International Journal of Fatigue 47 (2013) 8–17

Contents lists available at SciVerse ScienceDirect

International Journal of Fatigue

journal homepage: www.elsevier.com/locate/ijfatigue

Frequency-domain methods for a vibration-fatigue-life estimation – Application

to real data
Matjaž Mršnik, Janko Slavič, Miha Boltežar ⇑
Faculty of Mechanical Engineering, University of Ljubljana, Slovenia

a r t i c l e i n f o a b s t r a c t

Article history: The characterization of vibration-fatigue strength is one of the key parts of mechanical design. It is closely
Received 9 April 2012 related to structural dynamics, which is generally studied in the frequency domain, particularly when
Received in revised form 12 July 2012 working with vibratory loads. A fatigue-life estimation in the frequency domain can therefore prove
Accepted 17 July 2012
advantageous with respect to a time-domain estimation, especially when taking into consideration the
Available online 4 August 2012
significant performance gains it offers, regarding numerical computations. Several frequency-domain
methods for a vibration-fatigue-life estimation have been developed based on numerically simulated sig-
Keywords:
nals. This research focuses on a comparison of different frequency-domain methods with respect to real
Vibration-fatigue
Structural dynamics
experiments that are typical in structural dynamics and the automotive industry. The methods
Vibrational testing researched are: Wirsching–Light, the a0.75 method, Gao–Moan, Dirlik, Zhao–Baker, Tovo–Benasciutti
and Petrucci–Zuccarello. The experimental comparison researches the resistance to close-modes, to
increased background noise, to the influence of spectral width, and multi-vibration-mode influences.
Additionally, typical vibration profiles in the automotive industry are also researched. For the experiment
an electro-dynamic shaker with a vibration controller was used. The reference-life estimation is the rain-
flow-counting method with the Palmgren–Miner summation rule. It was found that the Tovo–Benasciutti
method gives the best estimate for the majority of experiments, the only exception being the typical
automotive spectra, for which the enhanced Zhao–Baker method is best suited. This research shows that
besides the Dirlik approach, the Tovo–Benasciutti and Zhao–Baker methods should be considered as the
preferred methods for fatigue analysis in the frequency domain.
Ó 2012 Elsevier Ltd. All rights reserved.

1. Introduction [3] and Miner [4]. The combination of rainflow counting and Palm-
gren–Miner has been tried and tested thoroughly and is generally
Fatigue is a common cause of failure in mechanical structures accepted as one of the best time-domain methods for a fatigue-life
and components subjected to time-variable loadings [1]. It is crit- estimation [1,6–9].
ical that fast and effective tools are available to estimate the fati- In reality, structures such as a car on a rough road or a wind tur-
gue life during the design process. Frequency-domain methods bine are exposed to random loads (e.g., road surface, wind speed).
for fatigue assessment aim to speed up the calculations substan- Such random loads can be viewed as the realization of a random
tially, as they define the loading process in the frequency domain. Gaussian process that can be described in the frequency domain
A well-established practice of fatigue assessment consists of a by a power spectral density [5], representing the spread of the
determination of the loading history, the identification of damag- mean square amplitude over a frequency range [10]. Operating
ing cycles [2] and an evaluation of the total fatigue damage by with a power spectral density proves especially beneficial when
aggregating the damage contributions [3,4] of the respective dam- working with complicated finite-element models where the calcu-
aging cycles [5]. lation of the frequency response is much faster than a transient dy-
In the time domain the identification of cycles in an irregular namic analysis in the time domain [6].
loading history is accomplished by means of a cycle-counting Long time histories can be simulated as samples of a random
method, usually employing the rainflow algorithm that was intro- process, defined in the frequency domain. However, this approach
duced by Matsuishi and Endo [2]. The accumulation of damage is is not an ideal design tool [11].
then carried out according to the hypothesis of linear damage A different approach is the development of frequency-domain
accumulation, which was independently presented by Palmgren methods for a fatigue assessment that offer a direct connection be-
tween the power spectral density and the damage intensity or the
⇑ Corresponding author. cycle distribution of loading. As most authors consider the rainflow
E-mail address: miha.boltezar@fs.uni-lj.si (M. Boltežar). method to be the most accurate, the frequency-domain methods

0142-1123/$ - see front matter Ó 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.ijfatigue.2012.07.005
 M. Mršnik et al. / International Journal of Fatigue 47 (2013) 8–17 9

Nomenclature

a peak amplitude a Weibull distribution parameter

a() best fitting parameters (Wirsching–Light method) ai(a2, a0.75) spectral width parameters
b() best fitting parameters (Wirsching–Light method) b Weibull distribution parameter
k S–N slope coefficient c maximum stress xmax to tensile strength Su ratio
C S–N curve constant C() the Euler gamma function
d eq. root for calc. of fatigue life (Zhao–Baker method) d Vanmarcke’s bandwith parameter
D damage intensity e spectral width parameter
DRFC rainflow damage intensity m0 expected positive zero-crossing frequency
f frequency mp expected peak occurrence frequency
f0 central frequency qWL empirical correction factor (Wirsching–Light method)
Df mode bandwidth at half-power points qZB correction factor (Zhao–Baker method)
G1  G3 Dirlik method parameters r2X random process variance
GXX one-sided power spectral density vk kth moment of cycle amplitude probability density
H(x) frequency response function U() standard normal cumulative distribution function
H(x) acceleration to stress frequency-transfer function W() approximation function (Petrucci–Zuccarello method)
m stress range mean W1  W4 approx. function coeff. (Petrucci–Zuccarello method)
mi ith moment of power spectral density a forced vibration acceleration profile
N number of cycles to failure [C] damping matrix
pp() probability density function of peaks {f} forced vibration vector
pa() cycle amplitude probability density function Has frequency-response function matrix (acc. to stress)
Q quality factor or Dirlik method parameter [K] stiffnes matrix
r stress range [M] mass matrix
re Goodman equivalent stress range s normal stress tensor
R Dirlik method parameter {x} degree-of-freedom vector
RH(t) high-frequency random process AL the a0.75 method
RM(t) intermediate-frequency random process DK Dirlik method
RL(t) low-frequency random process GM Gao–Moan method
s stress amplitude NB narrow-band method
Su tensile strength PZ Petrucci–Zuccarello method
SFF force excitation spectra TB1 Tovo–Benasciutti method
SXX power spectral density or displacement response spec- TB2 improved Tovo–Benasciutti method
tra WL Wirsching–Light method
T fatigue life estimate ZB1 Zhao–Baker method
Terr relative error of fatigue-life estimate ZB2 improved Zhao–Baker method
w weighting factor (Zhao–Baker method) AM typical automotive spectra
X random process BN increased background noise spectra
xm mean frequency (Dirlik method) CM close-modes spectra
xmax absolute maximum value of stress MM multi-mode spectra
Z normalized amplitude SW spectral width spectra

