You are on page 1of 9

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/227064167

Parametric structural optimization with respect to the multiaxial high-cycle


fatigue criterion

Article  in  Structural and Multidisciplinary Optimization · February 2007


DOI: 10.1007/s00158-006-0045-7

CITATIONS READS
29 139

2 authors, including:

Mirosław W. Mrzygłód
Opole University of Technology
34 PUBLICATIONS   165 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Mirosław W. Mrzygłód on 17 October 2014.

The user has requested enhancement of the downloaded file.


6th World Congresses of Structural and Multidisciplinary Optimization
Rio de Janeiro, 30 May - 03 June 2005, Brazil

Parametric Structural Optimization With Multiaxial High-Cycle Fatigue Criterion


M. Mrzygłód1, A. P. Zieliński2

(1) Cracow University of Technology, Institute of Rail Vehicles, Cracow, Poland, mrzyglod@mech.pk.edu.pl
(2) Cracow University of Technology, Institute of Machine Design, Cracow, Poland, apz@mech.pk.edu.pl

Abstract
Investigations on optimization of structures working in high-cycle load conditions were carried out and presented in the paper. The
work was concentrated on three principle areas: fatigue of material (with special regard to multiaxial criteria of high-cycle fatigue),
parametrical optimization of structures and application of the final element method (FEM).
The investigations and numerical implementation of several high-cycle criteria were made and the most convenient one for
optimization was selected. The main process of fatigue optimization was preceded by tests of methods of structural optimization and
preparation of tools for efficiency improvement of optimization algorithm. This stage includes preparation of software tools based on
evolutionary algorithms. Additionally, the decision variables were preselected by investigation of sensitivity of the objective function
on small increments of these variables. The work was illustrated by examples of optimization of mechanical structures working in
high-cycle load conditions. As observed in the computational examples, the proposed methodology of optimization allowed
effectively to lower the mass of the studied structure with maintaining its durability on an established level. The tools and fatigue
optimization methodology presented in the paper have universal character and can be applied to any case of a structure subjected to
high-cycle loads.
Keywords: structural optimization, high-cycle fatigue, multiaxial fatigue criterion, finite element method

1. Introduction
The analysis and optimization of complete life cycles of structural systems are recognized as one of key challenges in computational
mechanics [1]. The aspect of fatigue optimization began at early 70-ties and was focused mainly on shape optimization [2].
Contemporary investigations of fatigue optimization of structures encounter considerable difficulties. Large variety of fatigue
hypotheses make difficult decisions of choice of a unique fatigue criterion. Also adaptation of time history of loading is susceptible
to various interpretations. Finally, necessary methods of analysis of a problem (e.g. FEM) as well as complex algorithms of
optimization encounter a barrier of large time of computation. Hence, we can state that the development of methodology of structure
optimization with respect to fatigue damage seems to be highly justified.

2. Assumptions in fatigue optimization


Before the process of fatigue optimization several preliminary assumptions were made. It is possible to divide them into two groups
concerning:
I. Range of analyzed fatigue phenomena:
- high-cycle fatigue and infinitive fatigue life are taken into considerations,
- examined hypotheses concern arbitrary multiaxial state of loads,
II Way of modeling and level of simplifications:
- state of stresses and deformations in structures is modeled using the finite element method,
- particular components of loads have in-phase or reverse character,
- frequency of loads is considerably lower than the first eigenfrequency of studied structures
- inertial effects are not taken into account,
- real time load history is simplified to a form convenient for analysis and optimization,
- possible short time violations of established load regime are represented by a suitable safety factor.

3.Choice of fatigue criterion convenient for optimization


Experimental investigations of recent decades resulted in numerous hypotheses of high-cycle multiaxial fatigue (MHCF). In literature
many proposals of such criteria can be found [3-6]. From the point of view of applications to numerical optimization the most
convenient are two groups: criteria basing on stress state invariants and criteria using average stresses or deformations in an
elementary volume. For final selection of a suitable fatigue hypothesis the following three formulae were taken into consideration:
Sines [7], Crossland [8] and Dang Van [9]. The first two belong to the invariant group and the third one is based on the average value
concept. The components of these criteria are easy to obtain from programs of the FEM analysis.
The invariant formulae usually consist of quantities related to hydrostatic and octahedral stresses. The use of these hypotheses allows
to determine a point of initiation of fatigue cracks. However, orientation of the potential cracks with these criteria cannot be defined.
G. Sines [5,7] analyzed influence of different combinations of variable bending and torsion stresses on fatigue life of a structure. On
this base he formulated a criterion including an amplitude of octahedral shear stresses and an average value of normal hydrostatic
stresses. The Sines hypothesis shows satisfactory correlations with experimental investigations. It has the following form:

