.999E01 9.945E01 9.798E01 5.005E01 2.06 0.12 0.842E01 9.236E01 9.449E01 9.761E02 3.16 9.866E01 9.869E01 0.10: Numerical results for Weibull distributions.804E06 5.637E01 8.441E09 9.982E04 4.514E06 3.24 0.225E05 5.14 0.10 0.677E01 9.873E01 9.238E08 9.242E07 2.987E01 9.897E03 1.955E02 2.157E02 0.16 0.997E01 9.26 Pf (anal.08 4.734E02 2.
especially for the case of limit analysis. The Weibull distribution leads to much larger failure probabilities in shakedown analysis but to much smaller ones in limit analysis (figure 7.3 Limit analysis of cylindrical pipe under complex loading Beside the loading and material strength. Figure 7.9. compared with corresponding analytical solutions. it is well known that the load carrying capacity of shell structures is generally influenced by their initial imperfections which occurs during the manufacturing and construction stages such as variability of thickness. In this case SORM would give improved results.4b) Numerical probabilities of failure for limit and shakedown analyses for normal distribution are presented in table 7.01054σ r2 + σ s2 for shakedown analysis (7.3).10 shows the numerical results of FORM/SORM for the case that both random variables have Weibull distribution. In this example. the effect of thickness imperfection on the limit loads of a shell structure is examined. Besides normal and lognormal distributions.1μ r . s = 0.10267 μ r − μ s 0. The cylindrical pipe subjected to complex loading which was investigated in 123 . No analytical results are known.3 Probabilities of failure for normally and Weibull distributed variables 7. The numerical error now comes only from reliability analysis. Both random variables have standard deviations σ r . The limit state function is nonlinear after the transformation into the u space.β HL = 0. Table 7. Otherwise SORM gives slightly better results compared with FORM. It is shown that numerical results of FORM and SORM are very close to the exact ones. the Weibull distribution is frequently used for modelling loads and resistance. s .
σ g = ⎜ y ⎟ σ t2 + σ s2 r ⎝ r ⎠ 2 σy (7. The following loading cases are examined Internal pressure For this case. Z and defined by g ( X . The following geometrical and physical parameters are adopted: L = 2700mm. The mean and standard deviation of the limit state function can be calculated as follows ⎛σ ⎞ μg = μt − μ s . then the limit state function g (U) in the standard Gaussian space is a linear function. (7. the reliability index becomes β HL = μg = σg (σ / r ) μ − μ (σ / r ) σ + σ y t s 2 y 2 t .1 is considered here. The limit state function is a linear function of basic variables X .section 6. Note that σ y is the yield stress and not a standard deviation here. μ s and standard deviations σ t . Bending moment lim The exact plastic collapse limit moment is linearised by M b = 4 r 2σ y h (see (6. The limit state function is defined by g( X . Z ) = 4 r 2σ y Z − X . σ s respectively.2). (7. Analogously with (7.5) If both thickness and load random variables are supposed to be normally distributed with means μt . For this purpose. Z ) = σy r Z−X. r = 300mm.3).8) 124 .σ y = 120MPa . The magnitude of the internal pressure is the second basic variable X . the resistance R depends linearly on the realization h of the thickness basic variable Z . the exact plastic collapse limit pressure is given by plim = σ y h / r .7b) where mt . ms and δ t .7a) The limit state function becomes nonlinear if both basic variables are lognormally distributed. Thus. δ s are calculated as in (7.6) from which. only the loading and the thickness h of the pipe are modelled as random variables.1)). we obtain the exact reliability index for FORM β HL = log (σ y / r ) mt − log ( ms ) ( ) δ + δ s2 2 t (7. 2 s (7.
(7. Z and defined by g( X .9b) Torsion moment In this case the exact plastic collapse limit moment is given by M tlim = 2π r 2σ y h / 3 .10) The reliability indexes for both cases of normally and lognormally distributed random variables are obtained. = 2 δ t + δ l2 (7.The reliability indexes for both cases of normally and lognormally distributed random variables are obtained. The limit state function is a linear function of basic variables X . Z ) = 2π rσ y Z − X . (7. (7.13b) 125 . respectively β HL = 2π rσ y μt − μ s ( 2π rσ ) σ 2 y 2 t +σ (7. Z and defined by g ( X . respectively β HL = 4r 2σ y μt − μ s ( 4r σ ) σ 2 2 y 2 t +σ (7. The limit state function is a linear function of the basic variables X .11b) Axial load The exact plastic collapse limit load is given by Flim = 2π rσ y h .12) The reliability indexes for both cases of normally and lognormally distributed random variables are obtained. respectively 2 π r 2σ y μt − μ s 3 ⎛ 2 ⎞ π r 2σ y ⎟ σ t2 + σ s2 ⎜ ⎝ 3 ⎠ 2 β HL = (7.9a) 2 s β HL = log ( 4r 2σ y mt ) − log ( ms ) δ t2 + δ s2 .13a) 2 s β HL = log ( 2π rσ y mt ) − log ( ms ) δ t2 + δ s2 .11a) β HL ⎛ 2 ⎞ log ⎜ π r 2σ y mt ⎟ − log ( ml ) ⎝ 3 ⎠ . 3 (7. Z ) = 2 π r 2σ y Z − X .
330E04 6.356E01 9.999E01 Pf (SORM) 3.081E01 8.528E01 9.752E01 7.895E01 4.885E01 9.20 0.980E01 9.45 0.880E01 9.) 3.580E01 4.552E01 9.40 0.) 3.275E02 1.893E01 4.075E14 2. s Lognormal distributions: σ t .432 0.50 0.15 0.934E01 5.039E05 4.75 0.45 0.463E01 9.849E02 2.15 0.316E04 2.994E01 9.65 0.997E01 9.999E01 Pf (FORM) 3.869E02 2.969E01 9.40 0.30 0.791E12 4.488E10 5.139E01 8.1μt .60 0.799E01 9.55 0. s μ s / μt m3 ×10−6 kN Pf (anal.994E01 9.967E01 9.949E01 6.859E01 9.1μt .30 0.274E07 3.980E01 9.859E03 6.324E07 1.60 0.447E01 9.079E01 8.880E01 9.999E01 Pf (SORM) 3.117E04 2.999E01 μ s / μt m3 ×10−6 kN Pf (anal.999E01 Pf (FORM) 1.284E04 5.999E01 Pf (FORM) 3.117E04 1.927E01 5.976E02 1.719E01 5.860E01 9.466E01 9.783E02 2.45 0.30 0.704E03 6.279E08 5. s = 0.432E01 9.740E01 7.999E01 Pf (SORM) 1.1μt .35 0.50 0.973E01 9.20 0.360E10 5.897E01 9.30 0.891E01 4.997E01 9.175E12 3.70 0.997E01 9.017E02 1.565E01 9.013E02 2.25 0.15 0.333E01 9.946E01 6.880E01 9. s = 0.35 0.484E07 1.999E01 0.25 0.999E01 Pf (FORM) 3.030E14 2.596E01 4.000E01 6.980E01 9.179E14 1.55 0.582E01 4.160E14 1.999E01 Pf (SORM) 3.975E01 9.135E01 8.Table 7.428E09 3.950E01 6.408E01 9. s μ s / μt ×10−5 1 kN Pf (anal.720E13 1.798E01 9.175E12 3.975E01 9.432 0.085E01 8.15 0.596E01 4.787E01 9.60 0.999E01 μ s / μt ×10−5 1 kN Pf (anal.45 0.863E01 9.994E01 9.10 0.041E03 7.65 0.999E01 0.734E01 5.935E09 3.332E01 9.65 0.292E04 5.524E04 1.40 0.522E04 1.844E02 1.20 0.65 0.) 4.840E01 9.945E01 6.528E01 9.071E02 1.974E01 9.364E04 2.55 0.70 Table 7.231E14 2.55 0.473E07 4. s Lognormal distributions: σ t .50 0.888E01 4.935E09 3.542E01 9.996E01 9.20 0.076E05 4.257E08 5.995E01 9.651E06 6.393E08 5.000E01 7.35 0.718E13 2.997E01 9.981E01 9.973E01 9.700E06 6. s = 0.35 0.25 0.994E01 9.60 0.483E01 9.11: Numerical results and comparison for pressure case Normal distributions: σ t .792E01 9.25 0.500E01 9.1μt .378E10 5.551E01 9.996E01 9.901E01 9.649E02 2.282E05 4.000E01 6.964E14 1.277E07 3.896E01 9.14 0.994E01 9.967E01 9.972E01 9.729E03 6.871E03 6.50 0.70 0.629E02 2.083E01 8.75 126 .741E01 7.886E03 6.356E01 9.872E06 7.) 1. s = 0.753E01 7.981E01 9.840E01 9.70 0.000E01 7.849E02 1.724E13 1.449E01 9.12: Numerical results and comparison for bending moment case Normal distributions: σ t .40 0.284E07 1.
