1

TOOLS FOR A MULTIAXIAL FATIGUE ANALYSIS OF STRUCTURES SUBMITTED TO

RANDOM VIBRATIONS

X.Pitoiset, A.Preumont
Active Structures Laboratory, ULB - CP 165/42
Av F.D. Roosevelt 50, B-1050 Brusssels
Mail: [email protected], http://www.ulb.ac.be/scmero

A.Kernilis
Société Européenne de Propulsion, Division de SNECMA
Forêt de Vernon, BP 802 F-27 000 Vernon

ABSTRACT tigue assessment methods available can be applied in

the time domain (rainflow counting) as well as in the
frequency domain (spectral moment methods).
This paper proposes a frequency domain method to
estimate the high-cycle fatigue damage for multiaxial In a second part of this paper, a formulation of the
stresses caused by random vibration, directly from a multiaxial rainflow method in the frequency domain
spectral analysis. This approach is based on a new is proposed. The PSD of all linear combinations of
definition of the von Mises stress as a random pro- the stress tensor components is computed directly
cess. It is shown that the approach can be general- from the PSD-matrix. A set of uniaxial processes is
ized to include a frequency domain formulation of the thus obtained and the most damaging one is found.
multiaxial rainflow method for biaxial stress states.
The results obtained with both methods are com- Both methods applied in the frequency domain sup-
pared. Finally, this paper describes a procedure for ply a complete map of the fatigue damage in every
the generation of multiaxial stress tensor histories of element and are in excellent agreement.
a given PSD matrix. The objective of the previous approaches is not to
predict fatigue life very accurately, but are rather to
Key words: multiaxial high cycle fatigue, random be used to compare design alternatives and to iden-
vibration. tify the most critical locations in a structure submit-
ted to random loads. The designer may then wish to
analyse further the critical elements. For this pur-
1. INTRODUCTION
pose a multiaxial time history generation tool has
been developed and a procedure to generate a set of
High-cycle fatigue due to multiaxial random loads is time histories of the stress tensor that fit a given PSD
still the subject of ongoing basic research. Some mul- matrix is described. These stress histories can later
tiaxial damage models have been proposed during be used in a more accurate multiaxial fatigue life as-
the last decade and are based on stress, strain or en- sessment method based for example on the critical
ergy. For these methods, the full local stress or/and plane approach.
strain tensors are required. Such methods would be
very costly and time consuming to apply to each fi-
nite element of a finely discretized structure, and im- 2. UNIAXIAL GAUSSIAN STRESS
possible to apply in an early design stage. This paper
proposes a numerical procedure to estimate the high- The equivalent von Mises stress approach is based on
cycle fatigue damage for multiaxial random stresses, extensions of static yield theory and on a new defini-
directly from a spectral analysis. tion of the von Mises stress as a random process. It
The originality of the approach is that the equiva- assumes that fatigue damage under multiaxial load-
lent alternating uniaxial stress is constructed in the ing can be predicted by calculating an equivalent uni-
frequency domain by combining, for each frequency, axial stress or strain on which the classical uniaxial
the power spectral densities of the normal and shear random fatigue theory is applied.
stresses according to the quadratic von Mises rule. Rayleigh Approximation
The power spectral density of the equivalent process
is thus obtained. Consequently, all the uniaxial fa- Consider a material characterized by a determinis-
2

tic S-N curve of the form N S β = c, where S is It takes its name from the fact that it depends only
the constant amplitude of the alternating sinusoidal on the spectral moment m2/β . Although there is
stress and N is the number of cycles to failure; the no theoritical base to Equ.(6), it gives the correct
constant c and β are characteritics of the material dependence on both ω and σx and is equivalent to the
(5 < β < 20). We assume that the linear damage Rayleigh approximation for narrow-band processes.
theory apply. The single moment method is in closer agreement
with rainflow simulations for various spectral shapes,
According to the classical theory of random fatigue, including for bimodal PSD.
every positive maximum of amplitude a contributes
for one cycle to the damage, that is: Rainflow cycle counting :
In high-cycle uniaxial fatigue, the most accurate vari-
D = c−1 aβ (1) able for correlating damage is the stress amplitude of
closed hysteresis loops of the local stress-strain path
The contribution to the damage rate from the in- at the crack location site. In the elastic domain these
finitesimal range [a, a + da[ is given by: loops are flat. The proper way of extracting such cy-
cles is the rainflow counting method (N.E.Dowling,
1972). This counting is applied after Monte Carlo
∆(a)da = c−1 aβ ν0 p(a)da (2) simulations of stress time histories of a given spec-
tral content.
where ν0 p(a)da is the average number of cycles in the
range [a, a + da[. For a narrow-band process, p(a) is Definition : Let σ(t), 0 ≤ t ≤ T be a stress and
the Rayleigh distribution and the rate of maxima denote by Mi the peaks of σ(t) at time ti , ti ≤ ti+1
can be approximated by the central frequency ν0 . (see Figure 1). To define the rainflow cycles, each
Integrating from 0 to infinity, we get the well known local maximum Mi has to be paired with one par-
result : ticular local minimum mi , found as follows : from
a local maximum Mi with height u, say, one shall
try to reach above u (in the forward and backward
Z ∞ direction) with an as small downward excusion as
E[D] = ∆(a)da = c−1 σxβ ν0 2β/2 Γ(1 + β/2) (3) possible. The minimum m+ i = mi
RF C
, which repre-
0
sents the smallest deviation form the maximum Mi ,
is defined to be the corresponding rainflow minimum.
where σx is the RMS value of the stress and Γ(.) is Thus the rainflow cycle starting at Mi is (mRF C
, Mi )
i
the Gamma function. Equation (3) can be written (I.Rychlik, 1987).
alternatively in terms of the spectral moments as

