What is an acceptable Gage R&R (Repeatability and Reproducability) system?

[jaker]

What is an acceptable Gage R&R system? Do you have to have a study on each instrument or can in be done by family (calipers, ht gauges, etc.)? Does it have to be done on production parts or can it be done on gage blocks used for calibration?

Forgive my ignorance!!!

[This message has been edited by jaker (edited 06 June 2000).]

 

[Marc]

As I understand the requirement....

Typically QS auditors expect to see an R&R done with production parts on each instrument cited in the control plan.

However, if you are really MSA savvy and are ready to explain how you address each part of your uncertainty budget, you can use other methods. For example, if you have 100 sets of 6" calipers of the same model from the same manufacturer and they were maufactured within a 'reasonable' relative time frame (say - all were manufactured in 1997 and 1998) - you can make an arguement that all have nearly the same uncertainty in so far as the instrument its self goes. You had best have a study to prove this is, in fact, the case.

Yup - you can use gage blocks.

But - in both of these you had best be able and ready to explain how you address the uncertainty associated with surfaces being measured. If you use the calipers on an edge to edge measurement where they will be seated much like on a gage block, you will probably have a lower uncertainty factor than if you are measuring an 'unstable' part feature such as a diameter on a gently grooved feature. The fact that the instrument is stable and precise does not address several other parts of your uncertainty budget for that measurement.

The bottom line is if you do an R&R on the specific instrument at the station where the part is being measured, on the feature being measured on the part which is being measured, by the techincians who will/are actually be doing the measurement you are addressing all of the uncertainty budget and fulfilling the auditors paradigm (you are showing the auditor what s/he expects to see).

 

[David Bear - 2010]

I am also relatively new to Gage R&Rs.

During our last QS9000 audit, our auditor said his interpretation is that Gage R&Rs only need to be done on an "as needed" basis. The need occurs whenever a modifciation is made to a gage or if you have a new operator. Any requirements from your customer would override that frequency.

I have two questions I can't seem to get answered. Can someone help me with them?

1. If I add a new operator, can I add their measurements to an existing Gage R&R? ex. I already did a Gage R&R with 3 operators. Can I add a 4th operator and recalculate the variation by adding the newest measurements to the already existing measurements?

2. How many operators do I have to use and what does my sample size need to be? I have been told: a) 3rd edition requires 15 samples for a variability study and 50 samples for an attribute study, b) my company's intranet example uses 10 samples for variability and 30 samples for attibute, c) there are no set sample sizes - the sample size is based on what is reasonable (i.e. 50 samples with a pin gage is easy, 50 samples that take 20 minutes each is unreasonable) d) you must have 3 operators (all must be from production), e) you can use 2 or 3 operators f) all operators using the gage must be tested.

I am receiving a lot of conflicting opinions.

I appreciate any help!:thanks:

 

[lee01]

GRR on gauge blocks? This is a no no right?

The whole idea about a GRR is to ascertain the gauges' ability at measuring production parts in a production environment.

Measuring some gauge blocks at your desk all nice and comfortable, your bound to get better results as if your measuring production parts, say a large composite skin of an aircraft? You’ll be on the shop floor, you’ll be rushed, you could be stretching in some instances, the part could be un-washed, the part may be difficult to hold, all this provides an element of measurement uncertainty, which is exactly what GRR is trying to find out. That’s why it used production tolerances.

I have found this out by conducting GRR within an inspection facility at final inspection stages, the parts fully cleaned and the inspector sat at his desk, and during production by the random inspectors. The differences are great.

 

[bmccabe - 2006]

I echo Marc's comments.
Planning, rational, and MSA's are the way to go. GR&R's are pathology, MSA's forecast. In fact, I really despise R&R's, I think they're flawed and a waist of time: Especially true for automated equipment, where Appraiser/Operator error = 0 by definition. R&R’s will assign a substantial measurement error to the monkey who pushed the run button.

Two fundamental flaws in the math used in R&R’s are first: The assumption that Operator, Equipment, and Part variation exist simultaneously in every data set. And second, that it IS possible to reproduce the same measurement on a given object.

Try this for fun:
Conduct an R&R using 10 parts, 3 operators, and 3 measurements of each part. Compare the capability of two measuring devices: A six inch steel scale graduated in 1/32nds of an inch, and a set of calipers with .0005 graduations. Measure the length of the part, to see which device passes the GR&R.

Which do you think will pass? Now, what I didn’t tell you was the nature of the part we’re going to use for the test. Compare now, the same R&R methods on two different parts. First measure the thickness of a precision ground gage block. Second measure the length of a human hair.

Do you believe the results would change? Would you believe calipers fail both tests? Would you believe the steel scale passes both tests? Now, ask why.

 

[Miner]

I echo Marc's comments.
Planning, rational, and MSA's are the way to go. GR&R's are pathology, MSA's forecast. In fact, I really despise R&R's, I think they're flawed and a waist of time: Especially true for automated equipment, where Appraiser/Operator error = 0 by definition. R&R’s will assign a substantial measurement error to the monkey who pushed the run button.

