Documente Academic
Documente Profesional
Documente Cultură
1. Introduction
Many fatigue analyses from FE models use an elastic FEA for a unit applied load. The fatigue
software uses a description of the service load history to scale the results.
The finite element load case will consist of a linear elastic FEA solution for the stresses at each
node, calculated for a single applied load most conveniently a unit load. These results will be
written to the FEA results file as a step. At each node, the elastically-calculated stress tensor is
multiplied by the load history to give a time history of the stress tensor.
On the surface of the model, the fatigue software will calculate the time histories of the in-plane
principal stresses, and their directions. The time history of the principal stresses can be converted
into elastic-plastic stress-strains using a multiaxial cyclic plasticity model. This strain-time history
can be used in a strain-life fatigue calculation and the associated stresses can be used to apply a
mean stress correction. This procedure is repeated for each node on the model.
Components with multiple load directions can be analysed. Each load direction is modelled
separately in the FEA, and the fatigue software uses the principle of superimposition.
For some components, the sequence of stresses may be calculated in the FE analysis. For example,
an engine crankshaft FE analysis may model the stresses for each 5o of rotation through two or
three complete revolutions of the crankshaft. The fatigue software follows this sequence of
stresses.
The accuracy of the fatigue analysis results will depend on (i) the way in which the loading
information has been processed, (ii) the materials fatigue data, (iii) the FEA mesh and (iv) the
fatigue analysis algorithm. These four items are discussed in this paper.
2. Load processing
2.1
Sample frequency
The fatigue analysis requires an accurate description of the peaks and valleys in the load history.
Analogue signals must be sampled at an appropriate sample frequency. As an example, a sine
wave sampled at four times its frequency could produce many different sets of sampled values.
Two possible sets of samples are shown in Figure 1. In the upper example, the amplitude of the
signal has been determined correctly. In the lower example, the amplitude has been
underestimated, and a fatigue life calculated from these samples would very non-conservative.
Figure 2 shows the effect of sample frequency on the accuracy of the subsequent fatigue analysis.
Narrow band and broad band Gaussian random signals were used for this study. A sample
frequency of 100 points per cycle was used as a datum, and the effect of reducing this sample
frequency is shown in Figure 2. It can be seen that sampling at 10 times the signal frequency gave
calculated lives of 1.1 times the true value for a broad band signal, and up to 1.5 times the true
value for a narrow band signal. A sample frequency of 10 points/cycle is now widely used in
industry, as it offers a reasonable compromise between accuracy of analysis and quantity of data
(and hence analysis time).
2.2 Peak-valley extraction
Measured load histories can be truncated by extracting the peaks and valleys from the sampled
signal. Because real signals contain a large number of very small fluctuations, it may be
convenient to omit them during the peak/valley extraction. This process is known as cycle
omission, or gating (Figure 3).
The cycle omission criterion, or gate level, must be chosen with care. Many materials exhibit an
endurance limit stress amplitude under constant amplitude testing. Under variable amplitude
loading the endurance limit may disappear or its amplitude may be very much reduced [Conle,
1980],[ DuQuesnay, 1993]. Figure 4 shows a measured strain history from a truck steering arm
(upper signal), and the strain history that is produced if all the cycles smaller than the constant
amplitude endurance limit are removed. Fatigue testing using the truncated signal produced
fatigue lives which were 9 times longer than those produced using the full signal [Kerr, 1992].
Peak-valley extraction can also be carried out on multiaxial loading signals. In this case it is
necessary to retain the phase relationship between the signals. To do this, each time a peak or
valley occurs on one signal, the corresponding data points on the other signal are also retained.
The principle is illustrated in Figure 5. Gating to omit small cycles can be integrated into this
processing operation. The danger in this procedure is illustrated by considering the way these
signals are used in the fatigue analysis of a node in a finite element model.
(a) The unit load stress tensor for each node is multiplied by its corresponding load history, to
produce time histories of each stress tensor.
(b) The time histories of the stress tensors are superimposed.
(c) The time histories of the principal stresses are calculated.
(d) The damage parameter (for example the time history of the shear strains on a critical plane) is
calculated.
