Lecture Notes in Applied

and Computational Mechanics

Volume 33

Series Editors
Prof. Dr.-Ing. Friedrich Pfeiffer
Prof. Dr.-Ing. Peter Wriggers
Lecture Notes in Applied and Computational Mechanics
Edited by F. Pfeiffer and P. Wriggers
Further volumes of this series found on our homepage: springer.com

Vol. 33: Niesłony, A.; Macha, E. Vol. 20: Zohdi T.I., Wriggers P.
Spectral Method in Multiaxial Random Fatigue Introduction to Computational Micromechanics
146 p. 2007 [978-3-540-73822-0] 196 p. 2005 [978-3-540-22820-2]
Vol. 32: Bardzokas, D.I.; Filshtinsky, M.L.; Filshtinsky, L.A. (Eds.) Vol. 19: McCallen R., Browand F., Ross J. (Eds.)
Mathematical Methods The Aerodynamics of Heavy Vehicles:
in Electro-Magneto-Elasticity Trucks, Buses, and Trains
530 p. 2007 [978-3-540-71030-1] 567 p. 2004 [978-3-540-22088-6]
Vol. 31: Lehmann, L. (Ed.) Vol. 18: Leine, R.I., Nijmeijer, H.
Wave Propagation in Infinite Domains Dynamics and Bifurcations
186 p. 2007 [978-3-540-71108-7] of Non-Smooth Mechanical Systems
Vol. 30: Stupkiewicz, S. (Ed.) 236 p. 2004 [978-3-540-21987-3]
Micromechanics of Contact and Interphase Layers Vol. 17: Hurtado, J.E.
206 p. 2006 [978-3-540-49716-5] Structural Reliability: Statistical Learning Perspectives
Vol. 29: Schanz, M.; Steinbach, O. (Eds.) 257 p. 2004 [978-3-540-21963-7]
Boundary Element Analysis Vol. 16: Kienzler R., Altenbach H., Ott I. (Eds.)
571 p. 2006 [978-3-540-47465-4] Theories of Plates and Shells:
Vol. 28: Helmig, R.; Mielke, A.; Wohlmuth, B.I. (Eds.) Critical Review and New Applications
Multifield Problems in Solid and Fluid Mechanics 238 p. 2004 [978-3-540-20997-3]
571 p. 2006 [978-3-540-34959-4] Vol. 15: Dyszlewicz, J.
Vol. 27: Wriggers P., Nackenhorst U. (Eds.) Micropolar Theory of Elasticity
Analysis and Simulation of Contact Problems 356 p. 2004 [978-3-540-41835-1]
395 p. 2006 [978-3-540-31760-9] Vol. 14: Frémond M., Maceri F. (Eds.)
Vol. 26: Nowacki, J.P. Novel Approaches in Civil Engineering
Static and Dynamic Coupled Fields in Bodies 400 p. 2003 [978-3-540-41836-8]
with Piezoeffects or Polarization Gradient Vol. 13: Kolymbas D. (Eds.)
209 p. 2006 [978-3-540-31668-8] Advanced Mathematical and Computational
Vol. 25: Chen C.-N. Geomechanics
Discrete Element Analysis Methods 315 p. 2003 [978-3-540-40547-4]
of Generic Differential Quadratures Vol. 12: Wendland W., Efendiev M. (Eds.)
282 p. 2006 [978-3-540-28947-0] Analysis and Simulation of Multifield Problems
Vol. 24: Schenk, C.A., Schuëller. G 381 p. 2003 [978-3-540-00696-1]
Uncertainty Assessment of Large Vol. 11: Hutter K., Kirchner N. (Eds.)
Finite Element Systems Dynamic Response of Granular and Porous Materials
165 p. 2006 [978-3-540-25343-3] under Large and Catastrophic Deformations
Vol. 23: Frémond M., Maceri F. (Eds.) 426 p. 2003 [978-3-540-00849-1]
Mechanical Modelling and Computational Issues Vol. 10: Hutter K., Baaser H. (Eds.)
in Civil Engineering Deformation and Failure in Metallic Materials
400 p. 2005 [978-3-540-25567-3] 409 p. 2003 [978-3-540-00848-4]
Vol. 22: Chang C.H. Vol. 9: Skrzypek J., Ganczarski A.W. (Eds.)
Mechanics of Elastic Structures with Inclined Members: Anisotropic Behaviour of Damaged Materials
Analysis of Vibration, Buckling and Bending of X-Braced 366 p. 2003 [978-3-540-00437-0]
Frames and Conical Shells Vol. 8: Kowalski, S.J.
190 p. 2004 [978-3-540-24384-7] Thermomechanics of Drying Processes
Vol. 21: Hinkelmann R. 365 p. 2003 [978-3-540-00412-7]
Efficient Numerical Methods and Information-Processing Vol. 7: Shlyannikov, V.N.
Techniques for Modeling Hydro- and Environmental Elastic-Plastic Mixed-Mode Fracture Criteria
Systems and Parameters
305 p. 2005 [978-3-540-24146-1] 246 p. 2002 [978-3-540-44316-2]
Spectral Method in Multiaxial
Random Fatigue

Adam Niesłony Ewald Macha

•

With 57 Figures and 7 Tables

A. Niesłony
E. Macha
Opole University of Technology
Faculty of Mechanics
Department of Mechanics and Machine Design
ul. Mikolajczyka 5
45-271 Opole, POLAND
[email protected]
[email protected]

Library of Congress Control Number: 2007933178

ISSN 1613-7736

ISBN 978-3-540-73822-0 Springer Berlin Heidelberg New York

This work is subject to copyright. All rights are reserved, whether the whole or part of the material
is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broad-
casting, reproduction on microfilm or in any other ways, and storage in data banks. Duplication of
this publication or parts thereof is permitted only under the provisions of the German Copyright
Law of September 9, 1965, in its current version, and permission for use must always be obtained
from Springer. Violations are liable for prosecution under the German Copyright Law.

Springer is a part of Springer Science+Business Media

springer.com
© Springer-Verlag Berlin Heidelberg 2007

The use of general descriptive names, registered names, trademarks, etc. in this publication does not
imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.

Cover design: Künkel & Lopka, Heidelberg

Typesetting by the authors
Production: Integra Software Services Pvt. Ltd., India

Printed on acid-free paper SPIN: 12086764 45/3100/Integra 5 4 3 2 1 0

Preface

The monograph is devoted to the spectral method for determination of fatigue

life of machine elements and structures under multiaxial service loading. A
detailed review of literature on spectral methods that includes: publications
describing various relations between histories of loading and power spectral
density functions, some methods of fatigue life determination under Gaus-
sian loading and under multiaxial service loading has been presented in the
monograph. The monograph contains theoretical foundations of the spectral
method for fatigue life determination. The authors have discussed a rule of
description of random loading states with the matrix of power spectral den-
sity functions of the stress/strain tensor components. Some chosen criteria of
multiaxial fatigue failure being linear combinations of stress or strain compo-
nents on the critical plane have been analyzed. The proposed formula enables
to determine power spectral density of the equivalent history directly from
the components of the power spectral density matrix of the multidimensional
stochastic process. The assumptions and the procedure of determination of
basic relationships of the spectral method according to stress and strain ap-
proaches have been presented. The authors worked out equations determining
the fatigue life according to the spectral method using various linear hypothe-
ses of fatigue damage accumulation. The proposed algorithm of fatigue life
contains five blocks: 1 – determination of loading, 2 – determination of the
critical plane position for the assumed multiaxial fatigue failure criterion, 3
– determination of power spectral density of the equivalent stress or strain,
4 – determination of statistical parameters of the equivalent parameter re-
sponsible for fatigue damage, and 5 – fatigue life calculation according to a
suitable hypothesis of damage accumulation. In order to compare the results
of fatigue life estimations many simulation tests were performed with the use
of the spectral and the cycle counting methods. The comparison was done
in three stages: 1 – positions of critical planes determined with the variance
method in frequency and time domains, 2 – some chosen statistical parameters
of the equivalent histories (distributions of instantaneous values, distributions
of amplitudes after cycle counting and power spectral density functions), 3 –
VI Preface

lives calculated with the spectral and the cycle counting methods. The mono-
graph contains the results of the test for 18G2A steel under random bending,
torsion and combined bending with torsion with different correlations and lev-
els of torsional and bending loading components. The results from the test and
from the calculations with the use of spectral method and the cycle counting
method have been compared and discussed. It appears that both considered
methods give similar and satisfactory estimations of fatigue life. We hope that
problems discussed in the monograph are interesting for experienced structure
integrity engineers and give them possibility to broaden the knowledge about
the efficient methods of lifetime evaluation of machine elements and struc-
tures under multiaxial random loading. The book can be also recommended
to postgraduate and PhD students with an interest in fatigue of engineering
materials.
A large part of the editorial works connected with the book was made
during the Alexander von Humboldt Research Fellowships in Germany. Dr.
Adam Nieslony wants to thank the AvH Foundation for their support as
well as Prof. Sonsino from Fraunhofer Institute for Structural Durability and
System Reliability LBF in Darmstadt for scientific and personal care during
the stay in Germany. Authors also want to thank the Commission of the
European Communities for the financial support as part of the CESTI project
under the FP5 GROWTH Programme, contract No. G1MA-CT-2002-04058.
Special thanks are dedicated to friends from the Department of Mechanics
and Machine Design, Opole University of Technology, for the valuable help in
working the book out, particularly to Prof. Tadeusz L  agoda, Dr. Aleksander
Karolczuk and Dr. Henryk Achtelik.

Adam Nieslony
Ewald Macha

Department of Mechanics and Machine Design

Mechanical Engineering Faculty
Opole University of Technology
ul. Mikolajczyka 5
45-271 Opole, Poland
Contents

List of Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Spectral Methods for Fatigue Description . . . . . . . . . . . . . . . . . . 7

2.1 Random Gaussian Loading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.1.1 Narrow-Band Frequency Stress . . . . . . . . . . . . . . . . . . . . . . 10
2.1.2 Broad-Band Frequency Stress . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 Random Non-Gaussian Loads . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.3 Multiaxial Loading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3 Theoretical Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.1 Description of Random Stress and Strain States . . . . . . . . . . . . . 37
3.2 Multiaxial Fatigue Failure Criteria . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.3 Power Spectral Density Function of Equivalent History . . . . . . . 44
3.4 Amplitude Distribution in Spectral Methods . . . . . . . . . . . . . . . . 47
3.5 Spectral Method of Fatigue Life Determination . . . . . . . . . . . . . . 50
3.5.1 Fatigue Life Calculation Based on Characteristics (σa
– Nf ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.5.2 Fatigue Life Calculation Based on Characteristics (εa
– Nf ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4 Algorithm of Spectral Method for Evaluation of Fatigue

Life . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.1 Block 1 – Loading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.2 Block 2 – Determination of Critical Plane Position . . . . . . . . . . . 68
4.3 Block 3 – Calculation of Power Spectral Density Function of
Equivalent Stress or Strain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.4 Block 4 – Statistical Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.5 Block 5 – Calculation of Fatigue Life . . . . . . . . . . . . . . . . . . . . . . . 70
VIII Contents

5 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
5.1 Algorithm, Assumptions and Objectives of Simulations . . . . . . . 73
5.2 Input Quantities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.3 Analysis of Positions of Critical Planes . . . . . . . . . . . . . . . . . . . . . 83
5.4 Analysis of Equivalent Quantities . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.4.1 Random Histories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.4.2 Power Spectral Densities . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.5 Analysis of Fatigue Life . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

6 Experimental Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

6.1 Test Stands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
6.2 Material and Geometry of Tested Specimens . . . . . . . . . . . . . . . . 117
6.3 Testing under Narrow-Band Frequency Loading . . . . . . . . . . . . . 118
6.4 Testing under Broad-Band Frequency Loading . . . . . . . . . . . . . . 119
6.5 Comparison of Calculated and Experimental Fatigue Life . . . . . 120

7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
List of Symbols

A — coefficient derived from Wöhler fatigue characteristic,

b — fatigue strength exponent,
c — fatigue strain exponent,
D — damage,
E MPa Young modulus,
e [·] — expected value of an expression,
ˆ· — mean value, expected value of a variable,
f s−1 frequency,
f0 s−1 dominant frequency,
F N force,
G(f ) — one-sided power spectral density function,
G∗ (f ) — complex conjugate of the function G(f ),
h(τ ) — impulse response (transfer) function of physical system,
j — imaginary unit,
I — irregularity coefficient,
m — exponent of the Wöhler fatigue characteristic,
mk — k-th moment of spectral power density function,
M Nm moment of force,
M+ — expected number of peaks (local extremes) in a time
unit,
N0+ — expected number transitions of zero level with positive
slope in a time unit,
Nf cycles number of cycles to failure,
NG cycles number of cycles to failure for fatigue limit,
p — probability density,
P — probability,
r — correlation coefficient,
R(τ ) — correlation function,
sgn(·) — signum function,
t, τ s time,
2 List of Symbols

T s period,
σ MPa stress,
σ(t) MPa stress tensor,
ε(t) — strain tensor,
σaf MPa fatigue strain limit for tension-compression,
τaf MPa fatigue strain limit for torsion,
σf MPa fatigue strength coefficient for tension-compression,
εf — fatigue plastic strain coefficient,
ε — strain,
 — spectrum width parameter,
κ — kurtosis,
μ — covariance,
Γ (·) — gamma function,
Γ (· , ·) — incomplete gamma function,
ω rad · s−1 angular frequency (pulsatance),
Δ — range,
λ — coefficient including spectrum width,
γ — engineering shear strain.

Indices and Abbreviations

a – amplitude,
m – mean value,
ZC – cycle counting method,
SP – spectral method,
min – minimum,
max – maximum,
eq – equivalent or reduced value,
PSD – power spectral density,
FFT – fast Fourier transformation,
IFFT – inverse fast Fourier transformation,
NB – narrow-band frequency spectrum,
BB – broad-band frequency spectrum,
RC – range counting,
RF – rain flow,
RF C – rain flow counting.
1
Introduction

The question of material, machine and construction fatigue occurs prevalently

in machine, building, maritime, and aircraft industries both at the stage of
manufacturing and during operation. Some structural components suffer from
failure associated with fatigue in the conditions of cyclic or random multiaxial
loading. The fatigue process is particularly hazardous in aviation, power en-
gineering and multiple supporting structures (e.g. jack-up platforms, bridges,
towers, masts), in which the failure of a component causes huge material
losses, ecological disaster and is directly threatening to the life and health of
people. It is indicated [3, 11, 14, 63] that a number of random loading follow-
ing natural phenomena (sea waves, variable wind velocity, etc.) is Gaussian
in character with various spectral frequency widths. Therefore, this case is
particularly taken into account in this book.
Research throughout the world is devoted to an elaboration of a universal
method for determination of machine and component strength. A calculation
algorithm with an application to a wide spectrum of materials under random
loading, starting with a simple uniaxial state and involving complex multiaxial
loading states is attempted in the course of research worldwide. It could be
indicated that calculation algorithms in bibliography are divided into a group
of cycle counting method, in which loading is analysed in the time domain,
and a group of spectral methods in which power spectral density function
defined in the frequency domain is applied for the description of loading.
In the cycle counting methods, the assessment of fatigue related strength is
enabled by adequate criteria of multiaxial random fatigue, which are applied
for determination of the equivalent stress or strain histories [2, 36]. Subse-
quently, algorithms for extracting cycles of histories and familiar procedures
for accumulation of damage related failure stated for uniaxial random fatigue
are applied [22, 95]. The postulated algorithm of determination of fatigue life
by means of the spectral method applies the fatigue failure criteria and con-
sists in the reduction of the multiaxial loading state to an equivalent uniaxial
state in the frequency domain. By reduction of the multiaxial loading state
the multiaxial fatigue assessment criteria based on critical plane concept was
4 1 Introduction

used. For determination of the necessary critical plane position the methods of
variance or damage accumulation can be applied. Statistical parameters appli-
cable for a description of the amplitude distribution are derived directly from
the power spectral density function of equivalent stress or strain, obtained
during the reduction of multiaxial loading state to the equivalent uniaxial
one. As in the case of cycle counting methods, the fatigue life is determined
under an assumption of an adequate hypothesis of damage accumulation.
During literature studies it should be pointed out that in a number of
cases it is easier to define load in the frequency domain. It is often necessary
to account for the dynamics of the system which affects the loading of its
components. In the course of such analyses (conducted e.g. by means of the
Finite Element Method (FEM)), power spectral density function of stress
or strain are obtained, which could be directly used for the determination of
fatigue life by the spectral method. In the contrary case, it is necessary to refer
to the time domain, which makes calculation more complex and extends its
duration. Therefore the focus in the current book is on the comparison between
the cycle counting method and the spectral method by conducting computer
simulations devoted to the determination of fatigue life for various stress and
strain states and subsequently, a comparison of calculations with experimental
data. Chapter 2 is devoted to an overview of bibliography coverage of the
spectral methods for determination of fatigue life with a particular focus on
multiaxial Gaussian loading.
After the introduction in Chapter 3 of the theoretical essentials of random
stress and strain states, the criteria for multiaxial fatigue failure determi-
nation in the frequency domain [36, 54, 56] familiar from the literature are
stated. Subsequently the assumptions and a technique for the derivation of
spectral formulae involving fatigue characteristics (σa − Nf ) and (εa − Nf )
are stated. A number of various hypotheses for the fatigue damage accumula-
tion by derivation of fatigue life formula in accordance with Palmgren-Miner,
Haibach, Corten-Dolan and Serensen-Kogayev hypotheses are considered.
Chapter 4 is devoted to a description of the algorithm of spectral method
for evaluation of fatigue life under multiaxial random loading. A simple cal-
culation model is postulated for determination of the fatigue life of a specific
material in a plain stress state. The following Chapter focuses on a description
of the simulation along with a comparison between algorithms of determina-
tion of fatigue life by means of the cycle counting method and the spectral
method. An emphasis in the course of calculations is placed on a compari-
son between the particular algorithm blocks, such as the computed equivalent
values and the determined critical plane positions.
The experimental part of this work – Chapter 6 is devoted to fatigue test-
ing of 18G2A steel with a combination of bending with torsion under random
loading. The loading was Gaussian with a narrow-band frequency spectrum.
The spectral formula was derived for determination of fatigue life on the basis
of the Serenssen-Kogayev linear damage accumulation hypothesis. The com-
puted fatigue life was compared with experimental data. To the verification
 1 Introduction 5

of fatigue life calculation algorithm applied, experimental data was done by

Karolczuk, Nieslony, Achtelik and was conducted as a part of Ph.D. disser-
tations and papers of Department of Mechanics and Machine Design, Opole
University of Technology [47]. In Achtelik’s research, the stress history was
Gaussian with broad-band frequency spectrum where the specimens under
pure bending, pure torsion and proportional bending with torsion were tested.
2
Spectral Methods for Fatigue Description

2.1 Random Gaussian Loading

Material and structural component fatigue tests require over the test con-
ditions accurate control [10, 41, 60]. One of the methods involves measure-
ment of strain or forces occurring in actual structures during operation and
precise representation during laboratory testing. This method is, however,
labour and time consuming and is barely applicable with reference to ran-
dom non-stationary loading. Therefore, real history of loading is adapted by
the formation of standard loading, which is postulated for determination of
fatigue life of the investigated structural components. Sequences applied in
automobile industry: CARLOS [16, 87] and in construction: DABM, TR440,
BRE [105] (loading generated by wind in particular climatic zones) could be
listed here.
In a number of particular cases it is desirable to present the frequency
structure of loading or state a direct definition of loading in frequency do-
main [24, 25, 30, 63, 65, 70, 97]. The power spectral density (PSD) function
of loading is the most commonly applied one. A number of standard PSD
functions are postulated, similarly as in time domain. The issues are covered
widely by Pook [77], Bitner-Gregersen and Cramer [14] by a comparison of
several analytical PSD functions applied for the description of loading gen-
erated by the waving sea. However, an awareness of the frequency structure
does not give a sufficient description of loading. It is additionally necessary
to learn the probability distribution of instantaneous values of loading his-
tory. On the basis of observations it was remarked that a number of natural
phenomena directly affecting fatigue loading of machines and components are
normally distributed. Under this assumption the determination of amplitude
distributions in the statistical sense is possible, similar to the distribution
gained by means of the cycle counting algorithm [1]. This distribution, along
with material constants, constitutes input data for determination of life time
with hypotheses of fatigue damage accumulation. An adequate determination
8 2 Spectral Methods for Fatigue Description

of amplitude distribution1 constitutes a major problem to solve in the spectral

method of fatigue life determination.
Bishop and Sherratt [13] submit an algorithm for stress range estimation
directly from the power spectral density function. The function of range prob-
ability density, similarly as Rychlik [81, 82], is formally defined as a product of
the probability of three occurrences P1 , P2 and P3 constituting prerequisites
of determination of a cycle from random load. A graphical representation of
cycle definition is presented in Fig 2.1. k-range cycle probability density is
defined as follows:

2.0 
ip=∞
pk (k) = P1 (ip, ip − k) P2 (ip, ip − k) P3 (ip, ip − k) P (ip) , (2.1)
dk
ip=k+1

where: k – range expressed in terms of the number of classes with a de-

fined width dk,
ip – number which includes the first point of a determined cycle,
point 1 in Fig 2.1,
ip − k – class number which includes the second point of a determined
cycle, point 2 in Fig 2.1,
P1 – probability of occurrence that signal reaching point 1 crosses
the level of point 2 from below,
P2 – probability of occurrence that signal approaching point 2 from
point 1 following arbitrary trajectory must not exceed levels
of point 1 or 2,
P3 – probability of occurrence that signal approaching point 3,
which is placed above level of point 1, from point 2 must not
exceed level of point 2,
P (ip) – probability of occurrence of peak on the ip level.
It was indicated that the definition of a loading cycle leads to the establish-
ment of cycle probability density (2.1) by determination of functions of P1 ,
P2 , P3 occurrence probability and probability of P (ip) peak occurrence. In
order to determine the quantities, it is assumed that the time history is Gaus-
sian with arbitrary band frequency width. Subsequently, a two-dimensional
probability density function of adjacent extremes is applied, as postulated by
Kowalewski [38] for Gaussian processes

(xmax − xmin )
Pmin,max (xmin , xmax ) = 
4m0 I 2 2πm0 (1 − I 2 )
 2 
−xmin − x2max − 2xmin xmax (2I 2 − 1)
exp ,
8m0 I 2 (1 − I 2 )
(2.2)
1
Amplitude distribution is expressed as a probability density function of amplitude
or range for distinguished cycles of a random load history.
 2.1 Random Gaussian Loading 9

Point 3.

Point 1. dk

P3 occurrence
P1 occurrence Arbitrary transition from
Arbitrary transition point 2 to point 3.
from level below
point 2 (point 4) k
to point 1.

P2 occurrence
Arbitrary transition
from point 1 to point 2. dk

Point 2.

Point 4.
time

Fig. 2.1. Representation of cycle definition determined from random time history
by rain flow algorithm (Bishop and Sherratt [13])

where: xmin , xmax – local extremes – minimum and maximum,

∞ k
mk = Gxx (f )f df – k-th moment of power spectral density of pro-
0 cess X,
m2
I= √ – irregularity coefficient for process X.
m0 m4
In order to describe the relations between non-adjacent extremes, the
Markov chain2 is applied. On its basis, three transition matrices are devel-
oped for the determination of the probability of transition from one level to
another, from point 4 to 1, 1 to 2, 2 to 3, respectively (see Fig 2.1).
The derived range probability density function (2.1) is equivalent to the
same function determined by direct application of cycle counting by the rain
flow algorithm.
Etube et al. [23] postulate the determination of the power spectral density
function of stress in the components of a jack-up platform with a transmit-
tance function. In the course of analysis of an offshore platform this was largely
simplified by representation as a combination of a cylindrical base and a cube
as the platform. By the application of equation of forced vibration motion:

M ẍ(t) + C ẋ(t) + Kx(t) = F (t) , (2.3)

2
Application of Markov chain processes for calculation of fatigue strength is widely
covered by Sobczyk and Spencer [90] and Johannesson [34].
10 2 Spectral Methods for Fatigue Description

where: M , C, K – mass, damping, stiffness, respectively,

ẍ(t), ẋ(t), x(t) – acceleration, velocity and transition, respectively,
F (t) – time variable force
under an assumption of steady state, i.e. one in which due to damping the
transient decayed, a formula3 for spectral transmittance of the investigated
system is derived
1
H(ω) = 2
, (2.4)
(K − M ω ) + j (Cω)
which is applied for determination of power spectral density of dislocation
2
Gxx (f ) = |H(f )| Gyy (f ) , (2.5)

where: H(f ) – spectral transmittance in frequency domain,

Gxx (f ) – power spectral density of dislocation,
Gyy (f ) – power spectral density of loading (force).
In the course of structure analysis modified analytic function of Pierson-
Moskovitz power spectral density [71] is adapted
 
H 2 Tz −5 1 −4
Gyy (f ) = s 2 [(f − β)Tz ] exp − [(f − β)Tz ] , (2.6)
8π π

where: Hs , Tz – mean wave height and period – values taken from tables de-
pendent on forecasted sea state and platform position,
(f − β)– frequency diminished by correction coefficient depending on
type of platform structure.
On the basis of (2.4) it could be remarked that spectral transmittance
takes into account the effect of structure inertia, rigidity and damping and
is not dependent on loading. During comparison between function Gxx (f )
obtained experimentally from measurements of dislocation at several points
with functions determined from (2.5), a large degree of equivalence is obtained.
The huge advantage of the application of loading definition in frequency
domain is the possibility of application of finite element method for the deter-
mination of power spectral density of displacements, strain and stress taking
into consideration frequency transmittance of a system [25, 31, 63, 94].

2.1.1 Narrow-Band Frequency Stress

One of earliest papers devoted to spectral methods for determination of fatigue

strength is a paper by Miles from 1954 [59]. The author states an assump-
tion that probability density function of peak values of stress time history is
equal to amplitude probability density which could be distinguished in random
3
This operation is popular with deriving formulae applied in FEM dynamic anal-
yses [104].
 2.1 Random Gaussian Loading 11

time history. This assumption is only valid for narrow-band frequency histo-
ries with the following stages occurring subsequently: local maximum, zero
level crossing with negative inclination, local minimum and zero level cross-
ing with positive inclination,4 followed by further recurrences periodically. By
the application of Palmgren-Miner linear hypothesis of damage accumulation
and amplitudes approximation by means of Raleigh probability distribution
a formula for fatigue life is derived of the form
A
T =
, (2.7)
m m+2
M + (2m 0)
2
Γ
2

where: A = σam Nf – coefficient derived from Wöhler curve,

m  – Wöhler curve exponent,
+ m4
M = – expected number of peaks in a time unit,
m2
Γ (·) – gamma function.
The author skilfully adapted the excedance theory [80] developed by Rice
for determination of the character and parameters of amplitude probability
distribution. It needs to be stated that obtaining such simple relations was
possible only under an assumption that instantaneous values of stress history
are normally distributed over a narrow frequency band.
The paper by Perruchet and Vimont [69] is devoted to experimental ver-
ification of the Miles formula (2.7). The testing involved plain cylindrical
specimens made of AU4G1 aluminium alloys applied in aviation components.
The tests were conducted on a Schenk 100kN hydraulic machine controlled
by force. A random history was reproduced repeatedly from a magnetic tape
loop. Throughout the testing the force was recorded and was subsequently
subjected to spectral analysis.
Experimental results indicated good correlation with calculated fatigue
life. Random loading fatigue life was remarked to spread more than in the case
of cyclic loading. In the 1970s the testing conducted by Perruchet and Vimont
was considered innovative (random loading, magnetic tape data recording,
hydraulic test machine). Currently this research plays historical and cognitive
roles.
Another work devoted to experimental verification of formula (2.7) is a
paper by Clevenson and Stainer [19]. Investigations involved aluminium al-
loy 2023-T4, subject to static (tensile test) and dynamic testing for cyclic
and random tension-compression. Random loading was stationary Gaussian
with zero expected value. Testing included three PSD function forms while
the controlled parameter was standard deviation of stress. A solid specimen
with circular cross section (d = 12.7 mm in diameter) had a grooved notch
4
This type of narrow-band frequency loading is often called variable amplitude
loading.
12 2 Spectral Methods for Fatigue Description

(r = 1.57 mm in radius), which implied a theoretical stress concentration co-

efficient of 2.2. Fatigue test machine included an electromagnetic actuator. A
comparison of computed fatigue life with experimental data indicated equiva-
lent results for large stress levels. For small standard deviations of stress, the
calculations indicate exceeded life time, which is attributed to the fallibility
of the applied linear hypothesis of damage accumulation. A specific feature of
the investigated material involves a similarity of the surfaces of fatigue failure
for the case of cyclic and random loading.
The extension of the formula for fatigue life of narrow-band frequency into
cases of loading with two dominant angular frequencies ω1 and ω2 is proposed
by Sakai and Okamura [84]. The expected life time to material failure was
determined under an assumption that total damage is obtained by summation
of a component process with dominant angular frequency ω1 and the other
with dominant angular frequency ω2 with appropriate weight

21− 2 πA m1−m m1−m
m
1 0,1 0,2
T =
ω1 + ω2 , (2.8)
m + 2 ω1 + ω2 m2,1 m2,2
Γ
2

where: m0,1 , m2,1 – respective zero and second moment of power spectral den-
sity of component process with dominant angular fre-
quency ω1 ,
m0,2 , m2,2 – respective zero and second moment of power spectral den-
sity of component process with dominant angular fre-
quency ω2 .
Sakai and Okamura indicate that the model correlates well with results
gained by the cycle counting algorithm and damage accumulation, on the
condition that the two dominant angular frequencies ω1 and ω2 are sufficiently
distant. The proximity of the positions brings about interference, in which case
the assumption of weight summation of two independent processes is false.
Fu and Cebon [26] indicate that the Sakai and Okamura model does not
account for the amplitude magnification for a lower frequency process. By the
simplification of a random process with two dominant frequencies it could be
represented as a sum of two sinusoid histories with various frequencies. This
case is presented in Fig. 2.2, in which maximum amplitude of the process is
described as the sum of component parts, i.e. Amax = A1 + A2 .
This leads to a conclusion that an interaction occurs between component
histories which brings about a necessity of accounting for it in calculations.
Fu and Cebon apply the linear Palmgren-Miner hypothesis for damage accu-
mulation resulting from process y(t) = y1 (t) + y2 (t) with a formula

n1 n2 − n1 T
1 + (ω2 − ω1 ) σa 2 ] .
[ω1 σa m m
D= + = (2.9)
N1 N2 2πA
 2.1 Random Gaussian Loading 13

y(t) = A1 sin(ω1 t) + A2 sin(ω2 t)

y(t)
2(A1 + A2 ) ≈ 2A2

time t
A2 sin(ω2 t)

A1 sin(ω1 t)

Fig. 2.2. Representation of Fu and Cebon model assumptions [26]

where: n1 – cycles with amplitude σa1 = A1 + A2 ,

n2 − n1 – cycles with amplitude σa2 = A2 ,
N1 , N2 – cycles derived from Wöhler curve for stress amplitudes σa1
and σa2 , respectively.
By extension into the case in which component processes y1 (t) and y2 (t)
have a narrow-band frequency spectra, the formula is derived
⎡ ⎤
∞ ∞
T ⎣ pσa 1 (σa ) pσa 2 (σa )
D= ω1 dσa + (ω2 − ω1 ) dσa ⎦ , (2.10)
2πA σa−m σa−m
0 0

where the distribution of amplitudes σa1 = A1 + A2 is described by the

probability density function

1 −σa2
pσa1 (σa ) = exp
m0,1 m0,2 2m20,2
∞   (2.11)
 2
 σa y y2 y2
σa y − y exp − − dy ,
m0,2 2m0,1 2m0,2
0

and the distribution of amplitudes σa 2 = A2 is approximated by the Rayleigh

probability density function.
The final fatigue life formula can be derivered by the transformation of
formula (2.10) under the assumption that damage D = 1. Difficulties are
encountered associated with the separate treatment of component processes
y1 (t) and y2 (t) during the application of the formula. It is difficult to explicitly
determine the dominant frequencies and moments of the component processes
by familiarity with power spectral density of their sums.
14 2 Spectral Methods for Fatigue Description

A similar issue constitutes the focus of research conducted by Jiao in paper

[33]. A theoretical model for the determination of fatigue life is considered in
the circumstances of loading under a sum of two processes, a normal one with
narrow-band frequency and shock loading. Figure 2.3 represents a section of a
loading history in an ideal form. X1 (t) process is substituted with a sinusoidal
history, while XU1 (t) and XU2 (t) are narrow-band frequency processes with
dominant frequency higher then from X1 (t) process and damping.