s Lognormal distributions: σ t .984E01 9. s = 0.431E01 8.153E13 2.000E10 9. s = 0.884E01 4.854E01 4.584E04 2.585E01 5.427E05 1.15 5.998E01 9.737E01 9.522E01 9.585E09 5.35 2.444E02 1.309E05 1.998E01 9.866E01 4.811E13 3.496E03 1.000E01 7.989E01 9.914E01 5.000E01 7.758E01 9.522E01 9.1μt .55 9.2262 0.507E06 1.730E03 3.996E01 0.10 0.123E09 6.145E02 2.758E01 9.07 0.1μt .773E01 9.) 1.175 0.601E09 1.15 0.999E01 0.381E07 7.20 0. s Lognormal distributions: σ t .09 0.30 0.890E01 7.945E01 9.257E01 0.40 0.584E01 8.674E07 2.999E01 Pf (FORM) 1.910E01 9.998E01 9.60 0.845E01 0.990E01 9.009E01 4.525E10 1.845E01 5.803E13 0.890E01 7.169E01 9.909E11 4.821E02 2.928E01 5.999E01 Pf (SORM) 4.275 0.40 127 .127E13 3.545E01 9.35 0.50 0.009E01 4.383E14 5.Table 7.65 9.516E07 2.70 9.14: Numerical results and comparison for tension case Normal distributions: σ t .846E01 5.756E07 1.986E01 9.512E01 9.581E01 9.70 2.30 0.021E01 4.715E09 1.45 0.729E03 3.375 0.999E01 Table 7.983E09 1.20 9.512E01 9.35 0.895E01 9.15 0.985E01 9.20 0.427E05 1.421E02 1.989E01 9.40 5.431E01 8.737E01 9.919E01 9.999E01 0.354E01 9.39178 5.949E01 9.670E06 0. s = 0.913E11 5.126E01 9.980E01 9.25 0.1μt .30 2.869E01 7.272E01 9.585E09 0.609E01 9.832E12 7.132E01 5.142E03 3.257E01 9.832E12 5.14 1.497E13 1.60 9.999E01 9.584E04 2.999E01 0.325 0.017E03 2. s μ s / μt ×10−6 m kN Pf (anal.39178 0.13: Numerical results and comparison for torsion moment case Normal distributions: σ t .08 0.756E07 6.) 3.427E02 1.312E03 1.521E01 9.) Pf (FORM) Pf (SORM) 1.370E01 8.102E11 5.998E01 9.996E01 9.998E01 9.850E01 4.673E13 4.45 8.999E01 Pf (FORM) 2.670E06 1.021E01 0.846E01 5.127E13 3.889E02 2.120E01 2.125 0.821E02 2.558E01 9.313E03 1.796E03 3.718E13 1.25 0.06 0.889E02 0.000E01 4.987E01 9.017E03 0.10 0.984E01 9.895E01 0. s μ s / μt 1 ×10 kN −5 Pf (anal.65 0.999E01 μ s / μt ×10−6 m kN Pf (anal.752E13 1.105E08 1. s = 0.000E01 5.272E01 9.986E01 9.999E01 Pf (SORM) 1.40 0.414E14 6.991E09 1.) 2.545E01 9.10 0.904E01 9.980E01 0.429E01 8.599E01 9.998E01 9.910E01 9.15 0.522E01 0.999E10 9.25 7.982E01 9.05 0.50 9.30 0.55 0.997E01 9.429E01 0.530E07 2.1μt .35 0.250E04 6.945E01 9.20 0.25 0.126E01 9.924E02 2.790E09 1.999E01 Pf (SORM) μ s / μt 1 ×10 kN −5 Pf (anal.385E14 5.753E01 9.2262 0.031E12 3.882E01 7.999E01 Pf (FORM) 4.
It is shown that our results compare well with the analytical ones for all cases. both random variables are normally or lognormally distributed with standard deviations σ t .14 for both distributions and in figures 7. For each loading case. Numerical solutions of the limit analyses are compared with the analytical ones resulting from exact solutions.4 Comparison of the results for lognormally distributed variables Figure 7.5 Comparison of the results for lognormally distributed variables 128 .The probabilities of failure Pf are presented in tables 7.1μt . Figure 7.5 for lognormal distributed variables.47.117. s = 0. s .
those are sway mode. V .5 kN/m 2 . σ H =15 kNm μV = 40 kNm.7. There are three basic variables and their mean values and standard deviations are given by Horizontal load H Vertical load V Material strength σ y μ H = 50 kNm. a wellknown problem of having several failure modes is investigated. three failure modes caused by plastic hinge mechanisms are expected to occur.7). Consider the frame formed of three plates of figure 7.4 m . As in the original plane frame.6a which is generated by prolonging a plane frame in the third direction. σ σ =337. The plastic moment capacity M p at each section is then a random variable with mean value and standard deviation as follows μM = p σM p bh 2 μσ y = 134. σ V =12 kNm μσ = 3372. It is made of elasticperfectly plastic material and subjected to two uniform horizontal and vertical loads H . [1986]. frame mode and beam mode (figure 7.4 Folding shell subjected to horizontal and vertical loads In this example.6 Geometrical dimensions and FEmesh of the frame 129 .49 kNm 4 Hinge lines are thought to form at the end of elements (beam and columns) or at lines of load application.25 kN/m 2 y y The thickness of the shell is supposed to be constant h =0. Loads and limit plastic bending moment (material strength) are random variables which are assumed mutually independent and lognormally distributed. The loading and geometrical data were selected to match those of the plane frame included in the book of Madsen et al.9 kNm 4 bh 2 = σ σ y = 13. H V l=5 m b= 1m l=5 m l=5 m Figure 7.
461 = 1.V H V H Frame mode Sway mode H V Beam mode Figure 7.194] 4 T with β HL 4 = 3.083 was found firstly. If we 4 further place a bulge B4 ( u) T at u* 4 and continue the algorithm.421] with β HL 3 = 3.4 × 3.218 between the two design points is less than but close to the radius 4 r1 = 0.307 [0.4 .7 Three failure modes of the frame Numerical computation is carried out by using 300 quadrangular flat 4node shell elements as shown in figure 7.240 .776 0.925 0. Our numerical results of design points are presented in table 7. at this stage.667 0.59 0.47 0. u* and 5 3 3 5 u* − u* = 1.325 0.461 which corresponds to the Beam mode. the algorithm converges to the second design point u* = 3.24 [ 0.384 and r2 = 0.070 is found.461[ 0. Obviously it is also a spurious 5 design point since the distances u* − u* = 1.3 in order to find all the three design points as expected. Thus.083 = 1. Our search algorithm now 3 converges to u* = 3. The ‘barriers’ technique developed by Der Kiureghian and Dakessian is performed with γ = 0. This design point corresponds clearly to 2 T the Frame mode because both of loads have big contributions to the failure of the structure.6b.783 0.07 [ 0.083[ 0.24 between the two design points u* .4 × 3.307 . We continuously added a bulge B2 ( u) at u* and found the third design point 2 u* = 3. the point u* = 4. δ = 0. thus confirming that u* is a spurious design point.329] T with β HL1 = 3. * After adding a bulge B1 ( u) at u1 .296 of the bulges. The distance * u* − u1 = 1. u* are less than the radius 5 2 2 5 r3 = 0.024 0.4 × 3.028 between the two design points u* . The global design * point u1 = 3.24 = 1.15. This design point corresponds to the Sway mode since the effect of the horizontal load H is dominant.454] with β HL 5 = 4.944 0.457 0. we stop 130 .422] with β HL 2 = 3.233 of the bulge. 3 T Now we suppose to proceed further and place a bulge B3 ( u) at u* .