try to obtain a cycle distribution according to rainflow counting in In 2004 Benasciutti [22] compared a group of frequency-do-
the time domain [7,12,13]. main methods, i.e., Wirsching–Light [23], Zhao–Baker [7], Dirlik
The methods described here are concerned with the stationary [12], the empirical a0.75 [22], and Tovo–Benasciutti [8,13], and
Gaussian process. It is further divided into narrow-band and wide- found that the Tovo–Benasciutti method matches the accuracy of
band processes, of which the narrow-band allows for a straightfor- the Dirlik method in terms of numerically simulated power spec-
ward derivation of the cycle distribution, as pointed out by Lutes tral densities. A thorough study of fatigue evaluation for a multi-
and Sarkani [11]. For a wide-band process the relation of the peak axial random loading was made by Lagoda et al. [24]. Experiment
distribution and cycle amplitudes is much more complex. Several showed a good agreement with the fatigue-life estimate of the
empirical solutions (e.g., Dirlik [12] and Zhao–Baker [7]) have been Wirsching–Light method. Braccesi et al. [25] developed the Per-
proposed, but only very few completely theoretical solutions (e.g., centage Error Index (IEP) by which they compared the Dirlik,
Bishop [14]). The theoretical solution presented by Bishop is based Zhao–Baker, Fu–Cebon [26], Sakai–Okamura [27] and Bendat [28]
on the Markov process theory and is computationally intensive. methods on bi-modal triangular spectra. Results produced by the
However, it shows little improvement in accuracy over Dirlik, Dirlik method were again among the best and most consistent.
which is usually the preferred method [6]. Special case is a new method by Gao and Moan [21], which is based
Frequency-domain methods are used in different fields, such as on similar principles as bi-modal methods, but can be generalized
structural health monitoring [15], the FEM environment [16], mul- for a broad-band spectra.
ti-axial fatigue [17] and a non-normal process fatigue evaluation A subdomain group of frequency-domain methods, such as Sa-
with the use of a correction factor [18]. kai–Okamura [27] and Fu–Cebon [26], are aimed at evaluating bi-
Since the introduction of Dirlik and Bishop several comparison modal processes which can often represent two modal shapes of a
studies have been made [19,20,13,5,21]. These comparison studies structure response spectra. Recent work has been done in this field
show that the Dirlik method performs better than most of the by Low [29], demonstrating the exceptional accuracy that can be
other methods. attained using the bi-modal formulation. However, the focus of this
10 M. Mršnik et al. / International Journal of Fatigue 47 (2013) 8–17

paper is on frequency-domain methods that can be used blindly Another frequently used spectral parameter is Vanmarcke’s
across different broad-band random processes and thus bi-modal parameter d [34]:
methods are not explored in detail. qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
In this research, a theoretical and experimental comparison of d¼ 1  a21 ð4Þ
several frequency-domain methods is given. Some of those methods
are for the first time compared side-by-side. Existing studies already The expected peak occurrence frequency mp and the expected
deal with experimental validation of results, namely Lagoda et al. positive zero-crossing rate m0 are defined as:
[24,30] and Kim et al. [31]. In this research the focus is on perfor- rffiffiffiffiffiffiffi rffiffiffiffiffiffiffi
mance of selected frequency-domain-methods with different groups m4 m2
mp ¼ m0 ¼ ð5Þ
of spectra with the aim of finding a most accurate frequency-domain m2 m0
method(s) for blind use on different broad-band processes. They are imperative for determining the fatigue-life intensity.
Besides localized effects, the stress distribution of a real struc- The analytical derivation of both was described by Newland [33].
ture significantly relates to the structural dynamics [32]. The fre- The foundations for a frequency-domain approximation of a cy-
quency-domain methods are researched with regards to the cle distribution were set by Rice [35], who managed to analytically
structural dynamics properties: close-modes, increased back- define the probability density function of the peaks pp(a) based on
ground noise, number of modes and spectral width. Typical vibra- the power spectral density:
tion profiles used in automotive accelerated tests are also 0 1
qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
researched. For this purpose, different vibration profiles are devised,
1  a22 2r2 a1 2
a a
2
a2
B a a C
e X ð a2 Þ þ 2 e 2rX U@ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiA
2 2 2
which are varied according to one of the listed dynamic properties. pp ðaÞ ¼ pffiffiffiffiffiffiffi ð6Þ
This research is organized as follows: Section 2 gives the theo- 2prX rX rX 1  a 2
2
retical background of different frequency-domain methods. Sec-
tion 3 briefly presents some selected frequency-domain methods where a is the peak amplitude and rX and a2 are defined with Eqs.
for a fatigue-life evaluation and cycle-distribution estimation. In (2) and (3), respectively. U(z) is the standard normal cumulative
Section 4 the experimental set-up is explained, followed by a dis- distribution function:
cussion of the results in Section 5 and, finally, the conclusions Z z
1 t2
are given in Section 6. UðzÞ ¼ pffiffiffiffiffiffiffi e 2 dt ð7Þ
2p 1

