Comparison of Frequency-Domain Methods for a

Vibration-Fatigue-Life Estimation

J. Slavic, M. Mrsnik and M. Boltezar

Faculty of Mechanical Engineering, University of Ljubljana

Askerceva 6, 1000 Ljubljana, Slovenia
E-mail: janko.slavic@fs.uni-lj.si

ABSTRACT. Vibration fatigue analysis is a two part process. A structure model must
be built, be it in a simulated or real environment, which enables the study of stress loads
in material points. On the basis of those stress states, different methods are then used to
asses fatigue life. Quite a few frequency domain methods exist, that can operate on ob-
tained power spectral densities, to give the fatigue life estimate. The results differ among
methods and thorough analysis must be made, before such methods can be applied with
confidence. Comparison studies have already been made, but they compare different sets
of methods on different data. Based on specifics of structural dynamics responses, a new
approach to comparison has been devised. Different groups of spectra are aimed at evalu-
ating the sensitivity of the frequency domain methods on different changes in spectra, such
as changing the distance between modes or raising and lowering the level of background
noise. The accuracy is also tested on some typical accelerated-vibration test spectra,
often used in the automotive industry. To come as close as possible to real-world prob-
lems, stress data is measured in the form of electro-dynamic shaker vibration. Fatigue
life estimates are compared based on the value of fatigue life given by more conventional
approach in the time-domain. Conclusions show, that methods by Tovo and Benasciutti,
Zhao and Baker and by Dirlik give good results.

INTRODUCTION

For the purpose of vibration fatigue analysis, knowledge of structural dynamics and fa-
tigue must be combined. This paper researches the available methods that can be used to
evaluate the fatigue-life of a structure (or to find a critical point), once the stress response
in the nodes on the structure is known. A number of frequency domain methods for
fatigue life analysis have been collected and compared side-by-side, based on some ref-
erence estimate, that was calculated by a more conventional approach in the time domain
(using the Palmgren-Miner [1, 2] hypothesis and the Rainflow count [3]). The meth-
ods considered for comparison are: narrow-band approximation [4], empirical 0.75 [5],
Wirsching-Light [6], Gao-Moan [7], Tovo-Benasciutti (2 versions) [8, 9], Zhao-Baker (2

1
versions) [10] and Petrucci-Zuccarello [11].
Similar studies have already been made, but the choice of compared methods is differ-
ent, and so are the analyzed random responses. Aforementioned studies (e.g. by Gao and
Moan [7] and by Tovo and Benasciutti [9]) find the Dirlik method to be most accurate and
this is indeed one of the more popular methods available [12]. However, it is relatively
old, being devised in 1985 [13] and quite a few methods have been developed since.
Before relying on any of the frequency domain methods which have been developed
based on numerous numerically simulated spectra, the accuracy of results must be eval-
uated specifically on the spectra that are of interest. One should aim to chose the best
possible (i.e. most accurate) method for the specific application. In the course of this re-
search, the studied spectra are those exhibited by material nodes in a structure, responding
to some random vibration of the Gaussian distribution.
Different groups of carefully devised power-spectral-densities were analyzed using the
collected methods, a total of 27 spectra. Spectral shapes were determined based on typical
structural responses of structures and on accelerated-vibration-tests spectra, that are often
used in the automotive industry.
The data was produced by means of experiment, to resemble as close as possible a real-
world scenario. The rainflow counting method, developed by Matsuishi and Endo [3] was
used to calculate the reference time-domain fatigue-life estimate. To take into account the
effect of material (i.e. the S N curve parameters), pointed out by existing studies [14, 9],
three different sets of parameters, found in literature [11] were used.
Throughout this study, the basics of vibration fatigue are presented in the following
section. The experiment is described next. Concluding remarks are given in the end.

VIBRATION FATIGUE

When approaching fatigue-life evaluation of a structure, the stress distribution is first

calculated or measured. Modal analysis is a powerful tool that can aid in deducing the
response spectra in material nodes. The basics of this approach are presented in the fol-
lowing section, but for a detailed explanation, the reader is referred to [15].
Once the responses are known, they are characterized based on spectral moments.
More detailed explanation can be found in books by Shin and Hammond [16] or by New-
land [17]. Only very brief introduction is made in the following sections.

Structural Dynamics
Modal analysis can already be applied in FEM environment [18, 19, 20] and supple-
mented later with the experimental modal analysis. Structures are modeled as linear,
multi-degrees-of-freedom systems with second-order differential law [15]:

[M] {x} + [C] {x} + [K] {x} = {f} (1)

where [M] is the mass matrix, [C] is the damping matrix, [K] is the stiffness matrix, {x}
is the vector of degree- of-freedom and {f} is the excitation force vector. The frequency-

2
response-functions (FRF) connecting a material point on the structure to another point are
deduced. There are different FRFs, that can be calculated, but for fatigue-life evaluation,
the FRF from acceleration a (vibration load) to material stress s is of special interest. For
many nodes it takes the matrix form:

s = Has a (2)
where s is the stress tensor in the frequency domain, a is the forced vibration accelera-
tion profile and Has is the FRF matrix. Once the vibration load a and the transfer functions
are known, stresses s can be calculated and evaluated for fatigue.