X(t) X1 (t) XU 2 (t)

0
time t

XU 1 (t)

Fig. 2.3. Representation of a section of a loading history conducted by Jiao [33].

X1 (t) history for narrow-band frequency, XU 1 (t) and XU 2 (t) – shock faults

Under such assumptions, the following fatigue life formula was determined

2−m AT0 [1 − exp (−2πξm)]

T = ,
∞ T0 ∞ ∞

N0+ m z 2
r2m pR2 (r2 )dr2 + (z + αr2 ) z exp − pR2 (r2 )dzdr2 dt
μX1 2μX1
0 0 0 0
(2.12)
where: T0 – mean duration of shock load XU ,
ξ – damping coefficient,
pR2 (r2 ) – probability density of constituent ranges of shock loading.
A comparison of calculations with life time gained by means of the cycle
counting method indicate a large degree of equivalence. Such solutions are
applied in the course of fatigue life determination for ships with large dis-
placement. X1 (t) process defines stress resulting from motion while XU1 (t)
and XU2 (t) fault – from sea waves in the direction of the bow. In this case,
the application of the spectral formula enables quick fatigue analysis as a
result of sufficient knowledge of sea surface behavior [14, 24].
In the course of this spectral formula statement, a hypothesis of fatigue
damage accumulation is assumed. The linear Palmgren-Miner hypothesis is
 2.1 Random Gaussian Loading 15

applied [12, 59] along with the amplitude cycles while summing fatigue dam-
age. Liou et al. [42] postulate application of Morrow hypothesis5 [61] for ac-
counting for plastic strain work
 ni σa i
dM
D= , (2.13)
i
Nf i σmax

where: σmax – maximum amplitude in analysed stress history,

dM – exponent including the effect of plastic strain work on fatigue
damage.
dM coefficient describes sensitivity of materials to loading history (ampli-
tude distribution). In accordance with excedance theory in the case of Gaus-
sian history with narrow-band frequency the distribution of amplitudes ap-
proaches the Rayleigh distribution. The formula (2.13) could then be restated
in the following form
∞
dM
p(σa ) σa
D =N0+ T dσa
Nf (σa ) σmax
0
(2.14)
∞

N +T σam+dM +1 σ2
= 0 dM
exp − a dσa ,
Aμσ σmax 2μσ
0

where: μσ – variance of stress σ(t).

It needs to be remarked that the integral in the preceding equation is

calculated in the limit 0 to ∞, which indicates that damage accumulation
involves all the cycles with amplitudes value from 0 to +∞. In practice, how-
ever, during testing in a lab, a limit of stress amplitude σmax is determined
in order to prevent the fatigue test stand from damage. In this case all of the
values σa > σmax , which result from normally distributed stress history, are
converted into the maximum admissible ones. Upon consideration of the fact,
the damage formula takes the form

N0+ T m+dM
2
m + dM 2
σmax
D= dM
(2μσ ) Γ + 1,
Aσmax 2 2μσ
+ 
2


N0 T m−dM d M dM dM σmax
+ σ (2μσ ) 2
Γ +1 −Γ + 1, ,
A max 2 2 2μσ
(2.15)

where the first term of the equation represents damage resulting from ampli-
tudes up to the limit value (σa < σmax ) while the other one represents damage
5
The same hypothesis of fatigue damage accumulation is postulated by Corten
and Dolan [20, 95] but under different theoretical assumptions
16 2 Spectral Methods for Fatigue Description

resulting from maximum amplitude values (σa = σmax ). Liou et al. indicate
that formula (2.15) under the assumptions dM = 0 and σmax → ∞ could
be simplified, thus deriving the Miles formula (2.7). The analysed paper [42]
includes a number of remarks and formulae with practical application, which
constitute a basis for the modification of formula (2.15). One of the formulae
is
Nf = N0+ T , (2.16)
for determination of damage D in the function of cycle number Nf under
an assumption that the analysis involves narrow band frequency history. By
the approximation of amplitude by the Rayleigh distribution, the standard
deviation could be derived directly from the variance of stress history μσ
 
√ π
μσa = μσ 2 − . (2.17)
2
Following from that, formula (2.15) could be modified into
√
2

Nf μσa m+dM m + dM + 2 (1 − π/4)σmax

D= Γ ,
dM
Aσmax 1 − π/4 2 μσa
√
dM
Nf m−dM μσa
+ σ (2.18)
A max 1 − π/4

2


dM + 2 dM + 2 (1 − π/4)σmax
Γ −Γ , ,
2 2 μσa
where damage is determined on the basis of material constants A, m, dM and
two parameters associated with loading – amplitude variance μσa and max-
imum amplitude σmax . It is the most applicable solution, particularly desir-
able during fatigue testing where the quantities are the controlled parameters.
Maximum stress amplitude σmax is commonly a familiar parameter and re-
sults from maximum level on fatigue testing machine for safety purposes. If it
is unknown, it could be determined in accordance with the asymptotic theory
of extreme statistics
 
μσa  0.577
σmax = 2 ln Nf +  (2.19)
2 − π2 2 ln Nf

on condition that Nf is sufficiently large.

Experimental verification was performed on aluminium alloy 7075-T651.
The conducted fatigue testing in the conditions of constant strain amplitude
serve for the determination of fatigue characteristics. Random fatigue test-
ing is conducted under constant standard deviation of stress amplitude with
the Rayleigh distribution. The experimentally registered fatigue life was com-
pared with calculations on the basis of Eq. (2.18) and the cycle counting
method for various exponents dM = 0.0 (the Palmgren-Miner damage ac-
cumulation hypothesis) and dM = −0.25; −0.35; −0.45. The testing indicates
 2.1 Random Gaussian Loading 17

similarity of results between calculated fatigue life with experiments under the
application of the Morrow hypothesis (2.13) and exponent dM = −0.45. The
damage determined by means of the cycle counting method indicated lower
level and larger scatter in comparison to proposal (2.18). Although it is not
stated directly, the adaptation of the formula (2.18) to the spectral method
by the determination of stress history variance μσ = m0 from PSD function is
possible.

2.1.2 Broad-Band Frequency Stress

The assumption of narrow-band frequency stress history imposes considerable

restrictions on the application of the Miles formula due to the fact that ma-
jority of random loading have broad-band frequency [30, 77, 90]. A number
of proposals for the solution of the problem have been postulated.
Kowalewski [38] accounts for width frequency loading spectrum by an
introduction of irregularity coefficient I, which is determined on the basis
of power spectral density moments. The following formula for fatigue life is
proposed:
A
T =
. (2.20)
m m+2
I m M + (2m0 ) 2 Γ
2
Rajcher [79] postulates another algorithm of the determination of the ex-
pected number of cycles (peaks) in a time unit M + . His formula for life time
calculation takes the form:
A
T =   m2
, (2.21)
∞ 2 m m+2
2
G0 (f )f m df (2m0 ) Γ
0 2

G(f )
where: G0 (f ) = – normalised power spectral density function of stress.
m0

It could be easily remarked that the Rajcher formula for determination

of the expected number of peaks in time unit accounts for coefficient m of
the relation describing the Wöhler curve. This leads to the correction of the
parameter by accounting for characteristics (σa − Nf ) and results from Ra-
jcher conclusions based on experiment. Below is an suitable transformation of
formula for determination of the expected number of zero level transitions in
a time unit N0+ and Rajcher postulate which reveals their similarity
18 2 Spectral Methods for Fatigue Description
 ⎡ ∞ ⎤ 12
 ∞ ⎡∞ ⎤ 12
  2 2

 G(f )f df ⎢ G(f )f df ⎥
m2  0
N0+ = =  ∞ =⎢ 0
⎣ ∞
⎥ = ⎣ G0 (f )f 2 df ⎦ ,
⎦
m0 
G(f )df G(f )df 0
0 0

(2.22)
⎡ ∞ ⎤ m2
⎡ ⎤ m2 2
∞ G(f )f m df
⎢0 ⎥
⎣ G0 (f )f m df ⎦ =⎢ ⎥
2

⎣ ∞ ⎦ . (2.23)
0 G(f )df
0

Bolotin [15] states a similar solution postulating the replacement of the

expected number of peaks M + with the expected number of zero level transi-
tions with a positive slope N0+ . For narrow-band frequency stress history the
fatigue life is identical with the Miles formula (2.7). The Bolotin postulate is
additionally less susceptible to the noise included in the analysed stress his-
tory due to the application of lower order moments (m0 and m2 ). Similar to
Rajcher, Bolotin includes normalised power spectral density function in the
final formula
A
T =   12
. (2.24)
∞ m m+2
G0 (f )f 2 df (2m0 ) 2
Γ
0 2

Rice postulates the peaks probability density function of Gaussian history

with broad-band frequency in his paper [80],

   2

1 −x2 x x −x
p(x) = √  exp + I 1 + erf  exp ,
2πμ 2μ2 2μ −2
μ(2I − 2) 2μ
(2.25)
where: μ √ – variance of Gaussian random history,
 = 1 − I 2 – spectrum width parameter,
later quoted by Sobczyk and Spencer [90].
This distribution is applied by Lü and Jiao [44] and Chow and Li [18]
for determination of weighted average amplitude σaw in accordance with the
Palmgren-Miner linear hypothesis of damage accumulation
 m+2

 m1
  m+1 I m+2
σaw = 2μ √ Γ + Γ + IZ , (2.26)
2 π 2 2 2

where
 2.1 Random Gaussian Loading 19
∞  x m+1      
x 2 x
Z= erf(x) exp − d , (2.27)
I I I
0
x
2
erf(x) = √ exp(−t2 )dt . (2.28)
π
0

In order to simplify the calculations, in the further part of Chow and Li

paper [18], the transcendental function erf(x) called error function is replaced
with Maclaurin series. A similar approach is adopted by Lü and Jiao [44]
by replacement of the error function with two series, which results in gain-
ing better equivalence of results with approximated function erf(x) than the
calculations for the case of Maclaurin series.
The applicability of weighted average amplitude determined in this way for
strength calculation6 is doubted due to the fact that mean values of component
cycles are not accounted for. This course is criticised by Dowling [22] as the
various cycle counting methods are verified. The papers [18, 44] do not include
experimental verification of the postulated solutions.
Wirsching and Light [103] postulate a coefficient λ for correction of fatigue
life T derived from formula (2.24)
T
TBB = , (2.29)
λ
to obtain the fatigue life TBB for broad-band frequency loading.
This coefficient constitutes an empirical function related to spectrum width
parameter  and the Wöhler curve exponent m

λ(m, ) = a(m) + [1 − a(m)](1 − )b(m) , (2.30)

where a(m) = 0.926 − 0.033m and b(m) = 1.587m − 2.323 are experimentally
determined functions.
Determination of the final form of coefficient λ was conducted under fol-
lowing assumptions:
• considerations are devoted to high cycle fatigue regime of material de-
scribed with relation (σa − Nf ),
• linear damage accumulation in accordance with the Palmgren-Miner hy-
pothesis is assumed,
• stress amplitudes are determined by means of the rain flow algorithm,
• simulation is performed for 4 peculiar power spectral densities.
In the latter part of the paper [103], a sample calculation involves fatigue
life determination of the components of a rig. In this case non stationary
random loading was described in terms of 11 stationary states accounted for
6
Application of peak distribution of stress could be compared with the peak count-
ing method described in ASTM standard [1].
20 2 Spectral Methods for Fatigue Description

by analytic functions of power spectral density [11, 14]. Familiar with fraction
ti for each of 11 states fatigue life is determined

11
Ti
TBB = . (2.31)
λ t
i=1 i i

Larsen and Lutes [40, 45] postulate a spectral formula for determination
of damage for broad-band frequency loading in the form:
3m

2 2 m+1 m
2
D= Γ m2/m , (2.32)
2πA 2

where
∞
m
m2/m = f 2 G(f )df . (2.33)
0

It is called the single moment spectral method due to the characteristic

application of single moment of PSD function m2/m . In paper [40], a number of
simulations are performed for the purpose of making comparison with other
formulae for damage determination in accordance with proposals by Miles
(2.7), Wirsching and Light (2.29), Oritz and Chen [67]
3m
m−1
2
1
2 2 m+1 m
2
m2 m42
D= Γ m2/m   m2 , (2.34)
2πA 2 m2+2/m

where
∞
f 2+ 2 G(f )df .
m
m2+2/m = (2.35)
0

The analysis involves fatigue life determined on the basis of three various
single mode and two double mode analytic PSD functions. In order to com-
pare the calculated fatigue life with life gained by cycle counting method, the
stress history is generated directly from PSD function. In a majority of cases
more equivalent results are registered by the comparison of the single mo-
ment spectral method (2.32) and the cycle counting method in time domain
for damage determination than for any other models. An advantage of the
single moment spectral method is the simple formula and short calculations.
Chaudhury and Dover [17] remark that the generalised form of probabil-
ity density distribution of peaks (2.25) could be simplified to the Rayleigh
distribution for narrow-band frequency history (while  → 0) and to nor-
mal distribution for broad-band frequency history (while  → 1). Under an
assumption of the linear Palmgren-Miner damage accumulation hypothesis
weighted average amplitudes for this two extremes take the following form
 2.1 Random Gaussian Loading 21

 1
 m+2 m
σaw N B = 2 2μ Γ , (2.36)
2

 1
 1 m+1 m
σaw BB = 2 2μ √ Γ , (2.37)
2 π 2

where: σaw N B , σaw BB – weighted average amplitudes for narrow- and broad-
band frequency stress histories, respectively.
It is indicated that the weighted average amplitude calculated for the dis-
tinguished cycles from various stress histories resulting from the schemati-
sation by the rain flow and range counting algorithms is situated between
σaw N B and σaw BB extremes. Therefore it is postulated that the calculated
fatigue life should be derived from sums of both weighted average amplitudes
with the particular weight relative to frequency spectrum width. The weights
of particular components are determined on the basis of the Monte Carlo
simulation method with an assumption that fatigue life is equal to the life
determined by the damage accumulation of cycles determined through the
rain flow algorithm
A
T =   m  m+2 m + 1
3I m + 2
 , (2.38)
ti Mi+ 2 2m0 i √ Γ + Γ
i
2 π 2 4 2 i

where: ti – time fraction for i-th component process,

Mi+ – expected number of peaks in a time unit for i-th process,
m0i – zero moment (variance) of i-th process.
The general form provides for the determination of fatigue life for Gaussian
loading with arbitrary width of the frequency band. Additionally, similar as in
[103] if stress history displays non stationary character the authors postulate
the separation of the components stationary processes and individual treating
them. The final formula then constitutes a sum of the component damages.
Chaudhury and Dover solution [17] is criticised by Tunna [100], who re-
marks that mean values of component cycles of broad-band frequency stress
histories, are not accounted for. A coefficient for compensation for the in-
consistency is postulated. It is familiar [52, 48] that in many cases the mean
values of component cycles could be ignored in the process of accumulation,
if the expected value of total stress history σm ≈ 0.
Rychlik [83] proves applicability of the following inequality by assumption
of the Gaussian random history in the course of analysis

e [DRC ] ≤ e [DRF C ] ≤ e [DN B ] , (2.39)

where the particular expected damage is determined on the basis of amplitude

distribution derived from the range pair algorithm e [DRC ], rain flow algorithm
e [DRF C ] and approximation by the Rayleigh probability distribution e [DN B ].
22 2 Spectral Methods for Fatigue Description

Inequality (2.39) is applied by Benasciutti and Tovo [4, 5] who postulate

determination of e [DRF C ] as weighted sum of damages e [DN B ] end e [DRC ]

e [DRF C ] = WD e [DN B ] + (1 − WD )e [DRC ] . (2.40)

Benasciutti and Tovo in papers [4, 5, 99] present two postulations for
coefficient WD determination

α1 − α2
WD = min ,1 , (2.41)
1 − α1
 
(α1 − α2 ) 1.112 [1 + α1 α2 − (α1 + α2 )] e2.11α2 + (α1 − α2 )
WD = 2 , (2.42)
(α2 − 1)
mk
where: αk = √ for k = 1, 2 – coefficient describing the shape of PSD
m0 m(2·k)
function, for k = 2, α2 = I.
The expected damage e [DN B ] was determined by transformation of fatigue
life formula (2.7) postulated by Miles [59] to the following form

To + m m+2
e [DN B ] = M (2m0 ) 2 Γ , (2.43)
A 2

while e [DRC ] was derived from the Madsen formula [57]

∼ m−1 To m+2
m
+
e [DRC ] = I 2
M (2m0 ) Γ = I m−1 e [DN B ] . (2.44)
A 2

By the comparison of formula (2.43) with (2.44), the difference between

them is visible, which results from the addition of expression I m−1 to formula
(2.44). This fact is applied for transformation of the formula (2.40) to the
simplified form
!
e [DRF C ] = e [DN B ] WD + (1 − WD )I m−1 . (2.45)

It could be remarked that determination of damage under loading with

the arbitrary width of frequency band is reduced to the determination of
damage by the assumption of the Rayleigh amplitude distribution e [DN B ],
calculations of irregularity coefficient I and adoption of adequate weight co-
efficient WD .
The postulated thought Benasciutti and Tovo [4] model of damage deter-
mination is verified on the basis of two simulations [5]. The first involves a
comparison between the damage calculated under the postulated model (2.45)
with two models gained from literature – Wirsching [103] and Dirlik [11, 21],
which are often cited and considered as relatively accurate with regard to
damage determined by the cycle counting method with application of the
rain flow algorithm. In the course of comparative study the Gaussian random
loading is generated with irregularity coefficient I = 0.1. It is indicated that
 2.1 Random Gaussian Loading 23

the largest equivalence of results between the range counting method and
postulated formula (2.45) is observed for weight coefficient WD in accordance
with (2.42).
The other simulation involves determination of the fatigue life of a sim-
plified discrete model of a vehicle by consideration of wheels and suspension
separately. Roughness of the road is accounted for by modelling it as stochas-
tic process with the known power spectral density function. Finite parameter
discrete model enables derivation of the transition function defining the rela-
tion between loading and force in wheels and suspension. The functions are
applied for determination of the force history and power spectral density func-
tion directly on the basis of the applied loading. The calculation of fatigue life
presented in charts [5] by means of the spectral method and range counting
method indicate large equivalence.
On the basis of the conducted simulations a conclusion that postulated
method of fatigue life determination by means of spectral method is the most
accurate one could be drawn. The results correlate best with fatigue life gained
by cycle damage accumulation determined by the rain flow algorithm. How-
ever, as it was developed only recently it was not verified and must be ap-
proached very cautiously. It offers better results in comparison to the Dirlik
postulate [21]; however, the probability distribution of amplitude postulated
by Dirlik provide for the application of other damage accumulation hypothe-
ses in the course of fatigue life determination, which is not possible under
Benasciutti and Tovo postulations.
The issue of structure life time under wind loading is introduced by Holmes
[32]. He postulates determination of a simple formula accounting for the mean
wind velocity Û . The parameter is related to standard deviation of stress in
the investigated structural components
√
μσ = K Û n , (2.46)

where: K – coefficient determined on the basis of static structure analysis,

n – exponent relative to resonance response of a system.
The mean wind velocity Û constitutes a random variable. Its probability
density is described with the Weibull distribution as
   
Û −1 Û
pÛ (Û ) = exp − , (2.47)
c c

where:  – shape coefficient,

c – scale coefficient.
The application of the distribution for the approximation of real time data
it is remarked that values of shape coefficient  are approximately equal to 2.
In this case the Weibull distribution assumes a specific shape of the Rayleigh
distribution. By the assumption of this distribution of mean wind velocity,
24 2 Spectral Methods for Fatigue Description

the lower and higher limits Tlower and Thigher of the expected fatigue lifes are
determined. The lower limit is determined for narrow-band frequency loading
and higher – for broad-band loading
A
Tlower =
√ m m  mn + k
, (2.48)
+ mn
N0 2K c Γ +1 Γ
2 k

Tlower
Tupper = , (2.49)
λ
where: λ – empirical coefficient accounting for broad-band frequency loading,
following [103].
It is remarkable that the above formulae account for probability distribu-
tion parameters of the mean wind velocity Û , which accelerates simulation
of structures under various service conditions. Holmes in paper [32] does not
target the issue of the effect of mean stress on fatigue life. The postulated
solution accounting for wind direction as the weighted sum of damage from
all directions is under scrutiny


N
D= P (θi ) Di , (2.50)
i=1

where: P (θi ) – probability of mean wind from i-th direction.

Totally different assumptions are stated by Tovo [98] who postulates, sim-
ilar as Nagode and Fajdiga [64], the approximation of cycles determined by
the rain flow schematisation with multimodal Weibull distribution


β −1 
β 
k k
βi Δσ i Δσ i
p(Δσ) = pi (Δσ) = wi exp − , (2.51)
i=1 i=1
ϑi ϑi ϑi

where: k – number of component distributions named characteristic

forms,
wi – i-th weighted coefficient,
βi , ϑi – shape and scale coefficients for i-th characteristic form.
It is assumed that damage generated by a single cycle with a given range
Δσ constitutes a parameter for the determination of accuracy of cycle distri-
bution approximation. It is remarked that the postulated distribution matches
well enough for k = 1. Nagode and Fajdiga gain similar results for compo-
nent distributions k = 4. The numerical solution of a system of equations is
required for determination of coefficients β1 and ϑ1
 2.1 Random Gaussian Loading 25
⎧
⎪   &S
⎪
⎪ n+1 Ni Δσin+1
⎪
⎪ Γ + 1
⎪
⎪ ϑ1 
β1
 = i=1
⎪
⎪ &
⎪
⎪ Γ n
+ 1
S
⎨ β1 Ni Δσin
i=1 , (2.52)
⎪
⎪   &S
⎪
⎪ n+2
Γ β1 + 1
n+2
Ni Δσi
⎪
⎪
⎪
⎪ ϑ2   = i=1
⎪
⎪ 1 Γ n +1 &S
⎪
⎩ β1 Ni Δσin
i=1

where: S – number of stress range levels Δσ,

Ni – number of cycles with stress range Δσi ,
n = 3 – constant.
Weighted coefficient w1 in this case is equal to

&
S
Ni Δσin
1
w1 = i=1
 , (2.53)
NT ot ϑn1 Γ βn1 + 1

where: NT ot – sum of cycles determined by means of the rain flow algorithm

from random history.
The approximation of the amplitude or range distribution from random
loading constitutes a common technique of the formal recording and defini-
tion of parameters of fatigue life determination algorithms. A valuable sample
of this approach is represented by papers by Sobczykiewicz [91] and Oziem-
ski [68]. Service histories of excavator component loading are registered and
analysed. A linear combination of two theoretical probability distributions,
exponent distribution and extreme distribution for description of stress am-
plitude distribution is postulated.
Szala [95] put together a compilation of formulae of fatigue life calcu-
lation with the assumption of various amplitude distribution and damage
accumulation hypotheses in a book devoted to fatigue damage accumulation
hypotheses. The total number of the programmed load cycles to fatigue failure
is determined for Rayleigh and exponent distributions and right leg normal
distribution in the combination with the Palmgren-Miner and Haibach linear
damage accumulation hypotheses. The form of loading spectrum frequency is
accounted for by means of an adequate spectrum with parameter.
Unfortunately, the discussed postulations by Tovo, Nagode, Sobczykiewicz,
Oziemski and Szala find no direct application with regard to spectral methods
despite good approximation of amplitude probability distribution. The prob-
lem of the direct determination of parameters from stress history and range
counting in probability distributions is encountered. The parameter cannot
be determined if the input data only include power spectral density function.
26 2 Spectral Methods for Fatigue Description

Dirlik [21] postulates the approximation of amplitude distribution gained

by rain flow algorithm by means of an empirical formula derived from the
Monte Carlo simulation method. It is assumed that the zero moment m0 of
stress power spectral density function constitutes the most relevant parameter.
On the basis of 17 representative stress histories and their power spectral
density functions the standard error between the assumed model and real
amplitude distribution is minimised. The derived probability density function
of ranges of stress cycles takes the form
 
1 G1 −Z G2 Z −Z22 −Z 2
p(Δσ) = √ e Q + 2 e 2R + G3 Ze 2 , (2.54)
2 m0 Q R

Δσ 2(xm − I 2 )
where: Z= √ , G1 = ,
2 m0 1 + I2
1 − I − G1 + G21
G2 = , G3 = 1 − G1 − G2 ,
1−R
I − xm − G21 1.25(I − G3 + G2 R)
R= , Q= ,
1 − I − G1 + G21 G1

1
m1 m2 2 m2
xm = , I=√ .
m0 m4 m0 m4

According to authors [11, 12, 60, 62] more accurate amplitude distribution
approximation by rain flow algorithm [1, 81] is gained by application of this
formula in comparison to other approximations, and it can be applied for stress
histories with narrow- and broad-band frequency spectra. Unfortunately, the
complicated form of the function does not enable derivation of a simple life
time formula
1
T = ∞ (2.55)
p(Δσ)
M+ dΔσ
N (Δσ)
0

and the above integral must be calculated numerically.

Morrill et al. [60] apply probability distribution developed by Dirlik (2.54)
for the procedure of the determination of characteristic PSD function applied
in fatigue testing machines. The paper undertakes the issue of precise mapping
of service conditions during fatigue testing with the concurrent assumption
of their short duration. The decisive parameter during determination of char-
acteristic PSD functions for a specific loading type is damage calculated on
the basis of amplitude probability distribution (2.54) and the adoption of the
Palmgren-Miner hypothesis of damage accumulation. The functions are deter-
mined on the basis of histories of acceleration registered with car suspension
components during road tests in various conditions. The following relation is
 2.1 Random Gaussian Loading 27

applied
C
Gσ (fi ) = Gẍ (fi ) , (2.56)
fi4
where: Gσ (fi ) – discrete stress PSD function value for fi frequency,
Gẍ (fi ) – discrete acceleration PSD function value for fi frequency,
C – scalar constant,
in order to derive power spectral density of stress from spectral density of
acceleration. The authors postulate determination of constant C under the
finite element method or calibration method by measurement of acceleration
and strain in the investigated structural points. In the particular cases the
analytical determination of constant C would be difficult.
The registered acceleration histories subjected to the analysis are non-
stationary. By the statistical analysis of history sections, stationary sections
are distinguished and divided into groups with regard to statistical similarity.
Mean acceleration and variance variation in time are analysed. PSD func-
tion is determined for the distinguished sections. Weighted mean values are
calculated for individual groups
k 
 
Gi (f )Ti
Ĝ(f ) = , (2.57)
i=1
TΣ

&
k
where: TΣ = Ti – total time for a single group,
i=1
Ĝ(f ) – mean PSD function for single group,
Gi (f ) – PSD function for i-th section,
k – total number of stationary sections distinguished within
a single group,
Ti – length of i-th section,
which results in a characteristic PSD function for a group, Ĝ(f ). After de-
termination of power spectral density of stress Gσ (f ) on the basis of (2.56)
damage is determined on the basis of Dirlik postulate (2.54). The determined
damage and fundamental statistical parameters for distinguished stationary
history groups constitute the foundation for distinguishing groups for fatigue
lab testing. For the currently analysed example loading history the method
contributes to the reduction of testing duration from 1300 to 20 hours. In order
to trace the complex process of determination of specific PSD the following
algorithm steps are conducted:
• registration of loading (acceleration history),
• selection of data block length in seconds for statistical analysis,
• statistical analysis of particular blocks,
• definition of a finite group number with various statistical parameters,
• attribution of blocks with similar statistical parameters into appropriate
groups,
28 2 Spectral Methods for Fatigue Description

• determination of weighted mean PSD function in accordance with (2.57),

• conversion of weighted mean PSD acceleration function into stress PSD
function in accordance with (2.56),
• determination of PSD moments and damage by means of the Dirlik formula
for each group,
• selection of distinctive groups taking damage, mean value and variance of
stress.