The peak/valley procedure in Figure 5 therefore assumes that a peak or valley in the principal
strains will always coincide with a peak or valley in one of the load histories. In general this is far
from being true and serious errors in the calculated fatigue lives can be produced by peak-valley
extraction of multiaxial loading histories. The increase in processing speed can be dramatic, but
the potential errors are great. Safe Technologys fe-safe software does not peak-valley multiaxial
loading histories unless the user specifically requests it. A sensitivity analysis should always be
carried out to asses the effect on the calculated fatigue lives.
2.2
Figure 6 shows fatigue damage histograms for a fatigue analysis of the first 3 000, 30 000 and 300
000 cycles of a long signal. Although the calculated fatigue lives (adjusted for the different
lengths of signal) were very similar, the fatigue damage distribution for the shortest signal is
dominated by the largest few cycles. This is a characteristic of short signals. It is possible to
obtain quite adequate calculated fatigue lives from relatively short lengths of signals, but these
lives are much more dependant of the accuracy of measurement of the few largest cycles, and on
the statistical validity of their frequency of occurrence.
3. Materials data
Fatigue analysis requires the parameters for the relationship between strain amplitude and fatigue
life
De = s f (2 N )b + e (2 N ) c
f
f
f
2
E
where De
is the applied strain range
2 N f is the endurance in reversals
s f
e f
b
c
( )
where
n
e = s + s
E
K
E
is the cyclic elastic modulus
K
is the cyclic strain hardening coefficient
n
is the cyclic strain hardening exponent
This data is widely available for many commonly-used steels, aluminium alloys and cast irons
(see for example [Boller, 1987]). Where data is not available, approximation algorithms may be
used. The Seeger approximation can give acceptable estimates for materials properties. Two
examples, comparing the estimated and measured properties of a steel and an aluminium alloy, are
shown in Figures 7 and 8.
4. FE mesh effects
Fatigue cracks often initiate from the surface of a component. The accuracy of the surface stresses
therefore has a significant effect on the accuracy of the subsequent fatigue analysis. [Colquhoun,
2000] compared calculated fatigue lives for a forged aluminium suspension component, using a
preliminary and a final mesh, and found significant differences (Figure 9). The final mesh
produced fatigue lives which correlated very well with the results of a fatigue test of the
component with a calculated life to crack initiation of 27 000 miles, compared to a test life of 41
000 miles at which quite long fatigue cracks were discovered. A difference of less than 15%
between un-averaged and averaged nodal stresses is a reasonable criterion for defining an
adequate mesh density for fatigue analysis.
In selecting the parameter for analysis, possible options are integration point stresses (Gauss
points), elemental averaged stresses, nodal averaged stresses, or un-averaged nodal stresses.
Integration point and elemental averaged stresses do not normally give adequate estimates of the
surface stresses, and are not recommended.
With an adequate mesh, there should be little difference in the lives calculated from nodal
averaged stresses, or un-averaged nodal stresses. In practice, mesh density is rarely ideal, and
experience has shown that fatigue lives calculated from un-averaged nodal stresses correlate most
closely with test results. A recommended method of assessing mesh density is to compare fatigue
life contour plots, calculated from un-averaged nodal stresses, with different amounts of averaging
set in the contour plot software.
Uniaxial fatigue
The use of uniaxial fatigue methods to analyse biaxially stressed components can give very
optimistic life estimates. In [Devlukia, 1985] a welded steel bracket from a passenger car
subjected to multiaxial loading developed fatigue cracks at a life much shorter than that predicted
by uniaxial local strain fatigue analysis. The component had also been tested under two different
service duties and uniaxial analysis failed to reproduce the relative severity of the two duties.
[Bannantine, 1985] reported the following results from a multiaxial fatigue test programme (Table
1). The three specimens were (i) simple bending, (ii) in-phase bending and torsion and (iii) axial
and torsion loading with random phase relationship. Fatigue life predictions using uniaxial
methods were always non-conservative, with a predictions up to 19 times the achieved test life.
5.2
Early attempts to analyse biaxial fatigue were based on principal stresses, using a conventional SN curve. For a fatigue cycle, the stress range of Ds 1 , or the stress amplitude
Ds 1
, would be
2
used with a stress-life curve obtained by testing an axially loaded specimen. The (false)
assumption in this procedure is that the fatigue life is always determined by the amplitude of the
largest principal stress s 1 , and therefore that the second principal stress s 2 has no effect on
fatigue life.
Consider a simple circular shaft loaded in pure torsion. If
s 1,2 = t xy2
i.e. the maximum principal stress is equal to the torsion stress. A fatigue cycle of
produce a principal stress cycle of
t xy will
predicts that the fatigue strength in torsion is the same as the fatigue strength under axial loading.