J 2, a + κσ H , m ≤ λ (1)
where: κ and λ are material parameters,
J2 is the second invariant of stress deviator tensor:
1
J 2 ,a =
6
[ ]
(σ 1a − σ 2a )2 + (σ 2a − σ 3a )2 + (σ 3a − σ 1a )2 ’
1
and σ H ,m = [σ 1m + σ 2 m + σ 3m ] ,
3
σ1a , σ2a , σ3a are amplitude - type principal stresses; symbol "a ” means corresponding
amplitude of loads,
σ1m , σ2m , σ3m are mean - type principal stresses; symbol "m ” means average (mean) value
of loads.

A criterion formulated by B. Crossland [5,8] is very close to the Sines formula. A difference in approach of both researchers concerns
the hydrostatic stresses σH , which according to Crossland should be represented by their maximum value:

J 2,a + κσ H ,max ≤ λ (2)


where: σ H , max = σ H , a + σ H , m

Criteria based on average stresses in an elementary volume V take into consideration the average value of shear and normal stresses
in this volume. K. Dang Van [5,9] formulated his hypothesis observing local plastic deformations in microscopic scale, on a level of
crystallites. They can initiate micro-cracks even then, when a studied structure remains in macroscopic scale in a range of elastic
strains. According to Dang Van the fatigue damage appears in a definite time, when the combination of local shear stresses τ(t) and a
hydrostatic stress σH(t) cuts the borders of an admissible fatigue area. Initially complicated, the way of calculation of shear stress
amplitude became simplified in a new form of the Dang Van criterion [4]:

max [τ (t ) + κσ H (t )] ≤ λ (3)
A

where: A is the area of studied object,


σ 1 (t ) − σ 3 (t ) ,
τ (t ) =
2
1
σ H (t ) = (σ 1 (t ) + σ 2 (t ) + σ 3 (t ) ) .
3
For numerical implementation of the criteria the scripting language APDL of the program ANSYS® [10] was used. The APDL makes
possible to record current values of stress tensor components and a certain predefined quantity (e.g. equivalent stresses). A user
specifies vectors of variables as well as algebraic operations on these vectors. Therefore, these new endurance criteria can easily be
implemented.
To test the Sines, Crosland and Dang Van criteria suitable FEM models and analysis programs were prepared. As it resulted from the
numerical investigations, the Dang Van hypothesis turned out to be the most rigorous formulation. It leads to the most demanding
requirements ensuring, however, the highest safety level. More information on the above fatigue numerical tests can be found in the
work [11]. The Dang Van hypothesis is used in the next stage of the investigations concerning optimization of a structure working in
high-cycle load conditions.

4. Methods and tools of optimization


The Dang Van criterion (3) proposes the analysis with respect to variable time. For optimization the following pattern of reduction of
the load time history was proposed:
- determination of equivalent amplitude and mean load values from the real time history,
- calculation of five load levels, which would be checked in every single fatigue analysis (see Fig. 1).

a) b)
Fig. 1. Example of load time history q( t) (a) transformed to convenient for fatigue Dang Van criterion form (b)
The equivalent loads could be determined in different ways. In our case we decided for the extreme values ( Fig. 1), which would
increase the safety factor of the new, improved structures. The whole fatigue analysis of five load cases according to the proposed
rules was carried out in an automatic way.
The method of probabilistic search based on evolutionary algorithms was chosen for fatigue optimization. The evolutionary
algorithms conduct multidirectional searches protecting population of potential solutions and exchanging information between them.
Evolution of different solutions is guided by a fitness function, which reflects behavior of the objective function [12].
As a main tool of optimization the software package Evolutionary Optimization System (EOS)[12] was chosen. Certain
modifications in the algorithm of this program were necessary to apply it in the numerical optimization. A new additional procedure
enabling cooperation of the EOS with the external FEM program ANSYS® had to be implemented.
To decrease the number of decision variables, simple sensitivity investigation of the objective function was carried out. It delivered
information about sensitivity of the objective function on incremental values of the design variables [13]. The measure of sensitivity
S of the objective function was proposed as:

⎧ Q ( x k , g ( x k ) ) − Q ( x 0 , g ( x0 ) )
⎪ , when Q( x k , g ( x k ) ) < Q( x 0 , g ( x0 ) ) (4)
S=⎨ ∆xk