0269 131 . The corresponding approximations of failure mode correlations and joint failure mode probabilities are listed in table 7. Table 7. This sensitivity even increases very fast in the case of small probabilities of failure.194 0. It should be noted here that.325 0. not only the bending moment but also the compression force has an influence on the loadcarrying capacity of the structure.944 0.667 Mode global design point Sway mode local design point Frame mode local design point Beam mode spurious design point spurious design point  The linear approximation of the failure set is now constructed by the tangent hyperplanes at the three design points.17 and 7.0037 0. the SQP algorithm worked well in this case to seek all the optimal points. It is understandable since the stress state is now threedimensional. leading to a smaller failure probability of the system.422 0.024 0. the simple gradient search algorithm failed after adding the first bulge.0 0.307 4.783 0.16.589 2 0.0 0.421 0.0155 0.16: Failure mode correlations and joint failure mode probabilities (first order) Joint failure probabilities Failure mode correlations ρij Pij × 102 1 1 2 3 1.0 1 0.18. respectively.0037 2 0. The single and system failure probability and reliability indices are presented in table 7.925 0.0155 3 0.461 3. By comparing the differences between present results and those of Madsen et al. It oscillates around a foot of the bulge unstoppably.776 0.17 and in table 7.0301 0.589 0. Table 7.to search and assume that there are only three design points for this problem.18.899 1. in the tables 7.070 Step 1 2 3 4 5 a1 0.0301 0.899 3 0. compared with those of the original plane frame obtained by Madsen et al. It is shown that our two first failure probabilities of sway and frame modes are smaller than those of plane frame while the last one compares well with the solution of Madsen et al.883 1. It is shown that about 10% difference in the reliability indices β HL leads to about 330% difference in the failure probabilities Pf .15: Multiple design points and search steps a2 a3 β HL Nature 0. one can easily observe the sensitivity of the failure probability Pf as discussed previously.240 3.590 0.1023 0.457 0.470 0.454 3.329 0. Conversely.0598 0.083 3.883 0.
0166 0.240 2.155 0.336 FORM 0.0291 0.0598 0. 0.18: Reliability indices of the frame β HL Method Present * * u1 alone (sway) u* alone (frame) u* alone (beam) u1 .461 3.082 2. u* and u* 2 3 2 3 FORM SORM FORM SORM FORM SORM 0.467 Table 7.600 132 .199 0.710 FORM 2.083 3.786 3. 2.0269 0.589 3.Table 7.1026 0.140.880 3.9582.062 2.110 0. u* and u* 2 3 2 3 FORM SORM FORM SORM FORM SORM 3.17: Failure probability of the frame ( Pf × 102 ) Method Present * * u1 alone (sway) u* alone (frame) u* alone (beam) u1 .447 Madsen et al.725 3.322 0.1023 0.989 2.440 3.267 0.0283 Madsen et al.
Parametric studies carried out in . Some of them could be checked against twice the elastic limit because they appear to represent the alternating plasticity limit under the considered loading conditions. A numerically very effective method is achieved from the lesser computational cost by using shell elements compared with volume elements and by direct plasticity methods which achieve plastic solutions in the computing time of only several linear elastic steps. In the deterministic LISA procedure three failure modes of structures such as plastic collapse. The new European pressure vessel standard EN 134453 is based on perfectly plastic limit and shakedown analysis (LISA) [European standard. structural reliability analysis based on probabilistic LISA is also performed to establish a rational basis for the choice of safety factors. 200506] thus indicating the industrial need for LISA software for safe and reliable design of such structures. Additionally. The load limits are obtained here together with tests of mesh convergence and robustness. Line search is performed to improve the current kinematic solution. A special “smooth regularization method” was also used for overcoming the nondifferentiability of the objective function. The actual Newton directions are updated at each iteration by solving a purelyelasticlike system of linear equations which ensures automatically the kinematical condition of the displacements.8 SUMMARY Integrity assessment of pressure vessels and piping by means of direct plasticity methods has been a problem of great interest to many designers. To this purpose. the present thesis provides an effectively plastic analysis method for the integrity assessment of general shells. practical design codes often prescribe what kind of values to choose for the safety factor of the resistance and of the loads for a given problem since all resistance and loading variables are generally random. The direct method involves the use of the exact Ilyushin yield surface and the solution of the nonlinear constraint optimization problems is solved by Newton’s method in conjunction with a penalty method and the Lagrange dual method. It has been tested against experiments and several load limits which have been calculated in literature with different numerical methods using shell or volume elements. The proposed method appears to be capable of identifying reasonable estimates of the limit and shakedown load factors for a wide range of thin shell problems. especially in the design of industrial and nuclear power plants. Besides a deterministic LISA algorithm. low cycle fatigue and ratchetting are analyzed based upon an upper bound approach.
1998] has been performed to successively find the multiple design points of a reliability problem. then the use of FORM is sufficient and SORM is only necessary if the linearity is lost after the transformation into the u space. when they exist on the limit state surface.some examples show that different choices of optimization parameters have quite different effects on the behaviour of the algorithm. On the other hand. The development of a kinematic algorithm of probabilistic limit and shakedown analysis for general plate and shell structures using the First and Second Order Reliability Methods. The results are achieved just after several deterministic steps even if the starting points are chosen far from the actual design point. In this thesis. sensitivity analyses are obtained directly from a mathematical optimization with no extra computational cost. • 134 . If the limit state function is linear with the limit and shakedown analysis. the loading and strength of the material as well as the thickness of the shell are considered as random variables. which is based on the Sequential Quadratic Programming for finding the design point. This technique is based upon a ‘barrier’ method by constructing a bulge around previous found design points. even in the case of very small probabilities. Numerical examples were tested against literature with analytical methods and with a numerical method using volume elements. A nonlinear optimization was implemented. In the probabilistic LISA. Different kinds of distribution of basic variables are taken into consideration and performed with FORM/SORM for the calculation of the failure probability of the structure. The exact Ilyushin yield surface has not been used so far in LISA. Numerical example of a frame made of plates showed that it improves considerably the estimates of the system failure probability. The advantage of the method in structural reliability analysis is that the failure under cyclic loading is treated as a timeinvariant problem and it is applicable with incomplete data. To the author’s knowledge the above mentioned results contain the following new contributions which have never been published elsewhere: • The development of a kinematic algorithm of deterministic limit and shakedown analysis for general plate and shell structures using the exact Ilyushin yield surface. Secondorder bounds of the reliability of series system are then calculated based on the firstorder system reliability analysis. The procedure involves a deterministic limit and shakedown analysis for each probabilistic iteration. The present thesis is the first one dealing with reliabilitybased LISA for shells and using SORM. a technique which was developed by [Der Kiureghian and Dakessian. thus forcing the algorithm to seek a new design point. It is shown that the proposed method is capable of identifying good estimates of the failure probability. Practical experience showed that the existence of multimode failure (multiple design points) in component reliability analysis could give rise to large errors in FORM/SORM approximations of the failure probability. Nonlinear sensitivity analyses are also performed for computing the Jacobian and the Hessian of the limit state function.