2β/2 (β−1)/2 1/2

E[D] = c−1 Γ(1 + β/2)m0 m2 (4)
2π

where the spectral moments are defined as

Z ∞
ma = 2 ω a Φ(ω)dω (5)
0

and Φ(ω) is the double-sided PSD function. This

result can also be applied as an approximation for a Figure 1: Rainflow cycle
wide-band process (P.H.Wirshing and E.B.Haugen,
1973; S.H.Crandall and W.D.Mark, 1963).

Single Moment : Damage estimation : From this counting, each

stress cycle (σm , σa ), where σm is the mean stress
The following heuristic formula has been proposed by and σa is the stress amplitude, is tranformed into
Larsen and Lutes (L.D.Lutes and C.E.Larsen, 1990) an equivalent cycle of zero mean (0, σaeq ) according
as an alternative to Equ.(4): to the Goodmann diagram. Afterward, the damage
corresponding to each cycle is computed according to
the S-N curve, and the linear cumulative Palmgren-
2β/2
E[D] = c−1 Γ(1 + β/2)(m2/β )β/2 (6) Miner law is applied.
2π
 3

3. EQUIVALENT VON MISES STRESS where Φss (ω) is the PSD matrix of the stress vector.
Equivalently,
X
Fatigue cracks normally initiate at free surfaces, Φc (ω) = Trace{QΦss (ω)} = Qij Φsi sj (ω) (13)
where the stress state is biaxial. For such multiax- i,j
ial stress fields, the fatigue phenomenon is generally
regarded as being governed by a combination of the Note that, in the uniaxial case [where only sx 6= 0,
shear and normal stress acting on a critical plane. with a PSD ΦSx (ω) ], Equ.(13) supplies Φc (ω) =
It has also been reported in the literature (G.Sines ΦSx (ω). Equation (13) defines the equivalent alter-
and G.Ohgi, 1987) that the von Mises criterion corre- nating stress (von Mises stress) as the scalar ran-
lates fairly well with a large amount of experimental dom process whose PSD is obtained from the PSD
data for biaxial stress states with constant principal matrix of the stress components according to the von
directions. When the excitation is random, the prin- Mises quadratic combination rule. The von Mises
cipal directions can rotate continuously with time ; stress is then considered as an equivalent counting
the von Mises stress is proportional to the root mean variable taking into account the multiaxiality. It is
square of the shear stress over all planes. this scalar process that we propose to use in the sub-
sequent fatigue damage analysis according to a uni-
In this study, we propose to base a first estimate
axial prediction model for Gaussian processes such
of the fatigue life on the von Mises stress which
as the Rayleigh approximation, the Single Moment
is used as an equivalent uniaxial counting vari-
or rainflow simulations. The methodology developed
able (A.Preumont, 1994; A.Preumont and V.Piéfort,
for biaxial stress states can readily be extended to
1994).
triaxial stress states, it is formally the same, with
For a biaxial stress, the von Mises stress sc is defined different definitions of the stress vector and the Q
by the quadratic relationship matrix. A finite element formulation of the foregoing
method is presented in (A.Preumont and V.Piéfort,
1994).
s2c = s2x + s2y − sx sy + 3s2xy (7)
where sx , sy and sxy are the normal and tangential 4. FREQUENCY DOMAIN FORMULATION OF
stresses, respectively. Defining the stress vector as THE MULTIAXIAL RAINFLOW METHOD
s = (sx , sy , sxy )T , Equ.(7) can be rewritten

s2c = sT Qs = Trace{Q[ssT ]} (8) 4.1. Multiaxial Rainflow Counting

with   Let Y (t) be a random n-dimensional vector process ;

