How do you calculate reliabitliy factor (confidence interval)?

[Barry1492]

I have a set of samples with a large standard deviation for ultimate strength. We are going into fatigue testing and I need to calculate what load to use.

I have Norton's Machine Design as a reference but it only gives a table for steel (Sd=0.08 of mean) and shows:

90% at .897 of ult. str.
99% at .914
99.9% at .753

etc.


I'm currently looking through mil-hndbk-17 for the answer but it is VERY cumbersome.


Thanks for any help.

 

[CompositeGeek]

There are a number of resources you can tap into for help in this arena.

If you company employs a Quality Engineer AND that individual is a CQE, they have the necessary skills to help.

Otherwise, contact your local chapter of American Society of Quality (

http://www.asq.org)

and see if there is a local/regional source for help. Meaning an individual who has passed the CQE or CRE test.

You will also find a number of books on the subject that will help far more than reading Mil-hdbk-17.

Your request is pretty sopisticated, and the necessary calculations are very dependant on a number of factors that you have not described.

Good Luck



Composites and Airplanes - what was I thinking?

There are gremlins in the autoclave!

 

[SWComposites]

What is the purpose of the fatigue test? How many cycles and what load ratio? What is the material and what is the application?

 

[Barry1492]

I don't really know a lot of these answers. I'm kind of in between a technician who doesn't really know math and my boss who doesn't remember the formula and is very busy.

I know that it is a composite sandwich structure that is udergoing 3 point bending. The static fatigue tests showed a large standard deviation....i don't know the values.

The next series of tests is for fatigue. The idea is that we want the load to be high enough that it doesn't take very long to fail, but not so high that a large percentage of the specimens fail after only one cycle.

In the past they had used 80% of the mean as a standard, but our SD is so large that many specimens ult. str. is less than that value.


Personaly, I don't see why we can't just assume a normal distribution. I just went through the values for steel in Norton's book and compared them to the normal distribution and HUZZAH! that's all they were.....normal distribution percentiles.

 

[PanelGuy]

Actually, section 2.5.14 of vol 1 (pg90) isn't too bad a provides a place to start.

However, even hdbk 17 is quick to state that fatigue is variable and application dependent.

Knowing you have a large SD in your initially testing brings in good data actually. Was your mode of failure the same? Was the ply orientation consistent. Were the laminae all exactly the same? Any of these will throw your data totally off, BUT are indicative of problems in your production methods. Don't worry, everyone goes through this.

In any situation, the fatigue is application dependent anyway. Most imprtantly, you should start anywhere and see where it lead you. Start high and work you way down. Depanding on your matrix you will probably find a crossroads where on emoment you are failing fast, the next you are lasting forever.

"Analysis and performance of Fiber Composites" by Broutman and Agarwal has a great section that is detailed, but still vague. The new text from Tsai, Gay and Hoa barely addresses it other than to say fatigue is good because the discontinuities between brittle fibers and brittle matrices prohibit crack propagation.

At any rate, most handbooks you review have well defined tables on traditional materials. Big surprise. Composites are so varied, there is no one good place to start.

I would think 80% would be OK to start, but only after making sure of the failure modes. If you look at your samples onder a scope you should be able to determine whether the failure was in the cross ply or other, the failure had a defect(inclusion bubble) or the failure varied due to ply orientation. But, like I said above, these issues can oly help to understand what your process variability will do to the overall performance.

you should get a normal distribution if all of the processing variables are in control!

Good luck!