the present results may be safe to be applied in some practical engineering since some practical materials often exhibit strain hardening. It should be noted that in the present work we accepted some simplified hypotheses. the geometrically nonlinear effects may become important to very thinwalled structures. The implementation of a special technique to successively find all the multiple design points on the limit state surface allows to calculate the secondorder bounds of the reliability for series systems. small deformation and displacement models. However. i. in which both lower and upper bounds are obtained at each iteration. • • • 135 . On this basis. Analytical solutions for a simple problem were derived in order to compare with the numerical solutions. On the other hand. In this meaning limit and shakedown analysis based on the kinematic hardening may improve the present design procedures. The introduction and formulation of the reliability problem for the structural systems (series systems).• • • The introduction and formulation of the probabilistic LISA in which the thickness is considered as a random variable. Perform the deterministic and probabilistic LISA by the application of the primaldual theory. Extend the present work for probabilistic LISA to the real shell structures in which the thickness of the shell is considered a random field. In a similar way temperature fields and temperature dependent material strength can be considered.e. The sensitivity analysis of the limit state function versus load. we would propose here some possibilities for future work: • Extend the present work for both deterministic and probabilistic LISA to the real materials in which we take into account the bounded kinematic hardening of the material. Extend the present work for both deterministic and probabilistic LISA to the real behaviour of thin shell structures in which we consider the large deformation and large displacement hypotheses. The perfectly plastic material is sufficient with respect to the design codes. material strength and shell thickness variables. the random thickness is a function in space. such as perfectly plastic material.
Probabilistic Method in Structural Engineering. John Wiley & Sons. pp. and Christiansen E... Overton M. Secondorder cone and semidefinite representations of material failure criteria. Italy. D. pp. 22. 13.. 10461062. 19. International Center for Mechanical Sciences. Andersen K. M. R. (2007). (1971).org/10. 197204. 138144. Bolt S. London. D. D. and Papaioannou G. Udine. Plastic systems reliability by LP and FORM.. ASME J. Andersen K. Inc. In Baniotopoulos C. Bleich H. Odense University. and Greenstreet W. pp. Andersen K. Juli 78. 134 (2). (1998b). New York. Über die Bemessung statisch unbestimmter Stahlwerke unter der Berücksichtigung des elastischplastischen Verhaltens des Baustoffes. Comput. L. Appl. Comput. Determination of Shakedown Limit Load for a 90Degree Pipe Bend Using a Simplified Technique. Wien.. Saal H. 1990. Casciati F. Bauer H. (1993). Comput. Thessaloniki. Brno. pp. Optim. pp.. D. Conference on Nonsmooth Nonconvex Mechanics with Applications in Engineering (II. An efficient primaldual interiorpoint method for minimizing a sum of Euclidean norms. Pressure Vessel Technol. (2006). In Course on Reliability Problems: General Principles and Applications of Solids and Structures. 128 (4). Bazaraa M. 31. Denmark. 618624. and Overton. (1989). Augusti G. Bisbos C. Shakedown analysis of FEMdiscretized steel shell structures under the Ilyushin yield criterion. Sci. Carrying capacity of steel shell structures. Andersen K. (1998a). and Younan M. 11. Shetty C. pp. A Newton barrier method for minimizing a sum of Euclidean norms subject to linear equality constraints. CISM lecture notes. Computing limit loads by minimizing a sum of norms. Methods for structural reliability computations. Computers and Structures. (1932). E. L. C. Conf. Conn A. Journal of Optimization Theory and Applications. Bjerager P. 187196. Megahed M. D.. (1997). Pardalos P. pp. (2000). http://dx.. SIAM J.. Int.1007/s1095700792438 Bjerager P. Minimizing a sum of norms subject to linear equality constraints.doi. (1991).. SIAM J. Department of Mathematics and Computer Science at SDU. D. Bisbos C. . (2006). S. M.. M. 243262.. L. IMADA preprint PP199507. M. F. Y. (1995). Int. Experimental determination of plastic collapse loads for pipe elbows. D. and Christiansen E. pp. A. pp. Springer.. 275301.REFERENCES Abdalla H. 261267. Nonlinear programming Theory and Algorithms. Bauingenieur. Chapman and Hall.. Nozzles in the knuckle region of pressure vessels heads – strength and stiffness under various loading conditions. Baratta A. Sci. Sherali H. Christiansen E. 6579. (1984). In Proc.. Christiansen E. (ed.): II. ASME paper 71PVP37. NNMAE).
A product of EDF R&D. pp. Eng. Exact Ilyushin yield surface. (1980). 226238. pp. A. Desquines J. Chattopadhyay J. 163242. Crisfield M. 281288. Cen Z. ClosedForm collapse moment equations of Elbows under combined internal pressure and inplane bending moment. 2934. (1988). Elasticplastic behaviour of thin tubes subjected to internal pressure and intermittent highheat fluxes with applications to fastnuclearreactor fuel elements. 171200. (1952). Arts Sci.3 (2003). Chattopadhyay J. Inplane limit moment for an elbow lower bound analytical solution and finite element processing by elastic compensation method. pp. (1973). J. pp. 2. and Leckie F. in Eng. 211243. 11131131. and Wielgosz C. Sol. ST1. J. Report 51UC. Vol. J. 58. J. 30 (8). G.. Int. 509516. 110119. Proc. Bree J (1967).. Z. Structural Safety. Dutta B. Bridgman P.codeaster. (2000). SMiRT. & Piping. Carvelli V. (1993b). Nathani D. G. 22. 8587. (1997). (1997). Transport and Road Research Lab. France. pp.. (1967). Shakedown analysis of defective pressure vessels by a kinematic approach. and Brennan M. 431436. F. (1923).S. pp. pp. 20. C. Burgoyne C.. ASCE. K. 751764. Transactions of the ASME. and Kleiber M. Burgoyne C.. Clamart. Chapelle D. (1982). Meth. J. pp.. 30 (8).. and Corradi L. A. (2003). Bounds on the Reliability of structural systems. (1998). Eng. User’s manual. 16. Num. 18. Struct. pp. Journal of Applied Mechanics. 69. Journal of Strain Analysis. Num. (1993a). McGrawHill. Advances in Applied Mechanics. 137 . Casciaro R. 2. Calladine C. The finite element analysis of shells – fundamentals... The compressibility of thirty metals as a function of pressure and temperature.. Korea. Liu Y. Int. V. J.. and Cascini L. Calculation of elastoplastic rigidities using the exact Ilyushin yield surface. Int. pp. Struct. Studies in Large Plastic Flow and Fracture. W. New York. On a numerical approach to shakedown analysis of structures. http://www. pp. Bathe K. 122. (1999). Int. Venkatramana W. Dutta B. Acad... Multiple design points in first and secondorder reliability. Code_Aster 7. 93122. A. J. K. 71.org Cornell C.Borkowski A. Maier G. Science. Mech. Pres. pp. Limit analysis of curved tubes. Bridgman P. R. 55. pp. A finite element formulation of the rigidplastic limit analysis problem. Berlin: Springer. J. 3749.. (1999). (1974). and Brennan M. Journal of the Structural Division. Plancq D.. Sol. (1951). pp. Cocks A. 11331145. J. Ves. W. Proceedings of the 15th International Conference on Structural Mechanics in Reactor Technology.. and Kushwaha H. Meth. Computer methods in applied mechanics and engineering. Large deflection elastoplastic buckling analysis of plates using finite elements. and Kushwaha H. and Reissner E. Dakessian T. A. Am. Der Kiureghian A. 20632086. 40. Int. ASME J. A mixed formulation and mixed finite elements for limit analysis. S. pp. Feb. pp. Pressure Vessel Technol. Deformation bounds for cyclically loaded shell structures operating under creep condition. 93. Capsoni A. Limit Load of Elbows Under Combined Internal Pressure and Bending Moment. Clark R. Bending of curved tubes.. Archive of Applied Mechanics. K.