2. Theoretical background
2.2. Structural dynamics and modal analysis
To facilitate an understanding of frequency-domain methods
and their definitions, a brief introduction to the stochastic process The objective of the modal analysis is to establish a mathemat-
theory will be presented first. For a deeper understanding and an ical model of the structure. The structural dynamics and the re-
elaborate derivation the reader is referred to the reference works sponse can already be characterized during the design stage,
of Shin and Hammond [10] and Newland [33]. The basics of the usually with the help of FEM methods [16,32,36]. The linear struc-
modal analysis that can be used to determine the stress function tural dynamics is described in the frequency domain, which is why
at different material points of the vibrating structure are covered the frequency-domain methods for fatigue analysis are of great
next. Finally, a general approach to specifying the fatigue strength interest. Recent research [17] even suggests it is possible to evalu-
in the frequency domain is presented. ate more complicated multi-axial loadings in the same manner. So
as not to lose focus, here only the basics are given and the inter-
2.1. Properties of random processes ested reader is advised to read references [37,36].
Complex structures can be viewed as linear, multi-degree-of-
In the frequency domain the random loading of a random pro- freedom systems that are described by a system of second-order
cess X is defined by the power spectral density SXX(f), with f denot- differential equations [37]:
ing the frequency. It is customary to use a one-sided power €g þ ½Cfxg
½Mfx _ þ ½Kfxg ¼ ffg ð8Þ
spectral density GXX(f), defined on the positive half-axis only. The
statistical properties of a stationary process can be described by where [M] is the mass matrix, [C] is the damping matrix, [K] is the
the moments of the power spectral density. The general form for stiffness matrix, {x} is the vector of degrees of freedom and {f} is the
the ith spectral moment mi is given by: excitation force vector. Furthermore, through modal analysis the
Z transfer functions from one point on the structure to the other
1
can be deduced, mathematically or by experiment.
mi ¼ f i GXX ðf Þdf : ð1Þ
0 For a selected geometrical location on the structure the dis-
placement response spectra SXX(x) can be obtained by knowing
For a fatigue analysis the moments up to m4 are normally used.
the frequency-response function H(x) (phase information is trun-
The even moments represent the variance r2X of the random pro-
cated with PSD, thus jH(x)j2) and the force excitation spectra
cess X and its derivatives:
SFF(x) [37]:
r2X ¼ m0 r2X_ ¼ m2 ð2Þ
SXX ðxÞ ¼ jHðxÞj2 SFF ðxÞ ð9Þ
The spread of the process or the spectral width is estimated The response characteristics other than the displacements (e.g.,
using the parameter ai, which has the general form: stress, acceleration) can also be expressed. As mechanical struc-
m
i tures are usually exposed to a known acceleration profile, for a fa-
ai ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffi
ffi ð3Þ
tigue analysis the frequency-transfer function Has(x) from the
m0 m2i
acceleration a to the material stress s is of special interest. For mul-
The most commonly used a2 is the negative of the correlation ti-degree-of-freedom systems the response is dependent on multi-
between the process and its second derivative, as described by ple inputs and Eq. (9) is rewritten in the matrix form as:
Tovo [8]. It takes values from 0 to 1. The higher the value, the nar-
rower is the process in the frequency domain and vice versa. s ¼ Has a ð10Þ
 M. Mršnik et al. / International Journal of Fatigue 47 (2013) 8–17 11

where s is the normal stress tensor in the frequency domain, a is the DWL ¼ qWL DNB ð15Þ
forced vibration acceleration profile and Has is the frequency-re- NB
sponse function matrix. where D is the narrow-band approximation obtained with Eq.
(13) and qWL is defined as:
2.3. Fatigue-life estimation qWL ¼ aðkÞ þ ½1  aðkÞð1  eÞbðkÞ ð16Þ

A possible scenario in fatigue analysis is to numerically simu- with the spectral width parameter e being:
late the operating time until failure for a structural member with qffiffiffiffiffiffiffiffiffiffiffiffiffiffi
a given critical stress function in the frequency domain at selected e¼ 1  a22 ð17Þ
node(s) (10). Instead of the time-until-failure criteria, this research
and the best fitting parameters a() and b() dependent on the S–N
focuses on the damage intensity D, see Benasciutti and Tovo [13].
slope k:
The damage intensity estimates the damage per unit of time. It is
defined with: aðkÞ ¼ 0:926  0:033k bðkÞ ¼ 1:587k  2:323 ð18Þ
Z 1
1
D ¼ mp C sk pa ðsÞds; ð11Þ where the authors used the S–N slope k values of 3, 4, 5 and 6 for the
0 simulations.
where mp is the expected peak occurrence frequency, given by Eq. (5),
C and k are material parameters and pa(s) is the cycle amplitude prob- 3.3. a0.75 method
ability density function. The function pa(s) is the unknown and a key
factor considered by frequency-domain methods for fatigue analyses. The second correction method, named a0.75, is simple yet agrees
The evaluation of frequency-domain methods is carried out by a fairly well with the data from the numerical simulation according
comparison of the fatigue-life estimate T in seconds, which is ob- to Benasciutti and Tovo, who also propose this method [22]. It is
tained from the damage intensity D: based on the spectral parameter a0.75 and is given with:
1 DAL ¼ a20:75 DNB ð19Þ
T¼ ð12Þ
D NB
where D is the narrow-band approximation (13) and a0.75 is de-
fined with (3).
3. Frequency-domain methods
3.4. Gao–Moan method
Typical methods for fatigue analyses in the frequency domain
designed to deal with Gaussian random loadings are discussed in Recently, a trimodal method was presented by Gao and Moan
this section. While the narrow-band (NB) approximation is only [21] along with a generalized variant that can be applied to other
appropriate for narrow-band processes, others are also appropriate (wide-band) types of processes. The fatigue-damage intensity
for wide-band loading processes. These are the Wirsching–Light DGM is the sum of the high-frequency process RH(t), the intermedi-
method (WL) [23], the a0.75 method (AL) [22], the Gao–Moan ate-frequency process RM(t) and the low-frequency process RL(t)
method (GM) [21], the Dirlik method (DK) [12], both Zhao–Baker contributions, each are being narrow-banded and having a Ray-
methods (ZB1 and ZB2) [7], the Tovo–Benasciutti methods (TB1 leigh-distributed amplitude with the respective variances equal
and TB2) [8,13] and the Petrucci–Zuccarello method (PZ) [9]. The to rH, rM, rL:
a0.75 method and the Wirsching–Light method differ from the oth-
DGM ¼ DP þ DQ þ DH ð20Þ
ers in the way that they provide the correction factor to correct the
narrow-band approximation for wide-band processes. where RH(t) is treated as a narrow-band process and DP and DQ are
damage intensities due to the processes RP(t) and RQ(t),
3.1. Narrow-band approximation respectively:
RP ðtÞ ¼ RH ðtÞ þ RM ðtÞ
For a narrow-band process it is reasonable to assume that every ð21Þ
peak is coincident with a cycle and that, consequently, the cycle RQ ðtÞ ¼ RH ðtÞ þ RM ðtÞ þ RL ðtÞ
amplitudes are Rayleigh-distributed. The narrow-band expression The damage component DH is determined using (13), where m0H
was originally presented by Miles in 1956 [38] and is here defined is given by (5). Eq. (11) is used for DP and DQ , where the expected
for stress amplitudes: zero-crossing frequencies are used instead of the peak frequencies
pffiffiffiffiffiffiffiffiffiffik  k
 and are calculated with:
DNB ¼ m0 C 1 2m0 C 1 þ ð13Þ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 rM 
m0P ¼ m2H d2H þ m2M  ð22Þ
where m0 is the expected positive zero-crossings intensity, which is r2H þ r2M
very close to the peak intensity mp for a narrow-band process, C and qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
k are material fatigue parameters, m0 is the 0th spectral moment, a2 m0Q ¼ m2H d2H þ m2M d2M þ m2L
is determined with (3) and C() is the Euler gamma function, which 2 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 3
rH rM
is defined as:
62rL r2H þ r2M þ r2L  prH rM 2rH rM arctan rL pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ffi
r2H þr2M þr2L 7
Z 1 4 qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3 þ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi3 5
CðzÞ ¼ t z1 et dt ð14Þ 2 r2H þ r2M þ r2L 2 r2H þ r2M þ r2L
0
ð23Þ
3.2. Wirsching–Light method where dH and dM are Vanmarcke’s parameters from Eq. (4), calcu-
lated for the processes RH(t) and RM(t), respectively. The probability
Wirsching and Light [23] in 1980 used an additional spectral density functions for the sum of two or more random variables, also
width parameter a2 to correct the narrow-band approximation needed for calculating (11), are determined by means of the convo-
with the empirical factor qWL: lution integral:
12 M. Mršnik et al. / International Journal of Fatigue 47 (2013) 8–17