Fatigue-Life in Frequency-Domain
In fatigue analysis, the Palmgren-Miner [1, 2] hypothesis is used to sum the damages of
respective cycles, that have been identified beforehand using the Rainflow [3] counting
method. The accumulated damage can be written as a product of time T and damage
intensity D:
 Z 
1 k
D = T D = T p C s pa (s) ds (3)
0
where C and k are material parameters, that define the S N curve. p gives the
expected frequency of peaks and the peak amplitudes are given with probability density
function pa (s). These values, as it turns out, can be estimated using the moments of the
PSD response spectra [17]:
Z
mi = f i GXX ( f ) d f . (4)
0
r
m4
p = (5)
m2
where GXX ( f ) is a one sided PSD of the stress response, and its moments correspond
to the variance of the random process X2 and, for the higher moments, the variances of
the derivatives of the random process X2 .

X2 = m0 X2 = m2 (6)

The probability density function pa (s) is calculated differently, depending on the frequency-
domain method used, but in most cases it is derived empirically, because of the complexity
of the Rainflow counting rules that are applied in the time-domain.
A group of different frequency domain methods was compared side-by-side in this
study, namely: the narrow-band approximation [4], Wirsching-Light method (WL) [6],
the 0.75 method (AL) [5], the Gao-Moan method (GM) [7], the Dirlik method (DK) [13],
both Zhao-Baker methods (ZB1 and ZB2) [10], the Tovo-Benasciutti methods (TB1 and

3
 103
PSD [MPa2 /Hz] 103

PSD [MPa2 /Hz]

102

102

101

101
100
0 100 200 300 400 500 600 700 800 900 0 100 200 300 400 500 600 700 800 900
f [Hz] f [Hz]

Figure 1: The multi-mode (MM) group of Figure 2: The background noise (BN) group
spectra. of spectra.

TB2) [8, 9] and the Petrucci-Zuccarello method (PZ) [11]. The formulae for the con-
sidered methods are available in literature. All listed methods are aimed at analysis of
Gaussian processes.

EXPERIMENT

Real data was obtained using an experimental setup, by which the stress response spec-
tra were simulated directly on the electro-dynamic shaker. The collected data was then
processed using different frequency-domain methods and also with more standard time-
domain approach, for which it was assumed it is correct.

Response Spectra
The spectra, considered for the comparison of the accuracy of frequency-domain methods
were devised based on experience in structural dynamics analysis. Specifically the spec-
tra that characterize response in different material nodes were the focus of our research.
Producing each of the devised response spectra using a vibrating structure would be very
hard to achieve and since the response was already predicted, it was simulated on the
shaker directly, without any structure used.
Five different groups of response spectra were conceived, namely the multi-mode
(MM), background noise (BN), spectral width (SW), close modes (CM) and automo-
tive (AM) group. The spectra were varied by some parameter in each group, such as
the number of modes, level of background noise (see Figures 1 and 2) and distance be-
tween modes. This way, the effect of different structural dynamics phenomena on the
fatigue-life estimate accuracy was studied. The automotive group contained profiles, that
are often used in the automotive industry for accelerated-vibration tests, to validate the
fatigue strength of a part or component. The frequency range of each spectra was 10 to
1000 Hz.

Experimental set-up
The experimental set-up consisted of an electro-dynamic shaker, shaker controller, con-
ditioning amplifier and data acquisition hardware. The spectra were programmed into

4
 Figure 4: Electro-dynamic shaker with the
conditioning amplifier and the
Figure 3: Schematics of the experiment. accelerometer.

shaker controller and an accelerometer, mounted on the vibrating plate measured 5 min
time-histories of vibration, that were processed later on. The fatigue-life estimate was
found to converge well for a 5 min sample.

Results
The calculations were made for different materials, namely steel, aluminum and spring
steel. The relative error Terr of each frequency-domain method T XX was calculated in
reference to the time-domain approach estimate T RFC :

T XX TRFC
Terr = (7)
TRFC
After close inspection of the results it was clear, that methods by Tovo and Benasciutti,
Dirlik and by Zhao and Baker give results well above average. These methods are com-
pared side-by-side in Figures 5, 6 and 7. For each comparison a different pair of material
parameters was used: steel with k = 3.324, aluminum with k = 7.300 and spring steel
with k = 11.730 [11].