This is not supported by test data, as Figure 11 shows.
Figure 11 shows the results of fatigue tests on a commonly-used steel. It is clear that the torsion
fatigue strength is much lower than the axial fatigue strength - the allowable principal stress in
torsion is approximately 60% of the allowable axial stress. Calculating fatigue lives using
principal stress will clearly be grossly optimistic for torsion loading, and allowable torsion fatigue
stresses will be overestimated by a factor of 1/0.6 = 1.66. This could mean the difference between
identifying and missing a potential fatigue 'hot spot'. (In 1927, Moore reported that From the
quite considerable amount of test data available for fatigue tests in torsion the general statement
may be made that under cycles of reversed torsion the endurance limit for metals ranges from 40
per cent to 70 per cent of the endurance limit under cycles of reversed flexure [Moore, 1927]).
It has been shown over the past 20 years that principal stresses should only be used for fatigue
analysis of 'brittle' metals, for example cast irons and some very high strength steels. A fatigue
analysis using principal stresses tends to give very unsafe fatigue life predictions for more ductile
metals including most commonly-used steels and aluminium alloys.
5.3
This criterion proposes that fatigue cracks initiate on planes which experience the largest
amplitude of principal strain. The standard strain-life equation for unixial stresses is
De = s f (2 N )b + e (2 N ) c
f
f
f
2
E
where De is the applied strain range
2 N f is the endurance in reversals
2003 ABAQUS Users Conference
s f
e f
b
c
Replacing the axial strain with the maximum principal strain gives :
s f
De1
(2 Nf )b
=
2
E
e f (2 Nf ) c
The SAE multiaxial test programme [Tipton, 1989] used a 40mm diameter notched shaft with
5mm fillet radii, machined from SAE1045 steel. The specimens were tested under pure bending
loads, pure torsion loads, and combined bending-torsion with various proportions of bending and
torsion. The test results have been compared with life estimates made from measured strains at the
notch. The maximum principal strain criterion produced life estimates which were nonconservative, particularly at lower values of endurance, and the scatter was large (Figure 12).
Experience has shown that this criterion should be used only for fatigue analysis of brittle metals,
for example as cast irons and some very high strength steels.
5.4
Because the von Mises criterion provides an estimate of the onset of yielding, it has been
proposed as a criterion for fatigue life estimation.
The strain-life equation in terms of von Mises equivalent strain is
s f
De EFF
(2 Nf )b
=
2
E
e f (2 Nf ) c
e EFF = b (e1 - e 2 ) + (e 2 - e 3 ) + (e 3 - e1 )
The value of
2 0.5
b is chosen so that e EFF has the same value as the principal strain e1 for the
2
2
2
s EFF = 1 (s 1 - s 2 ) + (s 2 - s 3 ) + (s 3 - s 1 )
2
0.5
Ds EFF
= s f (2 N f ) b
2
A major problem with the practical application of von Mises criteria to measured signals is that
the von Mises stress or strain is always positive, even for negative values of stress or strain, and so
Rainflow cycle counting cannot be applied directly. Some approximations have been proposed,
such as to assign the sign of the largest stress or strain to the von Mises stress or strain, or
alternatively to assign the sign of the hydrostatic stress or strain to the von Mises stress or strain.
These are termed signed von Mises criteria. The different methods of determining the sign can
give significantly different life estimates.
The von Mises criteria correlate poorly with test data, particularly for biaxial stresses when the
two in-plane principal stresses change their orientation during the fatigue loading.
5.5
Brown-Miller criterion.
The Brown-Miller equation proposes that the maximum fatigue damage occurs on the plane which
experiences the maximum shear strain amplitude, and that the damage is a function of both this
shear strain g max and the strain normal to this plane, e N
s f
Dg max
De N
(2 N f ) b
+
= 1.65
2
2
E
1.75 e f (2 N f ) c
This formulation of the Brown-Miller parameter was developed by Kandil, Brown and Miller
[Kandil, 1982].
The Brown-Miller criterion is attractive because it uses standard uniaxial materials properties.
Figure 13 shows the results from the SAE test programme [Tipton, 1989]. In general, test results
and predictions agreed to within a factor of 3. The Brown-Miller criterion is widely accepted
for the analysis of most metals with the exception of very brittle metals such as cast irons.