1 −1/2 0
the n signals may be thought of either as load com-
Q =  −1/2 1 0  (9) ponents of external forces acting on the structure or
0 0 3
as components of the local stress or strain tensor at
After taking the expectation, a given point. The idea is to count rainflow cycles
on all linear combinations Yc (t) of the random vector
E[s2c ] = Trace{QE[ssT ]} (10) components of the form :
where E[SS T ] is the covariance matrix of the stress
n
vector, related to the PSD matrix of the stress vector X
by Yc (t) = ci Yi (t)
i=1
Z ∞ Pn
T where ci belong to a hypersphere i=1 c2i = 1. Prac-
E[SS ] = Φss (ω)dω (11)
−∞ tically, when the stress state is biaxial, the stress
components can be written under the form of the
From these equations, we can define the PSD Φc (ω) three dimension vector s = (sx , sy , sxy )T . A set of
of the equivalent von Mises stress as a frequency de- linear combinations
composition of its mean square value:
smrf (t) = c1 sx (t) + c2 sy (t) + c3 sxy (t) (14)
Z ∞
E[s2c ] = Φc (ω)dω can be chosen for values of ci such as
−∞
Z ∞
q
= Trace{QΦss (ω)}dω (12) c21 + c22 + c23 = 1 (15)
−∞
4

Figure 2: Map of the fatigue damage per time unit of the aluminium rectangular plate

thus defining a sphere (K.Dressler et al., 1995; where Φss (ω) is the PSD matrix of the stress vector.
A.Beste et al., 1992). Such a scalar process can be defined for every vector
c.
A frequency domain formulation
The damage is computed for each scalar process
After defining the two vectors c = (c1 , c2 , c3 )T and Φmrf , and the most critical combination is deter-
s = (sx , sy , sxy )T , Equ. (14) can be rewritten mined. Equation (20) shows that the formulation
of the equivalent von Mises stress method can be
smrf = cT s (16) easily generalized to include the multiaxial rainflow
method. Both formulations are formally similar,
The auto-correlation function of smrf (t) can be com- with different definitions of the Q matrix.
puted as follows:
5. APPLICATION

E[smrf (t + τ )smrf (t)] = cT E[s(t + τ )sT (t)]c (17)

The use of both methods is illustrated with a simply
supported rectangular aluminium plate (15.24cm ×
This equation is equivalent to :
30.48cm, e = 0.8mm) subjected to a band lim-
ited white noise random pressure field with per-
fect spatial coherence [Φpp (ω) = 1 P a2 sec/rad,
Rssmrf (τ ) = E[s(t)T ccT s(t + τ )] ωc = 6280 rad/sec]. The material constants are
= Trace{Q∗ E[s(t + τ )s(t)T ]} (18) c = 4.57 1055 , β = 6.09. The first three modes
(ω1 = 663 rad/sec, ω2 = 1061 rad/sec and ω3 =
with 1723 rad/sec) are within the bandwidth of the ex-
citation, although the second mode is not excited
(because it is anti-symmetrical). 200 shell elements
c21
 
c1 c2 c1 c3
have been used in the discretization.
Q∗ = ccT =  c1 c2 c22 c2 c3  (19)
c1 c3 c2 c3 c23 Results : Figure 2 shows the maps of relative dam-
age per unit time for the Single Moment method
Taking the Fourier transform of the autocorrelation respectively established according to the equivalent
function of Rssmrf (τ ), we get the PSD Φmrf (ω) of von Mises stress method and to the multiaxial rain-
the scalar process smrf : flow method. The gray scale refers to the logarithm
of the damage per unit time. Although the life pre-
dicted by both methods for a specific element are dif-
Φmrf (ω) = Trace{Q∗ Φss (ω)} ferent, we note that the localization of the damage
X
= Q∗ij Φsi sj (ω) (20) levels is similar. Figure 4 shows the PSD of the com-
i,j ponent stresses and that of the equivalent von Mises
 5

stress and of the multiaxial rainflow for the element

marked with a question mark in Fig.2. The similar-
ity of the two plots is striking. Note that Φc (ω) as
well as Φmrf (ω) always envelope the PSD’s of the
stress components.

For the selected element, Fig.3 shows the sphere with

a gray scale referring to the damage corresponding
to each linear combination of the stress components
as defined Equ.(14) and Equ.(15). This plot shows
how large the difference between maximum and min-
imum damage is and how the damage depends on the
projection (i.e. the elements of the Q∗ matrix).