Two fundamental flaws in the math used in R&R’s are first: The assumption that Operator, Equipment, and Part variation exist simultaneously in every data set. And second, that it IS possible to reproduce the same measurement on a given object.

Try this for fun:
Conduct an R&R using 10 parts, 3 operators, and 3 measurements of each part. Compare the capability of two measuring devices: A six inch steel scale graduated in 1/32nds of an inch, and a set of calipers with .0005 graduations. Measure the length of the part, to see which device passes the GR&R.

Which do you think will pass? Now, what I didn’t tell you was the nature of the part we’re going to use for the test. Compare now, the same R&R methods on two different parts. First measure the thickness of a precision ground gage block. Second measure the length of a human hair.

Do you believe the results would change? Would you believe calipers fail both tests? Would you believe the steel scale passes both tests? Now, ask why.

The steel scale would fail the very first test that should be performed, namely does the gauge have adequate discrimination. If it does not have adequate discrimination, STOP!, Do Not pass GO!, Do not perform any more MSA tests!, Get a gauge with adequate discrimination, then perform the MSA.

 

[bmccabe - 2006]

Yes; we all know the scale has insufficient resolution. You may have missed the "for fun" part, and my point:truce: . My point is, the R&R rewards inferior measuring devices. As the quantity (not quality) variability decreases, the greater the chances the R&R will pass. The R&R places absolutely no relevance to the significant digits (number of decimal places) in which the variability occurs. It needs some percentage of the data to have the exact same value. As a greater quantity of data points vary, the more likely it is that the R&R will not succeed. Up until you get the "insufficient variability" error message, and mistakenly believe your good to go.

 

[Jim Wynne]

I echo Marc's comments.
Planning, rational, and MSA's are the way to go.

GR&R and MSA are not mutually exclusive; GR&R is a type of MSA.
And: did you mean rationale, or rationalism?

GR&R's are pathology, MSA's forecast. In fact, I really despise R&R's, I think they're flawed and a waist of time

Properly executed GR&R analysis provides predictive value. What's flawed and a "waist" of time is GR&Rs improperly done, or properly done and improperly evaluated.

Especially true for automated equipment, where Appraiser/Operator error = 0 by definition. R&R’s will assign a substantial measurement error to the monkey who pushed the run button.

If a GR&R study is done in a situation where other forms of analysis would be more appropriate, the fault lies with the person who decided to use GR&R, not with GR&R itself. You wouldn't use a sledge hammer to drive a finishing nail and then, when the inevitable happens, opine that hammers shouldn't be used to drive nails.

Two fundamental flaws in the math used in R&R’s are first: The assumption that Operator, Equipment, and Part variation exist simultaneously in every data set. And second, that it IS possible to reproduce the same measurement on a given object.

GR&R makes neither of these assumptions.

Try this for fun:
Conduct an R&R using 10 parts, 3 operators, and 3 measurements of each part. Compare the capability of two measuring devices: A six inch steel scale graduated in 1/32nds of an inch, and a set of calipers with .0005 graduations. Measure the length of the part, to see which device passes the GR&R.

Which do you think will pass? Now, what I didn’t tell you was the nature of the part we’re going to use for the test. Compare now, the same R&R methods on two different parts. First measure the thickness of a precision ground gage block. Second measure the length of a human hair.

Do you believe the results would change? Would you believe calipers fail both tests? Would you believe the steel scale passes both tests? Now, ask why.

Building strawmen isn't my idea of fun. What you're suggesting is that if it's possible to find a certain set of circumstances under which GR&R (or any sort of statistical analysis, for that matter) might prove inefficacious, then GR&R will be inefficacious under all possible sets of circumstances.

 

[David Bear - 2010]

Properly executed GR&R analysis provides predictive value. What's flawed and a "waist" of time is GR&Rs improperly done, or properly done and improperly evaluated.

I agree with JSW05. GR&R, done properly appears to be an excellent tool.

Does anyone use GR&R to test operators? Should other methods be used instead? I want to ensure people doing the measuring really know what they are doing. A GR&R seems to be a good way to test their knowledge and capabilities.

 

[Miner]

Yes; we all know the scale has insufficient resolution. You may have missed the "for fun" part, and my point:truce: . My point is, the R&R rewards inferior measuring devices. As the quantity (not quality) variability decreases, the greater the chances the R&R will pass. The R&R places absolutely no relevance to the significant digits (number of decimal places) in which the variability occurs. It needs some percentage of the data to have the exact same value. As a greater quantity of data points vary, the more likely it is that the R&R will not succeed. Up until you get the "insufficient variability" error message, and mistakenly believe your good to go.

My point was that you should never perform a GRR study if the discrimination was not adequate. All statistical tools have basic conditions that must be met before the results have any meaning. GRR is no different.