Z 1
1  a2
pa;HM ¼ pa;H ðsÞpa;M ðt  sÞds ð24Þ w¼ qffiffiffi   ð31Þ
2
Z0
1
1 p C 1 þ 1b a1=b
pa;HML ¼ pa;HM ðsÞpa;L ðt  sÞds ð25Þ
0 and a and b are the Weibull parameters:
where pa,H, pa,M and pa,L are the cycle-amplitude probability densi- a ¼ 8  7a2 ð32Þ
ties of the respective narrow-band processes. Hermite integration

was used by Gao and Moan to calculate the distribution of the 1:1; a2 < 0; 9
b¼ ð33Þ
sum of the multiple Rayleigh random variables [21]. 1:1 þ 9ða2  0; 9Þ; a2 P 0:9
The trimodal approach is generalized to a general wide-band
process by splitting the power spectral density into three parts, According to Eq. (30) the Rayleigh part corresponds mainly to
according to the chosen criteria. The authors recommend splitting the large cycle amplitudes and the Weibull part to small ampli-
into three parts with equal variance, with each part then being tudes when looking at the cycle-amplitude distribution.
treated as one of the three modes. This same approach is used in For the spectral width parameter a2 < 0.13 the expression for
this paper, referenced with GM. Additionally, the accuracy of the the weighting factor w gives the incorrect results w > 1 and there-
Gao–Moan method could be improved by using a higher order, fore the Zhao–Baker method is not suitable for those cases [22].
multi-modal formulation of the method, such as a four- or five- Also, because the simulations were only concerned with the values
modal formulation [21]. of 2 6 k 6 6 the model should be re-evaluated to reach a good
agreement for higher values of the S–N slope k [7].
3.5. Dirlik method For small values of the S–N slope k (k = 3) the rainflow damage
is more closely related to a0.75 than a2. Zhao and Baker offered an
The Dirlik method [12], devised in 1985, approximates the cy- enhanced method of their calculation (ZB2) where a is calculated
cle-amplitude distribution by using a combination of one exponen- as:
tial and two Rayleigh probability densities. It is based on numerical a ¼ db ð34Þ
simulations of the time histories for two different groups of spec-
tra. This method has long been considered to be one of the best with d calculated as a root of:
and has already been subject to modifications, e.g., for the inclusion     rffiffiffiffi
3 3 1 p
of the temperature effect, by Zalaznik and Nagode [39]. The rain- C 1þ ð1  a2 Þd þ 3C 1 þ ðqZB a2  1Þd þ 3 a2 ð1
b b 2
flow-cycle amplitude probability density estimate is given by:

 qZB Þ
1 G1 Z G2 Z Z2 Z 2
pa ðsÞ ¼ pffiffiffiffiffiffiffi e Q þ 2 e 2R2 þ G3 Ze 2 ð26Þ ¼0 ð35Þ
m0 Q R
where Z is the normalized amplitude and xm is the mean frequency, and the correction factor qZB at k = 3 is determined by:

as defined by the author of the method [12]: 0:4154 þ 1:392a0:75 ; a0:75 P 0:5
 1
qZB jk¼3 ¼ ð36Þ
s m1 m2 2
0:28; a0:75 < 0:5
Z ¼ pffiffiffiffiffiffiffi xm ¼ ð27Þ
m0 m0 m4 A closed-form expression for directly calculating the fatigue-
damage intensity is defined with [22]:
and the parameters G1 to G3, R and Q are defined as:    

mp k k k
G1 ¼
2ðxm a22 Þ
G2 ¼
1a2 G1 þG21 DZB1=ZB2 ¼ mk=2
0 wab C 1 þ þ ð1  wÞ2k=2 C 1 þ ð37Þ
1þa22 1R C b 2
a xm G21 ð28Þ where the use of the ordinary approach is denoted with ZB1 and the
G3 ¼ 1  G1  G2 R ¼ 12a 2
2 G1 þG1
use of the enhanced method with ZB2.
Q ¼ 1;25ða2 G
G1
3 G2 RÞ

