Correlating Gages
Six Sigma – iSixSigma › Forums › Old Forums › General › Correlating Gages
- This topic has 10 replies, 6 voices, and was last updated 14 years, 5 months ago by
Dennis Craggs.
-
AuthorPosts
-
January 17, 2008 at 3:19 pm #49137
ncwalkerParticipant@ncwalkerInclude @ncwalker in your post and this person will
be notified via email.I have heard of several methods to correlate gages. For each method, I have some questions and then the general question of which one should I use.1) Isoplot – Measure 20 pcs representative of production range on two gages. Plot these results on an X and Y graph and look for alignment to a 45 degree line. That makes sense, BUT to what degree do you say it is good? If I do a linear regression, I would be looking for a slope of 1 and I would get an R. What is the accepable R? What if the R is really good (say .96) but the SLOPE is not 1? If the slope was 2.5 with a good R, I think I could mathematically adjust safely.2) Gage R&R Method – Run a Gage R&R but use each gage as an appraiser. This makes sense to me. I would assume the same Gage R&R hurdles would be used. But if I were doing this on a CMM, would I not have to do this on EVERY dimension that it output? Or just the ones that don’t agree? Or representative dimensions? (A couple of TPs, a couple of Diameters, etc…)3) Percent Difference Method – Measure a couple of parts (3??) on both gages and look at the percent difference between the results. This is a little nebulous. Percent difference of what? The tolerance? How close is good? This is appealing because you are not measureing so many parts, but it just doesn’t feel statistically accurate for the same reason. But we aren’t dealing with statistics, really, are we? I don’t need statistics to tell me that “This one is higher than that one …”4) One Fifth Tolerance Rule of Thumb. Same as above, but the agreement hurdle is that no measurement is different by more than one-fifth the total tolerance.Any help would be appreciated. I need two answers: 1) The right answer. 2) The answer that will make my customer’s SQE happy.ncwalker
0January 17, 2008 at 4:36 pm #167369
Mike CarnellParticipant@Mike-CarnellInclude @Mike-Carnell in your post and this person will
be notified via email.ncwalker,
When you are evaluating gage correlation what you want to know is how well do multiple gages make the same measurement. That is reproducibility and that is simple enough to do with a R&R study just as you described.
Good luck0January 18, 2008 at 12:51 am #167388Mike,
I sent you an e-mail requesting your current mailing address. I sent something your way (Harbor Way) and it was returned undeliverable.
Pete0January 18, 2008 at 2:31 am #167389
Mike CarnellParticipant@Mike-CarnellInclude @Mike-Carnell in your post and this person will
be notified via email.Pete,
Will send you the current address. I went back to the Hill Country.
Regards0January 18, 2008 at 6:27 am #167393
KrishnamParticipant@KrishnamInclude @Krishnam in your post and this person will
be notified via email.Hi
ncwalker
can you send me data on the following email and i will try to solove the same at earlist.
[email protected]
Regards,
Krishnam0January 18, 2008 at 11:24 am #167397Sounds like you are trying to answer a vague question. It’s like going into a pizza shop and just ordering a plain-ole pizza and then yelling because you didn’t get your mushrooms.
You can match the performance of 2 measurement tools with paired t-testing. (choice 3). Just my 2 cents, but that would be my first choice. Sample size depends on the difference you’d like to detect. Sounds like an easy question, but you’d be surprised how many times I’ve been looked at with a blank stare on that one.
R&R with machine substituted for operator is also a suitable choice but you lose visibility of the operator effects.
You could realy turn this into a science project, so I would start out simple.
HACL0January 18, 2008 at 11:31 am #167398I should have proofed my message first! Paired t-testing was not your 3rd option listed. You referenced percent difference.
My suggestion still holds, but measure a selection of parts on both machines and analyze with paired t-testing. (Not % difference)
Try hard to explore the preferred method with the customer before you embark on the study.0January 18, 2008 at 2:30 pm #167406
ncwalkerParticipant@ncwalkerInclude @ncwalker in your post and this person will
be notified via email.Krishnam,I don’t have any data yet. I am trying to define the test so I don’t waste time taking data I cannot use.ncw
0January 18, 2008 at 2:31 pm #167407
ncwalkerParticipant@ncwalkerInclude @ncwalker in your post and this person will
be notified via email.hacl,Let me see if I have this right….The paired t-test is for small sample sizes (because if I had a lot of them, I could use the tails of the normal distribution). I am going to measure parts on one CMM and measure the same parts on another CMM (Once?) maintaining traceability of sample numbers.Then, I line these up and calculate the differences for each part.I get my test value (T) from: avg of differences / variance of differences / sqrt(n)Then I use n and my confidence to look up my t-hurdle statistic (t). (Which is why you say the number of measurements drives how CLOSE I want to know).If T >= t, it means the measurements are different. If T= t, it means the measurements are different. If T= t, it means the measurements are different. If T
0January 18, 2008 at 2:39 pm #167411
ncwalkerParticipant@ncwalkerInclude @ncwalker in your post and this person will
be notified via email.Correction:Something funky happened with my greater thans and less thans,Here is the decision in words:If T is greater than or equal to t, it means the measurements are different and my CMMs do not correlate.If T is less than t it means the measurements are the same and my CMMs do correlate.
0January 18, 2008 at 9:20 pm #167469
Dennis CraggsParticipant@Dennis-CraggsInclude @Dennis-Craggs in your post and this person will
be notified via email.GR&R can be used with appraiser replaced by the gage, but then variation due to appraiser variation is lost. Traditional GR&R compares appraiser, part, and repeat measures. If you want to include more variables, consider using DOE with covariates. Covariates are study variables that are difficult to control, such as individual part values. DOE with covariate analysis basically considers the covariates to have a linear effect on the result. For a gage, you want to assess bias, accuracy, and repeatability over a range of values. Parts could be selected that span this range. Then, two or more gages can be compared as fixed or random factors. In Minitab, Stats -> ANOVA -> General Linear Model provides analysis capability. The statistical significance of any gage difference should be presented in the ANOVA table. The mean difference between individual gages can come from the analysis. Again, the experiment needs to be designed in a balanced manner and there is a linearity assumption for the covariates.
0 -
AuthorPosts
The forum ‘General’ is closed to new topics and replies.