Rackwitz R. A16: Guide for Defect Assessment and LeakBeforeBreak Analysis. ASME J. Goodall I. Biometrica. University of London. A general approximate technique for the finite element shakedown and limit analysis of axisymmetrical shells. Ditlevsen O. Annex 2: Annex B Direct route for design by analysis and Annex C Stress categorisation route for design by analysis. FORM Manual. 42. pp. Large deformations in plastically deformed curved tubes subjected to inplane bending. CEN European Committee for Standardization. A theoretical investigation of the yield point loading of a square plate with a central circular hole. et al.cimne. Research Division Report CEGBRD/B/N4312. Central electricity generating board. Dissertation. AdibiAsl R. 156169.. Q. http://www.. and Seshadri R. (1994). (1986). and Bjerager P. Meth. 127. 138 . Gaydon F. Ponter A. Gokhfeld D. S. Franco J. Ditlevsen O.Part 3: Design. Drubay B. for Num. (1954). pp. Int. Approximations to the probability integral and certain percentage points of multivariate analogue of Student’s tdistribution. (1983).. A. (1979). R. F.. R. 40 (19) pp. Structural Safety. Germany. Commissariat A L’energie Atomique. (1979). W. Neumann H. Franco J. pp. and Cherniavsky O.und Versagensverhalten von Rohrbogen. Rackwitz R. (2005). 34953513. A. Academic Press. Rapport DMT 96. 443448. Sequentielle Quadratische Programmierung für effektive Zuverlässigkeitsanalyse mit der Finite Element Methode. in Engng. pp. F.a). 2. Meth.2 (2002).es Gollwitzer S. Journal of Structural Mechanics. in Engng... UK. 7 (4). ASCE.b). Untersuchungen zum Geometrieeinfluß auf das Verformungs. Pressure Vessel Technol. S. Sobel M. France.. (1995).. (1988). R. 258260. (1955). Diplomarbeit. (1980). M..Diem H. Fachhochschule Aachen Campus Jülich. Fanous I. W. Abdo T. (1975).. Methods of structural system reliability. 661676. Narrow reliability bounds for structural systems.gid. Quadratic Limit States in Structural Reliability Theory. pp. Third Draft. Fiacco A. Journal of the Mechanics and Physics of Solids. pp. Unfired pressure vessels . München. New York. 35153536. J. for Num. Part 1: Theory and fundamental relations. Ponter A.. EN 134453:2002. Issue 14. (1978a). J. Sijthoff & Noordhoff. A general approximate technique for the finite element shakedown and limit analysis of axisymmetrical shells. PhD Thesis. R. 453472. Part 2: Numerical applications. 105. Dunny C. Int. A. A product of Center for Numerical Methods in Engineering (CIMNE). European Standard (200506). The Netherlands. J. 40 (19) pp.. Journal of Engineering Mechanics. User’s manual. Frieze P. Barcelona. Ultimate load behaviour of steel box girders and their components.096. McCrum A. (1997. Q. Limit analysis of structures at thermal cycling. Fiessler B. (1997. Z. Introduction to Sensitivity and Stability Analysis in Nonlinear Programming. RCP GmbH. 195229. Universität Stuttgart. GiD 7. Limit load analysis of pipe bends using the Rnode method.upc. 3.. (2005). Labor für Biomechanik. Dronia S.
336442. The Mises yield condition for rotationally symmetric shells. 93106. Pres. pp. Limit analysis of rotationally symmetric plates and shells. (1987). Vol. (1974). Dordrecht. 21. Staat M. (1979). et al. W.. Thesis. 4. Kruse W. Univ. 13501354.. The Rigidplastic analysis of symmetrically loaded cylindrical shells. 267284. No. C. Plastic analysis of structures. (1959). (1999). Hohenbichler M. 1. J. Griffiths J. (1978b). New York. pp. (1999).Ing. New light on first.. Dr. Jr. pp. (2000). (1961).). 193 (3). In: Stochastic Optimization Techniques. Structural Safety. Structural Safety. pp. 513. Rackwitz R. Hodge P. Jr. Hilsenkopf P. Int. Automatic generation of collapse mode equations. Limit and Shakedown Analysis with Uncertain Data. McGrawHill. Research Division Report RD/B/N4360. (1982). pp. 349358.und Einspielanalyse zur Bewertung der Sicherheit passiver Komponenten. pp. Structural Safety. G. Lind N. Nuclear Engineering and Design. Hohenbichler M. Berichte des Forschungszentrums Jülich.. and Staat M. In: Inelastic Analysis of Structures under Variable Loads: Theory and Engineering Applications (edited by Weichert D. G. R. (1983). Jr. Hohenbichler M. (2002). Hodge P. Jr. 21. ST7. RWTH Aachen. Sci. XVIII. Rackwitz R.. Vol. Tech. Int..org/10. Maier G. Prentice Hall. EM1. Englewood. 253267. pp. pp. 111121. No. Vessels and Piping. 4. App. Heitzer M. 33 (2). FEMcomputation of load carrying capacity of highly loaded passive components by direct methods. No. Grimmelt M. Springer Heidelberg. Firstorder concepts in system reliability. 269288. Numerical Methods and Technical Applications (edited by Marti K. Hodge P.. Academic Press.doi. J. http://dx. pp. Heitzer M.. Lower Bound Limit Analysis of Curved Tubes Loaded by Combined Internal Pressure and Inplane Bending Moment. Gorman M. 305311.. Benchmark Study on methods to determine collapse failure probabilities of redundant structures. Staat M. Traglast. Heitzer M. Journal of Engineering Mechanics Division. The effect of cracks on the limit load of pipe bends under inplane bending experimental study.1016/S00295493(99)001909 Heitzer M. J. 1. 100.. G. Hasofer A. UK. Hodge P. Approximate evaluation of the multinormal distribution function. 111128. Quarterly of Applied Mathematics. München. Heft 58. Vol.) Kluwer. I. M. ASCE. Reliability Analysis of ElastoPlastic Structures under Variable Loads. (1988). An exact and Invariant First Order Reliability Format. G. Berichte zur Zuverlässigkeitstheorie der Bauwerke. Central electricity generating board. pp. Schuëller G. New Jersey. Gollwitzer S. 107. (1954). Vol.Goodall I. 177188. Experimental Study of Behavior and Functional Capability of Ferritic Steel Elbows and Austenitic Stainless Steel ThinWalled Elbows. Lecture Notes in Economics and Mathematical Systems. J.. J. Struct. July. Div. ASCE. (1963). pp. (1981). (1981). 119130. E.and secondorder reliability methods.. Jül3704. 139 . Mech.
