Create S-N curve with FEA 1

[eltooon]

Hi, can anyone advise how to generate an S-N curve of a component using FEA?

Our company designed a product some years back, tested it in a resonance fatigue testing machine and generated an S-N curve.
Now we created a few variations of that product (similar geometry but different in size), and customers want S-N curves.
Putting each variation into a fatigue test is too costly for us.

Would FEA be able to help in this situation?




We make a living by what we get; we make a life by what we give.

 

[rb1957]

i'd expect that the "s" part of the s/n is a very simple calc, something like P/A.

you might use FEA to compare different components (to show that the new ones don't have stress peaks not in the original).

another day in paradise, or is paradise one day closer ?

 

[rstupplebeen]

I would create an FEA model of the existing design and tune the inputs such as, material properties, to match the existing test data. Then simulate the new designs. I hope this helps.

Rob Stupplebeen

http://www.optimaldevice.com

 

[corus]

SN curves are usually related to the material, or to a weld category. Your original SN curve must have taken into account both the material and the geometry of stress concentrations and/or welds etc. If you carry out an FEA on the original design then you could relate any future designs to the peak stress of the original design where it failed. A crude approximation would be to say that the stress cubed is related to the design life, so if you double the stress for example then you get an eighth of the life.

 

[rb1957]

i thought the idea was why to say that the old s/n curves are applicable to the new parts.

change of material would be problematic, unless you have good fatigue definition of the two materials so you can say (hopefully) that the old material is conservative. if you don't have "good fatigue definition" then you're screwed ! and have to retest.

minor geometry changes could be resolved with FEA ... showing that the stress peaks are similar.

another day in paradise, or is paradise one day closer ?

 

[rb1957]

"i thought the idea was why to say that the old s/n curves are applicable to the new parts." sigh ...

another day in paradise, or is paradise one day closer ?

 

[eltooon]

The product is machined from steel. There are some sharp corners, but no welding.


I'm giving a fictitious example to facilitate the discussion:

I use FEA and find out that:
The SCF of existing product is 3.
The SCF of new product (only changes in dimensions) is X, and the hot spot is in the similar location.

Now what can I comfortably derive if (1) X is close to 3, (2) X > 3 and (3) X < 3 ?

And what if I change the material, from 80ksi to 100ksi steel?

We make a living by what we get; we make a life by what we give.

 

[rb1957]

you have an s/n curve for the original fitting, probably based on a simple stress calc (rather than the stress peak revealed by FEA, possibly it is a P/N curve if the loading is simple). so you have to relate the new stress peak to the original stress peak, and the new basic stress calc to the original; and "new" and "original" are run at appropriate loads (maybe the same, maybe be different). The idea is to read into the original s/n curve with the new basic stress calc modified to account for the change in the stress peak. if the new stress peak is less than the original fttg, then i would conservatively neglect it ... the new fttg should have a longer fatigue life than the original one if the details have a lower Kt.

changing material can be very difficult to "hand wave". If you have a good fatigue definition of the two materials (s/n curves at several Kts) then you can build a story for adjusting your original s/n curve. if you have a good fatigue definition of the original material, then you should be able to use your s/n curve to derive a Kt.

another day in paradise, or is paradise one day closer ?

 

[realtime2]

As some of the other posts suggested you will need to find a
baseline material data set.
My approach to the problem would be to get tests on axial specimens tested
in strain control. They are widely available, for your materials with
80 and 100 ksi yields(?) probably the AISI website
barsteelfatigue.autosteel.org
is a good place to start looking.
If you cannot find an exact match pick something with similar carbon
level and hardness or ultimate. Alternatively get someone to do a baseline
test series. Typical cost is probably $15k.
The reason for knowing your base material properties is that in fatigue
with on-going plasticity the materials can cyclically harden or soften
and large SCFs will throw your local stress-strain response into the short
life plastic zone.

For example method: Pick up the data from an HSLA 80 here:
fde.uwaterloo.ca/Fde/Materials/Steel/Hsla/mergedSAE980.html
hit the "Send " button at the top. You should get back a page with 3 plots,
the one in lower right corner is a "Neuber Plot" of the axial (unnotched)
specimen data. On this plot place the S/N data points (stress Range!) from your previous
tests. The vertical numeric value/ratio between the two curves is your Kt or SCF
multipier. You can also plot your old FEA predictions on this plot.
The attached file is a schematic of the idea.

Then do same for a 100ksi (yield) material and using you FEA estimated SCF
draw in the expected nominal or notched line below the axial result. Be
sure to "cut off" your notched curve at yield *2 (stress range) as this
method presumes that you are not going to get full plastic behavior in your
notched specimens.

Whether or not your customers will accept this as proof of durability
is, of course, another problem.

