Not exactly the same topic, but sort of. Have a few questions about setting tolerances and control limits.
1. Our customers often put TBD on the print for physical properties of the material until after 25 batches. For each batch, 5 samples are taken and after 25, we are expected to recommend a specification based on Cpk = 1.33. I have a problem with this that I think is legitimate and trying to make the customer understand. First of all, the samples are not 25 groups of 5 from the same batch of material over time. If 100 samples are taken from the same batch, a nice normal curve is seen. However, calculating Cpk does not take into account sub-group to sub-group variation based on Rbar/d2 - subgroups being different batches. Alot of the variation is a mean shift from batch to batch and the statistics are being done on different populations, different batch of material, thru several independent processes that could affect the measurements. Not only are different batches (from which only 5 samples are taken) being lumped into the equation, but several different setups over time. The result is usually a non-normal distribution. The problem is, given a specification, if I were to sample lots of parts from a particular batch, the capability is fine because 'sigma' is tight, but if treated as if the data came from one batch (population) over the course of 25 batches, then on paper it looks terrible if Ppk is calculated, but so wonderful for Cpk that the specifications are made too tight. I understand that batch-to-batch variation needs to be eliminated, I am just saying that I don't think this is exactly right. We had Chrysler come in some time back, and they argued with our Tier 1 customer that statistical specifications could not be set for this application and that only go-no-go limits could be used.
Anyone know how to properly handle the statistics or do I have it all wrong?
2. Much like #1 above, but with regards to calculating control limits. We have a grinder that, once set up, is VERY capable with the variation taking up so very little of the tolerance. I could sample parts from now to eternity and Rbar would very small. On the next setup up after running a different part, the same capability would be seen. Using R-bar/d2 from the various setups, after the control limits are calculated, they are so tight that the operator has a very hard time just setting up between these limits because the formula doesn't not take into account different setups, which are different populations. However, we can't make new control charts for each new setup because of the short time involved. I've briefly looked at short-term SPC. Is this what I need, something else, or am I totally misunderstanding something?
Thanks.