Pol. M. Foundations of Plasticity. Thesis Forschungsberichte aus dem Fachbereich Bauwesen 9. and Kobes H. A. and Hill R. In Russian. 2: Applications of Statistics and Probability. A. Sci. application to structural reliability problems. 4957.. (1967). pp. V.F.Hsieh M. Methods of Structural Safety...O. 4.. Journal of Pressure Vessel Technology. Theory and Engineering Applications. Approximating the final relationship between the forces and moments of shell under the Mises plasticity condition. Kleiber M. Pipe Bend Behaviour at Load Levels Beyond Design. 74. Int. Maier G. Englewood Cliffs. 7475. Int.D. Integrity assessment of defective pressurized pipelines by direct simplified methods. (1995).. Moscow. An approximate Model for an elasticplastic pipe element under combined loading. pp. New York. 725744.. F. Krenk S. Mech. J. Kalisky S. 18. In: Proceedings of the ICASP7 conference. 2228. Stos. Gostekhizdat. Int. Koiter W. (1985). Université de Liège. pp. Balkema Rotterdam. 10951110. Vol. (1995). Pres. (1966). Arch. 4. Transaction of the ASME. Warsaw. 2. Madsen H. Mistry J. pp. (1992). Stokey W.. (2000). In Russian. A. Elsevier.. A. (2005). König J. H. Ves. Nozzles in the knuckle region of a torispherical head: limit load interaction under combined pressure and piping loads.E. Amsterdam.. General theorems for elastic plastic solids.C. Revised edition. Acad. Plasticity Theory. (1972). Acad. Sci. and Franzen W. NordHolland. Automatization of limit analysis calculations. On shakedown criterion. Pol. Sci. H. Liu Y.. Theory of shakedown of elasticplastic structures. Broookfield. Vol. (1986). & Piping. Ivanov G. Klingmüller O. Sci. G. Vol. Little G. Rapid analysis of plate collapse by live energy minimization. T.. (1987). König J. 5660. Proceedings of 13th International Conference on Structural Mechanics in Reactor Technology. Symp. Thèse de doctorat. Lubliner J. Maier G. 250260. Vessels and Piping. Carvelli V. On a new method of shakedown analysis. pp.).. University of California at Berkeley. (1977). Kussmaul K. K. 77.. 187198.. J. A shakedown theorem for temperature dependent elastic moduli. Amsterdam. Plasticity. A. König J. Uhlmann D. (1969). 19. (1975). pp. Etats limites des tuyauteries par la méthode des éléments finis et la programmation mathématique. pp. Lind N. pp. 227238. A. In: Progress in Solid Mechanics (edited by Sneddon I. G. A. Mébarki A. Techn. Bull. König J. Paris (edited by Lemaire M.. 161165.). 165221. (1948). PrenticeHall. 17. Larson L. Favre J. J. pp. Sér. Int. Proc. Sci. Meccanica. Shakedown of elasticplastic structures. pp. (1979). (1960).. (1997). Universität Essen Gesamthochschule. pp. 807815. Sér. Moffat D. Locci J. 6.. Anwendung der Traglastberechnung für die Beurteilung der Sicherheit von Konstruktionen. Plasticity. pp. N. König J. 140 . (1978). Inzhenernyi Zhurnal Mekhanica Tverdogo Tela. pp. Techn. Elsevier. (1969). L. Ilyushin A. Brazil. 165171. J. Bull. SMiRT... A. Pres. Diem H. Jospin R. Vol. 97. Shakedown theory in perfect elastoplasticity with associated flowlaws. Mech.
MIT Press. A/Solids. http://dx. 855878. (1991). Structural Reliability – Analysis and Prediction. J. Struct. ASME. Shakedown state simulation techniques based on linear elastic solution. pp. Borino G.. Analysis of cracked plates and shells using “metis” finite element model. Ponter A. Theoretical prediction of the strength of welded steel plates in compression. Theorie statisch unbestimmter Systeme aus ideal plastischem Baustoff. Fuschi P. 2025. R. B. 417433. Ponter A.. 140. Morelle P. pp. European Journal of Mechanics . Transactions of the 14th International Conference on Structural Mechanics in Reactor Technology (SmiRT 14). Methods Appl. Belgique. Amsterdam... (1973). Boston. Caddemi S.. D. pp. A matrix structural theory of picewiselinear plasticity with interacting yield planes. Thèse de Doctorat. Morelle P. Martin J. Cambridge. Panagiotopoulos P. Carter K. Limit load analysis of pipe bends under outofplane moment loading and internal pressure. Wiss. 19 (3). ASME J..).. Nguyen Dang H. ASME. A. 32–37. 183205. pp. Université de Liège. Melchers R. pp. Polizzotto C. (1936). In: Foundations of plasticity (edited by Sawczuk A. pp.b). (1997. (2000). (1975). 124. (edited by Smith.. 60. J. Mech. D. (1984). Convex and Nonconvex Energy Functions. pp.A/Solids. Carter K. 10. Ponter A. (1993. 7. NorthHolland. Eng. N. 195218.b). Comput. Moxham K.. Polizzotto C. A finite element formulation for shakedown problems using a yield criterion of the mean. Engrg. 423445. L. König J. pp. Mech. Eur. R. Appl. Mech. Akad. Birkhäuser. Lyon.. 5166. pp. (1970).a). Mech. Comp. 6 (1). (1997. On the condition to prevent plastic shakedown of structures: Part II – The plastic shakedown limit load. Finite Element in Analysis and Design. F. Mourad H. 621639. (2002).org/10. August 1722. 8.. M. (1985). S.. Structural shakedown analysis by dual finiteelement formulations. (1971). Nguyen Dang H. Shakedown problems for material with internal variables. Analyse duale de l'adaptation plastique des structures par la méthode des éléments finis et la programmation mathématique. On the condition to prevent plastic shakedown of structures: Part I – Theory. Inequality problems in Mechanics and Applications. pp. Shakedown limits for a general yield condition: implementation and application for a Von Mises yield condition. Fuschi P. Meth. Vol. Melan E. 60. 141 . R. E. (1987). Pressure Vessel Technol. Chichester. pp. (2004). J. (1990). A shakedown matrix theory allowing for work hardening and secondorder geometric effects. Morelle P. Ellis Horwood limited. 7079.. A.a). The calculation of limit load and shakedown boundaries by effective modulus method. (1989). Eng. (1976).. Mech. Tran T. In: Mathematical programming methods in structural plasticity. Sitzber. 259279. (6).doi. Cambridge University Report No CUED/CStruct/TR2. Plasticity. pp. Wien IIa 145.Maier G. S. F. 179192. S. Engelhardt M.1016/S0168874X(03)001173. Appl. 1519. France. Y. Maier G.). 40 (8). Plastic shakedown analysis. E. Meccanica. and Younan M. Appl. Polizzotto C. (1993. Nguyen Dang H.
(2000). 13. Birkhäuser. On incremental collapse of shells under cyclic loading. Ramachandran K. R. Systems Bounds: A critical study. Prager W.). Zuverlässigkeit und Lasten im konstruktiven Ingenieurbau.. Schittkowski K. Tech. AddisonWesley. Dissertation. (1951). pp. Nr. 345354. (2000). Studies on reliability of structural systems. Basel 1955. Solids. (19932004). (1978). Düsseldorf. Mech. 291301. ASME J. Ratcheting and shaekdown analysis of pressure equipment using elastoplastic FiniteElementAnalysis. Lecture Notes. 123128. In: Optimization Methods in Structural Design (edited by Eschenauer H. University of Bayreuth.Powell M. Mannheim. 142 . Vol. Phys. Schittkowski K. BI Wissenschaftsverlag. Olho N. HiriartUrruty J. Pressure Vessel Technol. A.M. pp. Department of Civil Engineering.Jr. implementation and test of a nonlinear programming algorithm. (1983a). pp. Lemarechal C. Wiley. Birkhäuser.). Rashedi M. Department of Computer Science. Case Western Reserve University. Civil Engineering Systems.. (1998a). D. (2006). (1983b). Theory of perfectly plastic solids. 14. (1984). Heft 29. In: Theory of the Thin Shells (edited by Niordson F. Report No.). Technische Universität München. Rackwitz R. New York. 197216. (1988). Schittkowski K. Optimization. SpringerVerlag. Moses F. Raleigh. Int. English edition: Introduction to plasticity. Noordhoff.G. (1959). International Series of Numerical Mathematics. Ohio. 3542.. pp. Boston. PhD Thesis. Berichte zur Rackwitz R. pp. Probleme der Plastizitätstheorie. Solving nonlinear least squares problems by a general purpose SQPmethod. 84. Rahman S.H.. Preiss R. Seitzberger M. (1983). Mech. A comparison of yield surfaces for thin shells. Univ. Y. In: Foundation of plasticity 2 (edited by Sawczuk A. R833. Report. Zowe J. (1969b). 328340. Robinson M. I. (1971). In: Trends in Mathematical Optimization (edited by Hoffmann K.. Mathematical Programming. Berlin. pp. (1969a). Theory. Algorithms for Nonlinear Constraints that Use Lagrangian Functions. Close bounds for the reliability of structural systems. Contributions to an Efficient Numerical Analysis of the Plastic Behaviour of ThinWalled Structures.). Sci. J. 17. 14. (2004). Prager W.. On the convergence of a sequential quadratic programming method with an augmented Lagrangian search direction. München. pp. Finite element analysis and related numerical schemes for ratchetting simulation. North Carolina State University. 247. 224248. 1. Hodge P. J. Schittkowski K. (1972). Zuverlässigkeitstheorie der Bauwerke. A. TU Wien. Groningen. (1978). 120. Sawczuk A.. NLPQLP20: A Fortran implementation of a sequential quadratic programming algorithm with distributed and nonmonotone line search . FortschrittBerichte VDI. Shalaby M.B. Cleveland. Limit loads for pipe elbows with internal pressure under inplane closing bending moment. Reihe 18. Limit analysis: The development of a concept.user’s guide. and Younan M. Evaluation of upper bounds to shakedown loads for shells. J. 295309. VDIVerlag. Sawczuk A.. pp. March. Prager W.