 

[eltooon]

Hi realtime2,

Thanks for your post. It's very helpful.

You also made a good point: acceptance from the customer. I foresee it's difficult to convince the customer. Do you know of any design code or specifications for such a problem?

We make a living by what we get; we make a life by what we give.

 

[realtime2]

Hi eltooon,

Tough question regarding customer acceptance of an analytical
projection. Probably it depends on the safety issue of what you
are proposing. If it is a component that is safety critical, such
as a brake system in a vehicle, testing is obligatory. If one doesn't
test a safety item and something goes awry in a customer's hands the lawyers
will hang the engineers out to dry.
If it is not safety critical they may accept the analysis, but perhaps
increase their factor of safety on the loads.
Probably the best compromise is to do some tests, at least at the
higher amplitudes of the fatigue curve.
Remember when testing to start at a high load level and then work your
way down the curve. Short tests that take hours are much cheaper than
runout tests that take weeks. People forget that and shoot to hit the
fatigue limits. If they miss and it does not fail one has nothing.

I don't think there is a design code for this Neuber stuff. The approach
to the problem is used by the ground vehicle industry; though each company
has different ways of defining the same thing.

Incidentally the Neuber equivalent Stress range curve is for "R= -1" or
fully reversed tension compression testing. If for buckling reasons you
have to do something like a R= Smin/Smax = 0.1 test you will need to
correct for mean stress when you plot the tests. If you are nominally
elastic the Neuber equivalent stress range can be estimated by
Srange_equiv. = 2.0 * sqrt (Smax * Sampl.)
Good luck.

 

[rb1957]

"If they miss and it does not fail one has nothing." .. well, not quite nothing; knowing that something doesn't fail in 10^x cycles of a load is only slightly less information than knowing that it fails. i mean 10^x (or the load) was picked for some reason.


another day in paradise, or is paradise one day closer ?

 

[realtime2]

Yes, you're right rb1957; I was being too glib. Perhaps a better
criterion may be: If your are building something like a machine of
which you only sell a few, and weight is not really that important, a
runout at worst load will show proof of durability. If the component
is something you will sell lots of, weight or $/lb becomes important
and one needs to know how far below the fatigue limit a runout is, and
revise the design for weight save; otherwise some competitor will offer
an improved lighter weight version.

 

[jagad5]

Actually, rb1957, in a fatigue test, a runout provides much less information than a fail, not slightly less. Statistical analysis of SN data is very complicated. Handling runouts is the hardest part. The random fatigue limit model (Pasual & Meeker, Technometrics, 1999) accounts for runouts using maximum likelihood estimation (MLE) statistics rather than least squares (LS). LS methods either include or ignore runouts. MLE methods explicitly account for a non-failure. In any test, continuing to failure is the preferable statistical event because the failure provides a concrete measure of the result while the best you can do with a survivor is to estimate when it might have failed, which is what the MLE actually does.

Doug

 

[rb1957]

i'd've given a run-out result the life experienced on test, which would have been the most anticipated going in and so should be acceptable for the product. if it became significant, then you could investigate the impact of these run-outs on the alaytical results and maybe end up saving "we need to run a longer test".

i guess what you're saying with the statistics is that from a population of X specimens, we had Y failures, which a safe/max/mean/min life of Z. and from the "cloak and dagger" of statistics you can refer this result (from Y tests) back to the population for a safe/mean life.

another day in paradise, or is paradise one day closer ?

 

[jagad5]

It depends on your reliability requirement and the quality of the component/material being tested. If you have a rare defect that leads to failure one time in 1000, you cannot run 10 tests that all survive the test and then declare you have demonstrated that the life will be equal to the duration, or even a fraction of the duration, of your test.

Doug

 

[rb1957]

and that'd be reflected in the statistics to create the safe life from the test result. it's highly unlikely that a limited number of test specimens would detect a 1/1000 production flaw, so you'd either ...
1) create the flaw in a test piece,
2) do some DTA with the flaw,
3) live with 1/1000 failures (if you can),
4) proof load each production piece to prove a flaw doesn't exist, or to set up a beneficial yield zone around the flaws,
5) or something else ?

another day in paradise, or is paradise one day closer ?

 

[jagad5]

I think we're in violent agreement.

Doug

 

[jagad5]

All you points are perfectly valid for the scenario I introduced of a failure do to a rare defect. But, if that 1 in a 100 failure is due to variation in loads, then running all tests to failure will give you much more insight than if all the tests survive. All surviving tests provides no information about inherent material scatter. Airplanes are designed to a 99/95 probability of exceedance/confidence level which is why MMPDS requires so much data to publish a design value (no less than 100 points for an A-Basis, and there are flaming hoops to jump through to use so little.)

Doug