Johannes Homan, MSc

Fatigue life prediction is a common task for engineers. How do we make

sure that we are doing the right thing? There are many codes and FE
tools available for fatigue analysis. Are those tools doing it right?
This ebook presents an overview of common mistakes made in fatigue
analysis.

1. Misunderstanding fatigue
Crack initiation and crack growth are driven by cyclic loads, constant
stresses will not cause fatigue crack initiation.

2. Confusing stress amplitude and stress range

In general, S-N curves are given with stress amplitudes vs. number of
cycles to failure. The stress amplitude is defined as half the stress range.
For welded joints, S-N curves are usually given with the stress range. An
error by factor 2 in stresses is therefore easily made.
Sometimes, the maximum stress level instead of the stress amplitude is
used in S-N curves.

3. Ignoring mean stress and residual stresses

Although fatigue life is determined by the cyclic character of stresses
(i.e. stress amplitude), the mean stress has an effect as well. An
increasing mean stress is unfavourable for the fatigue life, a decreasing
mean stress is favourable.
Note that residual stresses at the critical location (pre-stresses, assembly
stresses, etc.) have the same effect as mean stresses of the cycle.

May 2018 2 Doc. 2018-002/jh

 4. Applying mean stress correction on stresses
instead of fatigue limit
S-N curves are often given for a mean stress level of zero (R=-1). Mean
stress corrections (Gerber, Goodman, Schütz, FKM) are available to
estimate the fatigue data (actually the fatigue limit) for mean stress
levels other than zero. Note that the mean stress not only affects the
fatigue limit, but also the slope of the S-N curve. That implies that
applying a mean stress correction (MSC) on the applied stresses instead
of the fatigue data will give a different (i.e. incorrect) result in the
analysis. In case MSC applied on stresses does give the same result,
some other mistake has been made, e.g. not adjusting the S-N curve
slope.

5. Assuming welds behave like base material

Welding ruins the carefully built microstructure of the base material but
only locally at the weld. In the base metal the properties are determined
by the processing of the material. In the weld zone the material
processing is overruled by the welding process, resulting in a (fast)
solidified structure with decreased properties. The figure below shows
the large reduction in fatigue limit compared to base material.

May 2018 3 Doc. 2018-002/jh

Another reason for reduced fatigue strength in welds are welding defects
(porosity/voids, slag inclusions, incomplete fusion & penetration and
residual stresses).
The detrimental effect of welding is much larger than the effect of stress
concentrations. This is illustrated in the graph below, showing S-N curves
for an unnotched specimen, a notched specimen and a specimen with a
welded gusset.

6. Incorrect cycle counting

Just counting the consecutive cycles like in the figure below may lead to
a large overestimation of the fatigue life.

May 2018 4 Doc. 2018-002/jh

The best way to count cycles is “Rainflow Counting”. This method
roughly implies that small cycles are taken from large cycles and counted
separately, whereas the remaining large cycles are counted as well.

7. Underestimating small cycles

Small cycles (i.e. cycles with a small amplitude) lead to longer fatigue
lives than large cycles (i.e. cycles with a large amplitude). This fact may
lead to the incorrect conclusion that small cycles can be neglected.
However, in a spectrum the number of small cycles is often much larger
than the number of large cycles. If so, the small cycles do give a
significant contribution to the damage accumulation.
It is sometimes even thought that small cycles can be neglected,
resulting in a small number of large cycles, which situation then
erroneously is interpreted as LCF. Note that LCF corresponds with cyclic
plastic deformations, not with small number of occurrences of large
(elastic) cycles.

May 2018 5 Doc. 2018-002/jh

 8. Assuming simplified loads
When load spectra are compiled, often only functional loads are taken
into account. Disturbances (usually of stochastic nature) however are
often overlooked or underestimated. In the figure below an example is
given of a part of a spectrum with functional loads and disturbances
superimposed on the functional load. The disturbances lead to an
amplitude increase of the largest cycle (rainflow counting) and to an
increased number of small cycles.

9. Trusting the Miner rule

Damage accumulation according to the Miner rule is using this equation:
ni
N =1
i

With ni being the number of occurrences of cycle i and Ni being the

fatigue life that corresponds with the stress amplitude of cycle i. All
cycles above the fatigue limit have a finite life N and therefore give a
contribution to the damage accumulation. However, as soon as those
cycles have created some damage (micro crack) also cycles below the

May 2018 6 Doc. 2018-002/jh

fatigue limit will start contributing to damage accumulation. This
contribution is not included in the Miner rule. To account for the damage
accumulation by small cycles, some options are available. One option is
to ignore the fatigue limit and extrapolate the S-N curve below the
fatigue limit, i.e. applying Sak·N=c also for Sa<Sf. This option gives very
conservative results. Another option is the Haibach extrapolation, i.e.
using Sa2k-1·N=c for Sa<Sf. This option is rather arbitrary and not
accurate.

A better solution is the relative Miner rule:

n
N =D
The failure criterion D is then smaller than 1 an depends on the actual
load sequence. Unfortunately there is hardly any data available on D in
literature.
Another issue that make the Miner rule notoriously inaccurate is the
effect of notch root plasticity. High stress levels cause residual stresses at

May 2018 7 Doc. 2018-002/jh

the notch which have an effect on the actual mean stress at the notch in
the subsequent cycles.