3.7. Tovo–Benasciutti method

while a2 has already been defined with Eq. (3). The closed-form
expression for the fatigue-life intensity has been derived in the
In 2002, Benasciutti and Tovo [8,13] first proposed an approach
form:
where the fatigue life is calculated as a linear combination of the
pffiffiffik  

k k upper and lower fatigue-damage intensity limits. The final expres-

DDK ¼ C 1 mp m20 G1 Q k Cð1 þ kÞ þ 2 C 1þ ðG2 jRjk þ G3 Þ
2 sion for the estimated fatigue-damage intensity is:

ð29Þ DTB ¼ b þ ð1  bÞa2k1 a2 DNB ð38Þ
where b is unknown and is approximated with numerical simula-
3.6. Zhao–Baker method tion data. The narrow-band damage intensity DNB is defined with
Eq. (13), where the zero-crossing intensity m0 is replaced with the
This method was developed by Zhao and Baker in 1992 [7]. It peak intensity mp for wide-band processes. Similarly, the rainflow
combines theoretical assumptions and simulation results to give amplitudes’ probability density pa(s) estimate can also be written
an expression for the cycle distribution as a linear combination in the form of a linear combination of two probability densities [22].
of the Weibull and Rayleigh probability density function: Two different equations are suggested for determining the fac-
Weibull Rayleigh tor b in order to substitute into Eq. (38), both were proposed by
zfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflffl{ zfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflffl{2 Benasciutti and Tovo [8,13]:
b Z
pa ðZÞ ¼ wabZ b1 eaZ þ ð1  wÞZe 2 ð30Þ  
TB1 a1  a2
where Z is the normalized amplitude from Eq. (27) and w is the bapp ¼ min ;1 ð39Þ
1  a1
weighting factor defined by:
TB2 ða1  a2 Þ½1:112ð1 þ a1 a2  ða1 þ a2 ÞÞe2:11 a2 þ ða1  a2 Þ
bapp ¼ ð40Þ
ða2  1Þ2
 M. Mršnik et al. / International Journal of Fatigue 47 (2013) 8–17 13

TB1
where using the bapp from Eq. (39) will be referenced with TB1 and
TB2
using bapp from Eq. (40) with TB2.

3.8. Petrucci–Zuccarello method

This method was proposed in 2004 by Petrucci and Zuccarello

[9] who sought a connection between the moments of the proba-
bility density of the equivalent Goodman stress and the parameters
a1, a2, k and c, the last of these being the ratio:

xmax
c¼ ð41Þ
Su
Fig. 1. Experiment schematics.
where xmax is the absolute maximum value of the stress process and
Su is the material tensile strength.
As already mentioned, the Petrucci–Zuccarello method implic- considered 1 MPa).1 The response profile was simulated on a real
itly predicts equivalent stress ranges by calculating the Goodman electro-dynamical shaker and controlled with an industry-standard
equivalent stress range [40]: vibration controller. The proposed spectra, arranged into groups
(multi-mode, spectral width, etc.), were devised by the authors
r based on experience in the field of structural dynamics. The real sig-
re ¼ ð42Þ
1  Smu nal was then actually measured, by means of experiment, on an elec-
tro-dynamic shaker, controlled according to the aforementioned
where r is the original stress range, m is the range mean and Su is spectra. By this procedure one gains insight into frequency-domain
the tensile strength. The expression for calculating the damage- methods performance on spectra exhibiting certain structural
intensity estimate is: dynamics phenomena, while avoiding the cumbersome procedure
qffiffiffiffiffiffiffi of dealing with a real structure and controlling its response to
DPZ ¼ C 1 mp mk0 eWða1 ;a2 ;b;cÞ ð43Þ achieve the same goal.

where mp is defined with (5) and the spectral moment m0 with (1).
4.1. Experimental set-up
The approximation function W() [9]:

ðW2  W1 Þ The test site consisted of a personal computer for control and
Wða1 ; a2 ; k; cÞ ¼ ðb  3Þ þ W1 data acquisition, an acceleration sensor and electrodynamic shaker
6

2 4 equipment, see Figs. 1 and 2.

þ ðW4  W3  W2 þ W1 Þðk  3Þ þ ðW3  W1 Þ ðc  0;15Þ ð44Þ
9 3 Professional hardware and software with a control loop were
used to generate the excitation according to a given power spectral
depends on the material parameter k, the ratio c and the coefficients density in the range from 10 to 1000 Hz. At the same time, the
W1 to W4 given as polynomials: measured signal was also discretized with a DAQ module and
saved for later analysis.
W1 ¼ 1; 994  9; 381a2 þ 18; 349a1 The duration of one measurement was 5 min. The sampling fre-
þ15; 261a1 a2  1; 483a22  15; 402a21 quency was 10,000 Hz. The root-mean-square (RMS) value of every
W2 ¼ 8; 229  26; 510a2 þ 21; 522a1 random process run on the shaker was 20 MPa.
þ27; 748a1 a2 þ 4; 338a22  20; 026a21
ð45Þ
W3 ¼ 0; 946  8; 025a2 þ 15; 692a1 4.2. Simulated spectra
2 2
þ11; 867a1 a2 þ 0; 382a  13; 198a 2 1
A number of different power spectral densities were defined in
W4 ¼ 8; 780  26; 058a2 þ 21; 628a1
the industry-common frequency-range interval [10 Hz, 1000 Hz].
þ26; 487a1 a2 þ 5; 379a22  19; 967a21 They are separated into five groups: multi-mode (MM), increased
with a1 and a2 taken from Eq. (3). background noise (BN), spectral width (SW), close-modes (CM)
and typical automotive spectra (AM).
The quality factor Q (not to be confused with the Dirlik estima-
4. Experiment
tion parameter), also known as the amplification factor [37], is typ-
ically used in structural dynamics for a quantitative
In real life dynamical structures exposed to vibrations respond
characterization of the spectral peaks. It conveys information
according to their natural dynamics, and to avoid real-life (in appli-
about the sharpness of the resonance or damping, since Q = 1/
cation) experiments, accelerated tests are used. The load during
(2f) with f being the damping ratio. The value of Q was determined
accelerated vibration tests is based on real-life vibration loads,
based on [37]:
but these tests are usually accelerated with higher vibration ampli-
tudes. The accelerated vibration tests are usually made by using f0
electro-dynamical shakers, where a controlled excitation profile Q¼ ð46Þ
Df
in the frequency domain is used. According to Eq. (10) the excita-
tion profile is controlled at a selected point of the structure and this where f0 is the central frequency of the mode and D f = f2  f1 is the
results in a stress tensor frequency profile at all the other points. mode bandwidth at half-power points, i.e., the frequencies f1 and f2,
The transfer function from the excitation to the response differs
for each material point; so as not to focus on the structure, the 1
In a real damage estimation the transfer functions of two material points are
excitation acceleration profile is considered as the stress profile required and the damage estimation of these points is straightforward with respect to
and analyzed from the point of the damage estimation (1 m/s2 is Eq. (10).
14 M. Mršnik et al. / International Journal of Fatigue 47 (2013) 8–17