More recently, Chu, Conle and Bonnen [Chu, 1993] have shown improved correlation if the mean
shear stress is included, and have proposed the following extension to the Brown-Miller equation,
using a mean stress correction similar to a Smith-Watson-Topper correction
( s f ) (2 N ) 2b + 1.04 s e (2 N ) b+c
De
Dg
t max + N s N ,max = 1.02
f
f f
f
2
2
E
2
where
t max
and
Varvani-Farahani has further extended the Brown-Miller equation, by weighting the contribution
of the normal and shear stress/strains using the axial and torsion fatigue strength coefficients.
[Varvani-Farahani, 2000], [Varvani-Farahani, 2003].
s N ,m
1 +
s f
g
Dt max D max
t f g f
2
where
+ De N Ds N = f (2 N f )
sf e f
s N ,m
Dt max
Dg max
Ds N
De N
This equation has shown excellent correlation for constant amplitude loading where the two inplane principal stresses have the same amplitude but different frequencies, and it is being assessed
for random loading. However, it requires both axial and torsion fatigue test data.
6. Concluding remarks
This paper has given some guidelines to be followed when planning a fatigue analysis of a finite
element model. Many of the guidelines are set as defaults in fe-safe, allowing engineers with
relatively little fatigue experience to carry out successful analyses.
Processing speeds are also impressive. To give two examples: the fatigue analysis in fe-safe of a
700 000 element model (4-noded solid elements) containing two load steps, in a 3 GByte file,
took 35 minutes on a UNIX workstation. For an 8 GByte FEA results file containing 36 load
steps, the total fe-safe time for read-in, fatigue analysis of the 36 load steps in sequence, and
export of results, took 1 hour 15 minutes on a PC running Windows.
7. References
Bannantine J A, Socie D F. A variable amplitude multiaxial fatigue life prediction method.
Fatigue under biaxial and multiaxial loading, Proc. Third International Conference on
Biaxial/Multiaxial Fatigue, Stuttgart, 1989. EISI Publication 10, MEP, London.
Chu C-C, Conle F A and Bonnen J F. Multiaxial stress-strain modelling and fatigue life
prediction of SAE axle shafts. American Society for Testing and Materials, ASTM STP 1191,
1993 pp 37-54
8
(i)
(ii)
(iii)
PREDICTION
UNIAXIAL FATIGUE
600
5000
200
450
4000
1700
1000
53000
30000
19000
Figure 1. Possible samples from a signal sampled at four times the signal
frequency
10
Narrow band
Relative
life
Broad band
Points/cycle
Figure 3. Measured signal (top) and the same signal after peak-valley and cycle
omission (bottom)
11
Figure 4. Measured truck steering arm loading (top) and the same signal after
omitting cycles below the endurance limit (bottom)
12
0.00003
0.000025
0.00002
Dam age
0.000015
0.00001
0.000005
0
0
-1167
605
-568
1209
Range (me)
30
1814
629
2419
Range:uE
Mean
(me)
Mean:uE
1228
0.00016
0.00014
0.00012
0.0001
Damage
0.00008
0.00006
0.00004
0.00002
0
0
-1344
-648
703
48
1406
Range (me)
Range:uE
744
2109
2812
1440
Mean
(me)
Mean:uE
0.0014
0.0012
0.001
0.0008
Dam age
0.0006
0.0004
0.0002
0
0
-1417
721
-703
1443
Range (me)
Range:uE
11
2164
726
2885
1440
Mean
(me)
Mean:uE
Figure 6 Fatigue damage histograms from the first 3000 (top), 30000 (centre) and
300000 cycles of a long load history
2003 ABAQUS Users Conference
13
Figure 7. Actual and approximated strain-life curves (left), and cyclic stress-strain
and hysteresis curves (right) for SAE 1005 steel.
Figure 8. Actual and approximated strain-life curves (left), and cyclic stress-strain
curves (right) for 2014-T6 aluminium alloy.
14
60
50
40
30
20
10
0
1.0E+04
1.0E+06
1.0E+08
1.0E+10
1.0E+12
1.0E+14
1.0E+16
Figure 10. Principal stresses for a shaft under axial load and torsion load
15
1000
Axial stress
Torsional stress
100
1.0E+04
1.0E+05
1.0E+06
1.0E+07
Cycles
Figure 12. SAE notched shaft test results, principal strain theory
16
17