2.2 Random Non-Gaussian Loads

The literature devoted to random non-Gaussian loads in spectral methods is
scarce. One of the first ones is Liu and Hu paper [43], which is devoted to the
adaptation of damage DN B determined with the narrow-band frequency for-
mula (2.7) by accounting for coefficients λ and L. By reference to Wirsching
and Light paper [103], the parameter λ (2.30) is applied in order to revise
the recent definition of damage by accounting for frequency band width. Pa-
rameter L is introduced in order to account for deviation from Gaussian load
history. It is determined in the course of simulations by the Monte Carlo
method under following assumption

D [Y (t), m]
L(κ, m) = , (2.58)
D [X(t, κ), m]

where: D[·] – damage determined by means of the rain flow algorithm and
the linear Palmgren-Miner hypothesis,
Y (t) – stress history with non-Gaussian probability distribution,
X(t, κ) – Gaussian stress history determined by non-linear transforma-
tion of Y (t).
Finally, non-Gaussian loading damage with broad-band frequency spec-
trum is derived from the formula

D = λLDN B . (2.59)

Sarkani et al. [85, 86] remark that a number of random natural processes
are non-Gaussian and adopt a standard fourth central moment parameter
named kurtosis κσ for deviation from normal probability distribution history
(Fig. 2.4)
!
e (σ − σ̂)4
κσ = , (2.60)
μ2σ
where: e [·] – expected value of an expression,
σ, σ̂ – stress and mean stress, respectively,
μσ – variance of the stress.
 2.2 Random Non-Gaussian Loads 29

1
κY < 3
0.9 κY = 3
κY > 3
0.8

0.7

0.6
pY (Y )

0.5

0.4

0.3

0.2

0.1

0
−3 −2 −1 0 1 2 3
Y = Z(X)

Fig. 2.4. Variation of probability density function shape for various non-Gaussian
histories Y (t). Loading histories are derived by means of transformation function
Z (2.62). Coefficients β and n are incorporated in function β(n) = n + 2 for n =
(0; 0.2; 0.4; ...; 2) [85, 86]

In general, non-Gaussian history could be determined by means of non-

linear transformation of Gaussian history
Y (t) = Z [X(t)] , (2.61)
where: Y (t) – non-Gaussian history,
X(t) – Gaussian history,
Z[·] – non-linear transformation function.
A number of non-linear functions Z[·] could be postulated for performing
the transformation. Sarkani et al. [85, 86] postulated it in the form:
X + β [sgn(X)|X|n ]
Y = Z(X) = , (2.62)
C
which includes coefficients β and n to account for deviation from Gaussian
history. Constant C is introduced to normalise history Y and ensure mean
square values of Y and Gaussian history X to be equal. By means of the
linear Palmgren-Miner hypothesis, a damage formula for non-Gaussian history
is derived for an input kurtosis κσ , for which parameters β, n and C are
determined
30 2 Spectral Methods for Fatigue Description
∞

−σa2
m
σa (σa + βσan )
Dng = exp dσa , (2.63)
C m Aμσ 2μσ
0

where: A = σam N – fatigue characteristic,

μσ – variance of the stress history.
A comparison between the postulated theoretical solution and an experi-
ment was made. A welded joint under alternating bending loading was tested
for fatigue. The histories applied for investigation were determined by means
of transformation function (2.62) for three selected kurtosis values κσ = 3,
2 and 5. Four tests were performed for each of five various stress levels. The
comparison of calculations with experiment proves its accuracy and indicates
good applicability of the postulated calculation model for non-Gaussian loads
with narrow-band frequency spectrum.
Similar to Liu and Hu [43], Winterstein [24, 102] postulates application of
a coefficient for narrow-band frequency histories whose distribution is different
than Gaussian
m (m − 1) (κσ − 3)
LW = 1 + (2.64)
24
for correction of damage DN B determined by means of narrow-band frequency
formula (2.7). It needs to be remarked that coefficient LW takes the value 1
for Gaussian processes (κσ = 3). The final fatigue life formula takes the form

A
T =
 . (2.65)
m m+2 m (m − 1) (κσ − 3)
M + (2m 0)
2
Γ 1+
2 24

Lachowicz et al. [39] tested 10HNAP steel under broad-band frequency

loading and non-Gaussian distribution. Fatigue life calculation is performed in
time domain by application of the Palmgren-Miner hypothesis with coefficient
aP M = 0.5, which takes account of amplitudes below fatigue limit
⎧ ni
⎪
⎪
m for σai ≥ aP M σaf
k ⎨⎪ σaf
NG
D(To ) = σai , (2.66)
⎪
i=1 ⎪
⎪
⎩
0 for σai < aP M σaf

where: D(To ) – damage determined for observation time To ,

ni – cycle number with amplitude σai ,
σaf – fatigue limit.
Component amplitudes are determined by the rain flow algorithm. Con-
currently, calculations by means of spectral method in frequency domain are
performed upon the selection of three formulae from literature: Rajcher, Miles,
 2.3 Multiaxial Loading 31

Kowalewski. Due to non-Gaussian loading, a coefficient LL is applied for cor-

rection of this deviation, similar as in Liu and Hu [43]. The coefficient is
relative to kurtosis and is defined as follows
1
LL = . (2.67)
|1 + (κσ − 3)|

In the course of comparison of experimental data with calculation ones the

best equivalence of results was obtained for rain flow algorithm and Rajcher
fatigue life formula (2.21) including a correction coefficient LL .

2.3 Multiaxial Loading

The literature devoted to application of spectral methods with reference to
multiaxial random loading is scarce.
Kam and Dover [35] tackle the problem in the course of developing re-
search methodology for pipe joints, typical of drilling rig structures. The rigs
applied for crude oil mining on the shallow sea, are subjected to changeable
loading resulting from wind and sea waves. The loading was thoroughly in-
vestigated and is most commonly described in terms of a number of power
spectral density empirical functions [11, 14, 17, 32], whose shape is relative
to weather and wavy motion of sea surface. A generation of loading history is
developed for the purpose of fatigue testing. The Markov chain is applied for
the description of transitions from a single characteristic state to another7 .
The generator operates in a loop performing the following tasks:
• determination of power spectral density of stress for a particular charac-
teristic state; subsequent characteristic states are combined into Markov
chain,
• calculation of transmittance function for each of the states,
• calculation of filter function,
• random signal generation by withe noise filtration with normal probability
distribution.
A T-joint consisting of a major carrying pipe and an arm of smaller diam-
eter was tested. The arm with the pipe forms right angle. The carrying pipe
is fixed at both ends to the testing stand while loading is applied by means
of three hydraulic cylinders whose ends are attached to the arm of the tested
joint. The loading along centers of the arm with bending in two directions con-
stitute the combined fatigue loading. The actual measurements of structure
indicated considerable differences between this three service loadings (PSD
function shapes were different), although they where considerably correlated.
7
The issue of application of the Markov chain for separation of a non-stationary
process into a series of stationary subprocesses in order to count cycles is covered
by Ph.D. thesis by Johannesson [34].
32 2 Spectral Methods for Fatigue Description

The histories of loads with such properties are gained by the appropriate selec-
tion of phase shift of component harmonics during generation. Unfortunately,
the paper neither includes the results of fatigue life nor a postulate or proposal
of component life time calculations.
Leser et al. [41] postulate another solution of the problem of multiaxial
loading modeling for fatigue testing. It is assumed that loading history con-
stitutes a superposition of a stationary random process with the zero expected
value and random variables affecting variance and mean value of the modeled
process
x = x̂ + sx̄ , (2.68)
T
where: x = [x1 , . . . , xn ] – modeled n-dimensional random process,
T
x̂ = ⎡
[x̂1 , . . . , x̂n ] ⎤ – random variable affecting mean value x,
s11 · · · s1n
⎢ .. . . .. ⎥
s=⎣ . . . ⎦ – symmetrical matrix of n × n dimension for scaling
sn1 · · · snn x with regard to variance of particular dimension,
T
x̄ = [x̄1 , . . . , x̄n] – n-dimensional random stationary process with the
zero expected value.
In order to minimise the number of the required parameters for deter-
ministic description of mean value vector x̂ and matrix s the functions are
represented as finite Fourier series. It is possible due to the small variability
of the quantities in comparison to the dynamics of random variable x̄ for which
modelling ARMA (Auto Regressive Moving Average) model is applied [89].
With reference to the spectral methods Leser et al. remark that the awareness
of PSD function for components of multidimensional random variable x̄ and
coefficients from the finite Fourier series (x̂ and s modelling) is sufficient for
the adequate mapping of a non-stationary process.
An algorithm of fatigue life determination under multiaxial random load-
ing is first postulated by Macha [56]. It constitutes an extension of the postu-
lates of Miles [59], Kowalewski [38], Rajcher [79] and Bollotin [15]. It consists
in the application of power spectral density of the equivalent stress for deter-
mination of statistical parameters involved in the familiar fatigue life formulae.
PSD for the equivalent stress is determined by the application of the linear
criteria of multiaxial fatigue failure in frequency domain. As the criteria are
based on the notion of a critical plane, three methods of determination of its
location are distinguished:
• weight functions method [54, 55],
• variance method [9],
• damage accumulation method [51, 53].
Despite the fact that none of the above methods is presented from the
spectral point of view, a possibility is open for the instance of two latter ones.
The paper by Macha [56] could be considered typically theoretical, as it is
distinguished with a detailed presentation of the part devoted to the stress
 2.3 Multiaxial Loading 33

criteria of multiaxial random fatigue and their definition in frequency domain.

It is observable that Macha maintains the possibility of applying the method
for fatigue life calculations by means of strain criteria of multiaxial fatigue
failure.
Preumont and Piéfort [78] introduce a method for fatigue life determina-
tion under plain stress state by the application of equivalent stress in accor-
dance with the Huber-Mises-Hencky strength hypothesis
2 2 2 2
σeq = σxx + σyy − σxx σyy + 3σxy , (2.69)

where: σxx , σyy and σxy – components of stress tensor.

T
After defining vector σ = (σxx , σyy , σxy ) , the formula (2.69) could be
restated in accordance with the principles of matrix calculations
2
 
σeq = σ T Qσ = Trace Q[σσ T ] , (2.70)
⎡ ⎤
1 −0.5 0
where: Q = ⎣ −0.5 1 0 ⎦ – matrix of coefficients for Huber-Mises-
0 0 3 Hencky hypothesis under plane stress state,
σT – vector transposed to σ,
Trace{·} – sum of components of main diagonal of
square matrix.
On the basis of formula (2.70) the relation for expected values could be
stated !  !
2
e σeq = Trace Qe σσ T . (2.71)
Finally a formula for mean square value of equivalent stress is derived.
Mean square value could be determined directly from power spectral density
of equivalent stress
∞ ∞
2
!
e σeq = Geq (f )df = Trace {QGσσ (f )} df , (2.72)
0 0

where: Gσσ (f ) – matrix of autospectral and cross-spectral density functions

of stress vector σ.
On the basis of the preceding formulae the authors postulate a method
for determination of power spectral density function of the equivalent stress
directly from a matrix of spectral density function of stress vector σ

Geq (f ) = Trace {QGσσ (f )} = Qij Gσi σj (f ) , (2.73)
i,j

where: Qij – ij indexed coefficient of Q matrix,

Gσi σj (f ) – power autospectral density (i = j) or cross-spectral density
(i = j) functions of vector σ components.
34 2 Spectral Methods for Fatigue Description

It must be remarked that some values of coefficient matrix Q are equal

to zero. It is remarked that in the course of determination of power spectral
density function for equivalent stress (2.69), on the basis of formula (2.73) in
paper [78], the interaction between the components of stress σxx and σxy , and
also σyy and σxy is not taken into account, which constitutes a fault of the
method.
An extension of the above approach is presented by Potoiset et al. [72, 73,
75, 76] in the statement of power spectral density function of the equivalent
stress by the application of Matake and Crossland criteria. The criteria are
based on concept of a critical plane. Matake makes an assumption that a
critical plane is one for which the shear stress amplitude τaη is the largest.
The amplitude is defined as the radius of smallest circle circumscribing vector
for the shear stress τaη (t) in a plane with normal unit vector η̄ (see Fig. 2.5).

v
τaη = R

τaη (t)

τη (t)

τmη

Fig. 2.5. Determination of radius R of circle circumscribing vector for shear stress
τaη (t) on a plane with normal unit vector η̄ (based on [75])

The criterion based on the assumption takes the form

2σaf
τaη + − 1 (σmη + σaη )
τaf
≤ 1, (2.74)
τaf
 2.3 Multiaxial Loading 35

where: τaη – maximum shear stress amplitude in critical plane,

σaf , τaf – fatigue limit for tension-compression and torsion, respec-
tively,
σmη , σaη – mean value and amplitude of normal stress to critical plane,
respectively.
Preumont et al. solution is applied in a work by Sun end Wang [94] for
determination of fatigue life by means of finite element method under plane
stress state. The novelty involves the application of a linearized plate com-
ponent for determination of stress state components. The linear shape was
determined under an assumption that modal response of the analysed struc-
ture is Gaussian. Unfortunately, the work does not include an experimental
verification of the postulated model.
Grzelak et al. [28] conduct spectral analysis of selected multiaxial fatigue
failure criteria. Simulations involve the generation of full Gaussian stress ten-
sor components with broad-band frequency spectrum (maximum frequency
fmaxij = 160 Hz) and subsequently the application of strength criteria and
comparison of statistical parameters of equivalent stress history. Nine criteria
of multiaxial fatigue failure are analysed:
• criterion of maximum normal stress in a critical plane,
• criterion of maximum normal strain in a critical plane,
• criterion of maximum shear stress in a critical plane,
• criterion of maximum principal stress (Galileo hypothesis),
• criterion of largest normal stress (Ranky hypothesis),
• criterion of largest normal strain (Saint-Venant hypothesis),
• criterion of largest shear stress (Coulomb-Tresca-Guest hypothesis),
• Huber-Mises-Hencky hypothesis,
• Beltrame hypothesis.
The following conclusions are postulated:
1. Mathematical model of the multiaxial fatigue failure criterion stated as a
linear combination of stress tensor components allows to determine sta-
tistical parameters applicable during the estimation of life time,
2. Equivalent stresses gained by the application of linear criteria have the
same probability distribution function as stress tensor components applied
during simulations,
3. Stress tensor frequency band is retained in equivalent stress history only
for the case of the linear criteria.
The paper demonstrates that the determination of power spectral density
function of the equivalent stress for the case of linear multiaxial fatigue failure
criteria is possible directly form the matrix of autospectral and cross-spectral
density functions of stress tensor components.
36 2 Spectral Methods for Fatigue Description

Similar conclusions are drawn by L agoda and Macha [46]. The analysis fo-
cuses on frequency variations of equivalent histories gained by the application
of various multiaxial fatigue failure criteria. It is remarked that only for the
case of linear criteria the frequency characteristics of equivalent stresses are
not altered.
3
Theoretical Fundamentals

3.1 Description of Random Stress and Strain States

As mechanical loadings are observed it is remarked that they are often com-
bined of random quantities. Examples include wind pressure exerted on slen-
der building structures, hitting waves against obstacles, seismic vibrations, etc.
In order to formulate a mathematical model of such phenomena, the theory
of random functions is applied. Random functions including time parameter
t, which are named stochastic processes, are common. A set of a number n of
processes constitutes a vectorial, n-dimensional stochastic process
X(t) = [X1 (t), X2 (t), . . . , Xn (t)] . (3.1)
Under the general notion of loading, it induces random state of stress and
strain in a structural component. It could be described in terms of stress
tensor σ(t) and strain tensor ε(t), respectively, variable in time:
⎡ ⎤
σxx (t) σxy (t) σxz (t)
σ(t) = ⎣ σyx (t) σyy (t) σyz (t) ⎦ , (3.2)
σzx (t) σzy (t) σzz (t)
⎡ ⎤
εxx (t) εxy (t) εxz (t)
ε(t) = ⎣ εyx (t) εyy (t) εyz (t) ⎦ . (3.3)
εzx (t) εzy (t) εzz (t)
Under the assumption of material isotropy, stress tensor and strain tensor
are symmetrical matrices, in which
σxy (t) = σyx (t), σxz (t) = σzx (t), σzy (t) = σyz (t), (3.4)
εxy (t) = εyx (t), εxz (t) = εzx (t), εzy (t) = εyz (t) . (3.5)
It is remarked that under the above assumptions tensors (3.2) and (3.3)
could be described with application of their six components. Hence, six-
dimensional vectorial stochastic processes for stress and strain states can be
derived in the form
38 3 Theoretical Fundamentals

Xk (t) = σij (t) , (k = 1, . . . , 6; i, j = x, y, z) , (3.6)

Xk (t) = εij (t) , (k = 1, . . . , 6; i, j = x, y, z) , (3.7)

or

σ(t) = [σxx (t), σyy (t), σzz (t), σxy (t), σxz (t), σyz (t)] , (3.8)
ε(t) = [εxx (t), εyy (t), εzz (t), εxy (t), εxz (t), εyz (t)] . (3.9)

They describe stress and strain in materials on the basis of specific com-
ponent tensors related to the Cartesian co-ordinate system Oxyz.
In many applications random processes, whose probabilistic characteristics
are not variable under time axis shift, are encountered. Such phenomena are
modeled by means of so-called stationary stochastic processes. A stochastic
process X(t), t ∈ T , is strongly stationary or stationary in a strict sense if all
possible statistical distributions of component processes are independent of
an arbitrary time shift τ , provided that ti + τ ∈ T . For the current instance,
the relation of equality

pt1 , t2 , ..., tn (x1 , x2 , . . . , xn ) = pt1 +τ, t2 +τ, ..., tn +τ (x1 , x2 , . . . , xn , τ ) (3.10)

must be satisfied.
In practice it is difficult to verify if a physical process is stationary in a
strict sense, because the relation of equality (3.10) must be satisfied for each
component vector process and arbitrary time shift τ . Therefore, a wider class
of stationary processes is introduced for simplification purposes. A stochastic
process X(t), t ∈ T , is said to be weakly stationary or stationary in a wide
sense if its mean (expected) values are time invariant and all elements of cor-
relation matrix R(t1 , t2 ) depend only on the time difference τ = t2 − t1 , i.e.
R(t1 , t1 + τ ) = R(τ ). From the definition it follows that in order to consider a
vectorial stochastic process stationary, the fulfilment of the condition of sta-
tionary component content is not sufficient. Apart from that, the component
processes must be stationary mutually correlated.
If a random process X(t) is stationary and mean values X̂k and autocor-
relation functions RXk (τ ) of component processes Xk have equal statistical
properties, a multidimensional random process is named ergodic. For an er-
godic random process, the mean value and autocorrelation function of single
history are equal to the appropriate mean values in a set of histories, so
X̂k = X̂ and RXk (τ ) = RX (τ ). It must be remarked that only stationary
processes can be ergodic. In practice, random processes representing station-
ary physical phenomena are commonly ergodic. Due to this, in a majority of
instances, characteristics of a stationary random process could be sufficiently
determined on the basis of a single realisation. In the monograph it was as-
sumed that vectors of stress (3.8) and strain (3.9) are stationary and ergodic
processes.
In the correlation theory [6, 7], a stationary and ergodic vectorial process
(3.1) is described by means of a vector of expected values x̂ = [x̂1 , x̂2 , . . . , x̂6 ]
 3.1 Description of Random Stress and Strain States 39

and a matrix of correlation function Rx (τ ) or a covariance function μx (τ ),

where τ = t2 − t1 .
Under the assumption that a vectorial process X(t) is a six-dimensional
stationary and ergodic Gaussian process the joint probability density function
takes the form
 
1 1 −1
px1 ,...,x6 (x1 , . . . , x6 , τ ) = ' exp − x μx (τ ) x T
, (3.11)
6 2
(2π) |μx (τ )|
⎡ ⎤
μx11 (τ ) · · · μx16 (τ )
⎢ .. .. .. ⎥
where: μx (τ ) = ⎣ . . . ⎦ – covariance matrix of random vari-
μx (τ ) · · · μx (τ ) ables X1 , . . . , X6 ,
61 66
x = [x1 − x̂1 , . . . , x6 − x̂6 ] – row vector for variables x1 , . . . , x6
and expected values x̂1 , . . . , x̂6 ,
xT – x transposed vector,
|μx (τ )| – determinant of covariance matrix.
The awareness of the distribution of probability density of instantaneous
values and stationary character of random history does not give sufficient in-
formation about a stochastic process. However, for the case of stochastic pro-
cesses there is an instrument of spectral analysis parallel to harmonic analyses
of deterministic functions. The analysis is based on the power spectral den-
sity function and offers a possibility of gaining additional information about
frequency structure of a stochastic process.
For one-dimensional stochastic process X a two-sided power spectral den-
sity function SX (f ) is defined as Fourier transform of autocorrelation func-
tion1
∞
SX (f ) = RX (τ )e−j2πf τ dτ , (3.12)
−∞

where: SX (f ) – two-sided power spectral density function defined for fre-

quency range (−∞, +∞),
RX (τ ) – autocorrelation function.
Due to the difficulties in the interpretation of negative frequency, one-sided
power spectral density function GX (f ) is applied, in which the argument f
varies within a limit (0, +∞)

2SX (f ) for 0 ≤ f < ∞ ,
GX (f ) = . (3.13)
0 for f < 0 ,
However, in mathematical calculations the application of function SX (f ),
defined for the range (−∞, +∞) with an imaginary exponent in an integrand,
1
By replacement of exponent of Euler constant e in Fourier transform from
(−j2πf τ ) into (−jωτ ), two-sided power spectral density defined in angular fre-
quency domain SX (ω) is obtained, where SX (f ) = 2πSX (ω).
40 3 Theoretical Fundamentals

simplifies the analysis. It is important to apply both function forms appropri-

ately.
For the investigated example of a six-dimensional stochastic process (3.8)
or (3.9), one-sided power spectral density function takes the form of the Her-
mitian matrix G(f ), 6 × 6 of dimension
⎡ ⎤
G11 (f ) · · · G16 (f )
⎢ .. .. .. ⎥
G(f ) = ⎣ . . . ⎦. (3.14)
G61 (f ) · · · G66 (f )

The functions Gkl (f ) are defined for frequency f ≥ 0 and are equal to the
double value of two-sided power spectral density Skl (f )

2Skl (f ) for 0 ≤ f < ∞ ,
Gkl (f ) = (k, l = 1, . . . , 6) , (3.15)
0 for f < 0 ,

where: Gkk (f ), Skk (f ) – autospectral density functions of component pro-

cesses Xk (t),
Gkl (f ), Skl (f ) – cross-spectral density functions between component
processes Xk (t) and Xl (t).
Taking into consideration the fact that functions of power spectral density
are complex functions

Gkl (f ) = Re[Gkl (f )] + i Im[Gkl (f )] , (3.16)

where: Re[Gkl (f )] – coincident spectral density function, a real part of Gkl (f ),

Im[Gkl (f )] – quadrature spectral density function, an imaginary part
√ of Gkl (f ),
i = −1 – imaginary unit.
Since for the Hermitian matrices the relation Gkl (f ) = Re[Glk (f )] −
i Im[Glk (f )] is fulfilled, in order to characterise frequency structure of a ran-
dom stress or strain tensors the knowledge of 21 power spectral density func-
tions is required.

3.2 Multiaxial Fatigue Failure Criteria

Numerous criteria of multiaxial fatigue failure introduced in [53, 54] are based
on an assumption that material damage is induced by a combination of stress
and strain components operating on the critical plane. On the basis of the
assumptions, general function of material fatigue strength under multiaxial
random loading could be defined

S(t) = {Dij (t), Pn , Ck } , (3.17)

 3.2 Multiaxial Fatigue Failure Criteria 41

where: Dij (t) – components of stress or strain tensor; stochastic processes,

Pn – parameters for determination of critical plane position,
Ck – parameters characterizing a material.
The surface of limit state determining fatigue life under random multiaxial
state of stress (or strain) is described by the maximum value of strength
function
max {S(t)} , (3.18)
t

which is equivalent to fatigue strength of material under uniaxial cyclic load-

ing. On the basis of [9, 53, 54], two general multiaxial fatigue failure criteria
could be distinguished:

1. Generalized criterion of maximum normal and shear stresses in the critical

plane for long life time,
2. Generalized criterion of maximum normal and shear strains in the critical
plane for ling and short life times.
Assumptions for the two criteria can be written as:
1. Fatigue damage depends directly on normal stress ση (t) (normal strain
εη (t)) and shear stress τηs (t) (shear strain εηs (t)) in direction s̄ in the
critical plane with normal vector η̄;
2. Direction s̄ is consistent with the mean direction of maximum shear stress
max {τηs (t)} (maximum shear strain max {εηs (t)}) in the critical plane;
s s
3. In the limit state the maximum value of strength function (3.18), which is
a linear combination of stresses ση (t) and τηs (t) (strains εη (t) and εηs (t))
satisfies the following equations:
– in accordance with stress criterion

max {Bτηs (t) + Kση (t)} = F , (3.19)

t

– in accordance with strain criterion

max {bεηs (t) + kεη (t)} = q , (3.20)

t

where constants B and b are applied for the selection of particular criterion
form. Constants K, F , k and q refer to fatigue characteristics of materials
and are gained from uniaxial cyclic tests. The positions of unit vectors η̄
and s̄ are determined from mean directional cosines of principal stress or
principal strain axes ln , mn , nn , (n = 1, 2, 3).

At the time of the analysis of experimental results under multiaxial cyclic

and random loading it was observed that for the case of brittle materials the
fracture plane is perpendicular to normal stress with the highest amplitude
or variance. For plastic materials, the fracture plane takes one of two posi-
tions for which shear stresses reach maximum amplitude or variance [55]. By
42 3 Theoretical Fundamentals

selection of adequate constants B, K, F and b, k, q and specifying the critical

plane position, the particular forms of stress (3.19) and strain criteria (3.20)
are derived. The following multiaxial fatigue failure criteria could be distin-
guished and give the resulting formulae of equivalent damage parameters, i.e.
equivalent stress and strain [53, 54].
1. Criterion of maximum normal stress in the critical plane,
(B = 0, K = 1) – the critical plane is determined by the mean position
of the maximum principal stress σ1 (t)

σeq (t) = l12 σxx (t) + m21 σyy (t) + n21 σzz (t)
(3.21)
+ 2l1 m1 σxy (t) + 2l1 n1 σxz (t) + 2m1 n1 σyz (t) .

2. Criterion of maximum shear stress in the critical plane,

(B = 1, K = 0) – the critical plane is determined by the mean position
of one of two planes in which the maximum shear stress τ1 (t) occurs
     
σeq (t) = l12 − l32 σxx (t) + m21 − m23 σyy (t) + n21 − n23 σzz (t)
+ 2 (l1 m1 − l3 m3 ) σxy (t) + 2 (l1 n1 − l3 n3 ) σxz (t) (3.22)
+ 2 (m1 n1 − m3 n3 ) σyz (t) .

3. Criterion of maximum normal and shear stresses in the critical plane,

(B = 1, K = 0) – the critical plane is determined by the mean position
of one of two planes in which the maximum shear stress τ1 (t) acts
(
1  2 )
σeq (t) = l12 − l32 + K l12 + l32 σxx (t)
1+K
(  2 )
+ m21 − m23 + K m21 + m23 σyy (t)
(  2 )
+ n21 − n23 + K n21 + n23 σzz (t) (3.23)

+ 2 [l1 m1 − l3 m3 + K (l1 + l3 ) (m1 + m3 )] σxy (t)

+ 2 [l1 n1 − l3 n3 + K (l1 + l3 ) (n1 + n3 )] σxz (t)

+ 2 [m1 n1 − m3 n3 + K (m1 + m3 ) (n1 + n3 )] σyz (t) .

4. Criterion of maximum normal strain in the critical plane,

(b = 0, k = 1) – the critical plane is determined by the mean position of
the maximum principal strain ε1 (t)

εeq (t) = l12 εxx (t) + m21 εyy (t) + n21 εzz (t)
(3.24)
+ 2l1 m1 εxy (t) + 2l1 n1 εxz (t) + 2m1 n1 εyz (t) .

5. Criterion of maximum shear strain in the critical plane,

(b = 1, k = 0) – the critical plane is determined by the mean position of
one of two planes in which the maximum shear strain γ1 (t) occurs
 3.2 Multiaxial Fatigue Failure Criteria 43
1 2     
εeq (t) = l1 − l32 εxx (t) + m21 − m23 εyy (t) + n21 − n23 εzz (t)
1+ν
+ 2 (l1 m1 − l3 m3 ) εxy (t) + 2 (l1 n1 − l3 n3 ) εxz (t)
!
+ 2 (m1 n1 − m3 n3 ) εyz (t) ,
(3.25)

where: ν – Poisson’s ratio.

