I have searched the forums and not found an answer. (I have to say that for my protection :confused: )

I have a situation where there is a requirement that the strength of a weld shall be minimum 40 N

When testing these parts some do not break but the test stops with results of 69, 75 etc.
I need to show a capability for which can be used the normal CLU ((Avg-40)/3 std ) but in this case how do you calculate the average as some of the results are no break, the std is also not correct if you include the no break examples.

Maybe I can do something with the minimum values. I have not got a lot of data at the moment as this is a new process which we have only just started experimenting with.

Any one got any thoughts on the matter?

Howard -

Possibly the distribution is Weibull. I think that weld strength was the study that Wallodi Weibull used to develop his distributional ideas.



Possible Procedure for you:

1) Get a few more data points, if really critical to get true proportion, get lots or bootstrap the sample
2) Analyze resulting data for best fit assuming some are right censored. Minitab’s latest versions 13 and 14 have this capability in the reliability section, also other reliability software (or do it the old fashioned way on probability paper, if you feel the need to whip yourself :lmao:)
3) armed with the distributional parameters, calculate the probability below your spec limit and convert to DPMO with the standard "back conversion" to Cpk

I'm not much of a fan of capability one of the reasons is shown here, it is difficult to do at times when normality is not the norm :biglaugh: !

I have no clue what the CI from a Weibull generated Cpk is like, hence my caution in 1). I do know; one of my other concerns about Cpks is that the studies done normally, 50 to 100 parts, have CIs of 30% of mean in half width or more. You need 300 plus observations to bring the statistic to +/- 5% or less.

If you just need a "rough" estimate of the potential failures and if you have never seen a failure yet in your testing (I'm guessing from your post statement about minimum values, rather than saying failures), you could potentially use the rule of three, but if it is required that failure percentage be low (<< 1%), than that attribute approach will require LARGE samples to be convincing.

Hope this is helpful.