September 0507.. NIC Series Vol... 441. and Heitzer M. and Owen D. (1999).. and Baker M. A. Berlin. Heitzer M.. pp. Structural reliability theory and its applications.bibliothek. Design Criteria for Piping Components Against Plastic Collapse: Application to Pipe Bend Experiments.de/opus/volltexte/2005/79/ Staat M. Barcelona. Tran T. (2007a). 82 (1). N. and Heitzer M. (1973). Proceedings of 6th International Conference of Pressure Vessel Technology.org/10. (1998b). Report EUR 19020 EN.doi. (2001). and Nichols. Knabel J. Conf. 163170. DC USA. Direct Finite Element Route for DesignbyAnalysis of Pressure Components. In [Staat and Heitzer (2003a)]. August 1217. September 2426. In N. Petten. Spence J. (1982). pp. pp.fhaachen. A. Tran T. Calculation of load carrying capacity of shell structures with elastoplastic material by direct methods. Proc. Limit loads for pipe bends under inplane bending. July 712. http://opus. Transactions of the 14th International Conference on Structural Mechanics in Reactor Technology (SMiRT 14).. Hanoi. pp. For error corrections and remarks see: http://info. Beijing. The restricted influence of kinematic hardening on shakedown loads.ijpvp. and Younan M. Transactions SMiRT16 Structural Mechanics in Reactor Technology.. Kreißig R. (1997). Numerical Methods for Limit and Shakedown analysis – Deterministic and Probabilistic Problems. The Netherlands.2002. 15. http://www. Limit Loads for Pipe Elbows Subjected to Inplane Opening Moment and Internal Pressure. Oñate E. European Commission. J. 1. Staat M. Austria. (ed.fzjuelich. Some Achievements of the European Project LISA for FEM Based Limit and Shakedown Analysis. Bielawski G. 5th World Congress on Computational Mechanics. Paper PVP20021300. (2002). July 2630.org/10. ASME PVP Vol. Cengdian. eds. Int. Badie (ed. SpringerVerlag. http://dx. Kreißig R. PVP Vol. Finite element shakedown and limit reliability analysis of thin shells. Staat M. Pressure Vessels Piping. Touboul F.. Jülich.Shalaby M. eds. (2002). Y. The designbyanalysis manual.1016/j. 393399. August 1722. Nuclear Engineering and Design. J.177185. 1998. California.doi. 6167. Staat M. 151166. Proceedings of WCCM V.fhaachen.pdf Staat M. Vienna. Liu. 206 (23). (1989).bibliothek.bibliothek.de/opus/volltexte/2007/227/ . 2nd Int. Joint Research Centre. Heitzer M. E.) Computational Mechanics: Developments and Applications .) (2003a).. http://opus.. Lang H. 368. September 1115. Paper Presented at the 1998 ASME/JSME Joint Pressure Vessels and Piping Conference held at San Diego. pp. et al..de/opus/volltexte/2007/233/ 143 .htm ThoftChristensen P... Spain. http://dx. Staat M. on Pressure Vessel Technology. Vietnam. and Findlay G. (2001).iasmirt. J. France. N. LISA – a European project for FEMbased limit and shakedown analysis.fhaachen. Probabilistic limit and shakedown problems. Siemaszko A.de/nicseries/volume15/nicseriesband15. Proceedings of 9th International conference on computational plasticity. Washington. http://opus. pp. 7384. China. Wirtz K.006 Taylor N. Staat M. (2007b). Shakedown and limit reliabilitybased design.2004..at/IAA/news/dba1_engl. John von Neumann Institute for Computing. R. Lyon.04.ac. Limit and shakedown analysis for plastic safety of complex structures. (2003b). et al. Proceedings of the International Conference on Material Theory and Nonlinear Dynamics. ASME Vol.1016/S00295493(00)004155 Staat M. Heitzer M. COMPLAS IX.org/SMiRT14/B022 Staat M. http://www. Heitzer M.tuwien. (2005)..
Sci. Springer. Dordrecht.Tran T.org/10. Yan A. N. (1990). H. Det Norske Veritas. Maier G.doi. Nguyen D.10. (1996).. Dual Limit and Shakedown analysis of structures. 144 . Technische Mechanik. PhD Thesis. (2001). (2003). (eds. Int. Belgium. LTASFRAMATOME. M.compstruc. K. 193. (2007). Analysis of pressure equipment by application of the primaldual theory of shakedown. pp. (2007e). R. 213225. M.2008. ASCE. Maier G. Appl. Contribution to the direct limit state analysis of plastified and cracked structures. Comp. Load bearing capacity of thin shell structures with elastoplastic material by direct methods. In [Staat and Heitzer (2003a)]. Berlin. 116 (10)..02. Mech. Tran I. K. June.) Limit States of Materials and Structures: Direct Methods. pp. Yan A. Weichert D. Hachemi A. http://dx.org/10. (2007d). Nguyen Dang H. Probabilistische Einspielanalyse mit der Finite Elemente Methode und FORM/SORM. J. pp. Kinematical formulation of limit and shakedown analysis.) Kluwer. 757767. Mech.1016/j.891 Wang.. pp. T.2007. (2008). (2000).. M. in press. Veritas Report RDIV/2000483.. Structural Safety. Fachhochschule Aachen Campus Jülich.. Vu D. In: Inelastic behaviour of structures under variable repeated loads: Direct analysis methods (edited by Weichert D. in press. pp. K. J. Tran T. Engng.strusafe. 23 (3). Vu D. Communications in Num. Staat M. Wien NewYork.. Kreißig R. PhD Thesis. 229238. Wierskowski C. Belgium. 46634674. to appear. Ellingwood B. Report CL26.005 Tran T.. Engng... Labor für Biomechanik.. 30. (1997).doi. GrossWeege J. Staat M.. N. Yan A. T.. In: A. D. Diplomarbeit.1002/cnm. (1988).. M. Vu D. In: Inelastic Analysis of Structures under Variable Loads: Theory and Engineering Applications (edited by Weichert D.org/10. N. Université de Liège. Kreißig R.). K. Vu D. A primaldual algorithm for shakedown analysis of structure. B.S. http://dx. Corotis R. Oslo.. (2007c).. in press. Limit analysis of general pipe elbow subjected to complex loading. The numerical assessment of elasticplastic sheets under variable mechanical and thermal loads using a simplified towsurface yield condition. Tvedt L... Upper bound limit and shakedown analysis of thin shells using the exact Ilyushin yield surface. Probabilistic limit and shakedown analysis of thin plates and shells. Springer. Pham P.doi. Vu D. Université de Liège. K. Kreißig R. Vu D. Meth. Yan A.. Nguyen Dang H. Limit state sensitivity of structural frames subjected to cyclic forces. Germany. M. K. Meth. Staat M.003.. Weichert. in Engng. Norway. Struct. 233254. 28242841. Reliability analysis of inelastic shell structures under variable loads.. Two Secondorder Approximations to the Failure Probability. Staat M. Tran T.. N. http://dx.1016/j. (2005). Weichert D. Ponter. Academic Press. (2002). Direct finite element kinematical approaches in limit and shakedown analysis of shells and elbows. Computer & Structures. University of Liege. (1983)...R. Staat M.. (2004). Shakedown of thinwalled structures with geometrical nonlinear effects. Yan A. Tsau Yueh.. pp.