6. Criterion of maximum normal and shear strains in the critical plane,

(b = 1, k = 1) – the critical plane is determined by the mean position of
one of two planes in which the maximum shear strain γ1 (t) acts

εeq (t) = l1 (l1 + l3 ) εxx (t) + m1 (m1 + m3 ) εyy (t) + n1 (n1 + n3 ) εzz (t)
+ [l1 (2m1 + m3 ) + l3 m1 ] εxy (t) + [l1 (2n1 + n3 ) + l3 n1 ] εxz (t)
+ [m1 (2n1 + n3 ) + m3 n1 ] εyz (t) .
(3.26)

The linear form of the quoted multiaxial fatigue failure criteria enable
presentation of equivalent stress and strain in a more generalised form. Row
vector of coefficients is defined as

a = [a1 , . . . , a6 ] . (3.27)

Equivalent history could be determined by the sum of the products of

suitable component tensors xk (t) and coefficients ak dependent on the selected
criterion and the position of the critical plane
6

xeq (t) = ak xk (t) , (3.28)
k=1

where: ak – suitable criterion dependent coefficients (Table 3.1),

xk (t) – components of random stress or strain tensor.
The conclusion that equivalent history has the same type of probability
density function as the stochastic process (tensor) from which it was deter-
mined results from the above linear formula. In the case of Gaussian vector
process X(t) (3.11), the equivalent history has normal distribution expressed
with probability density function in the form:
 
1 − (xeq − x̂eq )2
pxeq (xeq ) =  exp , (3.29)
2πμxeq 2μxeq

where the expected value of the process is derived as follows

6

x̂eq = ak x̂k (3.30)
k=1
44 3 Theoretical Fundamentals

Table 3.1. Components of coefficient vector a for three particular stress and strain
related criteria.
Stress criteria
a (3.21) (3.22) (3.23)
2  (  2 )
a1 l12 l1 − l32 1
1+K
l12 − l32 + K l12 + l32
 2  (  2 )
a2 m21 m1 − m23 1
1+K
m21 − m23 + K m21 + m23
 2  (  2 )
a3 n21 n1 − n23 1
1+K
n21 − n23 + K n21 + n23
2
a4 2l1 m1 2 (l1 m1 − l3 m3 ) 1+K
[l1 m1 − l3 m3 + K (l1 + l3 ) (m1 + m3 )]
2
a5 2l1 n1 2 (l1 n1 − l3 n3 ) 1+K
[l1 n1 − l3 n3 + K (l1 + l3 ) (n1 + n3 )]
2
a6 2m1 n1 2 (m1 n1 − m3 n3 ) 1+K
[m1 n1 − m3 n3 + K (m1 + m3 ) (n1 + n3 )]
Strain criteria
a (3.24) (3.25) (3.26)
2 
a1 l12 1
1+ν
l1 − l32 l1 (l1 + l3 )
 2 
a2 m21 1
1+ν
m1 − m23 m1 (m1 + m3 )
 2 
a3 n21 1
1+ν
n1 − n23 n1 (n1 + n3 )
1
a4 2l1 m1 1+ν
(l1 m1 − l3 m3 ) l1 (2m1 + m3 ) + l3 m1
1
a5 2l1 n1 1+ν
(l1 n1 − l3 n3 ) l1 (2n1 + n3 ) + l3 n1
1
a6 2m1 n1 1+ν
(m1 n1 − m3 n3 ) m1 (2n1 + n3 ) + m3 n1

and variance
6
6  6

  
k−1
μxeq = ak al μxkl = a2k μxkk + 2 ak al μxkl . (3.31)
k=1 l=1 k=1 l=1

It results from formula (3.30) that if the expected values of the components
of process X(t) are equal to zero (x̂k = 0; k = 1, . . . , 6), the expected value
of the equivalent process x̂eq = 0. It must be remarked that the occurrence
of zero expected value in the equivalent history with zero expected values of
components of process X(t) constitutes an important result from the physical
point of view. The result is not obtained if multiaxial loading is reduced to
an uniaxial one with the application of non-linear fatigue failure criteria (e.g.
Huber-Mises-Hencky or Tresca strength hypotheses) [28].

3.3 Power Spectral Density Function

of Equivalent History
Frequency structure of the equivalent history of the damage parameter, which
is defined as a function of power spectral density G(f ) [6, 7], plays an impor-
 3.3 Power Spectral Density Function of Equivalent History 45

tant role during calculations of fatigue life. For the case of multiaxial loading,
power spectral density of the equivalent stress or strain Gxeq (f ) should be
determined on the basis of suitable criteria of fatigue failure. The simplest
technique consists in determination of PSD from the equivalent stress or strain
history xeq (t) by means of a numerical method. The objective of the current
chapter is to derive it directly from power spectral density matrices (3.14) by
the application of the theory of linear multi-input systems [6, 7].
In the investigated example, the equivalent stress or strain xeq (t) could be
interpreted as output signal from a six input physical system (see Fig. 3.1).
The system has six input signals xk (t) representing suitable tensor components
σij (t) or εij (t).

x1 (t) h1 (τ ) y1 (t)

x2 (t) h2 (τ ) y2 (t)

xeq (t)

x6 (t) h6 (τ ) y6 (t)

Fig. 3.1. Interpretation of damage parameter xeq (t) as output signal from a physical
system with impulse transfer functions hk (τ ), k = 1, . . . , 6 with input feed of signals
xk (t), k = 1, . . . , 6

For the investigated system the output signal xeq (t) is derived under for-
mula
6 ∞

xeq (t) = hk (τ )xk (t − τ )dτ , (3.32)
k=1 0

and its autocorrelation function

Rxeq (τ ) = e [xeq (t)xeq (t + τ )] . (3.33)
Power spectral density function Gxeq (f ) is determined with Fourier trans-
form of autocorrelation function Rxeq (τ ) [6, 7]
6 
 6
Gxeq (f ) = Hk∗ (f )Hl (f )Gkl (f ) , (3.34)
k=1 l=1
46 3 Theoretical Fundamentals

where: Hl (f ) – spectral transmittance function for l-th input (Fourier trans-

form of impulse transfer function hl (τ )),
Hk∗ (f ) – Hk (f ) coupled function.
For the investigated criteria of multiaxial fatigue failure, the spectral trans-
mittance of the system Hk (f ) for the particular input signals do not depend
on frequency f and are equal to the constant coefficients, i.e. Hk (f ) = ak .
Hence, power spectral density of the equivalent history Gxeq (f ), under an as-
sumption that random components of process X(t) are correlated, could be
derived from the formula
6 
 6
Gxeq (f ) = ak al Gkl (f )
k=1 l=1
6 6
(3.35)
 
= a2k Gkk (f ) +2 ak al Re [Gkl (f )] .
k=1 k<l

The above formula enables determination of power spectral density func-

tion of the equivalent history directly from the components of the power spec-
tral density matrix (3.14) by the reduction of the spatial state to the equivalent
uniaxial state in frequency domain [28]. It could be remarked that the form of
the formula (3.35) is similar to formula (3.31) of equivalent history variance.
It could have been anticipated, as PSD function reflects mean square values
of the harmonic components of the signal.
The formerly derived formulae could be restated more briefly by means of
a matrix notation, which enables result presentation in a more comprehensible
form. By the application of coefficient row vector a (3.27), covariance matrix
μx , and PSD matrix Gx (f ) (3.14) the formulae (3.31) and (3.35) can be
presented in the form
μxeq = aμx aT , (3.36)
Gxeq (f ) = aGx (f )aT . (3.37)
It should be remarked that in the course of determination of power spectral
density function of the equivalent history, auto- and cross-PSD function of
matrix Gx (f ), which is a complex matrix, is applied. As a consequence of
the summation (3.37), which results from matrix operations, the imaginary
parts of matrix component functions Gkl (f ) and Glk (f ) are subtracted and
the result takes the form of PSD of the equivalent history Gxeq (f ) in real
number domain. This is a property of Hermitian matrices [37] as illustrated
in the example.
Example: For a Hermitian matrix X (2 × 2 of dimension) and vector b (1 × 2 of
dimension), matrix operations according to (3.37) are performed

Z = bXbT .

Matrix X and vector b are defined as follows

 3.4 Amplitude Distribution in Spectral Methods 47
 
x11 x12
X= ,
x21 x22
!
b = b1 b2

Consequently the form is obtained

  
! x11 x12 b1
Z = bXbT = b1 b2 =
x21 x22 b2
 
! b1
b1 x11 + b2 x21 b1 x12 + b2 x22 =
b2
(b1 x11 + b2 x21 ) b1 + (b1 x12 + b2 x22 ) b2 =
b21 x11 + b1 b2 x21 + b2 b1 x12 + b22 x22 =
b21 x11 + b22 x22 + b1 b2 (x21 + x12 ) .

The sum (x21 + x12 ) from the preceding formula is the sum of components of Hermi-
tian matrix for which the relation ixkl = −ixlk is satisfied. The imaginary parts are
reduced, which provides for the restatement of the presented example of equation
as:
Z = bXbT = b21 x11 + b22 x22 + 2b1 b2 Re (x12 ) .

3.4 Amplitude Distribution in Spectral Methods

Service loading histories of machines are most commonly complicated random
histories. Under such circumstances the application of cycle counting methods
for the determination of material life time leads to the schematisation of ran-
dom loading. The procedure of schematisation offers a possibility of counting
amplitudes and mean values of cycles from random time history. Subsequently,
in the process of damage accumulation, the cycles are considered to result
from cyclic loading. The sequence-bound structure of the schematisation of
random histories makes it a considerable share of computational time. For
the reason, suitable theoretical probability distributions are often applied for
the description of loading peaks (local extremes), mean values and amplitudes
distributions [13, 91, 95]. The most important factor from the point of view of
material life time is distribution of peaks, which is associated with the issue
of excedance familiar from theory of random processes.
The distribution of peaks is determined here for a one-dimensional station-
ary process represented with stress history σ(t) with expected value σ̂ = 0
and fulfilling the condition dσ(t)/dt > 0. The expected value of frequency of
level crossing Na+ is computed for level x = a where a > 0 of function x(t)
(Fig. 3.2). For the events that 0 < x < ∞ and a− ẋdt ≤ x ≤ a, where ẋ is time
derivative, the function x(t) crosses level a from downwards at an instant τ ,
(for the case of upward sloping function). It is assumed that joint probability
density function p(x, ẋ) is defined. The probability of both events is following
48 3 Theoretical Fundamentals

x(t) upward sloping function x(t)

downward sloping function x(t)

0
τ τ + dt t

Fig. 3.2. A section of random history x(t)

∞ a
p(x, ẋ)dxdẋ . (3.38)
0 a−ẋdt

Under the above assumptions a formula for mean number of level x = a

crossings Na+ in time unit is derived, first quoted by Rice in 1945 [80]
∞
Na+ = ẋp(a, ẋ)dẋ . (3.39)
0

Under the assumption that loading is a random history x(t) with normal
probability distribution of instantaneous values, the distribution of the mean
number of a level crossing in second could be defined.
The function of probability density of a normal process takes the form

1 x2
p(x) = √ exp − . (3.40)
2πμx 2μx

It can be remarked that all derivatives of a normal random function have

also normal probability distributions:

1 ẋ2
p(ẋ) = √ exp − (3.41)
2πμẋ 2μẋ

and

1 ẍ2
p(ẍ) = √ exp − . (3.42)
2πμẍ 2μẍ
 3.4 Amplitude Distribution in Spectral Methods 49

Joint probability density function can be derived from the relationship



1 1 x2 ẋ2
p(x, ẋ) = √ exp − + . (3.43)
2π μx μẋ 2 μx μẋ

The variances of the random history x(t) and its first ẋ(t) and second ẍ(t)
time derivatives are included in the formulae above. They are derived from
power spectral density function Gx (f ):
∞ ∞ ∞
2
μx = Gx (f )df, μẋ = Gx (f )f df, μẍ = Gx (f )f 4 df . (3.44)
0 0 0

Upon consideration of formula for k-th moment of power spectral density

function
∞
mk = Gx (f )f k df , (3.45)
0

the expressions from (3.44) take the form:

μx = m0 , μẋ = m2 , μẍ = m4 . (3.46)

By substitution of (3.43) into (3.39) the mean number of crossings in a

time unit is derived for normal random history
√

μẋ a2
Na+ = √ exp − . (3.47)
μx 2μx

On the basis of (3.47) the approximated mean number of oscillations N0+

(number of zero level crossing in secound) can be determined for narrow-band
frequency history. To this end, the level must be assumed to take the value
a=0  
+ μẋ m2
N0 = = . (3.48)
μx m0
The distribution function of peaks (local extremes) is defined under for-
mula
N+
P (a) = P (x < a) = 1 − a+ for a > 0 (3.49)
N0
and, subsequently, probability density function

dP (a) 1 dNa+
p(a) = =− + for a > 0. (3.50)
da N0 da

By differentiation of the relation (3.47) in the formula above, the proba-

bility density function of peaks is defined as
50 3 Theoretical Fundamentals

a a2
p(a) = exp − . (3.51)
μx 2μx
The resulting distribution takes the form of the familiar Rayleigh distribu-
tion. It can be remarked that thus an analytical formula (3.51) is derived for
the description of amplitude distribution for the considered case. There is no
need to apply algorithmic methods of schematisation of random histories. The
advantages of the procedure include simple and condensed formula shape, ap-
plicability in the analytical calculation for fatigue life determination and the
reduction of computational time in comparison to various algorithms of the
cycle counting methods. The drawback consists in the limited applicability
due to the assumption that normal probability distribution of loading type
with narrow-band frequency spectrum is under consideration.

3.5 Spectral Method of Fatigue Life Determination

The characteristic feature of the considered spectral method of fatigue life
determination is application of the power spectral density function of stress
or strain for estimation of amplitude probability distribution. It should be re-
membered that the PSD function does not sufficiently define stress or strain
history and the information about the shape of probability distribution of the
instantaneous values is indispensable. In the course of deriving life time for-
mula, a hypothesis of the accumulation of fatigue damage is applied along with
standard fatigue characteristics of materials, similar as in the cycle counting
methods. The following subchapters present the assumptions and the pro-
cess of spectral formulae derivation. The formulae apply two various fatigue
characteristics, (σa − Nf ), which relates fatigue life to stress amplitude and
(εa − Nf ), which relates fatigue life to strain amplitude.

3.5.1 Fatigue Life Calculation Based on Characteristics (σa – Nf )

For cyclic loading with constant amplitude, damage is calculated from the
formula
n
D= , (3.52)
Nf
where: n – number of stress cycles with constant amplitude,
Nf – number of cycles to failure derived from Wöhler curve (σa − Nf ).
For variable amplitude loading the formula (3.52) takes the form


k k
ni
D= Di = , (3.53)
i=1 i=1
N fi

where: k – number of stress amplitude levels.

 3.5 Spectral Method of Fatigue Life Determination 51

For the case of random histories it is difficult to determine explicitly the

number of cycles with specific amplitude. For the purpose cycle counting algo-
rithms, such as rain flow, range pairs, or histeresis loop are applied; however,
they cannot be directly applied in the spectral methods. In order to overcome
the problem it is assumed that random history is an ergodic stationary history
with narrow-band frequency spectrum. For this case, local extrema (peaks)
could be considered as amplitudes of history components2 . If probability of
the occurrence of peaks is denoted as P (σa ), ni is the number of cycles with
amplitude σa i during observation time To , the formula for ni takes the form

ni = n(σa ) = M + To P (σa ), (3.54)

where: M + – expected number of peaks in a time unit.

By the application of formula (3.54) and replacing the summation from

(3.53) with an integral in the range 0 to ∞, the following expression for damage
in observation time To is derived


k ∞
+ p(σa )
D(To ) = Di = M T o dσa , (3.55)
i=1
Nf (σa )
0

where: p(σa ) – probability density function of peaks of stress his-

tory σ(t),
Nf (σa ) = Aσa−m – number of cycles derived from the Wöhler curve.
By consideration of the fact that material fatigue life T expressed in sec-
onds corresponds to the damage D(T ) = 1 from the proportion:
1 T
= , (3.56)
D(To ) To
the formula is derived
To 1
T = = ∞ . (3.57)
D(To ) p(σa )
M+ dσa
Nf (σa )
0

If probability density function of peaks p(σa ) is described in terms of the

Rayleigh distribution

σa σa2
p(σa ) = exp − , (3.58)
μσ 2μσ
the integral expression from (3.57) takes the form:
2
Algorithmic ranges method described in ASTM standard [1] is equivalent with
this approach.
52 3 Theoretical Fundamentals

σa −σa2
∞ ∞ exp ∞

p(σa ) μσ 2μσ 1 σa −σa2

dσa = dσa = exp σam dσa ,
Nf (σa ) Aσa−m A μσ 2μσ
0 0 0
(3.59)
where: μσ = m0 – variation of stress history σ(t).

By analysis of the expression we can remark that the mathematical struc-

ture is similar to gamma function, also known as the Eulerian integral of
second order [37]
∞
Γ (z) = e−t tz−1 dt . (3.60)
0

The integral can be replaced with the gamma function after mathematical
operations involving substitution of variables

σa2 1 σa
t= , σa = (t2μσ ) 2 , dt = dσa (3.61)
2μσ μσ
hence,
∞ m ∞
1 −t
m (2μσ ) 2 m+2
e (t2μσ ) 2
dt = e−t t 2 −1 dt . (3.62)
A A
0 0

Subsequently, the integral (3.59) is equal to

∞ m

p (σa ) (2μσ ) 2 m+2

dσa = Γ . (3.63)
Nf (σa ) A 2
0

By referring back to fatigue life formula (3.57), the known Miles for-
mula [59] is derived

1 A
T = =
. (3.64)
∞ m m+2
p(σa ) M + (2μσ ) 2
Γ
M+ dσa 2
Nf (σa )
0

For stress history with narrow-band frequency spectrum the expected num-
ber of peaks per second M + is equal to the number of zero level crossings with
positive slope N0+ or the dominant frequency f0 , i.e.

+ + m2
M = N0 = = f0 . (3.65)
m0

The fact is often applied by the authors of papers [38, 103] for the deter-
mination of the expected number of cycles per time unit on the basis of the
 3.5 Spectral Method of Fatigue Life Determination 53

expected number zero level crossings, since the parameter is less sensitive to
the noise occurring in the analysed signal.
During the derivation of fatigue life formula (3.64) it is assumed that ampli-
tudes with ranges (0, +∞) participate in the process of damage accumulation.
Damage accumulation is often applied only for amplitudes above a specified
value in accordance with the conviction that amplitudes below some limit do
not affect material damage considerably. The limit is associated with fatigue
limit σaf , and accumulation of damage is conducted in accordance with the
modified Palmgren-Miner hypothesis. The hypothesis is illustrated in Fig. 3.3.

1
A m
σa σa =
Nf

σa max

σaf

p(σa )
aP M σaf

Nf
p(σa )

Fig. 3.3. Damage accumulation in accordance with modified Palmgren-Miner hy-

pothesis. The bold line marks a section of fatigue characteristics participating in the
process of damage accumulation

Damage for the postulated hypothesis can be derived from formula (2.66)
in the cycle counting method. In the spectral method, the alteration of the
lower limit of integration in formula (3.55) must be accounted for during
derivation
k ∞
+ p(σa )
D(To ) = Di = M T o dσa , (3.66)
i=1
N f (σa )
aP M σaf

where: aP M – coefficient accounting for amplitudes below fatigue limit σaf .

By means of further transformations with substitutions from (3.61), the

final life time formula is derived in the form
54 3 Theoretical Fundamentals

A
T =  , (3.67)
m m + 2 a2P M σaf
2
M + (2μσ ) 2
Γ ,
2 2μσ

where
∞
Γ (z, a) = e−t tz−1 dt (3.68)
a

is the incomplete gamma function [37]. It can be remarked that the formula
(3.67) is a restatement of Miles formula with coefficient aP M = 0.
Haibach [29] postulates accounting for amplitudes below the fatigue limit
σaf by the modification of the Wöhler curve exponent during damage accu-
mulation (Fig. 3.4)
⎧ ni
⎪
⎪
m for σai ≥ σaf
⎪
⎪ σaf
⎪
⎪ NG
k ⎪⎨ σai
D(To ) = , (3.69)
⎪
i=1 ⎪
⎪ ni
⎪
⎪
mH for σai < σaf
⎪
⎪ σaf
⎩ NG
σai

where: D(To ) – damage determined for observation time To ,

ni – number of cycles with amplitude σai ,
σaf – fatigue limit,
mH = 2m − 1 – modifies sloping of fatigue characteristics for ampli-
tudes below fatigue limit.
By application of the probability density distribution function of amplitude
p(σa ), the conditional sum from equation (3.69) could be replaced with the
sum of two integrals

∞ σaf
+ p(σa ) p(σa )
D(To ) = M To dσa + M + To dσa , (3.70)
Nf (σa ) Nf H (σa )
σaf 0

where: Nf H (σa ) = Aσa−mH – number of cycles form Haibach chart for ampli-
tudes below fatigue limit (Fig. 3.4).
Under the assumption that amplitudes are defined by means of the
Rayleigh distribution, the integrals from (3.70) take the simplified form:
 3.5 Spectral Method of Fatigue Life Determination 55

1
A m
σa σa =
Nf
σa max

1
1
A 2m−1
σa =
Nf
σaf

Nf
p(σa )

Fig. 3.4. Fatigue accumulation in accordance with Haibach hypothesis [29]. Bold
line 1 marks a section of Wöhler curve taking part in damage accumulation for
amplitudes σa ≥ σaf . Line 2 marks modified characteristics with slope coefficient
mH = 2m − 1 for amplitudes σa < σaf

∞ 
m
2
p(σa ) (2μσ ) 2 m + 2 σaf
dσa = Γ , , (3.71)
Nf (σa ) A 2 2μσ
σaf

σaf mH 
 
2
p(σa ) (2μσ ) 2
mH + 2 mH + 2 σaf
dσa = Γ −Γ , ,
Nf H (σa ) AH 2 2 2μσ
0
(3.72)
mH
where: AH = NG σaf – constant derived from the Haibach chart.

By the application of the Haibach linear hypothesis of damage accumu-

lation the fatigue life formula is derived by the transformation of expression
(3.70) under the assumptions that D(To ) = 1 and To = T

A
T =  +
2
m m + 2 σaf
M + (2μσ ) 2
Γ ,
2 2μσ
(3.73)
AH

  .
2
mH mH + 2 mH + 2 σaf
M + (2μσ ) 2
Γ −Γ ,
2 2 2μσ
56 3 Theoretical Fundamentals

Corten and Dolan [20, 95] postulate the application of secondary fatigue
charts with a different exponent qCD = m for fatigue computation. The hy-
pothesis is illustrated in Fig. 3.5.

1
A m
σa σa =
Nf
σa max

1
σaf

Nf max Nf
p(σa )

Fig. 3.5. Damage accumulation in accordance with Corten-Dolan hypothesis. Bold

lines mark fatigue characteristics participating in damage accumulation process.
Various slope of fatigue characteristics depending on exponent qCD (qCD < m for
line 1, qCD > m for line 2) are gained

The secondary fatigue chart has a point of contact with Wöhler chart
for stress amplitude σa = σamax . Damage accumulation is performed for the
entire amplitude spectrum (0 < σa ≤ σa max ). Damage for the hypothesis is
expressed with the formula


k
qCD
ni σai
D(To ) = , (3.74)
i=1
Nf max σa max

where: σa max , Nf max – maximum amplitude of a cycle in realisation To and

the corresponding number of cycles to failure derived
from the Wöhler curve.
In order to derive spectral damage formula under the assumption that ran-
dom stress history σ(t) has normal distribution with narrow-band frequency
spectrum and by application of Corten-Dolan hypothesis, the sum in equation
(3.74) must be replaced with an integral
 3.5 Spectral Method of Fatigue Life Determination 57
∞
qCD
p(σa ) σa
D(To ) = M + To dσa , (3.75)
Nf max σa max
0

and the Rayleigh distribution may be applied (3.58) for the description of
amplitude probability density function p(σa )

σa σ2
∞ exp − a
qCD
μσ 2μσ σa
D(To ) = M + To dσa . (3.76)
Aσa−m
max σa max
0

After mathematical transformations performed similarly as for the deriva-

tion of (3.64), the damage formula takes the form
qCD

M + To (2μσ ) 2
qCD + 2
D(To ) = −m Γ . (3.77)
AσaqCD
max 2
Under an assumption of damage D(To ) = 1 for observation time equal to
fatigue life To = T , the following formula is obtained
−m
AσaqCD
T =
max
. (3.78)
+
qCD
2
qCD + 2
M (2μσ ) Γ
2
Serensen and Kogayev [51, 88, 95] postulate damage accumulation on the
basis of secondary fatigue chart. It is derived by modification of the Wöhler
curve by accounting for coefficient bSK (Fig. 3.6). After the adoption of linear
Serensen-Kogayev hypothesis, the damage formula takes the form:

1 
k
D(To ) = m
ni σai for σai ≥ aSK σaf , (3.79)
bSK A i=1

&
k
σai ti − aSK σaf
i=1
where: bSK = – Serensen-Kogayev coefficient character-
σa max − aSK σaf
izes amplitude spectrum of random
loading (applicable for bSK > 0.1,
1 &
k
σa max /σaf > 1 and σa max σai ti >
i=1
0.5),
aSK ∈ 0, . . . , 1 – coefficient accounting for amplitudes below fatigue
limited, for presentations aSK = 0.5 was adopted,
σa max – maximum amplitude of counted cycles,
ni
ti = k – frequency of levels σai occurrence for observation
&
ni time To .
i=1
58 3 Theoretical Fundamentals

1
A m
σa σa =
Nf

σa max
bSK = 1

2
σaf

1
bSK < 1
aSK σaf

Nf max Nf
p(σa )

Fig. 3.6. Serensen-Kogayev hypothesis of damage accumulation. The bold lines

represent fatigue characteristics participating in damage accumulation process (bSK
coefficient is higher for line 1 than for line 2)

By comparison of formulae (3.66) and (3.79) for damage determination

D(To ) a remark that the difference consists in coefficient bSK could be stated.
By application of the premises for the statement of damage formula (3.66),
the expression accounting for amplitudes below fatigue limit takes the form
∞
M + To p(σa )
D(To ) = dσa , (3.80)
bSK Nf (σa )
aSK σaf

∞
σa p(σa )dσa − aSK σaf
aSK σaf
bSK = . (3.81)
σa max − aSK σaf
The formula for fatigue life T and coefficient bSK could be restated in
the light of the spectral aspect under an assumption that amplitudes are
approximated by means of the Rayleigh distribution (3.51) [49]

A
T =  , (3.82)
m m + 2 a2SK σaf
2
M+ (2μσ ) 2
Γ ,
2 2μσ

where
 3.5 Spectral Method of Fatigue Life Determination 59
 
√ 3 a2SK σaf
2
a2SK σaf
2
2μσ Γ , exp − aSK σaf
2 2μσ 2μσ
bSK = . (3.83)
σa max − aSK σaf

3.5.2 Fatigue Life Calculation Based on Characteristics (εa – Nf )

The following formula is derived under the same assumptions as in the pre-
ceding subchapter, while fatigue characteristics (σa − Nf ) is now replaced
with the characteristics (εa −Nf ), as described by the Manson-Coffin-Basquin
equation, which relates the total strain amplitude to the number of cycles to
failure
σf
εa = εae + εap = (2Nf )b + εf (2Nf )c . (3.84)
E
Formula (3.57) was applied for the expression of damage as the functions
relating to strain amplitude εa
1
T = ∞ , (3.85)
p(εa )
M+ dεa
Nf (εa )
0

where: p(εa ) – probability density function of peaks of strain history ε(t)

with narrow-band frequency spectrum,
Nf (εa ) – function of cycle number on the basis of the Manson-Coffin-
Basquin equation.
The function Nf (εa ) could not be directly determined from the formula
(3.84). Therefore, fatigue life T must be computed from formula (3.85) by
means of a numerical method. In a number of cases the process of material
damage is not associated with plastic strain with reference to the large number
of cycles to failure [66]. In that case, plastic strain could be neglected in (3.84)
and the formula takes the form
σf
(2Nf )b ,
εa = εae = (3.86)
E
after the restatement of which the analytical expression of cycle number
Nf (εa ) takes the form
 1b
1 εa E
Nf (εa ) = . (3.87)
2 σf
By application of the function and deriving the formula further, similar as
in subchapter (3.5.1), the formula takes the form
1
(2με ) 2b
T = 
1b
, (3.88)
σ 1
M +2
f
Γ 1−
E 2b
60 3 Theoretical Fundamentals

where: με – variance of strain history ε(t).

Unfortunately the neglect of the plastic strain imposes limitations on the

applicability of the formula for determination of life time with reference to
low-cycle fatigue, in case of which plastic strain is larger in comparison to
elastic strain. In order to overcome the problem, it was postulated that fa-
tigue characteristics (εa − Nf ) as defined in formula (3.84) could be replaced
with a family of tangent sections to Manson-Coffin-Basquin chart in a double
logarithmic system. Fig. 3.7 presents three sample tangents and marking the
points of intersection. Under the assumptions the fatigue life formula takes
the form
1
T =  
ε , (3.89)
ns uk
p(ε a )
M+ dεa
Nf k (εa )
k=1 εlk

where: ns – number of tangent sections,

εlk , εuk – low and high limit of integration for k-th tangent,
Nf k (εa ) – function defining the number of cycles for k-th tangent sec-
tion from characteristics (εa − Nf ).

