Fatigue test

[dansoarr]

Hi I'm after some advice

If I have a component that must flex 50 times through its life (2 months), I have an automated test fixture that will flext the component and check for a failure, (true return position).

If I wanted to prove that the component woudl not fail in normal operation say to 1:10,000 at 95%, I belive I'd have to test 30,000 parts to 50 flex's.

Now I only have 10 parts available, although completly representative of the final part.

So can I test each of the 10 parts to 3,000 actuations, and deduce the same fact. i.e

Is it true or fair to say, To test to a failure rate 1:10,000 @ 95%, I could either :-
Test 30,000 components, once each Or
Test 10,000 components, 3 times each Or
Test 1,000 components, 30 times each Or
Test 100 components, 300 times each Or
Test 10 components, 3,000 times each Or
Test 1 component, 30,000 times each ?

Some help would be great.

Thanks

Arron

[David]

Hi Arron,

You can only divide up the test time and units as you mentioned if you assume an Exponential distribution. Now, using the Design of Reliability Tests (DRT) in Weibull++ 7 and assuming an Exponential distribution, the total accumulated test time for your given specifications is 1.4979E+8 (not 30,000). In this scenario, you could test 1 unit for 1.4979E+8 cycles or 10 units with 1.4979E+7 cycles for each unit.

If you do not want to assume an Exponential distribution, but instead a Weibull distribution then the requirement would be to test each of the 10 units for 27,367 cycles. This assumes a Beta = 2.

Additional information on the DRT can be found at http://www.weibull.com/hotwire/issue24/relbasics24.htm.

I hope this helps.

[Pantelis]

Arron, let me also add my two cents and also emphasize some of the points made by David.

First with an exponential distribution you are assuming a constant failure rate. This assumption is poor at best, but most importantly absolutely WRONG in your case.

The second option mentioned was to assume some increasing failure rate behavior (linear increase in this case with beta=2). This is better but again an assumption.

Now my question is why are you attempting to run this as a success test? You could run it as a life test, i.e. run a some samples to failure (assuming flexing them is not very time consuming) and analyze that data using standard LDA techniques.

Hope this helps.

[Abhay]

Curious to know is it justified to test the unit above its MTBF to demonstrate the reliability requirements with small sample size

[Pantelis]

Yes. Absolutely