Do any of you out there know the acceptable method of dealing with a single sided nominal on the Gage R&R Study? Example: How do you introduce the numbers for the tolerance upon say a profilometer? You are looking for a minimum finish requirement of a 32 finish. Anything better than that is acceptable, so do you use the 32 as the upper end of the tolerance and give a tolerance range of 0-32? How about geometric controls such as flatness, true position, roundness, etc.? Is developing an SPC history and demonstrating capability the way to go?



------------------
Dave S.

Since I asked this question and received no responses, I sought an answer from other sources and will share the response with you.
First, is there a special Gage R&R method for unilateral toleranced features? No. The methods are the same, but the interpretation of the data must be treated differently. "Unilateral features may produce "bounded data". Basically, that is any measurement data limited by an upper or lower value (e.g. runout, flatness, straightness, etc. where measurement values cannot be recorded lower than "0"). For example, a highly capable process could produce a diameter with a runout tolerance of .001" max that measures on the average .0002" with occasional measurement values up to .0012". This would be considered a capable process. If we tested the data it would not be normal (bell shaped curve)but appear skewed right.
When we violate the assumptions of any statistical tool, the tool becomes unpredictable and may indicate a false conclusion. What happens to unilateral data that is not normal, but treated or assumed to be normal? The measurement system appears unstable, the Gage R&R error is inflated. It makes the measurement system look worse than it is.
An advanced practitioner of GR&R understands the assumptions of the tool and remains a healthy skeptic (test the assumption). When you know that you are dealing with unilateral or skewed data - test it for normality. If it is normal data, proceed with the standard methods and evaluation techniques. If the data is not normal, you can usually transform it to act normal, then proceed evaluating the transformed data. If you can't transform, proceed with skeptic caution. You will rely on profound knowledge an practical experience with the measurement system.
Certain high-end measurement systems (circular geometry, surface finish, CMM, etc.) are best evaluated using a control chart method and measuring master artifacts on a regular schedule.

I want to thank you for taking the time to stop back and let us know what you found out. I assure you we all appreciate it!

With all due respect, that doesn't seem to answer the question...

I have a measurement situation with one-sided tolerance, the spec. is that the parts must be less than -20 (this is an electrical characteristic). We do 100% checking, so this is a "product control" situation in MSA terms. In this situation tolerance limits should be used in the calculations. How should I calculate the GRR%?

Time for me to get thrashed (once again :rolleyes: ) and learn.

Doesn't the tolerance only come in to play to give the "% Part Variation" in an R&R study? That should be a large number in a "good" R&R study. As I look at the output stats in a study, I don't see where the tolerance has any effect on the Repeatability or Reproducibility values.

I don't understand the original poster's response - what does normality of the data have to do with a Gage R&R? Even with a "naturally" bounded feature, if I measure flatness (runout, position, etc.) on a part three (or more) times, should I not get a fairly comparable reading each time? Regarding surface finish (where I believe he meant a maximum of 32), I can understand a little difficulty with interpretation due to the "nested" study as it is a "destuctive" test (that just sounds funny this morning - when you check a surface with a profilometer, you are actually improving the finish - not destroying it :bonk:. Sorry, it's been a rough day already!) .

It is a second time I see, someone wants mix dimension tolerance in MSA.
:mg:
If I understand MSA correct, You are also allowed to take only NG parts for analysis. See page 113 and 114 of MSA Manual 3rd Edition.

If I understand MSA correct, You are also allowed to take only NG parts for analysis. See page 113 and 114 of MSA Manual 3rd Edition.
Why would you think you only use no good parts for a Gage R&R? ("NG" means "No Good"?????)

From my point of view, MSA should tell me if operator performs reccurent measurements and it is not important what kind of part he checks.
So, that's why I put this extreme example with only Not Good parts.
For sure no one use this (I guess), but it shows what is really important in MSA.
I see, some people treat this like dimensional report in PPAP and it is a reason, they put here dimesion tolerance

From my point of view, MSA should tell me if operator performs reccurent measurements and it is not important what kind of part he checks.
So, that's why I put this extreme example with only Not Good parts.
For sure no one use this (I guess), but it shows what is really important in MSA.
I see, some people treat this like dimensional report in PPAP and it is a reason, they put here dimesion tolerance

Really?

MSA page 73: "For Process Controll situations where the measurement result and decision criteria determine process stability [...] (i.e. SPC) [...] the availability of samples over the entire process range becomes very important"

The reason is simple: you will use study variation to estimate the unknown process variation. If you select only good parts (or only bad parts) the variation among the damples will be less than the varaition of the process...