εa (Nf )

σf
εa (Nf ) = (2Nf )b + εf (2Nf )c
E
εak (Nf ) = Dk (2Nf )dk
tangency point
εa (Nf ) = εak (Nf )
εuk

εlk

Nf uk Nf sk Nf lk Nf

Fig. 3.7. Manson-Coffin-Basquin curve approximated with three tangent sections

in a double algorithmic system (details in text)
 3.5 Spectral Method of Fatigue Life Determination 61

Under the assumption that the k-th tangent to characteristics (εa − Nf )

is defined by the function

εak (Nf ) = Dk (2Nf )dk , (k = 1, . . . , ns) , (3.90)

where: Dk , dk – parameters,

the inverse function could be defined. It is applied for determination of the

number of cycles to failure in the range εlk , εuk 

d1
1 εa k
Nf k (εa ) = , εa ∈ εlk , εuk  . (3.91)
2 Dk
By consideration of the integral from formula (3.89) in the range εlk , εuk
and under the assumption that amplitude probability density function p(εa )
is described by means of the Rayleigh distribution (3.51) (random history
with normal distribution strain and with narrow-band frequency spectrum)
the formula assumes the form
εuk εuk 2

d1
p(εa ) εa −ε a Dk k
dεa = e 2με 2 dεa . (3.92)
Nf k (εa ) με εa
εlk εlk

In order to perform further transformations, the integrand expression may

be replaced by the following substitutions:
ε2a √ εa
t= , εa = t2με , dt = dεa ,
2με με (3.93)
ε2lk ε2uk
tlk = , tuk = .
2με 2με
After the substitution
εuk
d1 tuk
p(εa ) Dk k
− 2d1
dεa = 2 √ e−t t k dt . (3.94)
Nf k (εa ) 2με
εlk tlk

It could be remarked that the mathematical structure of the expression is

similar to the incomplete gamma function [37]
∞
Γ (z, a) = e−t tz−1 dt . (3.95)
a

Aware of the fact that

x2 ∞ ∞
f (x)dx = f (x)dx − f (x)dx , (x1 < x2 ) , (3.96)
x1 x1 x2
62 3 Theoretical Fundamentals

the integral in equation (3.94) is substituted with the difference of two incom-
plete gamma functions and by substitution into (3.85), the formula is restated
in the following form
1
T = ns
d1 

 . (3.97)
 D k k 1 1
2M + √ Γ 1− , tlk − Γ 1 − , tuk
2με 2dk 2dk
k=1

The simple and concise form, the possibility of easy conversion into a pro-
gram and the applicability in the determination of fatigue life with regard to
low and high cycle numbers constitute the combined advantages of the formula
(3.97). The disadvantage is the limited applicability due to the assumptions
concerning the character of random strain history. Formula (3.97) could be
generalised into the case of loading with arbitrary spectrum frequency band
by the application of correction coefficient λ (2.30), similar as in Wirsching
[103] for each tangent section.
During the application of the postulated formula the problems of adequate
selection of tangent section number Nf k (εa ) and determination of the param-
eters Dk and dk are encountered. From the simulations it results that in a
majority of instances the number of tangent sections need not exceed 50 in
order to ensure that life time by the method by formula (3.97) and by the
numerical method (3.85) are similar. Figures 3.8 to 3.11 present error result-
ing from the approximation of the Manson-Coffin-Basquin chart with various
numbers of tangent sections (k = 2, 3, 9, 50). The error is determined from
formula
Nf num − Nf k
Nf error = 100% , (3.98)
Nf num
which defines per cent of difference between the number of cycles deter-
mined by means of the numerical method (estimation accuracy Nf num (εa ) =
Nf (εa ) ± 2.2204 · 10−16 performed by means of fzero function from MATLAB
programme [58]) and the analytical application of k tangent sections charac-
teristics. It could be remarked that the estimation error is dependent on the
number of tangent sections, and does not exceed 1% for k = 50 (Fig. 3.11).
In order to present the effect of estimation error of the cycle number
on fatigue life, simulations are performed for the uniform and Rayleigh am-
plitude distributions in relation to the number of assumed tangent sections
k = 2, . . . , 100 (Fig. 3.12). The calculations apply the Palmgren-Miner hypoth-
esis of damage accumulation. It is remarked that for an arbitrary number of
tangent sections the resulting damage is higher than for the case of direct
determination by the Manson-Coffin-Basquin characteristics. This produces
safe underestimation of fatigue life. In this case the conclusion that tangent
section number k = 50 enables sufficient accuracy of calculation.
The parameters of k-th tangent section are derived from the adopted num-
ber of cycles Nf sk , which includes a point of tangency with the Manson-Coffin-
Basquinne chart (εa − Nf ). A solution to a system of equations provided with
 3.5 Spectral Method of Fatigue Life Determination 63

0
100 10

90
−1
80 10
εa (Nf )num
70
Nf error , %

−2
60 10

50
εa (Nf )k=2

εa
−3
40 10

30 Nf error

−4
20 10

10
−5
0 10
0 5 10 15 20
10 10 10 10 10
Nf
Fig. 3.8. Error of cycle number Nf determination for approximation of Manson-
Coffin-Basquin fatigue characteristics with k = 2 tangent sections

0
100 10

90

80
εa (Nf )num 10
−1

70
Nf error , %

−2
60 10
εa (Nf )k=3
50
εa

−3
40 10
Nf error
30
−4
20 10

10
−5
0 10
0 5 10 15 20
10 10 10 10 10
Nf

Fig. 3.9. Error of cycle number Nf determination for approximation of Manson-

Coffin-Basquin fatigue characteristics with k = 3 tangent sections
64 3 Theoretical Fundamentals

0
100 10

90
−1
80 εa (Nf )num ≈ εa (Nf )k=9 10

70
Nf error , %

−2
60 10

50

εa
−3
40 10
Nf error
30
−4
20 10

10
−5
0 10
0 5 10 15 20
10 10 10 10 10
Nf
Fig. 3.10. Error of cycle number Nf determination for approximation of Manson-
Coffin-Basquin fatigue characteristics with k = 9 tangent sections

0
10 10

9
−1
8 10
εa (Nf )num ≈ εa (Nf )k=50
7
Nf error , %

−2
6 10

5
εa

−3
4 10

3 Nf error
−4
2 10

1
−5
0 10
0 5 10 15 20
10 10 10 10 10
Nf
Fig. 3.11. Error of cycle number Nf determination for approximation of Manson-
Coffin-Basquin fatigue characteristics with k = 50 tangent sections
 3.5 Spectral Method of Fatigue Life Determination 65

1.01

1.009

1.008 Uniform amplitude distribution

1.007
D, damage

1.006

1.005

1.004
Rayleigh amplitude distribution
1.003

1.002

1.001

1
0 20 40 60 80 100
k, tangent number

Fig. 3.12. Damage in the function of tangent section number. Broken line marks
uniform amplitude distribution and full line – Rayleigh amplitude distribution

suitable relations
⎧ * *
⎪ * ∂εak (Nf ) **
⎨ ∂εa (Nf ) * =
∂Nf *Nf sk ∂Nf *Nf sk , (3.99)
⎪
⎩
εa (Nf sk ) = εak (Nf sk )
which served for the derivation of damage
σf εap (b−c) εae (c−b)
Dk = (2Nf sk ) εa + εf (2Nf sk ) εa , (3.100)
E
εae b + εap c
dk = , (3.101)
εa
where: Nf sk – cycle for which equation Nf k (εa ) = Nf (εa ) is fulfilled,
εa , εae , εap – amplitude of total, elastic, and plastic strain, respectively
derived from (3.84) for Nf = Nf sk .
Strain amplitude εuk and εlk is determined from formulae

d dk
d dk
Dk k+1 −dk Dk k−1 −dk
εlk = Dk , εuk = Dk . (3.102)
Dk+1 Dk−1
The application of the method for the determination of fatigue life must
involve the following stages:
66 3 Theoretical Fundamentals

• determination of tangent section number (recommended ns = 50),

• determination of point of tangency for every tangent section (total cy-
cle number Nf sk ) (logarithmic division in the range 0.5 to 1012 cycles is
recommended),
• derivation of strain amplitudes εa , εae , εap for cycle number Nf sk on the
basis of (3.84),
• derivation of coefficients Dk (3.100) and dk (3.101) for each tangent section,
• determination of points of intersection εlk and εuk for adjoin tangent sec-
tion (for extreme tangent sections εu 1 = D1 and εl ns = 0),
• calculation of variance of strain history με = m0 ,
• calculation of integration limits tlk and tuk (3.93),
• calculation of fatigue life T from formula (3.97).
4
Algorithm of Spectral Method for Evaluation
of Fatigue Life

For the case of multiaxial random loading the calculation of fatigue life consists
in the reduction of triaxial stress or strain states into an equivalent uniaxial
one by the application of suitable criteria of fatigue failure [53, 54, 75]. In
the spectral methods, a matrix of power spectral density functions (3.14) de-
fines stress or strain states. The reduction of multiaxial loading state may be
conducted directly on power spectral density functions. The power spectral
densities of equivalent stress or strain gained as a result of the application
of multiaxial fatigue failure criteria are subsequently employed in the same
way as the functions defined on the basis of uniaxial tension-compression test-
ing [28]. This approach implements the familiar and experimentally verified
methods of determination of fatigue life associated with simple loading states.
On the basis of the review of state-of-the-art in research and experience
an algorithm for the determination of fatigue life under multiaxial loading is
currently postulated with the application of the spectral method (Fig. 4.1).
The following subchapters include the description of the particular blocks
comprised in the algorithm.

4.1 Block 1 – Loading

The input data for the calculation of fatigue life are matrices of power spectral
density functions of random stress or strain tensor (3.14). The functions could
be derived by:
1. Measurement of strain in structural components under service conditions
by means of spectrum analyser.
2. Calculation of real or designed structures by means of the Finite Element
Method (FEM), Boundary Element Method (BEM) or Finite Difference
Method (FDM) by application of analysis of frequency response of a sys-
tem [30, 62, 74].
68 4 Algorithm of Spectral Method for Evaluation of Fatigue Life

Registration, generation or calculation

1. Gσkl (f ) or Gεkl (f )

Determination of critical plane

2. η̄, s̄

Calculation of power spectral density function

3. Gσeq (f ) or Gεeq (f )

Determination of power spectral density moments mk

4. and derivative statistical parameters

Calculation of fatigue life

5. TSP

Fig. 4.1. Algorithm of fatigue life determination by spectral method for the mul-
tiaxial random loading defined with a matrix of power spectral density functions of
stress or strain tensor components

3. Calculation of the PSD function of generated random sequences1 with

adaptable probabilistic characteristics corresponding to service conditions
or anticipated circumstances.
4. Selection of parameters for the definition of the shape of the empirical PSD
functions derived on the basis of a statistical elaboration of the identified
physical phenomena2 .
The input quantity defining the loading state also includes the assumed
probability distribution of instantaneous values of stress or strain tensor com-
ponents.

4.2 Block 2 – Determination of Critical Plane Position

The determination of the expected position of critical fatigue fracture plane
constitutes an important part of the algorithm of life time calculation. One
1
For the purposes of fatigue life the estimators of auto- and cross-spectral den-
sity functions are most commonly determined by application of the fast Fourier
transform with the Welch method [89, 101].
2
Bibliography refers to only two instances in which the functions are applicable:
longitudinal component of wind velocity [32] and sea surface rise [14, 30]. A
considerable limitation to the method is the missing statement of the technique
of generation of cross-PSD functions.
 4.2 Block 2 – Determination of Critical Plane Position 69

of the fundamental factors applied for the determination of the critical plane
is the stress or strain state of the material. The position is most commonly
determined by the definition of direction cosines ln , mn and nn (n = η, s)
of unit vectors η̄ and s̄ occurring in the criteria of multiaxial fatigue failure,
where η̄ is orthogonal and s̄ is tangent to the critical plane. Application of
the following two methods is postulated for the determination of the expected
position of the critical plane.
1. Damage accumulation method 3 , discussed in [50, 51] consists in the deter-
mination of fatigue damage for the possible positions of the critical plane
and the selection of the position for which the damage is at the maximum.
As a result, fatigue life is determined besides the position of the critical
plane.
2. Variance method. It is currently popular and offers good results [8, 9, 48].
Its application is not recommended for the instances when the compo-
nents of stress or strain tensor have different probability distributions of
instantaneous values [92]. In the variance method it is assumed that the
planes for which the equivalent stress or strain variance reaches maximum
in accordance with the selected criterion of multiaxial fatigue failure, are
critical to the material and one of them might involve fatigue failure. The
search for variance maximum μxeq for the analysed stress or strain state
consists in the seeking the maximum value of the expression
6 
 6
μxeq = ak al μxkl , (4.1)
k=1 l=1

where: ak , al = f (ln , mn , nn , cp ) – non-linear functions of directional

cosines (Table 3.1),
cp – material constants.
The elements μx kl of covariance matrix is calculated directly from random
histories of component stress or strain state, although they could also be
determined from the matrix of power spectral density (3.14)
⎡∞ ⎤

μx = Re ⎣ Gx (f )df ⎦ . (4.2)
0

In the case of the above methods, the successful application depends on the
selection of an appropriate criterion of multiaxial fatigue failure and discreti-
sation step of angles change unit vectors η̄ and s̄. In this book the method of
maximum variance was employed due to its versatile use both in the spectral

3
In the spectral method damage accumulation is not realised directly as itera-
tive process familar from the cycle counting method. Damage is calculated from
integral (3.57).
70 4 Algorithm of Spectral Method for Evaluation of Fatigue Life

and in the cycle counting methods for fatigue life determination. A compar-
ison of the two methods follows in the current book. Besides, the variance
method is easily transformable into a computer programme and the algo-
rithm is marked with a shorter calculation time in comparison to the method
of damage accumulation.

4.3 Block 3 – Calculation of Power Spectral Density

Function of Equivalent Stress or Strain

Block 3 is devoted to determination of power spectral density function of the

equivalent stress or strain from formula (3.35). The formula could only be ap-
plied on condition that the equivalent value is defined as a linear combination
of suitable components of stress or strain tensor according to the criteria of
multiaxial fatigue failure.

4.4 Block 4 – Statistical Parameters

The statistical parameters defining the history of equivalent stress or strain are
directly determined from moments of power spectral density of the equivalent
quantities
∞
mk = f k Geq (f )df . (4.3)
0

4.5 Block 5 – Calculation of Fatigue Life

The calculated fatigue life (3.57) is the time, commonly expressed in seconds,
of material or structural component failure with 50% probability. The spectral
method applies various formulae adopted for the particular characteristics of
fatigue loading [11]. The following groups are distinguished in accordance with
a general classification:
1. Spectral formulae for loading history with normal probability distribution,
1.a. narrow-band frequency [19, 42, 59, 69],
1.b. narrow-band frequency spectrum with an additional harmonic com-
ponent or impulse loading [26, 33, 84],
1.c. broad-band frequency spectrum [15, 17, 21, 38, 79, 100, 103],
2. Spectral formulae for loading history with probability distribution other
than normal,
2.a. narrow-band frequency spectrum [85],
2.b. broad-band frequency spectrum [39, 43, 86].
 4.5 Block 5 – Calculation of Fatigue Life 71

It is remarked that the suitable selection of spectral formula depends on

the loading character. In practice, the formulae are most commonly derived
for the particular loading types or structures (drilling rigs, wind power plants,
body plating, etc.). For the instances when the loading type (especially fre-
quency width) is not determined, the application of broad frequency spectrum
formulae is recommended (1.c. or 2.b.) due to the larger versatility. The ap-
plication of spectral formulae groups 1.a. and 1.c. in this book results from
the normal probability distribution of loading analysed in the simulation and
experimental tests presented in the following chapters.
5
Simulations

5.1 Algorithm, Assumptions and Objectives of

Simulations

A number of simulations of multiaxial random loading have been performed

for improved comprehension and analysis of spectral method. They serve for
comparison of the applied and known from literature cycle counting method
with the postulated generalisation of the spectral method. Upon the extensive
study of the algorithms of the two methods it is concluded that some calcu-
lation blocks include common features, which are subsequently established.
Figure 5.1 presents block diagrams of two methods and focuses on the com-
mon features of the methods. The results of calculations of the two algorithms
are compared in the distinguished blocks PWO1, PWO2 and PWO3.
The assumptions of the conditions for the simulations are established:
1. Generated stress or strain tensors have normal probability distribution for
narrow- and broad-band frequency spectra;
2. Suitable stress or strain tensor components have different coefficients of
cross-correlation (r = 1, r = −1 and r ≈ 0);
3. Analysed stress states:
• biaxial tension-compression,
• uniaxial tension-compression with torsion,
• biaxial tension-compression with torsion,
• complete, spatial stress state.
4. Analysed strain states:
• strain states derived from four above stress states on the basis of the
Hooke law.
The strain tensor is derived from the stress tensor on the basis of the
Hooke law in order to maintain the fundamental relations between strain
tensor components and due to the linear form of the formulae. The current
74 5 Simulations

Cycle counting method Spectral method

Generation Determination of Gσ (f ) or Gε (f )
1. σ(t) or ε(t)
1. on the basis of σ(t) or ε(t)

Determination of critical Determination of critical

2. plane position η̄, s̄
2. plane position η̄, s̄

Examination of equivalence
PWO1 of critical plane positions

Calculation of stress σeq (t) Calculation of power spectral

3. or strain εeq (t)
3. density Gσeq (f ) or Gεeq (f )

σeq (t) ←→ σeq

IF F T
(t) εeq (t) ←→ εIF
eq
FT
(f )
PWO2 Gσeq (f ) ←→ Gσeq (f )
FFT
Gεeq (f ) ←→ Gεeq (f )
FFT

Cycle counting with Determination of moments mk

4. rain flow algorithm 4. and derived parameters

Damage summation and calculation Calculation of fatigue life

5. of fatigue life TZC and IF F T
TZC
5. TSP

Comparison of fatigue life

PWO3 IF F T
TZC , TZC and TSP

Fig. 5.1. Algorithm of fatigue life calculation applied in simulations including blocks
(PWO1 – PWO3) for comparison between cycle counting and spectral methods

authors are familiar with the limitation to the assumptions and consider the
resulting values of strain as input quantities for simulations.
The computer simulations apply material constants for steel 18G2A [27],
which are summarised in Table 5.1.

Table 5.1. Fatigue and strength characteristics of steel 18G2A

E ν σaf m NG σf εf b c
GPa — MPa — cycles MPa — — —
210 0,3 271 7,2 2,735·106 782 0,693 –0,118 –0,41
 5.2 Input Quantities 75

In general, the objectives of the simulations include:

1. Testing of the effect of various strain or stress states on the calculated life
time;
2. Comparison of cycle counting method and generalised spectral method on
the basis of:
2.1 calculated positions of critical planes (PWO1),
2.2 calculated values of equivalent stress or strain and parameters of
power spectral density functions of the equivalent quantities (PWO2),
2.3 calculated fatigue life (PWO3).

5.2 Input Quantities

The input quantities in the simulation calculations include components of
stress (3.8) or strain state (3.9) along with material constants (Table 5.1) and
the essential strength characteristics. The generation time of random histo-
ries To = 682.67 s (409600 points under sampling frequency 600 Hz) is fixed
experimentally. Random histories are generated by means of random num-
ber generator with normal probability distribution, which are subsequently
filtered with the bandpass Chebyshev Type I filter [89]. The filter order is
set experimentally to four and frequency damping outside the to −0.5 dB.
Depending on the width of passband the resulting Gaussian histories have
narrow-band (passband fmin = 19.5 Hz and fmax = 20.5 Hz) and broad-band
frequency spectra (passband fmin = 5 Hz and fmax = 20 Hz). The output
signal obtained from generator was scaled to the variance of values μx = 1.
Figure 5.2 presents a section of narrow-band frequency history, the probability
density distribution function (bar chart) approximated with normal distribu-
tion (full line) and power spectral density function. A similar pattern describes
a section of broad-band frequency history (Fig. 5.3).
The resulting histories are scaled and included into a vector of stress tensor
components σ(t), which results in 24 representative stress states. They are
marked S1 to S24 and presented in Table 5.2. The corresponding strain states
(E1 to E24) are determined in accordance with Hooke law:
1
εxx (t) = [σxx (t) − ν (σyy (t) + σzz (t))] ,
E
1
εyy (t) = [σyy (t) − ν (σzz (t) + σxx (t))] ,
E (5.1)
1
εzz (t) = [σzz (t) − ν (σxx (t) + σyy (t))] ,
E
1+ν 1+ν 1+ν
εxy (t) = σxy , εxz (t) = σxz , εyz (t) = σyz .
E E E
A matrix of power spectral density function of stresses Gσ (f ) or strains
Gε (f ) is applied for loading description in the spectral method. In order
76 5 Simulations

a)
4

2
x(t)

−2

−4
0 0,5 1 1,5 2 2,5 3 3,5 4 4,5 5
time t, s
b) c)
0,4 1

0,8
0,3

Gx (f )
px (x)

0,6
0,2
0,4
0,1
0,2

0 0
−4 −2 0 2 4 0 10 20 30 40 50
x f , Hz

Fig. 5.2. A section of history with narrow-band frequency spectrum (a), including
probability density distribution function (b), and power spectral density function (c)

a)
4

2
x(t)

−2

−4
0 0,5 1 1,5 2 2,5 3 3,5 4 4,5 5
time t, s
b) c)
0,4 0,08

0,3 0,06
Gx (f )
px (x)

0,2 0,04

0,1 0,02

0 0
−4 −2 0 2 4 0 10 20 30 40 50
x f , Hz

Fig. 5.3. A section of history with broad-band frequency spectrum (a), including
probability density distribution function (b), and power spectral density function (c)
 5.2 Input Quantities 77

to ensure the equivalence of the conditions of loading with regard to spectral

method and cycle counting method the matrix of power spectral density func-
tions is determined directly from the vector of stress or strain state component
histories. The function csd from MATLAB [89] is applied for this purpose. It
estimates auto- and cross-power spectral density function by means of Welch
method. The length of the subsequent signal sections under the analysis is
determined by means of Hann type window of the length of 212 points. Each
section was subjected to trend and constant component removal. The smooth
estimator of power spectral density function resulted from the mean of 100
history sections including total of 100 × 212 points [6].
 78

Table 5.2: Characteristics of generated stress tensors applied in calculations

Stress Covariance μσ Description

state of tensor components σ(t) (MPa2 )
1 2⎡ ⎤ 3
4356 4356 0
S1 ⎣ σ(t) = [σxx , σyy , 0], biaxial tension-compression, complete corre-
4356 4356 0⎦
(NB )
5 Simulations

lation of components, rxx,yy = 1, i.e. proportional loading

⎡ 0 0 0⎤
4356 4356 0
S2 ⎣
4356 4356 0⎦ —”—
(BB )
⎡ 0 0 0⎤
3600 3960 0 σ(t) = [σxx , σyy , 0], assymetric biaxial tension-compression, com-
S3 ⎣3960 4356 0⎦ plete correlation of components, rxx,yy = 1, σyy component with
⎡ 0 0 0⎤ larger variance, i.e. proportional loading
3600 3960 0
S4 ⎣3960 4356 0⎦ —”—
⎡ 0 0 0⎤
3600 16 0
σ(t) = [σxx , σyy , 0], biaxial tension-compression, correlation of
S5 ⎣ 16 3600 0⎦
components rxx,yy ≈ 0, nonproportional loading
⎡ 0 0 0⎤
3600 49 0
S6 ⎣ 49 3600 0⎦ —”—
⎡ 0 0 0 ⎤
3600 −3600 0
σ(t) = [σxx , σyy , 0], biaxial tension-compression, negative correla-
S7 ⎣−3600 3600 0⎦
tion, rxx,yy = −1, i.e. proportional loading
0 0 0
 Table 5.2: (continuation)

1 2
⎡ ⎤ 3
3600 −3600 0
S8 ⎣−3600 3600 0⎦ —”—
⎡ 0 0⎤0
3600 0 1800 σ(t) = [σxx , 0, σxy ], uniaxial tension-compression with torsion,
S9 ⎣ 0 0 0 ⎦ complete correlation of components, rxx,yy = 1, i.e. proportional
⎡1800 0 900 ⎤ loading
3600 0 1800
S10 ⎣ 0 0 0 ⎦ —”—
⎡1800 0 900⎤
3600 0 −5
⎣ 0 0 0 ⎦ σ(t) = [σxx , 0, σxy ], uniaxial tension-compression with torsion,
S11
correlation of components, rxx,xy ≈ 0, i.e. nonproportional loading
⎡ −5 0 900⎤
3600 0 21
S12 ⎣ 0 0 0 ⎦ —”—
⎡ 21 0 900 ⎤
1764 1764 882 σ(t) = [σxx , σyy , σxy ], complete plain stress state, complete cor-
S13 ⎣1764 1764 882⎦ relation of components, rxx,yy = rxx,xy = ryy,xy = 1, i.e. propor-
⎡ 882 882 441⎤ tional loading
1764 1764 882
S14 ⎣1764 1764 882⎦ —”—
⎡ 882 882 441⎤
3600 3600 59
5.2 Input Quantities

⎣3600 3600 59 ⎦ σ(t) = [σxx , σyy , σxy ], complete plain stress state, rxx,yy = 1,
S15
rxx,xy = ryy,xy ≈ 0, i.e. nonproportional loading
59 59 900
79
 80

Table 5.2: (continuation)

1 2⎡ ⎤ 3
3600 3600 −4
S16 ⎣3600 3600 −4 ⎦ —”—
⎡ −4 −4 900 ⎤
576 576 576 576 576 576
5 Simulations

⎢576 576 576 576 576 576⎥

⎢ ⎥ σ(t) = [σxx , σyy , σzz , σxy , σxz , σyz ], spatial stress state, full com-
⎢576 576 576 576 576 576⎥
S17 ⎢ ⎥ ponents correlation, r = 1 for every component pair, i.e. propor-
⎢576 576 576 576 576 576⎥
⎢ ⎥ tional loading
⎣576 576 576 576 576 576⎦
⎡576 576 576 576 576 576⎤
576 576 576 576 576 576
⎢576 576 576 576 576 576⎥
⎢ ⎥
⎢576 576 576 576 576 576⎥
S18 ⎢ ⎥ —”—
⎢576 576 576 576 576 576⎥
⎢ ⎥
⎣576 576 576 576 576 576⎦
⎡576 576 576 576 576 576 ⎤
2916 −17 −3 −180 −16 −41
⎢ −17 2916 40 −166 45 −172⎥
⎢ ⎥ σ(t) = [σxx , σyy , σzz , σxy , σxz , σyz ], spatial stress state, compo-
⎢ −3 40 2916 −78 −34 85 ⎥
S19 ⎢ ⎥ nent correlation, rkl ≈ 0; k, l = xx, yy, zz, xy, xz, yz for k = l, i.e.
⎢−180 −166 −78 2916 130 −6 ⎥
⎢ ⎥ nonproportional loading
⎣ −16 45 −34 130 2916 −17 ⎦
−41 −172 85 −6 −17 2916
 Table 5.2: (continuation)

1 2⎡ ⎤ 3
2916 52 −67 4 −9 22
⎢ 52 2916 59 8 41 4 ⎥
⎢ ⎥
⎢ −67 59 2916 28 −6 59 ⎥
S20 ⎢ ⎥ —”—
⎢ 4 8 28 2916 −39 −9 ⎥
⎢ ⎥
⎣ −9 41 −6 −39 2916 39 ⎦
⎡ 22 4 59 −9 39⎤ 2916
900 810 720 630 540 450
⎢810 729 648 567 486 405⎥
⎢ ⎥ σ(t) = [σxx , σyy , σzz , σxy , σxz , σyz ], spatial stress state, full com-
⎢720 648 576 504 432 360⎥
S21 ⎢ ⎥ ponents correlation, r = 1 for every component pair, different
⎢630 567 504 441 378 315⎥
⎢ ⎥ variance of particular components, i.e. proportional loading
⎣540 486 432 378 324 270⎦
⎡450 405 360 315 270 225⎤
900 810 720 630 540 450
⎢810 729 648 567 486 405⎥
⎢ ⎥
⎢720 648 576 504 432 360⎥
S22 ⎢ ⎥ —”—
⎢630 567 504 441 378 315⎥
⎢ ⎥
⎣540 486 432 378 324 270⎦
⎡450 405 360 315 270 225 ⎤
3600 107 69 −19 −68 −89
⎢ 107 2916 50 −6 −8 −3 ⎥ σ(t) = [σxx , σyy , σzz , σxy , σxz , σyz ], spatial stress state, compo-
⎢ ⎥
⎢ 69 50 2304 52 −77 25 ⎥ nent correlation, rkl ≈ 0; k, l = xx, yy, zz, xy, xz, yz for k = l,
S23 ⎢ ⎥
⎢ −19 −6 52 1764 −27 32 ⎥
⎢ ⎥ different variance of particular components, i.e. nonproportional
5.2 Input Quantities

⎣ −68 −8 −77 −27 1296 12 ⎦ loading

−89 −3 25 32 12 900
81
 82

Table 5.2: (continuation)

1 2⎡ ⎤ 3
3600 5 −4 −10 26 12
⎢ 5 2916 −11 −6 −30 −7 ⎥
⎢ ⎥
⎢ −4 −11 2304 −7 4
⎢ 2 ⎥⎥
S24 ⎢ −10 −6 −7 1764 −5 1 ⎥ —”—
⎢ ⎥
5 Simulations

⎣ 26 −30 4 −5 1296 −8 ⎦
12 −7 2 1 −8 900

symbol for width band frequency spectrum, NB – narrow-band, BB – broad-band
The cases S1, S3, S5, ..., S23 included narrow-band frequency random histories, and the cases S2, S4, S6, ..., S24 – broad-band frequency
random histories. Simulation calculations were performed with the MATLAB [58] software
 5.3 Analysis of Positions of Critical Planes 83

5.3 Analysis of Positions of Critical Planes

The awareness of the position of the critical plane, more precisely, the direc-
tion cosines of vector normal η̄(lη , mη , nη ) and vector tangent to the plane
s̄(ls , ms , ns ) is indispensable for the determination of equivalent stress or
strain with application of multiaxial fatigue failure criteria discussed previ-
ously. In order to determine the position of the critical plane, the variance
method is applied (see Chapter 4.2). The simulations involve the criterion
of maximum normal stress in the critical plane (3.21) and maximum normal
strain in the critical plane (3.24). In the cycle counting method the covariance
matrix is computed directly from random histories in accordance with formula

μxkl = e [(xk − x̂k )(xl − x̂l )] , (k, l = 1, . . . , 6) , (5.2)

while in the spectral method from the matrices of power spectral density
functions in accordance with formula (4.2).
The simulations are applied for determination of the equivalent stress or
strain variance for many (20897) possible positions of the critical plane. For
the selected multiaxial fatigue failure criteria the awareness of vector normal
to plane η̄ is sufficient for the determination of its position. For the instance
above the directional cosines are generated in a way that ensures filling of a
half of the sphere describing the tip of the vector normal to the critical plane
(Fig. 5.4).

a) b)
z z

y y

φz nη
mη
φy η̄
η̄

φx
lη

x
x

Fig. 5.4. Half of the sphere describing the tip of the vector normal η̄ to the critical
plane in search of maximum variance of equivalent stress or strain

The suitable direction cosines are generated with the limitations resulting
from the equation '
lη2 + m2η + n2η = 1 . (5.3)
84 5 Simulations

Under the assumption that lη ∈ 0; 1 (a half of the sphere, as in Fig. 5.4)

it follows that:
+ ' ' ,
mη ∈ − 1 − lη2 ; 1 − lη2 ,
+ ' ' , (5.4)
nη ∈ − 1 − lη2 − m2η ; 1 − lη2 − m2η .