Zhang Y. G. (1995b). Zhang Y. Baltimore. Meth.. PhD Thesis. Zimmermann J. Eng. John Hopkins University. Eng. An iteration algorithm for kinematic shakedown analysis. 145 . 127. 575582.. Appl. (1995a). Maryland. Appl. 33. Comp. (1991). Mech. Comp. Mech. Comp. (1982). pp. 217226. An algorithm for plastic limit analysis. 333341. A variational principle and an algorithm for limit analysis of beams and plates. W. Meth. H. pp. G. Eng.. 126. J. Meth.Yang W.. Analysis of structural system reliability with stochastic programming. Mech. Lu M. Appl. pp.
.1) and σ = 2 xb xa ∫ (x − μ) 2 f ( x )dx (A.. variance σ 2 .4) and approach the actual values of the mean and variance of X as m increases indefinitely. +∞ ) when X follows a Gaussian probability. standard deviation σ . this range is ( −∞. ˆ μ= and ˆ σ2 = 1 m (k ) ∑x m k =1 (A. Then. respectively.. that f is known. respectively. The estimates of μ and σ m are. The mean and variance of X can be calculated from the PDF f or measurements of X . The Probability Density Function (PDF) and Cumulative Distribution Function (CDF) of X are denoted by f and F . μ = ∫ xf ( x )dx x a xb (A. When f is not available. μ 1 2 and σ 2 can be estimated from a set of m 2 measurements {x( ) . coefficient of variation v = σ / μ and shift parameter a .Appendix PROBABILITY DISTRIBUTIONS AND TRANSFORMATION TO THE STANDARD GAUSSIAN SPACE Let X be a random variable with mean μ . x( ) } of random parameter X .. x( ) .3) 2 1 m (k ) ˆ ∑( x − μ2 ) m − 1 k =1 (A. This parameter gives the smallest possible value of X . Probabilities and distributions that are available in our code are given in this appendix when available in closed form.. For example. Closed form results are also presented for the transformation from the standard Gaussian U with mean zero and unit variance to variable X. Suppose first.2) where the limits of integration x a and x b extend over the range of possible values of X .
8) PDF: f ( x) = Parameters: 2 ⎧ ⎪ ⎪ ⎡ ln( x ) − μ ⎤ ⎫ exp ⎨ −0. (A. 2πσ ⎣ σ ⎦ ⎪ ⎪ ⎩ ⎭ (A.σ 2 ) PDF: 2 ⎧ 1 ⎪ ⎪ ⎡x−μ⎤ ⎫ exp ⎨ −0. Var ( X ) = σ 2 . − ∞ ≤ x ≤ +∞ . E ( λ . (A.1.5 ⎢ f ( x) = ⎥ ⎬ .11) 3.5 ⎢ ⎥ ⎬.7) Transformation: X = μ + σU .9) E ( X ) = μ . shift a and coefficient of variation v = σ /( μ − a ) σ 2 = ln ⎡1 + v 2 ⎤ ⎣ ⎦ μ = ln( μ ) − 0. Exponential distribution.10) X = a + exp ( μ + σ U ) . x ≥ a 2π xσ ⎣ σ ⎦ ⎪ ⎪ ⎩ ⎭ 1 (A. Var ( X ) = 1/ λ 2 . a ) PDF: f ( x ) = λ exp {− λ ( x − a )}. N ( μ .12) 147 .5) Parameters: CDF: E ( X ) = μ . (A. LogNormal distribution (A. 2.6) where Φ( x) = −∞ ∫ x f (u )du . (A. Parameters: CDF: E ( X ) = a + 1/ λ . Var ( X ) = σ 2 ⎡x−μ⎤ f (t )dt = Φ ⎢ ⎣ σ ⎥ ⎦ F ( x) = −∞ ∫ x (A. Normal distribution.5σ Transformation: 2 . x ≥ a .
vx ) . a ≤ x ≤ b. b−a 2 (A. Gamma distribution. 0 y (A.16) (A.15) E ( X ) = ( a + b ) / 2. 1 . b−a (A. 0 ∞ (A. a ≤ x≤b.13) 1 λ Φ ( −U ) . Γ (k ) (A. Var ( X ) = k / v 2 F ( x) = where Γ ( k . G ( k . Transformation: X =a− 4. x ≥ a. y ) = ∫ exp ( −u ) u k −1du .14) PDF: f ( x) = Parameters: CDF: F ( x ) == Transformation: X = a + ( b − a ) Φ (U ) . vx ) PDF: f ( x) = v ( vx ) k −1 exp ( −vx ) . Var ( X ) = ( b − a ) /12 x−a .18) where the gamma function Γ ( k ) = ∫ exp ( −u ) u k −1du . x ≥ 0.F ( x ) = 1 − exp {−λ ( x − a )}.21) 148 . x ≥ a . Γ (k ) (A.20) Γ ( k . U [ a.17) 5.19) Parameters: CDF: E ( X ) = a + k / v. Uniform distribution. (A. b ] (A.
26) α ln − ln ⎡Φ (U ) ⎤ . 6. E ( X ) = a + ( v − a ) Γ (1 + 1/ k ) . ⎣ ⎦ { } (A. Extreme Type I Distribution PDF: f ( x ) = α exp −α ( x − v ) − exp ⎡ −α ( x − v ) ⎤ .23) { } 1/ k . ⎪ ⎝ v−a ⎠ ⎪ ⎩ ⎭ Transformation: X = a + ( v − a ) − ln ⎡ Φ ( −U ) ⎤ ⎣ ⎦ (A. x ≥ a . x ≥ a . Var ( X ) = π 2 / ( 6α 2 ) .25) { } (A. ⎣ ⎦ Transformation: X =v− 1 E ( X ) = v + ( 0. − ∞ ≤ x ≤ +∞ .Transformation: not available in closed form. Var ( X ) = ( v − a ) ⎡ Γ (1 + 2 / k ) − Γ 2 (1 + 1/ k ) ⎤ ⎣ ⎦ 2 CDF: ⎧ ⎛ x − a ⎞k ⎫ ⎪ ⎪ F ( x ) = 1 − exp ⎨ − ⎜ ⎟ ⎬.27) 149 . Weibull Distribution PDF: k ⎛ x−a⎞ f ( x) = ⎜ ⎟ v−a⎝v−a ⎠ Parameters: k −1 ⎧ ⎛ x − a ⎞k ⎫ ⎪ ⎪ exp ⎨ − ⎜ ⎟ ⎬.24) 7.577 / α ) . − ∞ ≤ x ≤ +∞ . Therefore a NewtonRaphson iteration scheme can be employed in order to obtain the inverse. ⎪ ⎝v−a ⎠ ⎪ ⎩ ⎭ (A. { } (A. ⎣ ⎦ Parameters: CDF: F ( x ) = exp − exp ⎡ −α ( x − v ) ⎤ .22) k > 0. (A.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue listening from where you left off, or restart the preview.