Fig. 2. The LDS V555 electro-dynamic shaker with the B& K accelerometer mounted
on the vibrating plate and the B& K Nexus charge-conditioning amplifier.
Fig. 4. Increased background noise.

Fig. 3. Multi-mode spectra. Fig. 5. Spectral width group of spectra.

corresponding to half of the peak amplitude in the frequency central frequencies fc of 150 and 700 Hz, respectively, were
domain. 100 Hz closer, as demonstrated in Fig. 6.
The multi-mode spectra are designed to research the influence In accelerated vibration tests, the structure is fixed to the elec-
of multiple modal frequencies on the fatigue analysis. Initially, tro-dynamical shaker and excited with a given profile. Because the
there is only one dominant mode, and with each subsequent structural response natural modes discussed above are developed
experiment one additional mode is added (up to a maximum four). for the frequency region without natural dynamics, the excitation
This is accomplished by superimposing the modal peaks, as shown profile is linearly translated to the fatigue loads. Therefore, we
in Fig. 3. The modes are positioned at the central frequencies fc of investigated three realistic industry spectra that are typical for
200, 400, 600 and 750 Hz with amplitudes of 650 MPa2/Hz. the automotive industry, shown in Fig. 7.
The increased background noise group was selected to research The ranges of the spectral-width parameter a2 for each group
the effect of the background noise on the estimation accuracy in are given in Table 1.
the frequency domain. For this group two modes are always the
same: the left mode (see Fig. 4) has a quality factor Q of 5.00 with
4.3. Results
a central frequency fc = 150 Hz and the right mode Q = 17.96 with
fc = 600 Hz. In each new experiment, the amplitude of the back-
The fatigue-life estimates for the frequency-domain methods
ground noise was increased by 20 MPa2/Hz until the smaller peak
were compared to the life estimate in the time domain using a
is hidden completely, see Fig. 4.
combination of the rainflow count and the Palmgren–Miner
The spectral width group tested how the spectral width affects
hypothesis, TRFC, which in this study is assumed to be an exact ref-
the accuracy of the different spectral methods. In each new exper-
erence value. The relative error is calculated using:
iment the spectral width was broader, as shown in Fig. 5. The ini-
tial mode with a central frequency of fc = 430 Hz and factor T XX  T RFC
Q = 8.96 expands by 50 Hz in each direction during each concur- T err ¼ ð47Þ
T RFC
rent experiment (up to 450 Hz wide).
The close-modes spectra were used to determine the effect of where TXX is a life estimate, as calculated using one of the
the close modes on the fatigue-life estimation. In each iteration frequency-domain methods. Because of the significant impact the
the peaks with a quality factor of Q equal to 6.00 and 24.00 and S–N slope k has on the accuracy of the fatigue-life estimation, three
 M. Mršnik et al. / International Journal of Fatigue 47 (2013) 8–17 15

Table 3
Percentage of relative errors for different margins for k = 3.324.

Rel. error
<0.05 <0.1 <0.2 <0.5
NB 0 0 21 75
WL 18 25 62 80
AL 0 0 27 87
GM 0 0 39 100
DK 86 100 100 100
ZB1 42 91 100 100
ZB2 65 96 100 100
TB1 0 31 80 100
TB2 100 100 100 100
PZ 0 0 0 0
% Of results

Table 4
Fig. 6. Close-modes spectra.
Percentage of relative errors for different margins for k = 7.3.

Rel. error
<0.05 <0.1 <0.2 <0.5
NB 0 0 0 71
WL 14 19 56 80
AL 0 0 2 75
GM 0 0 0 100
DK 15 37 83 100
ZB1 7 22 69 100
ZB2 20 25 64 100
TB1 0 4 33 100
TB2 40 71 100 100
PZ 7 25 25 38
% Of results

Table 5
Percentage of relative errors for different margins for k = 11.76.

Rel. error
<0.05 <0.1 <0.2 <0.5
Fig. 7. Typical automotive spectra.
NB 0 0 0 22
WL 6 18 41 87
AL 0 0 0 35
Table 1 GM 0 0 0 14
Ranges of a2 parameter for each test group. DK 7 13 24 100
MM BN SW CM AM ZB1 0 0 12 90
ZB2 0 7 13 68
a2 0.49–0.80 0.81–0.92 0.80–0.93 0.73–0.87 0.40–0.48 TB1 0 0 0 90
TB2 15 24 69 100
PZ 3 7 9 30
% Of results
different material parameters, presented by Petrucci and Zuccarello
[9], were used and the signal was normalized according to a suitable
RMS value. The data is presented in Table 2.
The rainflow ranges must be corrected for the cycle means with ative to complete (300 s) time-history fell below 1% at around
(42) before determining the relative error of the Petrucci–Zucca- 100 s long time-history.
rello life estimate with Eq. (47). For real-life applications it is very important that the damage-
The number of cycles (and half-cycles) counted in the time his- estimation method does not rely on the properties of the response
tory of a single random process ranged from 1.4  105 to 2.6  105. spectra (e.g., narrow band, close modes, etc.); therefore, in this re-
The stability of fatigue-life estimate at mentioned cycle counts was search we are trying to find the best methods in the sense that they
checked. The relative error of an estimate for a signal fragment rel- perform well, regardless of the test group being analyzed. For the
materials in Table 2, the results in Tables 3–5 show the percentage
of life estimates for each method that are inside a certain margin of
error, taking into account all the used spectra. For instance, for 21%
Table 2
Material parameters [9] and RMS for the fatigue-life calculation.
of samples the NB estimation method’s relative error is less than
0.2, according to Table 3.
C (MPak) k Su (MPa) RMS (MPa)
Because they exhibited the best performance, the improved
Steel 1.934  1012 3.324 725 5 Tovo–Benasciutti (TB2), Dirlik (DK), Zhao–Baker (ZB1) and im-
Aluminum 6.853  1019 7.300 446 10 proved Zhao–Baker (ZB2) methods were selected for a detailed
Spring steel 1.413  1037 11.760 1850 100
analysis; for each groups sample analysis see Figs. 8–10.
16 M. Mršnik et al. / International Journal of Fatigue 47 (2013) 8–17

