samples for R&R: do I have to keep track 'behind the scenes'?

I'm getting ready to start running a gage R&R, I've only had to do it once about 10yrs ago

I've got my 10pcs & 3 operators, having them measure the pcs twice on different days

Do i have to keep track of which part is which?
Or is it a completely randomized free-for-all for all six data sets that will be generated?

My understanding is one of the things R&R looks at is measurement difference on the same part for individual users and also between users. Yes/No?
If I don't keep things straight 'behind the curtain' how can this happen?

ie if I give them to 1st person tagged "1-10", does the 2nd person have to go in the same order?

For the second trial, if I mark parts "a-j", mixing up the order from the first trial, do I have to keep track of which part is which from the first trial?

i.e. if part #1 from first trial is tagged as part "c" for the second trail and so on, do I have to reconcile which is which when I'm entering the data into my spreadsheet?


Am I making this too difficult?
Please set me straight, I'm second guessing myself and it's driving me batty

 

Thanks Miner, I've been driving myself crazy on this one. You explanation is the clearest of anything I've seen all week.

 

Going to add:
First time through, go ahead and go 1 through 10.
Subsequent times, you want the order to be different.

In statistics we say "block what you can't randomize and randomize what you can't block." That's fancy for saying do what you can to eliminate external effects. Without knowing what your gage is - how do you know that the gage won't "warm up" and start measuring differently over time? Especially if in an R&R study, they can go part, part, part in rapid succession as opposed to some process feeding the gage a part every 5 minutes.

If you measure 1 through 10 in a row in the same order, you will hide this possible source of noise. So you scramble the order.

But what Miner says is critical. While the warming up thing can happen, a far greater contributor to a bad study is letting one operator make his repeat measurements all at the same time. In fact, when he is on pass 2, you should give him a brand new data sheet to record his results. Think about it. Your operator measures and gets 0.250 on pass 1 and writes this down. He comes back later, and on the same part he gets 0.200 which is pretty different. He goes to write this down and sees he wrote 0.250 last time. He is going to say "one of these isn't right" and pollute the system, remeasuring until he gets a "better" result that agrees more simply because he knows what he just got.

You aren't after the most accurate measurement. You are after - how well does this gage work? In production, he's not going to have prior knowledge. He's going to be given a part and told to decide good or bad.

 

Thanks for insight NC

We are qualifying a new CMM. It's a smaller version of what we already have, and is running the same probes, programs etc.
I ran a couple trial parts and they are within .0002 between cmms on what I've checked so far
(Majority of our parts are +/-.010)

IMO this is a waste of time, but we have a semicon customer who is requiring it to be correlated to old one using R&R

 

Got it. That helps me understand.

In truth, I don't think your customer has it correct. An R&R study has nothing to do with correlating two gages.

Let's think bathroom scales. Suppose you have two scales that are identical. But I put a magic, invisible 20 lb weight on one of them. If we feed them 10 different adults repeatedly for an R&R, BOTH of them will pass the R&R. The results will be repeatable and reproducible. Except on one of them, all the results will be 20 lbs heavier. The one I have toyed with WILL PASS the R&R. And the results of the two different R&Rs will NOT give you any indication if they are correlated or not. Certainly, at the end of your R&R study, you'd see one scale was out of calibration. And to determine this, you'd have to go get a known weight of some sort and figure out which one was closer. (Or, you'd notice that one that was empty, was showing 20 lbs when it should have said zero ;-)

The example is simplistic, but I'm trying to demonstrate that correlation is a DIFFERENT thing.

If I were trying to correlate CMMs, I'd pick out some of the more difficult dimensions. Which would be things like:
1) Partial circles you were getting diameters on
2) Something with very tight tolerances
3) Anything where you are projecting - pick the longest projection on the given part
4) A couple of true positions - BUT also report the coordinates. It's more important the coordinates correlate than the actual true position result.
5) A dimension where the alignment system has been redefined the most times from the first alignment system.

Look at your part that way, select 6 to 10 dimensions (or, choose all of them, if your good at getting exports into Excel) and THEN pick 6 parts. Pick your 6 parts carefully. You'd really like a few of them to be out of spec on some of these dimensions. You want a range.

Measure the 6 on CMM A. Measure the 6 on CMM B. Take the delta of each of these measurements and compare this difference to the tolerance of the feature. If it's less than 20% different to available tolerance, you are pretty correlated.

But again, just performing a Gage R&R on the second CMM doesn't mean it is correlated to the first. It just means the second one is repeatable and reproducible.

So ... If you DO HAVE a Gage R&R on the first CMM that demonstrates the system is OK, which is: the CMM itself + the program + plus the selection of tips + the fixturing. I think you could just correlate the second to the first. If the second isn't repeatable, you'll see this in the correlation. The disagreement of the measurements will be all over the place.

 

Very interesting, I'm hitting 4/5 of your suggestions.
I'm working with a part called out by the customer that has several TP callouts and has a co-ord system realignment during the CMM routine.

Problem is we run everything to median on tolerances and I'm not going to have a heck of a lot of variation, which I know will make the R&R results marginal.

Expensive parts, lots of setup steps.
I can't have production run 'bad' parts for my R&R/qualification study.
We are too busy trying to get parts out the door, production would laugh themselves silly if I asked.

Maybe I can artificially tighten up the tolerances to tweak the R&R results.
And my customer is ALWAYS correct, infallible they are. Yup. (/sarc)

 

Bev:
Got several of the Wheeler papers on the topic to read...found them on Miner's blog
Now I just need to get left alone for more than 10min
I'll take a look at your reference papers too.

Thanks to all for the guidance, just have to dig into it and digest.

 

Do you have any setup or scrap parts? Customer rejects? The correlation does not have to be on good parts.

If you want to check a box with a customer, you COULD do the R&R but do NOT report the % Study Variation numbers. Those are going to be bad because your parts are so similar. Just report the % Tolerance numbers. (Which is what you really want anyway). % Study Variation answers the question "can I tell the difference between the parts I have studied." % Tolerance answers the question "can this gage be used to discriminate against the tolerance." That's the question the customer wants answered - can this new gage check effectively against the tolerance?

BevD above isn't wrong, but the right answer doesn't always satisfy the customer.

 

Everything I have is from stock, no setups/rejects hanging around.
I was looking at some examples, tweaking tolerance limits to see what happens, but I think the % tolerance is probably the path I'll take.
I had used those results on an old trial I did in '09
THnx again!

 

I'm with you. But hows that quote go? "You can lead a man to knowledge, but you can't make him think."

In my industry, if I don't pick my battles carefully, I will work myself to death. Isn't worth it. (I'm not suggesting I do anything unsafe or bad. I just don't always have the time to train my customer. It's more important I train my suppliers).

 

Now one more question - I’ve been asked to run the third iteration of a GR&R 1 month after the first two runs and I may even need to use different parts. It’s an ECU and the measurement is on a test system that automatically collects data. The GR&Rs with two runs are under 15%, so mostly good but I’m not comfortable sharing the part variation with only two runs when it should have been three per the AIAG standard.

I’m being told by production that this is a waste of time since the environmental concerns (
Atmospheric, generally but other things like processes going on in parallel). I agree with them that the best practice is to do the entire GR&R within a week but preferably a day, but I don’t think it’s a requirement and they’re just digging in their heels to avoid repeating the work. What are your thoughts?

 

Our test engineer ran 10 parts only twice and now the question is if the same thing parte could be added to the existing
GR&R one month later for the same 10 parts as the first two runs. ECUs and uncased.

 

It’s okay - I’m moving on since it’s too confusing to describe. Thanks all.