CONCLUSIONS

Different frequency-domain methods for fatigue-life evaluation have been compared side-
by-side in the course of this research. Real data was used, obtained by experiment and
carefully constructed response spectra were used in comparison.
The comparison was made based on the reference estimate, provided by the time-
domain approach. Although the Palmgren-Miner hypothesis is not exact in the sense that
it disregards the importance of the sequence and frequency with which the loading cycles
occur, it can still be used as a good measure of accuracy of frequency-domain methods
based on its wide-spread use. The main point is actually the ability to recover cycle
amplitudes directly in the frequency-domain.
Data, gathered by the application of frequency-domain methods shows, that a narrow
group of methods is especially suitable for fatigue-life evaluation. The Tovo-Benasciutti,

5
 0.6 TB2 DK ZB1 ZB2

0.4

Rel. error 0.2

0.0

0.2

0.4

0.6

Figure 5: Comparison of relative errors for k = 3.324.

0.6 TB2 DK ZB1 ZB2 0.6 TB2 DK ZB1 ZB2

0.4 0.4

0.2 0.2
Rel. error
Rel. error

0.0 0.0

0.2 0.2

0.4 0.4

0.6 0.6

Figure 6: Comparison of relative errors for Figure 7: Comparison of relative errors for
k = 7.300. k = 11.73.

Dirlik and Zhao-Baker methods should be the preferred choice as they proved to be most
accurate on the spectra used in this research.
The most accurate method for the typical automotive-industry accelerated test pro-
files was the improved Zhao-Baker method, consistently providing a conservative fatigue-
damage estimate. Overall, the improved Tovo-Benasciutti and Dirlik methods were most
accurate. Results are a bit different from those presented by Benasciutti and Tovo [5, 9],
where Dirlik was deemed most accurate.
The observed effect of the S N curve parameter k was expected. Studies by Buoyssy
et al. [14] and by Benasciutti and Tovo [5] came to similar conclusions. For high values
of parameter k, the errors and inconsistencies in results grow bigger.

6
REFERENCES

1. Palmgren, A. (1924). VDI-Zeitschrift 68, 339341.

2. Miner, M. (1945). J. Appl. Mech. 12, A159A164.

3. Matsuishi, M. and Endo, T. (1968). Paper presented to Japan Society of Mechanical

Engineers, Fukuoka, Japan .

4. Miles, J.W. (1965). J. Aeronaut. Soc. 21, 753762.

5. Benasciutti, D. (2004). Fatigue analysis of random loadings, Ph.D. thesis, University

of Ferrara, Italy.

6. Wirsching, P.H. and Light, M.C. (1980). J. Struct. Div. ASCE 106, 15931607.

7. Ghao, Z. and Moan, T. (2008). Int. J. Fatigue 30, 19441955.

8. Tovo, R. (2002). Int. J. Fatigue 24, 11371147.

9. Benasciutti, D. and Tovo, R. (2005). Int. J. Fatigue 27, 867877.

10. Zhao, W. and Baker, M.J. (1992). Int. J. Fatigue 14, 121135.

11. Petrucci, G. and Zuccarello, B. (2004). Fatigue & Fract. Eng. Mater. & Struct. 27,
11831195.

12. Halfpenny, A. (1999). in: DAMAS 99: Damage Assessment of Structures, volume
167-1 of Key Engineering Materials, (Eds.) M.D. Gilchrist, J.M. DulieuBarton and
K. Worden, pp. 401410.

13. Dirlik, T. (1985). Application of Computers in Fatigue Analysis, Ph.D. thesis, The
University of Warwick.

14. Bouyssy, V., Naboishikov, S.M. and Rackwitz, R. (1993). Struct. Saf. 12, 3557.

15. Maia, N.M.M. and Silva, J.M.M. (1997). Theoretical and Experimental Modal Anal-
ysis, Research Studies Press Ltd, Hertfordshire.

16. Shin, K. and Hammond, J.K. (2008). Fundamentals of signal processing for sound
and vibration engineers, John Willey & Sons, Ltd, Chichester.

17. Newland, D.E. (1987). An introduction to Random vibrations and spectral analysis,
Longman Scientific & Technical, Harlow.

18. Slavic, J., Opara, T., Banov, R., Dogan, M., Cesnik, M. and Boltezar, M. (2011).
in: Automotive engineering for improved safety, pp. 117125, Belgrade: Yugoslav
Society of Automotive Engineers.

7
19. Cermelj, M. and Boltezar, M. (2006). J. Sound Vib. 298, 10991112.

20. Bishop, N.W.M. and Sherrat, F. (2000). Finite Element Based Fatigue Calculations,
NAFEMS Ltd, Farnham.