The relations presented above enable a graphical representation of the

variance of the equivalent quantity in the function of the critical plane position
determined with two directional cosines of vector normal to the plane. The
value of the third direction cosines is determined from formula
'
lη = 1 − m2η − n2η . (5.5)

Figures 5.5 to 5.16 present variance charts of equivalent stress and strain,
respectively in the function of two directional cosines mη and nη (projection of
half sphere on a plane). The grey shade marks the value of variance respective
for the scale quoted to the right of the particular charts. The triangle marks
the locations for which case variance reach maximum values determined in
a conventional way in time domain. The square marks the maximum values
of variance determined directly from the matrices of power spectral density
functions.
It is remarked that maximum variance for plain stress state (cases S1 to
S16) occurs only when the critical plane coincides with X-Y plane, i.e. for
π
φz = . A similar distribution of maximums is distinguished for strain (cases
2
E1 to E16), although the cases involve spatial strain states.
In several cases the algorithm of seeking the maximum of a variance pro-
vides more than single solution. The justification of the phenomenon can be
twofold. The first is easily discernible in Fig. 5.5 and Fig. 5.11. The reason for
the phenomenon is associated with constant variance of equivalent quantity
as a result of the selection of components of stress tensor and its values. For
instance, during loading under biaxial tension-compression with component
correlation rxx,yy = 1 (cases S1 and S2) the variance method indicates that
the occurrence of the position of the critical' plane for every mη ∈ −1, 1
with equal probability, while nη = 0 and lη = 1 − m2η . Following the analy-
sis of mathematical forms of the applied multiaxial fatigue failure criteria the
other justification for the occurrence of several variance extremes μσeq and
μεeq could be identified. In several cases (e.g. S7, S8, E3, E4) under the selec-
tion of suitable variances of stress or strain tensor components, the functions
μσeq and μεeq demonstrate periodical qualities. As a result of equal variance
in the positions it could be presumed that fatigue failure initiate in one of the
determined planes. The selection of the plane is random and depends on the
quality of material.
It is noticeable that the shapes of the determined functions μσeq and μεeq
in time domain coincide with the shapes of adequate functions determined by
 5.3 Analysis of Positions of Critical Planes 85

means of spectral method. Apart form that, the maximum values are equal.
Differences are observed between the spectral method and cycle counting
method only in several cases. They result from numerical errors in the course
of covariance matrix calculation by the methods. It is particularly discernible
in the cases when tensor components have similar variances (S20, S24). In the
cases the small differences during estimation of covariance matrixes with the
two methods result in large difference of the estimated positions of critical
planes.
Table 5.3 summarises the direction cosines of vector normal to the critical
plane for which the variance of equivalent stress or strain reach the maximum.
For the purpose of the simplification of data analysis the angles of vector
normal η̄ with co-ordinate axes x, y, z are defined.
86 5 Simulations
2 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 4000 1 4000
S1 S1

0.5 0.5 3000

nη 0 2000 nη 0 2000

−0.5 −0.5 1000

−1 0 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη 2
mη 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 4000 1
S2 S2 4000

0.5 0.5 3000

nη 0 2000 nη 0 2000

−0.5 −0.5 1000

−1 0 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
σeq ,
2
(MPa ) μSP
σeq , (MPa2 )
1 4000 1
S3 S3

0.5 0.5 3000

nη 0 2000 nη 0 2000

−0.5 −0.5 1000

−1 0 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
σeq ,
2
(MPa ) μSP
σeq , (MPa2 )
1 4000 1 4000
S4 S4

0.5 0.5 3000

nη 0 2000 nη 0 2000

−0.5 −0.5 1000

−1 0 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.5. Variance charts of equivalent stress μσeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by stress tensors (S1–S4)
 5.3 Analysis of Positions of Critical Planes 87
2 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S5 S5
3000 3000
0.5 0.5

2000 2000
nη 0 nη 0

1000 1000
−0.5 −0.5

−1 0 −1 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη 2
mη 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S6 S6
3000 3000
0.5 0.5

2000 2000
nη 0 nη 0

1000 1000
−0.5 −0.5

−1 0 −1 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη 2
mη 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S7 S7
3000
3000
0.5 0.5

2000
nη 0 nη 0 2000

1000 1000
−0.5 −0.5

−1 0 −1 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη 2
mη 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S8 S8
3000 3000
0.5 0.5

2000 2000
nη 0 nη 0

1000 1000
−0.5 −0.5

−1 −1 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.6. Variance charts of equivalent stress μσeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by stress tensors (S5–S8)
88 5 Simulations
2 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S9 S9
4000
4000
0.5 0.5
3000
nη 0 nη 0
2000 2000

−0.5 −0.5
1000

−1 −1 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη 2
mη 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S10 S10
4000 4000
0.5 0.5

nη 0 nη 0
2000 2000
−0.5 −0.5

−1 0 −1 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη 2
mη 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S11 S11 3000
3000
0.5 0.5
2000
2000
nη 0 nη 0

1000 1000
−0.5 −0.5

−1 0 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
σeq ,
2
(MPa ) μSP
σeq , (MPa2 )
1 1
S12 S12
3000 3000
0.5 0.5

2000 2000
nη 0 nη 0

1000 1000
−0.5 −0.5

−1 0 −1 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.7. Variance charts of equivalent stress μσeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by stress tensors (S9–S12)
 5.3 Analysis of Positions of Critical Planes 89
2 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S13 S13
3000
3000
0.5 0.5

2000
nη 0 2000 nη 0

−0.5 1000 −0.5 1000

−1 0 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη 2
mη 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S14 S14

3000 3000
0.5 0.5

nη 0 2000 nη 0 2000

−0.5 1000 −0.5 1000

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
σeq ,
2
(MPa ) μSP
σeq , (MPa2 )
1 1
S15 4000 S15 4000

0.5 0.5

nη 0 nη 0
2000 2000

−0.5 −0.5

−1 0 −1 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη 2
mη 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S16 4000 S16 4000

0.5 0.5

nη 0 nη 0
2000 2000

−0.5 −0.5

−1 0 −1 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.8. Variance charts of equivalent stress μσeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by stress tensors (S13–S16)
90 5 Simulations
2 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S17 S17
4000 4000
0.5 0.5
3000 3000
nη 0 nη 0
2000 2000

−0.5 −0.5
1000 1000

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη 2
mη 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S18 S18
4000 4000
0.5 0.5
3000 3000
nη 0 nη 0
2000 2000

−0.5 −0.5
1000 1000

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
σeq ,
2
(MPa ) μSP
σeq , (MPa2 )
1 1
S19 S19
5000
4500
0.5 0.5

nη 4000 nη 4000
0 0

3500
−0.5 −0.5
3000
3000
−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
σeq ,
2
(MPa ) μSP
σeq , (MPa2 )
1 1
S20 S20
4500 4500
0.5 0.5

4000 4000
nη 0 nη 0

3500 3500
−0.5 −0.5

3000 3000
−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.9. Variance charts of equivalent stress μσeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by stress tensors (S17–S20)
 5.3 Analysis of Positions of Critical Planes 91
2 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S21 S21

3000 3000
0.5 0.5

nη 0 2000 nη 0 2000

−0.5 1000 −0.5 1000

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη 2
mη 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 1
S22 S22

0.5 3000 0.5 3000

nη 0 2000 nη 0 2000

−0.5 1000 −0.5 1000

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
σeq ,
2
(MPa ) μSP
σeq , (MPa2 )
1 3500 1
S23 S23

0.5 0.5 3000

3000
nη 0 nη 0
2500
2500
−0.5 −0.5

2000 2000
−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη 2
mη 2
μZC
σeq , (MPa ) μSP
σeq , (MPa )
1 3500 1
S24 S24 3500

0.5 0.5
3000 3000
nη 0 nη 0

−0.5 2500 −0.5 2500

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.10. Variance charts of equivalent stress μσeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by stress tensors (S21–S24)
92 5 Simulations

μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E1 E1
4
4
0.5 0.5
3
3
nη 0 nη 0
2 2

−0.5 1 −0.5
1

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E2 E2
4 4
0.5 0.5
3 3
nη 0 nη 0
2 2

−0.5 1 −0.5 1

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E3 E3 5

4
0.5 0.5 4

3 3
nη 0 nη 0
2 2
−0.5 −0.5
1 1

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1 5
E4 E4
4 4
0.5 0.5
3 3
nη 0 nη 0
2 2
−0.5 −0.5
1 1

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.11. Variance charts of equivalent strain μεeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by strain tensors (E1–E4)
 5.3 Analysis of Positions of Critical Planes 93

μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1 8
E5 E5

0.5 6 0.5 6

nη 0 4 nη 0 4

−0.5 2 −0.5 2

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E6 8 E6 8

0.5 0.5
6 6

nη 0 4
nη 0 4

−0.5 2 −0.5 2

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E7 E7

0.5 10 0.5 10

nη 0 nη 0
5 5
−0.5 −0.5

−1 0 −1 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−7 μSP
εeq x 10
−8

1 1
E8 E8

0.5 1 0.5 10

nη 0 nη 0
0.5 5
−0.5 −0.5

−1 0 −1 0
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.12. Variance charts of equivalent strain μεeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by strain tensors (E5–E8)
94 5 Simulations

μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1 14
E9 12 E9
12
10
0.5 0.5
10
8
8
nη 0 6
nη 0
6
4
−0.5 −0.5 4
2 2
−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1 12
E10 12 E10

10 10
0.5 0.5
8 8
nη 0 6
nη 0 6

4 4
−0.5 −0.5
2 2

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E11 E11
8
0.5 6 0.5
6
nη 0 4 nη 0
4

−0.5 2 −0.5 2

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E12 E12

0.5 6 0.5 6

nη 0 4 nη 0 4

−0.5 2 −0.5 2

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.13. Variance charts of equivalent strain μεeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by strain tensors (E9–E12)
 5.3 Analysis of Positions of Critical Planes 95

μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E13 E13
6
0.5 0.5 6

4
nη 0 nη 0 4

2 2
−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E14 E14
6 6
0.5 0.5

4 4
nη 0 nη 0

2 2
−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E15 E15
6
6
0.5 0.5

4 4
nη 0 nη 0

2 2
−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E16 E16
6 6
0.5 0.5

nη 4 nη 4
0 0

−0.5 2 −0.5 2

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.14. Variance charts of equivalent strain μεeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by strain tensors (E13–E16)
96 5 Simulations

μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E17 E17 10
10
0.5 8 0.5 8

6 6
nη 0 nη 0
4 4
−0.5 −0.5
2 2

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E18 E18
10 10
0.5 8 0.5
8

nη 0 6 nη 0 6

4 4
−0.5 −0.5
2 2

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−7 μSP
εeq x 10
−7

1 1
E19 E19
1.5 1.5
0.5 0.5

nη 0 nη 0

−0.5 1 −0.5
1

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−7 μSP
εeq x 10
−7

1 1.5 1 1.5
E20 E20

0.5 0.5

nη 0 nη 0

1 1
−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.15. Variance charts of equivalent strain μεeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by strain tensors (E17–E20)
 5.3 Analysis of Positions of Critical Planes 97

μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E21 E21
6 6
0.5 0.5

nη 4 nη 4
0 0

−0.5 2 −0.5 2

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E22 E22
6 6
0.5 0.5

nη 4 nη 4
0 0

−0.5 2 −0.5 2

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E23 E23

0.5 0.5 8
8
nη 0 nη 0

−0.5 −0.5 6
6

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη
μZC
εeq x 10
−8 μSP
εeq x 10
−8

1 1
E24 E24

0.5 0.5
8 8

nη 0 nη 0

−0.5 −0.5
6
6
−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
mη mη

Fig. 5.16. Variance charts of equivalent strain μεeq (mη , nη ) for conducted simula-
tions. The cases of loading defined by strain tensors (E21–E24)
 98

Table 5.3: Directional cosines of normal vector η̄(lη , mη , nη ) to critical plane determined by the maximum variance method

Case of [lη , mη , nη ]ZC [φx , φy , φz ]ZC [lη , mη , nη ]SP [φx , φy , φz ]SP

loading — degree — degree
1 2 3 4 5
S1 [0.035 -0.999 0] [88 178 90] [0.848 -0.53 0] [32 122 90]
[0.545 -0.839 0] [57 147 90] [1 -0.017 0] [1 91 90]
5 Simulations

[0.988 -0.156 0] [9 99 90] [1 0.017 0] [1 89 90]

[0.999 -0.035 0] [2 92 90] [0.998 0.07 0] [4 86 90]
[1 -0.017 0] [1 91 90] [0.995 0.105 0] [6 84 90]
[1 0.017 0] [1 89 90] [0.993 0.122 0] [7 83 90]
[0.999 0.035 0] [2 88 90] [0.616 0.788 0] [52 38 90]
[0.995 0.105 0] [6 84 90] [0.292 0.956 0] [73 17 90]
[0.857 0.515 0] [31 59 90]
[0.545 0.839 0] [57 33 90]
[0.292 0.956 0] [73 17 90]
[0.035 0.999 0] [88 2 90]
E1 [0.857 0.515 0] [31 59 90] [0.848 -0.53 0] [32 122 90]
[0.995 0.105 0] [6 84 90]
[0.978 0.208 0] [12 78 90]
[0.891 0.454 0] [27 63 90]
[0.857 0.515 0] [31 59 90]
S2 [0.848 -0.53 0] [32 122 90] [0.208 -0.978 0] [78 168 90]
[0.934 -0.358 0] [21 111 90] [0.998 0.07 0] [4 86 90]
[0.99 -0.139 0] [8 98 90] [0.978 0.208 0] [12 78 90]
[0.999 -0.035 0] [2 92 90] [0.891 0.454 0] [27 63 90]
[1 -0.017 0] [1 91 90] [0.208 0.978 0] [78 12 90]
[1 0.017 0] [1 89 90]
 Table 5.3: (continuation)

1 2 3 4 5
[0.999 0.035 0] [2 88 90]
[0.993 0.122 0] [7 83 90]
[0.99 0.139 0] [8 82 90]
[0.978 0.208 0] [12 78 90]
E2 [0.208 -0.978 0] [78 168 90] [0.848 -0.53 0] [32 122 90]
[0.208 0.978 0] [78 12 90]
S3 [0 -1 0] [90 180 90] [0 -1 0] [90 180 90]
[0 1 0] [90 0 90] [0 1 0] [90 0 90]
E3 [0 -1 0] [90 180 90] [0 -1 0] [90 180 90]
[0 1 0] [90 0 90] [0 1 0] [90 0 90]
S4 [0 -1 0] [90 180 90] [0 -1 0] [90 180 90]
[0 1 0] [90 0 90] [0 1 0] [90 0 90]
E4 [0 -1 0] [90 180 90] [0 -1 0] [90 180 90]
[0 1 0] [90 0 90] [0 1 0] [90 0 90]
S5 [0 -1 0] [90 180 90] [1 0 0] [0 90 90]
[0 1 0] [90 0 90]
E5 [1 0 0] [0 90 90] [0 -1 0] [90 180 90]
[0 1 0] [90 0 90]
S6 [1 0 0] [0 90 90] [0 -1 0] [90 180 90]
[0 1 0] [90 0 90]
E6 [1 0 0] [0 90 90] [0 -1 0] [90 180 90]
[0 1 0] [90 0 90]
5.3 Analysis of Positions of Critical Planes

S7 [0 -1 0] [90 180 90] [0 -1 0] [90 180 90]

[1 0 0] [0 90 90] [1 0 0] [0 90 90]
[0 1 0] [90 0 90] [0 1 0] [90 0 90]
99
 Table 5.3: (continuation)
100

1 2 3 4 5
E7 [0 -1 0] [90 180 90] [0 -1 0] [90 180 90]
[1 0 0] [0 90 90] [1 0 0] [0 90 90]
[0 1 0] [90 0 90] [0 1 0] [90 0 90]
S8 [0 -1 0] [90 180 90] [0 -1 0] [90 180 90]
5 Simulations

[1 0 0] [0 90 90] [1 0 0] [0 90 90]
[0 1 0] [90 0 90] [0 1 0] [90 0 90]
E8 [0 -1 0] [90 180 90] [0 -1 0] [90 180 90]
[1 0 0] [0 90 90] [1 0 0] [0 90 90]
[0 1 0] [90 0 90] [0 1 0] [90 0 90]
S9 [0.921 0.391 0] [23 67 90] [0.927 0.375 0] [22 68 90]
[0.921 0.391 0] [23 67 90]
E9 [0.921 0.391 0] [23 67 90] [0.921 0.391 0] [23 67 90]
S10 [0.927 0.375 0] [22 68 90] [0.927 0.375 0] [22 68 90]
[0.921 0.391 0] [23 67 90] [0.921 0.391 0] [23 67 90]
E10 [0.927 0.375 0] [22 68 90] [0.927 0.375 0] [22 68 90]
[0.921 0.391 0] [23 67 90] [0.921 0.391 0] [23 67 90]
S11 [1 0 0] [0 90 90] [0.999 -0.035 0] [2 92 90]
E11 [1 0 0] [0 90 90] [1 0 0] [0 90 90]
S12 [1 0.017 0] [1 89 90] [1 0 0] [0 90 90]
E12 [1 0.017 0] [1 89 90] [0.999 0.035 0] [2 88 90]
S13 [0.707 0.707 0] [45 45 90] [0.707 0.707 0] [45 45 90]
E13 [0.707 0.707 0] [45 45 90] [0.707 0.707 0] [45 45 90]
S14 [0.707 0.707 0] [45 45 90] [0.707 0.707 0] [45 45 90]
E14 [0.707 0.707 0] [45 45 90] [0.707 0.707 0] [45 45 90]
S15 [0.707 0.707 0] [45 45 90] [0.707 0.707 0] [45 45 90]
 Table 5.3: (continuation)

1 2 3 4 5
E15 [0.707 0.707 0] [45 45 90] [0.707 0.707 0] [45 45 90]
S16 [0.707 0.707 0] [45 45 90] [0.707 0.707 0] [45 45 90]
E16 [0.707 0.707 0] [45 45 90] [0.707 0.707 0] [45 45 90]
S17 [0.579 0.579 0.574] [55 55 55] [0.579 0.579 0.574] [55 55 55]
E17 [0.579 0.579 0.574] [55 55 55] [0.579 0.579 0.574] [55 55 55]
S18 [0.579 0.579 0.574] [55 55 55] [0.579 0.579 0.574] [55 55 55]
E18 [0.579 0.579 0.574] [55 55 55] [0.579 0.579 0.574] [55 55 55]
S19 [0.574 0.524 0.629] [55 58 51] [0.56 0.523 0.643] [56 58 50]
E19 [0.551 0.563 0.616] [57 56 52] [0.616 0.555 0.559] [52 56 56]
S20 [0.553 0.59 0.588] [56 54 54] [0.567 0.591 -0.574] [55 54 125]
E20 [0.579 0.579 -0.574] [55 55 125] [0.584 0.602 -0.545] [54 53 123]
S21 [0.633 0.578 0.515] [51 55 59] [0.633 0.578 0.515] [51 55 59]
E21 [0.633 0.578 0.515] [51 55 59] [0.633 0.578 0.515] [51 55 59]
S22 [0.633 0.578 0.515] [51 55 59] [0.633 0.578 0.515] [51 55 59]
E22 [0.633 0.578 0.515] [51 55 59] [0.633 0.578 0.515] [51 55 59]
S23 [0.956 0.292 -0.017] [17 73 91] [0.929 -0.317 -0.191] [22 108 101]
E23 [0.857 0.515 0.017] [31 59 89] [0.57 0.815 0.105] [55 35 84]
S24 [0.999 0.017 -0.035] [2 89 92] [0.913 -0.407 0.017] [24 114 89]
E24 [0.874 0.485 -0.017] [29 61 91] [0.913 -0.406 -0.035] [24 114 92]
5.3 Analysis of Positions of Critical Planes
101
102 5 Simulations

5.4 Analysis of Equivalent Quantities

Block 3 of algorithm for fatigue life calculation by means of the cycle count-
ing method (Fig. 5.1) is consists in the computation of histories of equivalent
stress σeq (t) or strain εeq (t), whereas the Block 3 of algorithm of the spectral
method, the power spectral density Gσeq (f ) or Gεeq (f ) is computed. They
are determined by application of multiaxial fatigue failure criteria in time and
frequency domains, respectively. A direct comparison between the quantities
is impossible due to the differences between them. Therefore, random history
FFT
σeq (t) is derived from power spectral density Geq (f ) and power spectral
density GF eq
FT
(t) is derived from equivalent stress σeq (t). Strain history is
treated in an analogous manner. The resulting values are subsequently com-
pared. In order to present the sequence of calculation leading to the results,
block PWO2 (diagram in Fig. 5.1) is elaborated. A unique case involving
stress history in Fig. 5.17 is presented.

PWO2 a) σeq (t) a) Geq (f )

b) FFT b) IFFT

c) GF
eq
FT
(f ) c) IF F T
σeq (t)

Comparison Comparison
d) GF FT
(f ) and Geq (f )
d) IF F T
σeq (t) and σeq (t)
eq

Fig. 5.17. Algorithm of comparison of results of equivalent stress history calculation

in time domain and power spectral density in frequency domain

5.4.1 Random Histories

IF F T
σeq (t) history is computed with the inverse fast Fourier transform (IFFT)
method. It consists in the application of algorithm of inverse fast Fourier
transform for the generation of random histories on the basis of amplitudes of
component harmonics and their phase shifts. The amplitudes of component
harmonics Ai are derived directly from the estimator of power spectral density
function, '
Ai = 2Geq i Δf , (5.6)
 5.4 Analysis of Equivalent Quantities 103

where: Ai – amplitude of i-th harmonic component,

Geq i – i-th value of estimator of power spectral density function,
Δf – distance between the successive discrete values of estimator of
power spectral density – frequency interval,
while the phase shift is determined by the generation of random numbers with
uniform distribution in the range 0, . . . , 2π). The bibliography includes other
methods of random history generation on the basis of power spectral density
[70, 77]. The current method is selected due to the opposition to power spectral
density determination by fast Fourier transform (FFT). The only drawback
of the method is the absence the phase shift of component harmonics, which
could not be retrieved while they are required during generation.
IF F T
A comparison of random histories σeq (t) and σeq (t) is carried out in
several stages. At first both random histories are examined for equivalence
of probability density functions. For the reason the Kolmogorov-Smirnov [96]
significance test of two random variables is performed for probability density
function under the assumption of the significance level αKS = 0.05. The
maximum of the difference of distribution function of the analysed variables
is tested. It could be remarked that the test is very sensitive, as it is sufficient
to increase the value of one of the tested variables by 5% for the negative
verification of the hypothesis of equality of probability distributions. The test
results prove the lack of premises for the rejection of the hypothesis concerning
the equivalence of distributions in the analysed cases.
Subsequently, a comparison is made for the amplitude distributions of
IF F T
random histories σeq (t) and σeq (t) obtained from the rain flow algorithm
[1]. Amplitudes from the random histories are compared, and the Kolmogorov-
Smirnov test [96] for the equivalence of distribution functions is performed just
as in the previous instance. The testing indicates that there are no reasons for
the rejection of the hypothesis of the equivalence of both amplitude probability
distributions.

5.4.2 Power Spectral Densities

Power spectral density functions are computed from the history of equivalent
stress by FFT algorithm in a similar manner as from the matrices of power
spectral density of stress and strain states components (Block 1 of simulation
algorithm, Fig. 5.1). A comparison is made between the moments of spectral
density mk , as they constitute the foundation for the determination of sta-
tistical parameters applied in spectral formulae for the calculations of fatigue
life. In the course of calculations, identical values of power spectral density
estimators of equivalent quantities are obtained in all cases. It could have
been anticipated as the determination of the value of equivalent quantity in
time or frequency domain is equivalent from the theoretical point of view. The
comparison additionally indicates sufficient accuracy of numerical calculations
and rejected the possibility of essential error in the further analysis.
104 5 Simulations

5.5 Analysis of Fatigue Life

Block PWO3 of the algorithm involves the comparison of the results of the
IF F T
calculated fatigue life. Symbols TZC and TZC mark the fatigue life calcu-
lated with the cycle counting method, with a difference that for the case of the
latter one the method of the inverse fast Fourier transform is applied for the
generation of the equivalent stress and strain histories (the histories registered
in block PWO2 are applied). In the cycle counting method the Palmgen-Miner
linear hypothesis of damage accumulation is applied
To To
TZC = =  nk , (5.7)
D(To )
Nk
k

where: To – observation time, loading history under cycle counting,

D(To ) – damage induced by loading during time To ,
k – number of levels of loading amplitude,
nk – number of cycles of amplitudes σak or εak ,
Nk – number of cycles determined from fatigue characteristics
(σa − Nf ) or (εa − Nf ) for k-th amplitude of stress or strain.
Fatigue life results calculated under the spectral method are marked with
TSP . For the case when loading is defined under the matrix of power spectral
density of stress, the formula postulated by Bollotin [15] is applicable while for
loading defined under the power spectral density of strain the formula (2.24)
is applied. The coefficient λ (2.30) from Wirsching [103, 43] is introduced for
the modification of computed fatigue life in order to account for the effect of
loading frequency spectrum band.
The fatigue life results are presented in Table 5.4 and a graphical inter-
pretation with a bar chart is provided in Fig. 5.18 and 5.19.
Table 5.4: Variance of equivalent quantities and fatigue life calculated by cycle counting and spectral
methods

IF F T
Case of μσeq , (MPa)2 μGσeq , (MPa)2 TZC , s TZC ,s TSP , s
loading and μεeq and μGεeq

1 2 3 4 5 6
S1 4356 4297 24858958 38013155 22986009
E1 4,84·10−8 1,042·10−7 67234162 52705192 52264751
S2 4356 4396 38594548 36551919 43735392
E2 4,84·10−8 9,825·10−8 99235609 94869021 124754163
S3 4356 4017 17723845 48991741 29353764
E3 5,224·10−8 1,098·10−7 47789773 44796705 46533835
S4 4356 4278 32428800 35294916 47584615
E4 5,224·10−8 1,065·10−7 80026925 77494113 101684312
S5 3600 3717 26934505 40972965 38730016
E5 8,393·10−8 1,713·10−7 20480584 18811746 17565585
S6 3600 3719 73905786 65152148 80267796
E6 8,864·10−8 1,775·10−7 25498783 24941948 33775807
S7 3600 3946 53898992 41783054 31213463
E7 1,38·10−7 2,736·10−7 6728184 7543379 6775990
S8 3600 3588 82609952 82569362 90363553
E8 1,38·10−7 2,73·10−7 10390489 10394199 14080962
S9 5245 4870 8956385 20632502 14660695
5.5 Analysis of Fatigue Life

E9 1,315·10−7 2,985·10−7 7149073 5590522 5722448

S10 5245 5362 18064104 18824753 21449129
E10 1,315·10−7 2,545·10−7 11248256 12324201 16220851
105
 Table 5.4: (continues)
106

1 2 3 4 5 6
S11 3600 3362 50156704 80516261 55567416
E11 8,163·10−8 1,912·10−7 19687579 15037548 13995446
S12 3600 3632 81414440 77522673 86324041
E12 8,165·10−8 1,642·10−7 30868660 29893642 39199964
5 Simulations

S13 3969 3758 47994169 57803042 37250262

E13 7,29·10−8 1,618·10−7 27614976 18486880 19815017
S14 3969 4044 52856090 49816318 58751088
E14 7,29·10−8 1,45·10−7 38580171 38135528 51519241
S15 4526 4500 26571009 32924271 19509929
E15 7,502·10−8 1,423·10−7 24825354 26817960 26163490
S16 4588 4532 32338550 29428628 38878267
E16 7,631·10−8 1,551·10−7 34370143 33930882 44757172
S17 5184 5183 16691575 18471462 11692869
E17 1,175·10−7 2,233·10−7 9956680 11011236 10157073
S18 5184 5185 17247805 21997937 24257814
E18 1,175·10−7 2,399·10−7 13960217 13644716 18154371
S19 5037 5455 12468718 13051272 9755663
E19 1,667·10−7 3,324·10−7 4644424 5223805 4640663
S20 4931 4947 22656140 28141801 28511227
E20 1,552·10−7 3,104·10−7 8308393 7909212 10941757
S21 4025 3943 37229857 49623582 31272921
E21 7,673·10−8 1,546·10−7 23095024 21397256 21837196
S22 4025 4123 49061226 46406646 55077998
E22 7,673·10−8 1,545·10−7 35015209 35168103 44793658
S23 3645 3431 37503145 88615708 51673197
 Table 5.4: (continues)

1 2 3 4 5 6
E23 9,62·10−8 1,861·10−7 14529498 14788695 14744152
S24 3602 3628 74645472 69489813 87332584
E24 9,229·10−8 1,866·10−7 23189441 22747671 30278142
5.5 Analysis of Fatigue Life
107
 7
108

x 10
10
TZC
S8 IF F T
9 S12
TZC S23 S24
TSP
S6 S11
8
5 Simulations

6 S13 S14
S7 S22

and TSP , s
5 S3 S4 S21
S2 S5

IF F T
4 S1 S16
S15

TZC , TZC
3 S20
S18
S9 S10
2 S17
S19
1

0
Random stress state type
IF F T
Fig. 5.18. Comparison of fatigue life calculated by cycle counting TZC and TZC and spectral methods TSP
 7
x 10
14
TZC
E2 IF F T
TZC
12 TSP

E4
10

and TSP , s
E1

IF F T
6
E3 E14
E16 E22

TZC , TZC
4 E12
E6
E15 E24
E13
E21
E5 E11 E18
2 E10 E23
E8
E17 E20
E7 E9 E19
0
Random strain state type
5.5 Analysis of Fatigue Life

IF F T
Fig. 5.19. Comparison of fatigue life calculated by cycle counting TZC and TZC and spectral methods TSP
109
110 5 Simulations