Let me be more specific about my problem.

MSA page 73: "For Product Control situations where the measurement result and decision criteria determine conformance or non conformance to the specification (i.e. 100% inspection or sampling) [...] the assessment of the measurement system is based on the feature tolerance (i.e. %GRR to TOLERANCE)"

this is in line with page 116: "If the analysis is based on tolerance instead of process variation [...] %EV, %AV, %GRR and %PV are calculated by substituting the value of tolerance divided by six in the denominator of the calculations in place of the total variation TV"

This is clear to me. But what if we have no tolerance range, because there is only one tolerance limit? Maybe I go and check the SPC topics, they should face the same problem when calculating CPK...

MSA page 73: "For Process Controll situations where the measurement result and decision criteria determine process stability [...] (i.e. SPC) [...] the availability of samples over the entire process range becomes very important"


:o I have to :read:
Yes, You are right. My understanding of MSA was "to check measurement system only and not dimension tolerance".

....

This is clear to me. But what if we have no tolerance range, because there is only one tolerance limit? Maybe I go and check the SPC topics, they should face the same problem when calculating CPK...

I've seen two views about this. One says that %Tolerance cannot be defined in case of single-sided tolerance (like Cp) and that you should compare GRR with Process variation that you get from ongoing SPC syudies.

The other view is that you can calculate the value of tolerance from process mean to the available specification and use the constant 3 instead of 6 (or 2.575 if you are using the 'traditional' 5.15 to represent the full spread of error)

Comments anyone?

I am bumping this up again as I do not think we have a definitive answer.
It is possible to use part to part variance instead of the tolerance method and so there exists one solution.
If I have a requirement for say max 20, then can I say that the tolerance is 20?
If I have a requirement for say min 20, then can I use trlal and error to decide on a tolerance that is reasonable in relation to my measuring uncertainty?
Any other comments after a year to think about it

I've seen two views about this. One says that %Tolerance cannot be defined in case of single-sided tolerance (like Cp) and that you should compare GRR with Process variation that you get from ongoing SPC syudies.

The other view is that you can calculate the value of tolerance from process mean to the available specification and use the constant 3 instead of 6 (or 2.575 if you are using the 'traditional' 5.15 to represent the full spread of error)

Comments anyone?

GRR as %Study Variation has taken the forefront as a more important measure of the gauge's suitability for process control or statistical analysis. If this number is acceptable and the process is capable of meeting tolerances, the gauge will automatically be acceptable for part acceptance to tolerance.

If the gauge will not be used for process control purposes, but only for product acceptance purposes is the only situation where GRR %Tolerance becomes important (unless the customer insists, of course).

The method described by Atul would yield differing results depending on a shift in the average. A better method for determining acceptability would be to develop a Gauge Performance Curve as described in the MSA manual. While it does not give you a single tidy number to base your decision upon, it does clearly show your Alpha/Beta risks as you approach the spec.

Thank you

From Minitab R14 Help:

%Tolerance: Percent of tolerance for each component.

If the tolerance (Upper spec - Lower spec) is given, percent tolerance is calculated by dividing the Study Var for each component by the specified tolerance.

If only one specification limit is given, percent tolerance is the one-sided process variation divided by the one-sided tolerance. The one-sided process variation is Study Var divided by 2. The one-sided tolerance is the absolute value of the given specification limit subtracted from the average of all the measurements.

I think the MSA book is vague enough to cause all these unnecessary (should be if they were not vague) questions :( . I've been fighting this problem and have figured out what MiniTab does with unilateral (at least checks like flatness, runout, etc.) tolerances. This is in the Minitab help if someone needs further clarification. Take the metric (EV, AV, or GRR) and do the following:

%Tolerance(unilateral) = ((EV*6)/2)/(abs(AvgOfAllSamples - Tolerance))

I'm a programmer working for a Gage Company that has been trying to get our software to calculate GRR's as the AIAG MSA 3rd Edition manual specifies. I've downloaded a free trial of MiniTab to verify my results. If you believe Minitab calculates Gage Studies in accordance with the MSA, which I do, it is probably your best source for how to do the calculations, via the help files and creating samples. The only thing I really had to dig for is how to compute the p-values for the ANOVA method, but after many hours on the internet I did finally solve that problem. If your using Excel this is a non-issue due to the fact the FDIST and FINV function will compute the necessary values.:)

OOOOOPS, I missed Atul response - basically says same thing.