Fatique calculations is highly statisticaly. The objects geometry plays just as large a role as the material itself. As far as I can remember in one formula half of the tensile strength is used, multiplied by up to six emperical and statistical factors. These are for surface ruffness, object size, relaibility, temperature, stress consentrations (geometry) and corrosion. For these factors there are tables, graphs and simple emperical relations to calculate each. Try to get hold of the book: 'Mechanical Engineering Design' by Joseph E. Shigley. It has most of the tables, graphs and relations for the factors in it. The book is also great for basic mechanical design aplications.
Some steel tables does include a unmodified fatique limit stress, noted as Se. This can be used instead of half the tensile strength.
Finding the fatigue properties of materials is quite tricky unless they are welded, because the crack initiation phase is difficult to predict. If they are welded, then one assumes they already contain flaws and calculating fatigue performance is easier as less information is required. Characterisation is then dependant on the type of joint configuration used.
If you want to know about the basics of fatigue in welded joints you will find some useful information on my web site
Mechanical Metallurgy by Dieter also has a good basic chapter on fatigue testing and the ins and outs. Other than that maybe the TMS website has some books for materials engineers. There are pifalls to be aware of regarding fatigue testing, and this is also why it is a highly researched topic. For example, be wary of handbook values for Al-Si casting alloys. The fatigue life is strongly dependent on porosity, which is an alloy-process variable.
The Conway & Sjodahl book is out of print but definitely worth the price. MIL-5 is the material properties bible for stress analysts working on defense projects.
To create a statistically acceptable fatigue design curve, you perform the following steps (Excel functions in parentheses):
1) Linear regression of S-N data, with S as the "X" and N as the "Y" (m=Slope(Y,X) & b=Intercept(Y,X))
2) This gives you an equation that looks like Log(N)= b + m*log(S)
3) Compute the standard error of the estimate (STEXY(Y,X)
4) Your design curve is Log(N) = b + m*log(S) - k*see
5) k is a function of how "safe" you want to be. You can set k to 3, giving you a "-3sigma curve." It is better to use k as a function of the number of specimens you tested. There is more detail about that approach in the MIL-5 reference cited above.
Correction to my previous post:
Usually, X is log(N) and Y is log(S). You can try other transformations, but as I described the process, these should be used.
Doug