Another technique of fatigue life comparison is presented in Fig. 5.20 ÷

IF F T
5.23. It is remarked that life time TZC calculated on the basis of power
spectral density of the equivalent stress is overestimated in comparison to life
time TZC for narrow-band frequency spectrum histories (Fig. 5.20). For broad-
IF F T
band frequency spectrum histories, the life time TZC fluctuate around TZC
with maximum deviations up to ±9%. It could be remarked that it is as-
sociated with the generation of random histories by the inverse fast Fourier
transform from power spectral density. It results from random determination
of phase shifts of the successive harmonic components. Therefore, variations
of random histories are registered depending on the number of component
harmonics with considerable amplitudes and their phase shifts. The applied
generation method is particularly sensitive for the cases of broad-band fre-
quency spectrum of generated histories and correlated tensor components.
Additionally, the scatter of the results is affected by the large difference of
number of component cycles in narrow- and broad-band frequency spectrums.
The life time calculated in accordance with the generalized spectral method
correlates well with life time calculated by the cycle counting method. It
should be remarked that the application of various spectral formulae for fa-
tigue life depends on the type and character of the loading (two fatigue char-
acteristics and coefficients λ).
 5.5 Analysis of Fatigue Life 111

8 T IFFT
10 ZC
TSP
NB frequency
spectrum

Fig. 5.20.
, TSP , s

Comparison
of fatigue life
IFFT

7 IF F T
10 TZC and
TZC

TSP with life

calculated by
cycle counting
3 method TZC .
Cases of
3 loading defined
by stress
6 tensors with
10 narrow-band
6 7 8
10 10 10 (NB) frequency
TZC , s spectrum

IFFT
8 TZC
10
T
SP
BB frequency
spectrum

Fig. 5.21.
, TSP , s

Comparison
of fatigue life
IFFT

7 IF F T
10 TZC and
TZC

TSP with life

calculated by
cycle counting
3 method TZC .
Cases of
3 loading defined
by stress
6
tensors with
10 broad-band
6 7 8
10 10 10 (BB) frequency
TZC , s spectrum
112 5 Simulations

8 T IFFT
10 ZC
TSP
NB frequency
spectrum

Fig. 5.22. , TSP , s

Comparison
of fatigue life
IFFT

IF F T 7
TZC and 10
TZC

TSP with life

calculated by
cycle counting
method TZC . 3
Cases of
loading defined 3
by strain
tensors with 6
narrow-band 10
6 7 8
(NB) frequency 10 10 10
spectrum TZC , s

IFFT
8 TZC
10
T
SP
BB frequency
spectrum
, TSP , s

Fig. 5.23.
Comparison
IFFT

of fatigue life 7
10
TZC

IF F T
TZC and
TSP with life
calculated by
cycle counting 3
TZC . Cases of
loading defined 3
by strain
tensors with 6
broad-band 10
6 7 8
(BB) frequency 10 10 10
spectrum TZC , s
6
Experimental Studies

6.1 Test Stands

Fatigue testing was performed on two stands under the combination of bend-
ing with torsion MZGS100L and MZGS200L as presented in Figs 6.1 and 6.2.
The stand MZGS100L is a device designed for fatigue testing of standard
specimens made of various materials. The major components include a frame
with a rotary table, a lever and an electromagnetic actuator. Under the elec-
trical current the vibrations of the actuator coil generates moment of force M
in the tested specimen over the arm. The specimen is fixed at one end to the
lever and at the other side is attached to the holder placed at a rotary table
(Fig. 6.3 a). Through a suitable rotation over the angle αM in the limit of
0 ÷ π/2 the moment M of the tested specimen could be resolved into two
component moments-bending Mσ and torsional Mτ . Two extreme cases could
be distinguished: when the specimen axis is parallel to the axis of the lever-in
which case the specimen loading is limited to bending moment M = Mσ and
the case in which the axes of the specimen and lever form a right angle and
a distinct case of the torsion is involved M = Mτ . Therefore, for the case of
MZGS100L stand the tests could involve an arbitrary combination of bend-
ing and torsional moments including correlation coefficient rMσ ,Mτ = 1, i.e.
proportional loading.
Apart from that, the test stand includes a personal computer with in-
put/output card and power amplifier for power supply to electromagnetic ac-
tuator. The diagram of the entire stand is presented in Fig. 6.4. The computer-
controlled time variable command signal with the voltage range of ±1 V is
generated. The signal is conveyed to the amplifier; and subsequently, the am-
plified signal is fed into the electromagnetic actuator. The stand includes a
limit switch for the control of maximum deflection of the arm. As the preset
limit is exceeded the amplifier trips and test timer switches off. As the oc-
currence of fatigue failure induces a rapid and large loss of specimen rigidity
(inclination of the lever increases), a limit switch is applied for the determi-
nation of experimental fatigue life Texp .
114 6 Experimental Studies

Fig. 6.1. Test stand MZGS100L (PC uplink from below)

Fig. 6.2. Test stand MZGS200L

 6.1 Test Stands 115

a) b)
specimen

αM
rotary
table αM

Mτ
Mσ
specimen
holder arm
M
M

Fig. 6.3. Rotary table of MZGS100L stand with fixed specimen (a) and graphical
representation of moment of force M resolved into bending moment Mσ and torsional
moment Mτ of specimen (b)

PC + card power amplifier frame and actuator

Fig. 6.4. Diagram of connections of components of MZGS100L stand for fatigue

testing of specimens under proportional bending and torsion

The measurement of strain at the lever with a strain gauge indicates the
instantaneous values of moment M . The voltage at strain bridge of measuring
card is scaled on the basis of equation of calibration line. The calibration is
performed by the static method by application of weight one after another to
the lever and recording the respective voltage. On the basis of the length of
the lever the relation between voltage and moment M is determined.
The test stand MZGS200L is a device designed for fatigue testing of ma-
terial specimens under the non-proportional bending and torsional moments
with arbitrary and independent random histories rMσ ,Mτ = −1 ÷ 1. In con-
trast to the MZGS100L stand it is equipped with two independently controlled
and supplied electromagnetic actuators. The specimen is fixed to the table and
there is no possibility of rotation (Fig. 6.5a). The bending moment Mσ and
torsional moment Mτ are equal to the moments from a parallel and perpen-
116 6 Experimental Studies

dicular lever to the axis of the specimen (Fig. 6.5b). The diagram of the stand
is presented in Fig. 6.6.

a) b)
stationary
table specimen π/2

Mτ

specimen
holder
Mσ
π/2 arm Mσ

arm Mτ

Fig. 6.5. Specimen fixing on table of MZGS200L stand with the angle π/2 between
arms (a) and graphical interpretation of bending moment Mσ and torsional moment
Mτ of specimen (b)

PC + card power amplifiers frame and actuators

Fig. 6.6. Diagram of connections of components of MZGS200L stand for fatigue

testing of specimens under non-proportional bending and torsion

The operation and control of the MZGS200L stand is similar as in the

case of MZGS100L. The differences result from the necessity of control by
means of two electromagnetic actuators and the measurement of the signals
from two arms simultaneously. Each of the actuators is attached with a limit
switch. The connection of the switches ensures both amplifiers to switch off
 6.2 Material and Geometry of Tested Specimens 117

in the case when one of the arms exceeds the maximum level of inclination.
The bending and torsional arms are statically calibrated and the obtained
characteristics served for the determination of adequate moments.

6.2 Material and Geometry of Tested Specimens

Fatigue testing is performed on smooth round specimens made of steel 18G2A.
A ground bar is selected as the representative input material for testing. As
a result of the assumption that fatigue testing involves marketed structural
material (in particular with reference to cast quality) there was not prelim-
inary material selection process. Chemical composition is presented in Table
6.1. The quality of the material complies with the Polish standard of steel
products PN-86/H-84018.

Table 6.1. Chemical composition of steel 18G2A (%)

C Mn Si P S Cr Ni Cu Fe
0.21 1.46 0.42 0.019 0.046 0.09 0.04 0.17 rest

The specimens are made in accordance with the current standard PN-
76/H-04326. Grinding ensures an adequate quality of the mid-surface. The
minimum specimen diameter is established with reference to particular tests
in the range 6.4-7.5 mm. The shape and dimensions of the specimens are
presented in Fig. 6.7. The essential fatigue and strength characteristics of
steel 18G2A are presented in Table 5.1 in the chapter 5 devoted to computer
simulation.
R 45

Ø8h7

12.9

30 30

90

Fig. 6.7. Shape and dimensions of specimens in fatigue test

118 6 Experimental Studies

6.3 Testing under Narrow-Band Frequency Loading

Fatigue investigation under narrow-band Gaussian loading is performed on
test stands MZGS100L and MZGS200L. The testing uses a generator of
narrow-band frequency history applied during computer simulation. A section
of history and the fundamental characteristics are summarized in Fig. 5.2 in
the chapter 5. Testing is performed under various value of correlation coeffi-
cients rσ,τ between the history of normal stress from bending σ(t) and shear
stress from torsion τ (t) and various variance of μσ and μτ . The results are
presented in Table 6.2.
Histories of normal stress σ(t) and shear stress τ (t) are nominal quantities
derived from formulae:
Mσ (t)
σ(t) = , (6.1)
Wx
Mτ (t)
τ (t) = , (6.2)
W0
πd3
where: Wx = – section modulus of bending,
32
πd3
W0 = – section modulus of torsion,
16
d – specimen diameter.
For the case of stand MZGS200L the histories of bending moment Mσ (t)
and torsional moment Mτ (t) are direct measure on the appropriate levers. For
the case of stand MZGS100L the measurement involves the resulting moment
M (t).On the basis of the resulting moment the determination of histories of
normal and shear stresses are possible with input of angle αM :
M (t) cos(αM )
σ(t) = , (6.3)
Wx
M (t) sin(αM )
τ (t) = . (6.4)
W0
The fatigue testing under bending, torsion and proportional bending and
torsion with correlation coefficient of components rσ,τ = 1 are performed en-
tirely on test stand MZGS100L. The selection of the stand provides with a
guarantee of constant correlation coefficient and the simple structure ensures
reliability and controllability. The stand MZGS200L serves for performing
fatigue testing under nonproportional bending and torsion with correlation
coefficient rσ,τ = 1. For the purposes a two-channel generator of random his-
tories with controlled correlation coefficient is prepared. The testing takes
advantage of the fact that two various histories x1 (t) and x2 (t) on genera-
tor output exhibit lack of correlation, i.e. rx1 ,x2 ≈ 0. By summation of the
histories under an appropriate relation a resulting one x3 (t) is obtained in
accordance with the relation
 6.4 Testing under Broad-Band Frequency Loading 119

x3 (t) = x1 (t) + ar x2 (t) , (6.5)

where: x1 (t) and x2 (t) – non-correlated histories rx1 ,x2 ≈ 0,

ar – coefficient for determination of correlation.
The following steps involve the selection of ar coefficient in a way that
ensures the required correlation of histories x1 (t) and x3 (t), which are scaled
for the same variance

μx1
x3 (t) = [x1 (t) + ar x2 (t)] . (6.6)
μx3

Eight fatigue tests were performed and marked with symbols as follows:
N01 – fatigue testing under pure bending, MZGS100L stand, angle between
axis of specimen and lever αM = 0◦ , 11 specimens,
N02 – fatigue testing under pure torsion, MZGS100L stand, angle between
axis of specimen and lever αM = 90◦ , 5 specimens,
N03 – fatigue testing under nonproportional bending with torsion, MZGS200L
stand, correlation coefficient rσ,τ ≈ 0, relation of standard deviation
√ √
μσ / μτ ≈ 2, 6 specimens,
N04 – fatigue testing under nonproportional bending with torsion, MZGS200L
stand, correlation coefficient rσ,τ ≈ 0, relation of standard deviation
√ √
μσ / μτ ≈ 1, 5 specimens,
N05 – fatigue testing under proportional bending with torsion, MZGS100L
stand, angle between axis of specimen and lever αM = 45◦ , correlation
√ √
coefficient rσ,τ = 1, relation of standard deviation μσ / μτ ≈ 2, 8
specimens,
N06 – fatigue testing under proportional bending with torsion, MZGS100L
stand, angle between axis of specimen and lever αM = 63.5◦ , correla-
√ √
tion coefficient rσ,τ = 1, relation of standard deviation μσ / μτ ≈ 1,
5 specimens,
N07 – fatigue testing under nonproportional bending with torsion, MZGS200L
stand, correlation coefficient rσ,τ ≈ 0.5, relation of standard deviation
√ √
μσ / μτ ≈ 1, 4 specimens,
N08 – fatigue testing under nonproportional bending with torsion, MZGS200L
stand, correlation coefficient rσ,τ ≈ 0.5, relation of standard deviation
√ √
μσ / μτ ≈ 2, 5 specimens.
The details of standard deviation of histories σ(t) and τ (t) and results of
fatigue life are summarised in Table 6.2.

6.4 Testing under Broad-Band Frequency Loading

Testing under wide band frequency spectrum Gaussian loading was performed
on MZGS100L by Achtelik [47]. A section of the registered history along with
 120 6 Experimental Studies

a)
40
σ(t), MPa

20

−20

−40
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
time t, s
b) c)
0.05 1000

Gσ (f ), MPa2 s
pσ (σ), MPa−1

0.04 800

0.03 600

0.02 400

0.01 200

0 0
−40 −20 0 20 40 0 20 40 60 80
σ, MPa f , Hz

Fig. 6.8. A section of stress history with broad-band frequency (a) with its proba-
bility density distribution function (b) and power spectral density function (c)

the probability density function and power spectral density are presented in
Fig. 6.8.
The fatigue testing is divided into three groups marked with symbols as
follows:
A01 – fatigue testing under pure bending, angle between axis of specimen
and lever αM = 0◦ , 16 specimens,
A02 – fatigue testing under proportional bending with torsion, angle between
axis of specimen and lever αM = 45◦ , relation of standard deviation
√ √
μσ / μτ ≈ 2, 16 specimens,
A03 – fatigue testing under pure torsion, angle between axis of specimen and
lever αM = 90◦ , 16 specimens.
The details of standard deviation of histories σ(t) and τ (t) and results of
fatigue life are summarized in Table 6.2.

6.5 Comparison of Calculated and Experimental Fatigue

Life

The fatigue life is computed with the cycle counting and spectral methods.
The calculation is performed similarly as in the simulations with the exception
of blocks PWO1, PWO2, PWO3 applied for comparison. The diagram of the
algorithm of life time computation is presented Fig. 6.9.
 6.5 Comparison of Calculated and Experimental Fatigue Life 121

Cycle counting method

Registration of
1. σ(t) and τ (t)

Determination of critical plane

2. position lη , mη , nη

Calculation of equivalent stress

3. σeq (t)

Cycle counting by
4. rain flow algorithm

Summation of damage and calculation

5. of fatigue life TZC

Spectral method

Calculation of Gσ (f )
1. on the basis σ(t) and τ (t)

Determination of critical plane

2. position lη , mη , nη

Calculation of
3. Gσeq (f )

Determination of moments mk
4. and coefficient bSK

Calculation of fatigue life

5. TSP

Fig. 6.9. Algorithm of fatigue life computation by cycle counting and

spectral methods
122 6 Experimental Studies

In order to simplify further calculations it is assumed that the consider-

ations involve plain state of stress for which case the vector of tensor com-
ponents takes the form σ(t) = [σ(t), 0, τ (t)]. Random histories of nominal
stresses resulting from bending σ(t) and torsion τ (t) determined from (6.1)
and (6.3) constitute the basis for the calculations. The modified criterion of
maximum normal stress in the critical plane [48] is applied for the determi-
nation of the history of equivalent stress, which takes the form
σaf
σeq (t) = lη2 σ(t) + 2lη mη τ (t) . (6.7)
τaf

The modification involves the consideration of the difference of levels of

fatigue limits of bending σaf and torsion τaf . For the plain state of stress
the position of the critical plane could be defined ' as a function of a single
direction cosines, as for nη = 0 we obtain mη = 1 − lη2 . It is a remark,
which assists in the calculations and reduces its duration. Table 6.2 presents
the value of direction cosines lη of vector normal to the critical plane defined
by the maximum variance method. The determination of critical plane in time
and frequency domains leads to identical results.
The equivalent stress history established from (6.7) is submitted to cycle
counting process with the rain flow algorithm. The derived cycles serve for
damage accumulation by the linear Serensen-Kogayev hypothesis (3.79) [51,
88, 95]. The fatigue life formula is derived in the form

To bSK A
TZC = for σai ≥ aSK σaf . (6.8)
k
m
ni σai
i=1

Up to the present the Serensen-Kogayev hypothesis of damage accumu-

lation has not been applied in the spectral method for correlation of exper-
imental data under random loading with broad-band frequency spectrum.
Therefore, it is necessary to derive a formula for fatigue life TSP under the
assumption that loading history has normal distribution with broad-band fre-
quency spectrum. The general notation for this case is following
1
TSP = TN B , (6.9)
λ
where: λ – coefficient for accounting the effect of width of band frequency
of loading on fatigue life [103],
TN B – fatigue life under the assumption of narrow-band frequency
spectrum and Serensen-Kogayev damage accumulation hypoth-
esis (3.82).
However, the fact that bSK coefficient depends on the variance of the
equivalent history μσ and expected maximum of cycle amplitude σa max must
 6.5 Comparison of Calculated and Experimental Fatigue Life 123

be considered. The quantities are different for each of the performed tests;
therefore, bSK coefficient is determined separately for each case. The diffi-
culty consists in the determination of maximum of cycle amplitude σa max
directly from power spectral density of equivalent stress history. In practice,
the maximum amplitude of a cycle has a limited value. In the investigated
case, following a number of laboratory tests and statistical processing the
following formula is derived
√
σa max = 3.73 μσ . (6.10)

The results of fatigue life calculations by the cycle counting method TZC
(6.8) and spectral method TSP (6.9) are summarised in Table 6.2.
Table 6.2: Results of fatigue testing under combination of bending with torsion and calculated fatigue life
124

√ √ 
Symbol I rσ,τ μσ μτ μσ /μτ Texp TZC TSP lη
and number MPa MPa s s s

1 2 3 4 5 6 7 8 9 10
N01 1 0.99 – 150.6 – – 13560 16605 14777 0
2 0.99 – 127.3 – – 55500 54882 48154 0
3 0.99 – 149.4 – – 9660 17520 15583 0
6 Experimental Studies

4 0.99 – 113.4 – – 195720 125910 109170 0

5 0.99 – 133.2 – – 43260 39591 34879 0
6 0.99 – 134.4 – – 43080 37275 32857 0
7 0.99 – 134.4 – – 35340 37275 32857 0
8 0.99 – 139.8 – – 21240 28140 24911 0
9 0.99 – 111.1 – – 316200 145560 126042 0
10 0.99 – 133.0 – – 33660 40215 35407 0
11 0.99 – 128.1 – – 62100 52367 45942 0
N02 1 0.99 – – 66.0 – 448020 267681 230396 0.707
2 0.99 – – 63.0 – 653760 369358 316066 0.707
3 0.99 – – 63.0 – 226800 369358 316066 0.707
4 0.99 – – 64.0 – 352200 324421 277995 0.707
5 0.99 – – 65.0 – 219360 285198 245131 0.707
N03 1 0.99 -0.01 96.8 54.2 1.79 191880 193105 194656 0.883
2 0.99 -0.01 96.8 54.2 1.79 144240 193105 194656 0.883
3 0.99 -0.01 118.3 66.2 1.79 30600 46724 46797 0.883
4 0.99 -0.01 118.3 66.3 1.78 35100 46724 46797 0.883
5 0.99 -0.01 106.5 59.6 1.79 118980 98309 99046 0.883
 Table 6.2: (continuation)

1 2 3 4 5 6 7 8 9 10

6 0.99 -0.01 104.0 58.2 1.79 141300 116763 117660 0.883

N04 1 0.99 -0.01 57.9 56.2 1.03 223320 470001 435709 0.766
2 0.99 -0.01 57.9 56.2 1.03 416400 470001 435709 0.766
3 0.99 -0.01 57.9 56.2 1.03 295740 470001 435709 0.766
4 0.99 -0.01 63.7 61.8 1.03 168360 237306 219352 0.766
5 0.99 -0.01 57.9 56.2 1.03 303660 470001 435709 0.766
N05 1 0.99 1 108.8 54.4 2 20700 13739 11906 0.866
2 0.99 1 108.0 54.0 2 22800 14517 12581 0.866
3 0.99 1 107.4 53.7 2 20580 15073 13056 0.866
4 0.99 1 96.3 48.1 2 41820 32628 28040 0.866
5 0.99 1 100.0 50.0 2 25020 24965 21526 0.866
6 0.99 1 89.2 44.6 2 81720 56314 48072 0.866
7 0.99 1 92.0 46.0 2 92280 45113 38549 0.866
8 0.99 1 79.8 39.9 2 380220 124908 105339 0.866
N06 1 0.99 1 61.8 61.8 1 134700 41552 35563 0.788
2 0.99 1 61.8 61.8 1 155520 41552 35563 0.788
3 0.99 1 61.8 61.8 1 99600 41552 35563 0.788
4 0.99 1 61.8 61.8 1 148740 41552 35563 0.788
5 0.99 1 61.8 61.8 1 167040 41552 35563 0.788
N07 1 0.99 0.5 71.7 69.6 1.03 48540 32955 30944 0.766
6.5 Comparison of Calculated and Experimental Fatigue Life

2 0.99 0.5 64.3 62.4 1.03 205140 70376 66190 0.766

3 0.99 0.5 63.4 61.5 1.03 72600 77879 73365 0.766
125
 Table 6.2: (continuation)
126

1 2 3 4 5 6 7 8 9 10

4 0.99 0.5 63.4 61.5 1.03 119700 77879 73365 0.766

N08 1 0.99 0.5 118.5 64.0 1.85 20940 12817 13073 0.848
2 0.99 0.5 100.7 54.4 1.85 86820 37959 38935 0.848
3 0.99 0.5 92.3 49.8 1.85 399000 69310 71383 0.848
4 0.99 0.5 101.3 54.7 1.85 133200 35816 36708 0.848
6 Experimental Studies

5 0.99 0.5 100.8 54.4 1.85 78180 37959 38935 0.848

A01 1 0.78 – 240.9 – – 5700 5312 3335 0
2 0.78 – 240.9 – – 4920 5312 3335 0
3 0.78 – 240.9 – – 6480 5312 3335 0
4 0.78 – 189.6 – – 17880 12856 9803 0
5 0.78 – 189.6 – – 17160 12856 9803 0
6 0.78 – 189.6 – – 14520 12856 9803 0
7 0.78 – 142.2 – – 19860 44097 34716 0
8 0.78 – 142.2 – – 19560 44097 34716 0
9 0.78 – 142.2 – – 27960 44097 34716 0
10 0.78 – 93.1 – – 98520 267684 256509 0
11 0.78 – 93.1 – – 114360 267684 256509 0
12 0.78 – 93.1 – – 173820 267684 256509 0
13 0.78 – 93.1 – – 218940 267684 256509 0
14 0.78 – 69.1 – – 330000 1476985 1339508 0
15 0.78 – 69.1 – – 330000 1476985 1339508 0
16 0.78 – 69.1 – – 330000 1476985 1339508 0
A02 1 0.78 1 170.4 85.2 2 3060 2132 1151 0.875
 Table 6.2: (continuation)

1 2 3 4 5 6 7 8 9 10

2 0.78 1 170.4 85.2 2 1920 2132 1151 0.875

3 0.78 1 170.4 85.2 2 3900 2132 1151 0.875
4 0.78 1 134.0 67.0 2 6600 6557 4710 0.875
5 0.78 1 134.0 67.0 2 10560 6557 4710 0.875
6 0.78 1 117.6 58.8 2 13980 14381 10072 0.875
7 0.78 1 117.6 58.8 2 15480 14381 10072 0.875
8 0.78 1 117.6 58.8 2 18900 14381 10072 0.875
9 0.78 1 100.6 50.3 2 39360 31716 24860 0.875
10 0.78 1 100.6 50.3 2 26040 31716 24860 0.875
11 0.78 1 100.6 50.3 2 29880 31716 24860 0.875
12 0.78 1 100.6 50.3 2 22320 31716 24860 0.875
13 0.78 1 83.4 41.7 2 108300 96521 73569 0.875
14 0.78 1 83.4 41.7 2 143700 96521 73569 0.875
15 0.78 1 65.8 32.9 2 330000 309305 291157 0.875
16 0.78 1 65.8 32.9 2 300000 309305 291157 0.875
A03 1 0.78 – – 97.9 – 5100 21498 9250 1
2 0.78 – – 97.9 – 6000 21498 9250 1
3 0.78 – – 86.2 – 10920 43686 22064 1
4 0.78 – – 86.2 – 15960 43686 22064 1
5 0.78 – – 86.2 – 18600 43686 22064 1
6.5 Comparison of Calculated and Experimental Fatigue Life

6 0.78 – – 77.1 – 43740 83505 51868 1

7 0.78 – – 77.1 – 35160 83505 51868 1
8 0.78 – – 77.1 – 41100 83505 51868 1
127
 Table 6.2: (continuation)
128

1 2 3 4 5 6 7 8 9 10

9 0.78 – – 67.6 – 103200 222159 132326 1

10 0.78 – – 67.6 – 159000 222159 132326 1
11 0.78 – – 67.6 – 62400 222159 132326 1
12 0.78 – – 62.8 – 117300 317010 226201 1
13 0.78 – – 62.8 – 276180 317010 226201 1
6 Experimental Studies

14 0.78 – – 62.8 – 205320 317010 226201 1

15 0.78 – – 62.8 – 300000 317010 226201 1
16 0.78 – – 57.8 – 330000 596652 401620 1

I – irregularity coefficient,
r – correlation coefficient,
√σ,τ
μσ – standard deviation of stress resulting from bending,
√
μτ – standard deviation of stress resulting from torsion,
Texp – experimental fatigue life,
TZC and TSP – fatigue life calculated with cycle counting and spectral method, respectively,
lη – direction cosines of vector normal to critical plane in relation to x axis of specimen.
 6.5 Comparison of Calculated and Experimental Fatigue Life 129

The comparison of calculated fatigue life TZC and TSP with experiment
results is presented in Figs 6.10–6.16. The quantities on the axes of the figures
are presented in a logarithmic order. The dashed lines mark the scatter band
with coefficient 3 which indicates that the points within the band result from
the comparison of two quantities for which the equation 1/3 < y/x < 3
is satisfied. The diagonal (full line) defines an ideal case for which the two
compared quantities y/x = 1.
For the case of narrow-band frequency loading under pure bending (N01),
pure torsion (N02) in Fig. 6.10 and nonproportional bending with torsion
with correlation coefficient rσ,τ ≈ 0 (N03 and N04) in Fig. 6.11 the results are
contained in the acceptable scatter band. In Figs 6.12 and 6.13 a comparison
is made between the experimental and computed results for correlation coef-
ficient rσ,τ = 1 (N05 and N06) and rσ,τ ≈ 0.5 (N07 and N08). For the latter
case the fatigue life calculated with the cycle counting method and spectral
method is lower with reference to experimental life.
The results of fatigue life for loading with broad-band frequency spectrum
are compared with experiments in Fig. 6.14 and 6.16 For this cases the best
equivalence of results is obtained for the case of life calculated by the spectral
method (A03 – pure torsion).
During fatigue testing under the combination of bending with torsion per-
formed for various ratios of standard deviation μσ /μτ ≈ 1 and ≈ 2 the
effect of the ratio on the difference of calculated TZC , TSP and experimental
life Texp is not observed.
The fatigue life results TZC and TSP calculated for narrow-band frequency
scpectrum (N01–N08) are similar with the maximum difference up to 16%. For
the histories with broad-band frequency spectrum (A01–A03) the maximum
differences amount to higher values (up to 43%), which could be associated
with the application of misfiled coefficient λ accounting for the frequency
width spectrum of stress.
In general, the spectral method tends to underestimation of fatigue life
in comparison to the cycle counting method. It is commonly considered as
an advantage of the method [93] due to the estimation of shorter fatigue life.
Therefore, the term safety error is used.
130 6 Experimental Studies

6
10
TZC , N01
TSP , N01
TZC , N02
TSP , N02

Fig. 6.10. TZC , TSP , s 5

Comparison of 10
fatigue life cal-
culated by cycle
counting TZC
and spectral TSP
methods with 3
experimental life 3
Texp under pure
bending N01 end
4
pure torsion N02 10
4 5 6
(narrow-band fre- 10 10 10
quency spectrum) Texp , s

6
10
TZC , N03
TSP , N03
TZC , N04
TSP , N04

Fig. 6.11.
TZC , TSP , s

Comparison of 5
fatigue life cal- 10
culated by cycle
counting TZC
and spectral TSP
methods with
experimental life 3
Texp under nonpro-
portional bending 3
with torsion N03
4
and N04 (rσ,τ ≈ 0) 10
4 5 6
(narrow-band fre- 10 10 10
quency spectrum) Texp , s
 6.5 Comparison of Calculated and Experimental Fatigue Life 131

6
10
TZC , N05
TSP , N05
TZC , N06
TSP , N06 3

3
Fig. 6.12.
TZC , TSP , s

5 Comparison of
10 fatigue life cal-
culated by cycle
counting TZC
and spectral TSP
methods with
experimental life
Texp under pro-
portional bending
with torsion N05
4
10 and N06 (rσ,τ = 1)
4 5 6
10 10 10 (narrow-band fre-
Texp , s quency spectrum)

6
10
TZC , N07
TSP , N07
TZC , N08
TSP , N08 3

3
Fig. 6.13.
TZC , TSP , s

Comparison of
5 fatigue life cal-
10 culated by cycle
counting TZC
and spectral TSP
methods with
experimental
life Texp under
nonproportional
bending with
torsion N07 and
4
10
4 5 6
N08 (rσ,τ ≈ 0.5)
10 10 10 (narrow-band fre-
Texp , s quency spectrum)
132 6 Experimental Studies

6
10

T , A01
ZC
TSP , A01

5
10

TZC , TSP , s
Fig. 6.14.
Comparison of
fatigue life cal- 10
4

culated by cycle
counting TZC
and spectral TSP 3
methods with ex- 3
perimental life Texp
3
under pure bending 10
3 4 5 6
A01 (broad-band 10 10 10 10
frequency spectrum) Texp , s

6
10

TZC , A02
TSP , A02
5
10
Fig. 6.15.
TZC , TSP , s

Comparison of
fatigue life cal-
culated by cycle
counting TZC 4
and spectral TSP 10
methods with
experimental life 3
3
Texp under pro-
portional bending
with torsion A02
3
(rσ,τ = 1) (broad- 10
3 4 5 6
band frequency 10 10 10 10
spectrum) Texp , s
 6.5 Comparison of Calculated and Experimental Fatigue Life 133

6
10

TZC , A03
TSP , A03
5
10
TZC , TSP , s

4
10

3
3

3
10
3 4 5 6
10 10 10 10
Texp , s

Fig. 6.16. Comparison of fatigue life calculated by cycle counting TZC and spec-
tral TSP methods with experimental life Texp under pure torsion A03 (broad-band
frequency spectrum)
7
Conclusions

1. A generalised spectral method for fatigue life calculation of materials un-

der multiaxial random loading defined in frequency domain expressed in
terms of power spectral density functions is postulated.