5. Discussion

An ideal damage-estimation method for use in a design process

should be consistent across different spectra and different slopes of
the S–N curve. It would give results that are close or equal to those
given by a time-domain approach and preferably be conservative
when not accurate.
Based on Tables 3–5, a smaller number of frequency-domain
methods were selected, i.e., ZB1, ZB2, DK and TB2, and analyzed
further. Here, a brief assessment of the accuracy for each spectral
group is made and the best-performing method is chosen, based
on accuracy and the consistency of the fatigue-life estimates.

5.1. Multi-mode

For k = 3.324 the best results are obtained for the bi-modal and
tri-modal spectra. For fewer modes, the chosen methods tend to
overestimate the fatigue life, while for more modes they become
conservative, although the relative error is still very small. The esti-
mate is less exaggerated for higher values of k and becomes con-
Fig. 8. Comparison of the relative errors for the selected methods when k = 3.324.
servative at k = 11.76 with all the methods. TB2 is the most
accurate method in this group and manages to give a relatively
close estimate, even with an S–N slope of 11.760. However, only
DK however, gives a conservative result for every experiment.

5.2. Increased background noise

The TB2 method does overestimate, in a few cases, for the BN

group of spectra, but it is also much more accurate than the other
three methods for all the different S–N slopes k. For k = 3.324 the
estimates are exceptionally accurate, as is the case with all five
groups of spectra. Even though DK is strictly conservative, TB2 is
still the recommended method for this group, and with a much
better accuracy.

5.3. Spectral width

Estimates for the slope k = 3.324 are again very good for all four
methods. However, with higher k some inconsistencies arise that
affect each method. TB2 is recommended in this group because
of its accurate and always-conservative estimates.

Fig. 9. Comparison of the relative errors for the selected methods when k = 7.3. 5.4. Close-modes

The best-performing methods overall, DK and TB2, exaggerate

the fatigue life when the modes are farther apart, but with a small
relative error. ZB1 and ZB2 do offer a conservative estimate at all
times, but with a large relative error, especially at high values of
k. It is clear that all four methods face inconsistencies with modes
that are positioned far apart. TB2 is again the recommended meth-
od, but one must be aware of the overestimation problems.

5.5. Typical automotive spectra

The automotive spectra ZB1 and ZB2 stand out as the best-per-
forming methods. ZB1 only slightly overestimates the fatigue life,
and achieves better accuracy than ZB2, which is always conserva-
tive. Only at the highest value of the slope k do these methods fall
behind DK and TB2 in in terms of accuracy.

6. Conclusion

This theoretical and experimental research on the frequency-

domain methods for vibration-fatigue estimation focuses on the
Fig. 10. Comparison of the relative errors for the selected methods when k = 11.76. questions typical for the structural dynamics (i.e., close modes,
 M. Mršnik et al. / International Journal of Fatigue 47 (2013) 8–17 17