Figure 3: Sphere damage for the selected element Figure 4: Power spectral densities of the stress com-
ponents, the von Mises stress and the multiaxial rain-
6. STRESS TENSOR HISTORY flow

For critical elements identified with a frequency do-

main analysis, the structural analyst may wish to U (ω) its eigenvectors. Λ(ω) is real and diagonal,
perform more detailed analyses for which he needs while the orthogonality of the eigenvectors implies
representative time histories of the stress vector si . that U (ω) is unitary (U H U = I).
Stress-based multiaxial random fatigue life predic-
The diagonal components of Λ(ω) characterize fully
tion methods have been developed and are now avail-
independent stochastic processes. The rotation ma-
able for industrial applications (e.g. (B.Weber et al.,
trix U (ω) defines the reference frame where the in-
1997)). The stress tensor histories must have the
dependence is effectively achieved at the frequency
proper frequency content [described by the diago-
ω.
nal components of Φss (ω)] and the proper cross-
correlations [described by the off-diagonal compo- The diagonal matrix Λ(ω) can be transformed into
nents of Φss (ω)]. The FFT generation of time histo- a vector process with independent components. If
ries of a given spectral content is a classical problem Z(ω, T ) denotes the Fourier transform of this ran-
which is described in textbooks (A.Preumont, 1994). dom vector of finite duration T , it satisfies the rela-
In this section, we discuss its extension to a vector tionship:
process of given PSD matrix Φss (ω). Recall that
Φss (ω) is Hermitian and semi positive definite. For
each frequency, Φss (ω) can be diagonalized according
to 1
Λ(ω) = lim E[Z(ω, T ), Z H (ω, T )]. (22)
T →∞ 2πT
Φss (ω) = U (ω)Λ(ω)U H (ω) (21)

where Λ(ω) contains the eigenvalues of Φss (ω) and Combining with Equ.(21) we get:
6

REFERENCES

1 A.Beste, K.Dressler, H.Kötzle, W.Krüger, 1992,

Φss (ω) = lim E[U (ω)Z(ω, T ), Z H (ω, T0 )U H (ω)]
T →∞ 2πT Multiaxial Rainflow - A consequent continuation
(23) of Professor Tatsuo Endo’s work, Butterworth-
Heinemann Ltd.
This relationship indicates that if Z(ω, T ) is con-
structed to be the Fourier transform of the vector A.Preumont, 1994, Random Vibration and Spectral
process with independent components of PSD ma- Analysis, Kluwer Academics Publishers.
trix Λ(ω),
A.Preumont, V.Piéfort, 1994, Predicting random
high cycle fatigue life with finite elements, Jour-
X(ω, T ) = U (ω)Z(ω, T ) (24) nal of Vibration and Acoustics, 116:245–248.

B.Weber, J.C.Clément, B.Kenmeugne, J.L.Robert,

is the Fourier transform of a vector process of PSD 1997, On a global stress-based approach for fatigue
matrix Φss (ω). The FFT generation of time histories assessment under multiaxial random loading, in
can thus be applied for each vector component. Conference ”Engineering against Fatigue”, Univ.
of Sheffield.

G.Sines, G.Ohgi, 1987, Fatigue criteria under com-

bined stresses or strains, Trans. ASME, J. of En-
gineering Materials and Technology, 103:82–90.
I.Rychlik, 1987, A new definition of the rainflow
counting method, Int J. Fatigue, 9:119–121.
K.Dressler, V.B.Köttgen, H.Kötzle, 1995, Tools for
fatigue evaluation of non-proportional loading, in
Proceedings of Fatigue Design’95, Helsinki, Fin-
land.

L.D.Lutes, C.E.Larsen, 1990, Improved spectral

method for variable amplitude fatigue prediction,
J. Struct. Div, ASCE, 116:1149–1164.

N.E.Dowling, 1972, Fatigue failure prediction for

complicated stress-strain histories, Journal of Ma-
Figure 5: Tensor time history of the selected element terials, 7:71–87.

P.H.Wirshing, E.B.Haugen, 1973, Probabilistic de-

sign for random fatigue loads, Journal of Engineer-
7. CONCLUSION ing Mechanics Division, ASCE, 99:1593–1607.
S.H.Crandall, W.D.Mark, 1963, Random Vibration
This paper discusses two methods for direct finite
in Mechanical Systems, Academic Press, N.Y.
element estimation of fatigue damage from spectral
data. The first method is based on the definition
of the von Mises stress as a uniaxial random pro-
cess for which the PSD function is computed. The
second one is a frequency domain formulation of the
multiaxial rainflow method. Both methods lead to
a similar implementation, although the multiaxial
rainflow requires the scanning of a large number of
linear combination of the stress components. The
two approaches give similar damage maps, allowing
the designer to localize the most critical elements.
The paper also describes a method for the genera-
tion of multiaxial stress tensor histories of given PSD
matrix, in order to perform a more accurate analysis
using methods in the time domain.