⎩ 0 , when Q( x k , g ( x k ) ) ≥ Q( x 0 , g ( x0 ) )
Q( xi , g ( xi )) = Q f ( xi ) + Qg ( g ( xi )) , i = {1,2 ...m} (5)

Q f ( xi ) = W ( xi )
W0
(6)

Q f ( x0 ) = 1
Qg ( g ( xi )) = ( )
g ( xi ) 2 β
g
(7)
where:
⎧ x1 ⎫ ⎧ x1 ⎫
⎪x ⎪ ⎪ x2 ⎪
⎪ 2⎪ ⎪ ⎪
⎪⎪ .. ⎪⎪ ; ⎪⎪ .. ⎪⎪ ,
x0 = ⎨ ⎬ xk = ⎨ ⎬
⎪ xk ⎪ ⎪ xk + ∆xk ⎪
⎪ .. ⎪ ⎪ .. ⎪
⎪ ⎪ ⎪ ⎪
⎩⎪ xm ⎭⎪ ⎩⎪ xm ⎭⎪
Q - dimensionless, unconstrained objective function,
Qf - basic objective function,
Qg - penalty function,
W0 - reference value of objective function,
g (xi) - state parameter ( in our case fatigue equivalent stress),
g - reference value of state parameter ( in our case reverse torsion fatigue life ),
β - chosen positive number (relatively large).

To improve efficiency of the evolutionary method, the optimization algorithm with the following steps was proposed:
I. Random determination of a vector { xi } of decision variables in the solutions space (for example i = 1…12; K = 1000 cases),
II. Selection of N starting points with the smallest values of the objective function (N = 5),
III. Sensitivity investigation of the objective function for {xi} variables in the selected points,
IV. Selection of a subset { xm } from { xi } design variables for optimization,
V. Optimization process using evolutionary algorithms for { xm } variables,
VI. Result of optimization - the optimal vector of the decision variables.

5. Example of fatigue optimization


A suspension arm (Fig. 2a, 3a) was chosen as one of the optimization examples. In a vehicle this element is subjected to high-cycle
load conditions (the load scheme is presented in Fig. 2b [14]). In the example the mass of the structure will be optimized with the
Dang Van fatigue equivalent stress as a state parameter.
For the examined element the Multi-Body Simulations (MBS) were carried out to obtain the load time history of close to standard
work conditions [14]. Basing on this simulation the history of 12 components of loads were determined.
The optimized suspension arm is built from two thin sheet drawpieces assembled by welding. Hence, the shell-type finite elements
(SHELL63) were used in the FEM modeling. This allowed to obtain acceptable results in a relatively small computational time.
Using the APDL scripting language the parametric FE model of the arm was coded. Because of limited possibilities of modeling of
the ANSYS® program, the original form of the suspension arm had to undergo certain simplifications (fig. 3b). The process of
simplification was carried out to obtain results on the safe side for the studied object.
a) b)
Fig. 2. Car suspension with optimized part colored in blue (a) and scheme of loads acting on optimized suspension arm (b) [14]

a)
b)

Fig. 3 Suspension arm (a) and its FE model with loads and boundary conditions (b)

The Dang Van fatigue subroutine was added to the program of FE analysis. Also the sequences of automatic recording of the state
parameter and the objective function to the text file were included. In the convergence tests the results of calculations with three
meshes of different density nad similar character of pattern (Fig. 3b) were compared. For further analysis and optimizations the mesh
with Ne = 10616 elements was accepted. For coarser mesh (Ne = 5214) the von Mises equivalent stress was 4.5 % lower, and for
Ne = 20218 about 14.8 % higher. It was decided that such a difference in results is admissible to conduct the process of optimization.
However, it should be noted that the optimized part must be reexamined with the finer mesh to determine a real safety factor.
The load time history was transformed according to assumptions in the Dang Van criterion. The five cases from the equivalent time
history (see Fig. 1b) were calculated for all twelve load components. Moreover, to find the limit value of the state parameter (the
Dang Van equivalent stress), a preliminary fatigue analysis for an original shape of the suspension arm was done.
The optimization task was formulated in the following form:
- initial 12 decision variables were selected (see Fig. 4):
x = [x1, x1 x1 ... xn] , n = 12 (8)

Remark: All dimensions, which were chosen as the design variables, could be changed without disturbance of cooperation
between the suspension arm and the other parts of the vehicle suspension. It was assumed that the number of variables would
reduced after sensitivity investigation.
- the decision variables were limited by their upper and lower bounds:

xi ≤ xi ≤ xi , (i = 1,2,3, ... ,12)

- one state variable (maximum equivalent stress according to the Dang Van criterion) was considered, with the upper limit:

max g i ( x ) ≤ g i , ( i = 1 )
A

- the mass of the arm was chosen as the main objective function W(x).