2. The generalised spectral method is based on linear strain and stress criteria
of fatigue failure which enable the reduction of multiaxial stress state to
the equivalent uniaxial one.

3. In the equivalent uniaxial stress state fatigue damage is accumulated with

use of standard material characteristics determined in stresses for high-
cycle fatigue regime and in strains for high- and low-cycle fatigue life
regimes.

4. Calculations and analyses of various stimulated random stress and strain

states indicate that:
4.1. The application of linear criteria of multiaxial fatigue failure for
the determination of fatigue life in frequency domain by the spec-
tral method gives practically identical results as the cycle counting
method in time domain.
4.2. Equivalent results are gained by the variance method for the deter-
mination of the expected critical plane position based on random
histories and the power spectral density functions.

5. Fatigue testing of 18G2A steel specimens under proportional and non-

proportional random bending with torsion under narrow- and broad-band
frequency spectra and zero expected values indicate that:
5.1. Fatigue life calculated with the generalised spectral method, whose
power spectral density of the equivalent stress is carried out by means
of the modified criterion of maximum normal stress in the critical
plane and damage accumulation is determined in accordance with
136 7 Conclusions

Serensen-Kogayev hypothesis, extends into the acceptable scatter

band with coefficient 3 determined by experimental life.
5.2. Similar fatigue life is gained from the generalised spectral method
and the range counting method. The latter one is more time con-
suming.

6. The postulated spectral method offers considerable prospects of its ap-

plicability in association with the finite element method for mapping life
time of structural materials under multiaxial service loading.
References

1. ASTM E 1049–85 (Reapproved 1997): Standard practices for cycle counting in

fatigue analysis, in: Annual Book of ASTM Standards, Vol. 03.01, Philadelphia
1999, pp. 710–718
2. Banvillet A., L agoda T., Macha E., Nieslony A., Palin-Luc T., Vittori J.-F.:
Fatigue life under non-Gaussian random loading from various models, Interna-
tional Journal of Fatigue, Vol. 26, No 4, 2004, pp. 349-363
3. Benasciutti D.: Fatigue Analysis of Random Loadings, PhD Thesis, Department
of Engineering, University of Ferrara, Italy, 2005, pp. 254
4. Benasciutti D., Tovo R.: Modelli di previsione del danneggiamento a fatica per
veicoli in regime stazionario ed ergodico, Associazione Italiana per I’Analisi
delle Sollecitazioni (AIAS), XXXI Convegno Nazionale – 18- 21 Settembre 2002,
Parma, p. 10
5. Benasciutti D., Tovo R.: Spectral methods for lifetime prediction under wide-
band stationary random processes, Cumulative Fatigue Damage, Seville, 2003,
p. 5
6. Bendat J.S., Piersol A.G.: Random Data. Analysis and Measurement Proce-
dures, John Wiley and Sons Inc., New York, 1976
7. Bendat J.S., Piersol A.G.: Engineering Applications of Correlation and Spectral
Analysis, John Wiley and Sons Inc., New York, 1980
8. Bȩdkowski W., Lachowicz C., Macha E.: Predicted fatigue fracture plane ac-
cording to variance method of shear stress under random triaxial stress state,
Failure Analysis – Theory and Practice, Proceedings of the 7th European Con-
ference on Fracture (ECF 7), Budapest 1988, Ed. E. Czoboly, EMAS U.K.,
Vol. I, pp. 281–283
9. Bȩdkowski W., Macha E.: Fatigue fracture plane under multiaxial random load-
ings – prediction by variance of equivalent stress based on the maximum shear
and normal stress, Mat. –wiss. u. Werkstofftech., Vol. 23, 1992, pp. 82–94
10. Bı́lý M., Bukoveczky J.: Digital simulation of environmental processes with re-
spect to fatigue, Journal of Sound and Vibration, Vol. 49, No. 4, 1976, pp. 551–
568
138 References

11. Bishop N.W.M.: Spectral method of estimating the integrity of structural compo-
nents subjected to random loading, Handbook of Fatigue Crack Propagation in
Metallic Structures, Ed. An. Carpinteri, Elsevier Science B.V., 1994, pp. 1685–
1720
12. Bishop N.W.M., Sherratt F.: Fatigue life prediction from power spectral density
data. Part 2: recent developments, Environmental Engineering, Vol. 2, 1989,
pp. 11–19
13. Bishop N.W.M., Sherratt F.: A theoretical solution for the estimation of ”rain-
flow“ ranges from power spectral density data, Fatigue and Fracture of Engi-
neering Materials and Structures, Vol. 13, No. 4, 1990, pp. 311–326
14. Bitner- Gregersen E.M., Cramer E.H.: Uncertainties of load characteristics and
fatigue damage of ship structures, Marine Structures, Vol. 8, 1995, pp. 97–117
15. Bolotin V.V.: Random Vibration of Elastic Systems, Martinus Nijhoff Publish-
ers, Hague, The Netherlands, 1984, p. 480
16. Bruder T., Heuler P., Klätschke H., Störzel K.: Analysis and synthesis of stan-
dardized multi-axial load-time histories for structural durability assessment,
Proc. of the Seventh Int. Conf. on Biaxial/Multiaxial Fatigue and Fracture,
Berlin, Germany, pp. 63–77
17. Chaudhury G.K., Dover W.D.: Fatigue analysis of offshore platforms subject
to sea wave loadings, International Journal of Fatigue, Vol. 7, No. 1, 1985,
pp. 13–19
18. Chow C.L., Li D.L.: An analytical solution for fast fatigue assessment under
wide-band random loading, International Journal of Fatigue, 1991, pp. 395–404
19. Clevenson S.A., Stainer R.: Fatigue Life under Random Loading for Several
Power Spectral Shapes, NASA TR R- 266, Langley Research Center, 1983,
p. 18
20. Corten H.T., Dolan T.L.: Cumulative fatigue damage, International Conference
on Fatigue of Metals, ASME and IME, London, 1956, p. 235–246
21. Dirlik T.: Application of Computers in Fatigue Analysis, PhD Thesis, UK:
University of Warwick, 1985
22. Dowling N.E.: Fatigue failure prediction for complicated stress- strain histories,
Journal of Materials, Vol. 7, No. 1, 1972, pp. 71–87
23. Etube L.S., Brennan F.P., Dover W.D.: Modelling of jack- up response for
fatigue under simulated service condition, Marine Structures, Vol. 12, 1999,
pp. 327–348
24. Folsø R.: Spectral fatigue damage calculation in the side shells of ships, with
due account taken of the effect of alternating wet dry areas, Marine Structures,
Vol. 11, 1998, pp. 319–343
25. Friis Hansen P., Winterstein S.R.: Fatigue damage in the side shells of ships,
Marine Structures, Vol. 8, 1995, pp. 631–655
26. Fu T.-T., Cebon D.: Predicting fatigue lives for bi- modal stress spectral densi-
ties, International Journal of Fatigue, Vol. 22, 2000, pp. 11–21
27. Gasiak G., Pawliczek R.: Calculation of fatigue life of specimens made of 18G2A
steel under variable bending and torsion loading with different stress ratios,
 References 139

Scientific Papers of Technical University of Opole, Vol. 66, No. 268, Opole
2001, pp. 23–44, (in Polish)
28. Grzelak J., L
 agoda T., Macha E.: Spectral analysis of the criteria for multiaxial
random fatigue, Mat. –wiss. u. Werkstofftech., Vol. 22, 1991, pp. 85-98
29. Haibach E.: Betriebsfestigkeit, Verfaren und Daten zur Bauteilberechnung, VDI,
Düsseldorf, 1989, p. 141
30. Hartt W.H., Lin N.K.: A proposed stress history for fatigue testing applicable
to offshore structures, International Journal of Fatigue, Vol. 8, No. 2, 1986,
pp. 91–93
31. Heuler P., Birk O., Beste A.: Auf dem Weg zur virtuellen Festigkeit im Fahrzeug-
bau, Mat. –wiss. u. Werkstofftech., Vol. 32, 2001, pp. 369–376
32. Holmes J.D: Fatigue life under along- wind loading — closed- form solutions,
Engineering Structures, Vol. 24, 2002, pp. 109–114
33. Jiao G.: A theoretical model for the prediction of fatigue under combined Gaus-
sian and impact load, International Journal of Fatigue, Vol. 17, No. 3, 1995,
pp. 215–219
34. Johannesson P.: Rainflow Analysis of Switching Markov Loads, Mathematical
Statistics, Centre for Mathematical Sciences, Lund Institute of Technology,
Sweden, Doctoral Theses in Mathematical Sciences 1999:4, ISBN 91-628-3784-
2, p. 195
35. Kam J.C.P., Dover W.D.: Mathematical background for applying multiple axes
random stress histories in the fatigue testing of offshore tubular joints, Inter-
national Journal of Fatigue, Vol. 11, No. 5, 1989, pp. 319–326
36. Karolczuk A., Macha E.: A review of critical plane orientations in multiaxial
fatigue failure criteria of metallic materials, International Journal of Fracture,
Vol. 134, 2005, pp. 267–304
37. Korn G.A., Korn T.M.: Mathematical Handbook , 2nd ed., McGraw- Hill Book
Company, New York, 1968
38. Kowalewski J.: On the Relationship between Component Life under Irregularly
Fluctuating and Ordered Load Sequences, Part 2 , DVL Report 249, 1963
39. Lachowicz C.T., L agoda T., Macha E.: Comparison of analytical and algorith-
mical methods for life time estimation in 10HNAP steel under random loadings,
FATIGUE ’96, Proc. of the Sixth Int. Fatigue Congress, (6IFC) Berlin, Eds G.
Lütjering and H. Nowack, Pergamon 1996, Vol. 1, pp. 595–600
40. Larsen C.E., Lutes L.D.: Predicting the fatigue life of offshore structures by the
single- moment spectral method, Probabilistic Engineering Mechanics, Vol. 6,
No. 2, 1991, pp. 96–108
41. Leser Ch., Juneja L., Thangjitham S., Dowling N.E.: On multi- axial random
fatigue load modeling, SAE Technical Paper Series, No. 980696, Reprinted from:
Advancements in Fatigue Research and Applications (SP- 1341), 1998, pp. 93–
106
42. Liou H.Y., Wu W.F., Shin C.S.: A modified model for the estimation of fatigue
life derived from random vibration theory, Probabilistic Engineering Mechanics,
Vol. 14, 1999, pp. 281–288
140 References

43. Liu H.J., Hu S.R.: Fatigue under nonnormal random stresses using Monte-
Carlo method , FATIGUE’ 87, Eds. R.O. Ritche and E.A. Starke Jr. EMAS
(U.K.), Vol. III, 1987, pp. 1439–1448
44. Lü P.M., Jiao S.J.: An improved method of calculation the peak stress distri-
bution for a broad-band random process, Fatigue and Fracture of Engineering
Materials and Structures, Vol. 23, 2000, pp. 581–586
45. Lutes L.D., Larsen C.E.: Improved spectral method for variable amplitude fa-
tigue prediction, Journal of Structural Division, Vol. 116, No. 4, 1990, pp. 1149–
1164
46. L
 agoda T., Macha E.: Transformation of stress frequency in criteria for multiax-
ial random fatigue, Fracture Behaviour and Design of Materials and Structures,
(ECF 8), Ed. D. Firrao, EMAS (U.K.), Vol. III, 1990, pp. 1311–1316
47. L
 agoda T., Macha E., Achtelik H., Karolczuk A., Nieslony A., Pawliczek R.:
Multiaxial Random Fatigue of Machine Elements and Structures, Part IV, En-
ergy Approach to Fatigue Life Including the Stress Gradients, Studies and
Monographs 139, Technical University of Opole ,Opole 2002, 102 ps, (in Polish)
48. L
 agoda T., Macha E., Dragon A., Petit J.: Influence of correlations between
stresses on calculated fatigue life on machine elements, International Journal
of Fatigue, Vol. 18, No. 8, 1996, pp. 547–555
49. L
 agoda T., Macha E., Nieslony A.: Fatigue life calculations by means of the
cycle counting and spectral methods under multiaxial random loading, Fatigue
and Fracture of Engineering Materials and Structures, Vol. 28, 2005, pp. 409-
420
50. L
 agoda T., Macha E., Nieslony A., Müller A.: Fatigue life of cast irons GGG40,
GGG60 and GTS45 under variable amplitude tension with torsion, The Archive
of Mechanical Engineering, Vol. XLVIII, 2001, pp. 55–69
51. L
 agoda T., Macha E., Nieslony A., Müller A.: Comparison of calculation and
experimental fatigue lives of some chosen cast irons under combined tension
and torsion, Proc. ECF13, Application and Challenges, San Sebastian 2000
(Spain), Eds. M.Fuentes et al., Elsevier, Abstract Volume pp. 256–256, CD-
Rom, 6 ps
52. L
 agoda T., Macha E., Pawliczek R.: The influence of the mean stress on fatigue
life of 10HNAP steel under random loading, International Journal of Fatigue,
Vol. 23, No. 4, 2001, pp. 283–291
53. Macha E.: Mathematical Models of the Life to Fracture for Materials Subjected
to Random Complex Stress Systems, Scientific Papers of the Institute of Mate-
rials Science and Applied Mechanics of Wroclaw Technical University, No. 41,
Wroclaw 1979, 99 ps, (in Polish)
54. Macha E.: Generalization of Strain Criteria of Multiaxial Cyclic Fatigue to
Random Loading, VDI Reihe 18, Nr 52 VDI-Verlag, Dusseldorf 1988, 102 ps
55. Macha E.: Simulation investigations of the position of fatigue fracture plane in
materials with biaxial loads, Mat. –wiss. u. Werkstofftech., Nr 20, 1989, Teil I,
Heft 4/89, pp. 132–136, Teil II, Heft 5/89, pp. 153–163
 References 141

56. Macha E.: The spectral method of fatigue life calculation under random mul-
tiaxial loading, Physicochemical Mechanics of Materials, Vol. 32, No. 3, 1996,
pp. 86–96
57. Madsen H.O., Krenk S., Lind N.C.: Methods of Structural Safety, Prentice-
Hall, 1986
58. MATLAB Function Reference: Volume 2: F – O, Copyright 1988–2002 by The
MathWorks, Inc., Version 6, 826 ps
59. Miles J.: On the structural fatigue under random loading, Journal of Aeronau-
tical Science, Vol. 21, No. 11, 1954, pp. 753–762
60. Morrill J.H., Achatz T., Khosrovaneh A.: An application for fatigue damage
analysis using power spectral density from road durability events, SAE Technical
Paper Series, No. 980689, Reprinted from: Advancements in Fatigue Research
and Applications (SP- 1341), 1998, pp. 55–62
61. Morrow J.D.: The effect of selected subcycle sequences in fatigue loading his-
tories, Random Fatigue Life Prediction, ASME Publications, PVP 72, 1986,
pp. 43–60
62. MSC/FATIGUE User’s Guide: Vibration Fatigue Theory, MSC.Software Cor-
poration, 2002, 1360 ps
63. Naess A., Yim S.C.S.: Stochastic response of offshore structures exited by drag
forces, Journal of Engineering Mechanics, Vol. 122, No. 5, 1996, pp. 442–448
64. Nagode M., Fajdiga M.: A general multi- modal probability density function
suitable for the rain- flow ranges of stationary random processes, International
Journal of Fatigue, Vol. 20, No. 3, 1998, pp. 211–223
65. NASA Preferred Reliability Practices: Spectral Fatigue Reliability, Johnson
Space Center, Guideline No. GD- AP- 2303, 1993, pp. 1–5
66. Nieslony A., Macha E., L  agoda T.: Generalization of spectral method of fa-
tigue life determination under multiaxial random loading with the application
of strain criteria, Engineering Machines Problems, Vol. 20, 2002, pp. 119–126,
(in Polish)
67. Oritz K., Chen N.K.: Fatigue damage prediction for stationary wideband pro-
cesses, Proc. Fifth Int. Conf. on Application of Statistic and Probability in Soil
and Struct. Engrg, 1987
68. Oziemski St.: Analysis of Spectrum of Loading for Fittings of Excavators, Mono-
graphs CPBP 02.05., Warsaw University of Technology, Warsaw 1990, ps 59
69. Perruchet C., Vimont P.: Résistance à la fatigue des matériaux en contraintes
aléatoires, L’Aéronautique et l’Astronautique, No. 47, Vol. 4, 1974, pp. 71–81
70. Petrucci G., Zuccarello B.: On the estimation of the fatigue cycle distribution
from spectral density data, Proc. Inst. Mech. Engrs, Vol. 213, Part C, 1999,
pp. 819–831
71. Pierson W.J., Moskowitz L.: A proposed spectral form for fully developed wind
seas based on the similarity theory of SA Kitaigorodskii, J. Goephys Res,
Vol. 69, No. 24, 1964, pp. 5181–5190
142 References

72. Pitoiset X., Kernilis A., Preumont A., Piéfort V.: Estimation du dommage en fa-
tigue multiaxiale de structures métalliques soumisis à des vibrations aléatoires,
Revue Française de Mécanique, 1998, pp. 201–208
73. Pitoiset X., Preumont A.: Spectral methods for multiaxial fatigue analysis of
metallic structures, International Journal of Fatigue, Vol. 3, 2000, pp. 541–550
74. Pitoiset X., Preumont A., Kernilis A.: Tools for multiaxial fatigue analysis of
structures submitted to random vibrations, European Conference on Spacecraft
Structures, Materials and Mechanical Testing, Braunschweig, Germany, 4–6
November 1998, (pdf version)
75. Pitoiset X., Rychlik I., Preumont A.: Spectral methods to estimate local multi-
axial fatigue failure for structures undergoing random vibrations, Fatigue and
Fracture of Engineering Materials and Structures, Vol. 24, 2001, pp. 715–727
76. Pitoiset X., Rychlik I., Preumont A.: Spectral formulation of multiaxial high-
cycle fatigue criteria for structures subjected to random load , 6th Int. Conf. on
Biaxial/Multiaxial Fatigue & Fracture, Ed. Manuel Moreira de Freitas, Vol. I,
Lisboa, Tech. Univ. Portugal, 2001, pp. 297–304
77. Pook L.P.: Spectral density functions and the development of Wave Action Stan-
dard History (WASH) load histories, International Journal of Fatigue, Vol. 11,
No. 4, 1989, pp. 221–232
78. Preumont A., Piéfort V.: Prediction random high- cycle fatigue life with finite
elements, Journal of Vibration and Acoustics, Vol. 116, 1994, pp. 245–248
79. Rajcher V.L.: Gipoteza spektralnogo sumirovania i jejo primenenia dla oprede-
lenia ustalostnoj dolgovečnosti pri dejstvii slučajnych nagruzok , Trudy CAGI,
Č. 1134, Moskwa 1969, 3–38, (in Russian)
80. Rice S.O.: Mathematical analysis of random noise, Bell System Technical Jour-
nal, 23, 1944, pp. 282–332 and 24, 1945, pp. 46–156, or in: Wax N. Selected
Papers on Noise and Stochastic Processes New York, 1954
81. Rychlik I.: A new definition of the rainflow cycle counting method , International
Journal of Fatigue, Vol. 9, No. 2, 1987, pp. 119–121
82. Rychlik I.: Rainflow cycles in gaussian loads, Fatigue and Fracture of Engineer-
ing Materials and Structures, Vol. 15, No. 1, 1992, pp. 57–72
83. Rychlik I.: On the ‘narrow- band’ approximation for expected fatigue damage,
Probabilistic Engineering Mechanics, Vol. 8, 1993, pp. 1–4
84. Sakai S., Okamura H.: On the distribution of rainflow range for Gaussian
random processes with bimodal PSD, JSME International Journal, Series A,
Vol. 38, No. 4, 1995, pp. 440–445
85. Sarkani S., Kihl D.P., Beach J.E.: Fatigue of welded joints under narrowband
non- Gaussian loadings, Probabilistic Engineering Mechanics, Vol. 9, 1994,
pp. 179–190
86. Sarkani S., Michaelov G., Kihl D.P., Beach J.E.: Fatigue of welded steel joints
under wideband loadings, Probabilistic Engineering Mechanics, Vol. 11, 1996,
pp. 221–227
87. Schütz D., Klätschke H., Steinhilber H., Heuler P., Schütz W.: Standardized
Load Sequences for Car Wheel Suspension Components, LBF- Report No. FB-
191, Darmstadt, 1990, 32 ps
 References 143

88. Serensen S.V., Kogayev V.P.: Fatigue life of machines components with regard
to failure probability under nonstationary variable loading, Vestnik Masinos-
troenia, 1, 1966, pp. 7–12, (in Russian)
89. Signal Processing Toolbox: User’s guide, Copyright 1988–2002 by The Math-
Works, Inc., Version 6, 1036 ps
90. Sobczyk K., Spencer B.F.: Random Fatigue: From Data to Theory, Academic
Press, Inc., New York, 1992, ps 240
91. Sobczykiewicz W.: Design of Work Attachment of Earth Moving and Construc-
tion Machines Against Fatigue Failure with Respect to Manufacturing Process
and Operation Conditions, Monographs CPBP 02.05., Warsaw University of
Technology, Warsaw 1990, s. 99, (in Polish)
92. Sonsino C.M.: Limitation in the use of RMS- values and equivalent stresses
in variable amplitude loading, International Journal of Fatigue, Vol. 11, No. 3,
1989, pp. 142–152
93. Sonsino C.M., Kaufmann H., Grubišić V.: Transferability of material data for
the example of random loaded forged truck stub axle, SEA Technical Paper,
No. 970708, 1997, pp. 1–22
94. Sun J.Q., Wang X.: Fatigue analysis of non- linear structures with von Mises
stress, Journal of Sound and Vibration, Vol. 245, No. 5, 2001, pp. 947–952
95. Szala J.: Hypotheses of Fatigue Damage Accumulation, University of Technol-
ogy and Agriculture, Bydgoszcz 1998, ps 175, (in Polish)
96. Symbolic Math Toolbox: User’s guide, Copyright 1993–2000 by The Math-
Works, Inc., Version 2, p. 266
97. Tang J., Ogarevic V., Tsai C.- S.: An integrated CAE environment for
simulation- based durability and reliability design, Advances in Engineering
Software, Vol. 32, 2001, pp. 1–14
98. Tovo R.: A damage- based evaluation of probability density distribution for rain-
flow ranges from random processes, International Journal of Fatigue, Vol. 22,
2000, pp. 425–429
99. Tovo R.: Cycle distribution and fatigue damage under broad-band random load-
ing, International Journal of Fatigue, Vol. 24, 2002, pp. 1137–1147
100. Tunna J.M.: Fatigue life prediction for gaussian random loads at the design
stage, Fatigue and Fracture of Engineering Materials and Structures, Vol. 9,
No. 3, 1986, pp. 169–184
101. Welch P.D.: The use of fast fourier transform for the estimation of power
spectra: a method based on time averaging over short, modified periodograms,
IEEE Trans. Audio Electroacoust. Vol. AU–15, 1967, pp. 70–73
102. Winterstein S.R.: Nonlinear vibration models for extremes and fatigue, ASCE,
Journal of Engineering Mechanics, Vol. 114, No. 10, pp. 1772–1790
103. Wirsching P.H., Light M.C.: Fatigue under wide band random stresses us-
ing rainflow method , Journal of Structural Division, Vol. 106, No. ST7, 1980,
pp. 1593–1607
144 References

104. Zienkiewicz O.C., Taylor R.L.: The finite element method: solid and fluid me-
chanics, dynamics and non-linearity, Vol. 2, 4-th ed., McGraw-Hill, London,
UK,1991
105. Xu Y.L.: Fatigue damage estimation of metal roof cladding subjected to wind
loading, Journal of Wind Engineering and Industrial Aerodynamics, Vol. 72,
1997, pp. 379–388
Index

n-dimensional random process, 32 covariance function, 39

n-dimensional stochastic process, 37 covariance matrix, 83
criteria of multiaxial fatigue failure, 40
algorithm of fatigue life determination, coefficient vector, 44
68 frequency domain, 46
aluminium generalised form, 43
2023-T4, 11 modified, 122
7075-T651, 16 particular forms, 42
AU4G1, 11 critical plane, 32, 68
assymetric biaxial tension-compression, expected position, 69
78 cross-correlation coefficient, 73

Benasciutti, 22 damage accumulation hypothesis by

bending moment, 115 Corten and Dolan, 56
biaxial tension-compression, 78 Haibach, 54
Bishop, 8 Morrow, 15
Bolotin, 18 Palmgren and Miner, 12, 30, 53
broad-band frequency loading, 17, 20 Serensen and Kogayev, 57
broad-band frequency spectrum, 76 determinant of covariance matrix, 39
directional cosines, 41
Cebon, 12 Dirlik, 26
Chaudhury, 20 Dolan, 56
coefficient dominant frequency, 52
I, 9 Dover, 20, 31
L, 28
LL , 31 electromagnetic actuator, 113
λ, 19, 28 equivalent history, 43
bSK , 57 equivalent stress, 33
W D , 22 ergodic random process, 38
complete plain stress state, 79 error function, 19
correlation theory, 38 Etube et al., 9
Corten, 56 Eulerian integral of second order, 52
Corten-Dolan hypothesis, 56 expected number of peaks in a time
coupled function, 46 unit, 11
146 Index

expected number of zero level transitions Macha, 32, 35

in a time unit, 17 Manson-Coffin-Basquin characteristic,
expected value of the equivalent process, 59
44 Matake, 34
Matlab, 77
fast Fourier transform, 68 matrix of correlation function, 39
fatigue characteristic, 30 mean wind velocity, 23
fatigue life, 11 Miles, 10
fatigue limit, 30 Miles formula, 52
fatigue strength, 40 modified criterion of maximum normal
Fourier transform, 39 stress in the critical plane, 122
frequency domain, 102 moment of force, 115
Fu, 12 Morrill, 26

gamma function, 52 narrow-band frequency loading, 11

Gaussian process, 39 narrow-band frequency spectrum, 76
generation of random histories, 75 negative correlation, 78
non stationary random loading, 19
Haibach, 54 non-linear fatigue failure criteria, 44
Haibach hypothesis, 55 non-linear functions of directional
Hann type window, 77 cosines, 69
Hermitian matrix, 40 non-linear transformation, 29
Holmes, 23 non-proportional bending and torsion,
Hook law, 75 115
Hu, 28 nonproportional loading, 78
Huber-Mises-Hencky hypothesis, 33 normal distribution, 43
normal stress from bending, 118
impulse transfer functions, 45
incomplete gamma function, 54 Okamura, 12
inverse fast Fourier transform, 102
irregularity coefficient, 9 passband, 75
issue of excedance, 47 Piéfort, 33
Pitoiset, 34
Jiao, 14 power autospectral density, 33, 40
joint probability density function, 39, power cross-spectral density, 33, 40
47 power spectral density, 9, 10, 45
k-th moment, 9
Kam, 31 matrix, 40
Kogayev, 57 normalised, 17
Kolmogorov-Smirnov significance test, one-sided, 39
103 smooth estimator, 77
Kowalewski, 8, 17 two-sided, 39
kurtosis, 28 Preumont, 33
probability density function, 13, 24, 26,
Lachowicz, 30 39
Leser, 32 normal, 43
life time, 12 peaks, 18
Light, 19 range, 8
linear multi-input systems, 45 Rayleigh, 21
Liu, 28 two-dimensional, 8
 Index 147

Weibull, 23 Rajcher, 17
proportional bending and torsion, 115 Sakai and Okamura, 12
proportional loading, 78 Sarkani et al., 29
Winterstein, 30
rain flow algorithm, 9, 21 Wirsching and Light, 20
Rajcher, 17 spectral transmittance, 10
random Gaussian loading, 7 spectral transmittance function, 46
random non-Gaussian loading, 28 standard loading, 7
random number generator, 75 stationary process, 38
random state of steel
strain, 37 10HNAP, 30
stress, 37 18G2A, 74
range pair algorithm, 21 stochastic process, 37, 38
Rayleigh distribution, 50 strain tensor, 37
amplitude, 65 stress tensor, 37
Rychlik, 8, 21
tensor, 37
Sakai, 12
sampling frequency, 75 test stand MZGS100L, 114
Sarkani, 28 test stand MZGS200L, 114
Serensen, 57 time domain, 102
Serensen-Kogayev hypothesis, 57 torsional moment, 115
shear stress from torsion, 118 Tovo, 22, 24
Sherratt, 8
single moment spectral method, 20 uniaxial tension-compression with
six-dimensional vectorial stochastic torsion, 79
processes, 37 uniform amplitude distribution, 65
spatial stress state, 80
spectral damage formula, 56 variable amplitude loading, 11
spectral formula by variance charts of equivalent stress and
Benasciutti and Tovo, 22 strain, 84
Bolotin, 18 variance of equivalent history, 46
Chaudhury and Dover, 21 vectorial process, 38
Dirlik, 26
Fu and Cebon, 13 Wöhler curve, 11, 13
Holmes, 24 exponent, 11
Jiao, 14 modification, 57
Kowalewski, 17 Weibull, 23, 24
Larsen and Lutes, 20 weighted average amplitude, 18
Liou et al., 15 Welch method, 77
Liu and Hu, 28 wind loading, 23
Miles, 11 Winterstein, 30
Oritz and Chen, 20 Wirsching, 19