noise background, mode numbers and spectral width) in the [7] Zhao W, Baker MJ. On the probability density function of rainflow stress range
for stationary gaussian processes. Int J Fatigue 1992;14(2):121–35.
automotive accelerated test. In the theoretical background the
[8] Tovo R. Cycle distribution and fatigue damage under broad-band random
well-established and also recent methods are presented and dis- loading. Int J Fatigue 2002;24(11):1137–47.
cussed: Dirlik, Tovo–Benasciutti (2 versions), Wirsching–Light, [9] Petrucci G, Zuccarello B. Fatigue life prediction under wide band random
Zhao–Baker (2 versions), Petrucci–Zuccarello, empirical a0.75 and loading. Fatigue Fract Eng Mater Struct 2004;27(12):1183–95.
[10] Shin K, Hammond JK. Fundamentals of signal processing for sound and
the Gao–Moan method. Some of the methods have been compared vibration engineers. John Willey & Sons, Ltd.; 2008.
with each other, while this research compares all the methods [11] Lutes LD, Sarkani S. Random vibrations: analysis of structural and mechanical
(including very recent ones) side by side. Furthermore, this re- systems. Elsevier; 2004.
[12] Dirlik T. Application of computers in fatigue analysis. Ph.D. thesis. The
search focuses on the real experimental data obtained with an University of Warwick; 1985.
electro-dynamical shaker experiment. The vibration profiles typi- [13] Benasciutti D, Tovo R. Spectral methods for lifetime prediction under wide-
cal in structural dynamics and in an automotive-industry acceler- band stationary random processes. Int J Fatigue 2005;27(8):867–77.
[14] Bishop NWM. The use of frequency domain parameters to predict structural
ated test are analyzed. Thus a group of best performing methods fatigue. Ph.D. thesis. University of Warwick; 1988.
is selected, which can be applied to a general broad-band spectra. [15] Papadimitriou C, Fritzen CP, Kraemer P, Ntotsios E. Fatigue predictions in
At the same time, an estimate of the error one can expect is entire body of metallic structures from a limited number of vibration sensors
using kalman filtering. Struct Control Health Monit 2011;18(5):554–73.
provided. [16] Slavič J, Opara T, Banov R, Dogan M, Česnik M, Boltežar M. Development of
The frequency-based methods are compared to the time-do- next-generation vibration-fatigue software and catia integration for the
main rainflow method. Twenty-eight different fatigue loads (five automotive industry. Yugoslav Society of Automotive Engineers: Belgrade;
2011. p. 117–25.
different groups of loads) were experimentally compared. If all
[17] Cristofori A, Benasciutti D, Tovo R. A stress invariant based spectral method to
28 loads are evaluated, then the improved Tovo–Benasciutti is estimate fatigue life under multiaxial random loading. Int J Fatigue
found to be the best method, followed by the improved Zhao–Ba- 2011;33(7):887–99.
ker and Dirlik methods. These findings are in contrast to the find- [18] Braccesi C, Cianetti F, Lori G, Pioli D. The frequency domain approach in virtual
fatigue estimation of non-linear systems: the problem of non-gaussian states
ings of Benasciutti and Tovo [22,13], who obtained better results of stress. Int J Fatigue 2009;31(4):766–75.
with the Dirlik method. [19] Bishop NWM, Hu Z, Wang R, Quarton D. Methods of rapid evaluation of fatigue
Overall, the Dirlik, Zhao–Baker and Tovo–Benasciutti fre- damage on howden HWP330 wind turbine; 1993.
[20] Bouyssy V, Naboishikov SM, Rackwitz R. Comparison of analytical counting
quency-domain methods are all very consistent when the material methods for gaussian processes. Struct Saf 1993;12(1):35–57.
fatigue parameter k (S–N slope) is relatively low (k  3). With stee- [21] Ghao Z, Moan T. Frequency-domain fatigue analysis of wide-band stationary
per slopes the error increases. Similar conclusions were already gaussian processes using a trimodal spectral formulation. Int J Fatigue
2008;30(10–11):1944–55.
made by Bouyssy et al. [20] and Benasciutti and Tovo [22]. [22] Benasciutti D. Fatigue analysis of random loadings. Ph.D. thesis. University of
By analyzing selected load groups it was found that the im- Ferrara, Italy; 2004.
proved Tovo–Benasciutti method gives the most accurate results [23] Wirsching PH, Light MC. Fatigue under wide band random stresses. J Struct Div
ASCE 1980;106(7):1593–607.
if the increased background noise spectrum is increased or if the [24] Lagoda T, Macha E, Nieslony A. Fatigue life calculation by means of the cycle
spectral width is increased. For the close-modes group and for counting and spectral methods under multiaxial random loading. Fatigue Fract
the multi-mode group the improved Tovo–Benasciutti method Eng Mater Struct 2005;28(4):409–20.
[25] Braccesi C, Cianetti F, Lori G, Pioli D. Fatigue behaviour analysis of mechanical
was found to give a non-conservative damage estimation, while
components subject to random bimodal stress process: frequency domain
the Dirlik method always gave a conservative estimation; however, approach. Int J Fatigue 2005;27(4):335–45.
for the higher values of the material fatigue parameter k the Dirlik [26] Fu T, Cebon D. Predicting fatigue lives for bimodal stress spectral densities. Int
estimation error (compared to the rainflow time-domain method) J Fatigue 2000;22(1):11–21.
[27] Sakai S, Okamura H. On the distribution of rainflow range for gaussian random
is up to 50%. The improved Zhao–Baker method was found to be loading processes with bimodal psd. JSME Int J 1995;38(4):440–5.
the most accurate (giving a conservative fatigue-damage estima- [28] Bendat J. Probability functions for random responses. In: NASA report; 1964.
tion) for the typical automotive-industry accelerated test profiles. [29] Low YM. A method for accurate estimation of the fatigue damage induced by
bimodal processes. Probab Eng Mech 2010;25(1):75–85.
This research showed that the improved Tovo–Benasciutti and [30] Lagoda T, Macha E, Nieslony A. Comparison of the rain flow algorithm and the
the improved Zhao–Baker methods should also be considered to- spectral method for fatigue life determination under uniaxial and multiaxial
gether with the Dirlik method. In general, the improved Tovo– random loading (conference paper). In: Fatigue testing and analysis under
variable amplitude loading conditions, vol. 1439. ASTM Special Technical
Benasciutti method gives best results, while the improved Zhao– Publication; 2003, p. 544–56.
Baker method gives the best results for the tested vibration profiles [31] Kim CJ, Kang YJ, Lee BH. Experimental spectral damage prediction of a linear
typical in the automotive industry. elastic system using acceleration response. Mech Syst Signal Process
2011;25(7):2538–48.
[32] Čermelj M, Boltežar M. Modelling localised nonlinearities using the harmonic
References nonlinear super model. J Sound Vib 2006;298(4–5):1099–112.
[33] Newland DE. An introduction to random vibrations and spectral
[1] Rychlik I. Fatigue and stochastic loads. Scand J Stat 1996;23(4):387–404. analysis. Longman Scientific & Technical; 1993.
[2] Matsuishi M, Endo T. Fatigue of metals subject to varying stress. Paper [34] Vanmarcke EH. Properties of spectral moments with applications to random
presented to Japan Society of Mechanical Engineers, Fukuoka, Japan; 1968. vibration. J Eng Mech Div ASCE 1972;98:425–46.
[3] Palmgren A. Die lebensdauer von kugellagern. VDI-Zeitschrift [35] Rice SO. Mathematical analysis of random noise. Bell Syst Tech J 1944;23(3).
1924;68:339–41. [36] Bishop NWM, Sherrat F. Finite element based fatigue calculations. NAFEMS
[4] Miner M. Cumulative damage in fatigue. J Appl Mech 1945;12:A159–64. Ltd.; 2000.
[5] Petrucci G, Zuccarello B. On the estimation of the fatigue cycle distribution [37] Maia NMM, Silva JMM. Theoretical and experimental modal analysis. Res Stud
from spectral density data. Proc Institut Mech Eng Part C – J Eng Sci Pre; 1997.
1999;213(8):819–31. [38] Miles JW. On structural fatigue under random loading. J Aeronaut Soc
[6] Halfpenny A. A frequency domain approach for fatigue life estimation from 1965;21:753–62.
finite element analysis. In: Gilchrist MD, DulieuBarton JM, Worden K, editors. [39] Zalaznik A, Nagode M. Frequency based fatigue analysis and temperature
DAMAS 99: damage assessment of structures, vol. 167-1 of Key Engineering effect. Mater Des 2011;32(10):4794–802.
Materials; 1999, p. 401–10. [40] Goodman J. Mechanics applied to engineering. Longman, Green & Co.; 1954.