Fig. 4. Decision variables initially accepted to optimization process; variables selected by sensitivity investigation marked in red
color
According to the earlier proposed procedure, investigation of sensitivity of the objective function on increments of the decision
variables was conducted. In the first stage of investigation K=1000 random iterations were calculated to choose five initial points.
Additionally one heuristic point was also selected. For all of the starting points investigation of the sensitivity measure
S was made for 12 decision variables. The increments ∆xk were determined by division of the admissible range of a particular
variable in M = 100 segments. The values ∆xk = ( x k − x k )/ 100 were defined as an increment in a small neighborhood and 5∆xk as
an increment in a large distance [10]. The results of the research is presented in Tab.1. The variables, for which the objective function
shows considerable sensitivity (S ≥ 1.0 ) are marked in red.
To the further optimization process 5 from 12 decision variables were selected. The chosen parameters got at least 4 values of the
sensitivity measures above 1.0 (See Tab.1a). The whole procedure of the sensitivity investigation took ca 2 [h] on a single computer
with a P-IV processor. Taking into account that the full optimization process using all the twelve variables would take above 400 [h],
the application of the procedure of elimination of the less important variables can be considered as justified. However, it could be
noted that, after the first run of the optimization algorithm, the elimination procedure should rather be repeated for all the twelve
variables (see Tab. 1b).
For the evolutionary algorithm the following starting parameters were accepted:
- size of population J = 350
- crossover parameter pc = 0.7
- mutation parameter pm = 0.4
- number of generations Lg = 70
As a result of the optimization ca 11.3% decrease of mass of the studied structure was obtained. The total time of computations
amounted to about 180 hours. Fig. 5 presents initial and optimal forms of the investigated structure. The parameters of the
evolutionary optimization are listed in Fig. 6. The subsequent stages of evolution of the suspension arm are there also presented (last
improvement of the objective function was obtained in 53rd generation).
The whole fatigue optimization process of the examined object was completed by a checking computation for a different set of
control parameters of the evolutionary algorithm (J=500, Lg=32, pc=0.7, pm=0.4). The obtained result (W = 5777.02) was a little
worse than the primary optimum (see Fig. 6), which confirms the correctness of EOS setup parameters and effectiveness of the whole
process.

Table 1. Results of investigation of sensitivity S (def. (4)); S ≥ 1.0. is marked in red (a)
Sensitivity measure after first stage of optimization ( b)
a) b)
Result after
Random Random Random Random Random
Decision Heuristic first
point point point point point
variables point optimization
No. 1 No. 2 No. 3 No. 4 No. 5
run
X1 = l2 1.335 0.024 1.838 0.003 0.016 0.016 0.000
X2 = l3 0.558 0.099 1.759 0.012 0.161 0.161 0.044
X3 = l4 34.528 1.751 70.207 0.217 9.140 9.140 1.411
X4 = l5 3.925 5.880 35.041 0.000 30.952 30.952 0.177
X5 = l6 0.717 0.019 3.146 0.016 0.033 0.033 3.866
X6 = l7 4.538 0.132 0.370 0.081 0.189 0.189 0.000
X7 = l8 0.000 0.000 0.000 0.000 0.000 0.000 4.188
X8 = b4 33.902 0.998 105.517 0.000 7.261 7.261 1.039
X9 = b7 0.360 0.077 0.278 0.010 0.000 0.000 1.566
X10 = g1 53.409 8.196 127.552 1.632 33.026 33.026 5.503
X11 = g2 0.022 0.053 2.981 0.083 0.000 0.000 3.121
X12 = g4 7.073 0.228 10.839 0.433 2.409 2.409 12.484

Fig. 5. Comparison of initial (upper) and optimal (bottom) arrangement of structure


Generation 1

Design parameters:
l4=0.002752 [m]
l5=0.055171
b4=0.096032
g1=0.016801
g4=0.002021
State parameter:
DV Eqv Stress =
19476660.0 [Pa]
Objective function:
F = 6026.863 [g]