10. Using fixed S-N curve slope for local

stresses
The common approach for fatigue life analysis is using S-N curves based
on nominal stresses (i.e. average stresses in the net section). Depending
on the severity of the stress concentration (Kt) in the structure, a S-N
curve is then selected that corresponds with the actual structure. In the
diagram below a few S-N curves are shown for different Kt values. The
larger the Kt, the lower the fatigue limit and the steeper the S-N curve
(smaller value of k in the S-N curve equation (Sak·N=c).

In case the fatigue analysis is performed using local stresses, the S-N
curves based on nominal stresses cannot be applied, instead S-N curves
based on local stresses must be used. In the diagram below, such a set
of S-N curves is shown. These curves are derived from the S-N curves
based on nominal stresses shown above, by multiplying the nominal
stresses by Kt to obtain the local stresses. The S-N curves based on local
stresses show different fatigue limits and different gradients for different
Kt values.

May 2018 8 Doc. 2018-002/jh

Always be aware that a local stress can be any combination of global
stress and Kt. Additionally, the gradient depends on the stress ratio as
well.

11. Using S-N data for LCF

In the LCF (Low Cycle Fatigue) region, fatigue is driven by cyclic strains.
The material behaviour in that region is dominated by plastic
deformation, non-linear relation between strains and stresses. An S-N
curve describes fatigue in the HCF (High Cycle Fatigue) region, i.e.
elastic material behaviour and a linear relation between strains and
stresses.

May 2018 9 Doc. 2018-002/jh

 12. Using -N data for HCF
For HCF, sometimes the elastic part of the -N curve is used. It is argued
that because of the linear elastic material behaviour in that region, the
-N curve is in fact a S-N curve. Since the strains are local strains, and
the -N curve is based on unnotched specimens, this approach leads to
the same problems as described in section 10.
Furthermore, the position of the elastic part of the -N curve is partly
determined by the fatigue strength coefficient ’f. This coefficient is
derived from a static tensile test and has nothing to do with high cycle
fatigue.

13. Using equivalent stresses like Von Mises

Fatigue damage is propagating perpendicular to the largest principal
stress range, therefore this stress range determines fatigue behaviour.
Principal stresses in tension in the other directions have hardly any
influence on the crack growth, these stresses do not affect the shear

May 2018 10 Doc. 2018-002/jh

stress in the activated slip planes. In case of shear (2D stress state with
bi-axiality ratio of, or close to, -1), fatigue data for shear should be used.
3D stress states are hardly ever relevant for fatigue, crack always start at
a free surface. Even with sub-surface initiation it can be argued that the
stress state is 2D, cracks start at inclusions or voids.
If there is a 2D or 3D stress state with varying largest principal stress
direction, it is thought sometimes that using equivalent stresses is
attractive. This is not the case, equivalent stresses have no direction and
certainly not a varying one. The best approach would be the “Critical
Plane Approach”, i.e. analysing different crack growth directions (planes)
and using for each plane the stress components perpendicular to that
plane. Note that such an approach must be performed twice; viz. also for
shear stresses.

14. Using scatter data from laboratory tests

for actual service applications
Fatigue always shows a lot of scatter. This scatter is caused by (1)
variability in material properties, (2) variability in surface conditions
(process/manufacturing), (3) variability in loading, and (4) variability in
environment. S-N data obtained from laboratory tests only cover the first
two causes. If the S-N data just exists of the fatigue limit being a
function of the tensile strength, only variability of material properties is
taken into account.
The standard deviation (log-normal distribution of N) for laboratory S-N
data is lower than 0.25 (often between 0.05 and 0.15). For scatter in

May 2018 11 Doc. 2018-002/jh

service, typical values for the standard deviation range from 0.25
to 0.35.
So if an analysis based on laboratory scatter data gives a fatigue life with
a certain probability of failure, this result is inaccurate.

15. Ignoring surface conditions

Surface conditions have a large impact on fatigue behaviour but are
easily neglected. S-N curves are often established using nicely
manufactured specimens. Actual production of your components may be
completely different. This difference should be taken into account.

May 2018 12 Doc. 2018-002/jh

Do you want to know more about fatigue
analysis?

In this book, guidelines are given for fatigue life estimations

and analysis. A known quote from Arthur Schopenhauer is:
“Wo das Rechnen anfängt, hört das Wissen auf”. One may
translate this to: “Where calculating starts, understanding
ends”. The objective of this book is to give the reader
background information and knowledge on fatigue analysis
methods (i.e. understanding the calculation). It is therefore
neither a text book nor a recipe book, but the missing link
in-between.

https://www.fatec-engineering.com/publications/

About Johannes Homan

Johannes has a background in aerospace engineering and has specialized himself in
fatigue and damage tolerance. After his graduation at Delft University of Technology
(faculty of Aerospace Engineering) Johannes started working at Fokker Aircraft as
Fatigue & Damage Tolerance specialist.
In 1996 Johannes started his own firm: Fatec Engineering and works in this firm
since then. During that period, Johannes was also active as researcher and assistant
professor at Delft University of Technology (faculty of Aerospace Engineering) and as
researcher at GTM Advanced Structures.
https://www.linkedin.com/in/johanneshoman/

Rubenslaan 127
2661 RV Bergschenhoek
The Netherlands
Phone: +31 10 5299713
Email: johannes.homan@fatec-engineering.com
https://www.fatec-engineering.com

May 2018 13 Doc. 2018-002/jh