Generation 15

Design parameters:
l4=0.007348
l5=0.046694
b4=0.101848
g1=0.016525
g4=0.001913
State parameter:
DV Eqv Stress =
19927160.0
Objective function:
F = 5840.646

Generation 30

Design parameters:
l4=0.003388
l5=0.042541
b4=0.098660
g1=0.016017
g4=0.001934
State parameter:
DV Eqv Stress =
19865100.0
Objective function:
F = 5778.229

Generation 41

Design parameters:
l4=0.002862
l5=0.050876
b4=0.090470
g1=0.016033
g4=0.001921
State parameter:
DV Eqv Stress 19999510.0
Objective function:
F = 5756.380

Generation 53

Design parameters:
l4=0.001481
l5=0.043338
b4=0.096282
g1=0.016340
g4=0.001915
State parameter:
DV Eqv Stress =
19945180.0
Objective function:
F = 5713.026

Fig.6. Stages of evolutionary optimization


The sensitivity investigation of objective function for the found optimum point was repeated for the full set of 12 decision variables.
As we can see in Tab. 1b, the measure S in general decreased. However, 8 from 12 variables still have this measure above 1.0. This
suggests, that the optimization process could be carried out again, now for a different variable set. On the other hand, the found
optimum vector could be treated as a starting point for a standard gradient procedure. It is obvious, that for this kind of optimization,
reaching the global minimum is rather difficult, however, proceeding in the way described above we can find a point to this
minimum very near. It should also be noted, that the parallel computing used by the authors to accelerate calculations presented in
this paper [11] was an effective and very convenient tool in optimization algorithms.

6. Conclusions
In the paper a certain new approach to optimization of a structure working in multiaxial high-cycle load conditions is presented. The
proposed algorithm applies the MHCF criteria, which are in good agreement with experimental investigations. With the proposed
strategy of adaptation of the load history, preselections of the decision variables and application of the evolutionary algorithms this
fatigue optimization methodology can easily be adopted to more complex structural problems.
Further authors’ investigations of the proposed algorithm will take into account other types of cyclic failure, especially low-cycle
fatigue.

7. References
1. Bathe K. J., The Key Challenges in Computational Mechanics, IACM No 14,2003
2. Latos W., Życzkowski M., The optimum of rotating shaft for combined fatigue strength, Applied Mechanics Vol XXI (7-8),
pp341-351, PAN, 1973
3. Garud Y. S., Multiaxial fatigue: a survey of the state of the art, Journal of Testing and Evaluation, Vol 9, pp165-178, 1981
4. Ballard P., Dang Van K., Deperrois A., Papadopoulos Y. V., High cycle fatigue and a finite element analysis, Fatigue & Fracture
of Engineering Materials & Structures, Vol.18 , pp397-411, 1995
5. Papadopoulos I.V., Davoli P., Gorla C., Filippini M., Bernasconi A., A comparative study of multiaxial high-cycle fatigue criteria
for metals, Int. Journal of Fatigue, Vol 19, pp219-235, 1997
6. Dang Van K., Papadopoulos I.V.(Eds.), High Cycle Metal Fatigue, From Theory to Applications, C.I.S.M. Courses and Lectures
N° 392, Springer 1999
7. Sines G., Behaviour of metals under complex static and alternating stresses, in ‘Metal Fatigue’ , pp145-169, McGraw Hill, New
York, 1959
8. Crossland B., Effect of large hydrostatic pressures on the torsional fatigue strength of an alloy steel. In: Proceedings of the
International Conference on Fatigue of Metals, Institution of Mechanical Engineers, pp138–49, London, 1956
9. Dang Van K., Griveau B., Message O., On a new multiaxial fatigue limit criterion: theory and application, Biaxial and multiaxial
fatigue, EGF 3. Mechanical Engineering Publications, London, pp479-496, 1989
10. Ansys Inc., Release 8.1 Documentation ,2004
11. Mrzygłód M., Parametrical optimization of structures working in high-cycle load conditions, Cracow University of
Technology, PhD Thesis, 2005
12. Osyczka A., Evolutionary Algorithms for Single and Multicriteria Design Optimization, A Springer-Verlag Company, 2001
13. Zieliński A. P., Sanecki H., Karaś M., Effectiveness of the Trefftz method in different engineering optimization procedures,
Computer Assisted Mechanics and Engineering Sciences , Vol. 8, pp. 479-493, 2001
14. Walczak St., Analysis of Dynamic Loads Occurring in Different Types of Independent Car Suspension Systems, Cracow
University of Technology, PhD Thesis, 2003

View publication stats

You might also like