The customer isn't always right, is he?

Our customer (a large automotive supplier) is requesting that we do a "Cross Method" gage R&R on a device that is collecting variable data that we chart for SPC. Additionally, he says we need to have some know failures in the samples. Am I wrong or is he talking about the Cross-Tab method in the MSA? And that's for attribute data only, right?

He's communicating his requests through my boss who doesn't know a lot about gage R&R's, and then I get the directive to do it. Hopefully, I'll get to be in on the phone call today and ask him.

If anyone has heard of a cross method for variable data, please let me know ASAP.

Thanks

The customer isn't always right, is he?

Our customer (a large automotive supplier) is requesting that we do a "Cross Method" gage R&R on a device that is collecting variable data that we chart for SPC. Additionally, he says we need to have some know failures in the samples. Am I wrong or is he talking about the Cross-Tab method in the MSA? And that's for attribute data only, right?

He's communicating his requests through my boss who doesn't know a lot about gage R&R's, and then I get the directive to do it. Hopefully, I'll get to be in on the phone call today and ask him.

If anyone has heard of a cross method for variable data, please let me know ASAP.

Thanks
In general, "bad" parts are expected to be used only in attributes studies, because you are supposed to provide evidence that the system in question will alwaysdiscriminate between conforming and nonconforming parts for the characteristic in question. Unfortunately, and in my opinion, the AIAG 3rd Edition MSA manual took a complicated subject and made it more so. I think it would be a good idea for you to speak directly with the person who's communicating the requirements. There's no telling what might be getting lost in translation. You also might want to have a look at page 73 of the 3rd Edition MSA manual which describes "Preparation for a Measurement System Study."

ScottyWM -

Your customer is asking for nothing more than the "standard" GR&R to be ran. That is where the parts and operators are "crossed" in the model with each other. The treatments (operator, part) cross over into the response. The replicates are used to estimate all the effects and variance components for factor or treatment individually.

The most common alternative is called "nested" and is often used for destructive testing or in batch industries. It has the trials "nested" inside of samples. That is, the replicates are used to obtain an average response and that is then used by the model to estimate the effects.

It's hard to express in words. But if you should look in any good DOE book, it will talk about this and will likely show diagrams that will illustrate completely. A GR&R is nothing more than a DOE. The effects are usually not of importance, only the variance components.

Thanks Dave, that sounds sort of like what we did. I hope that's what the customer was talking about, but somehow I doubt it. It really sounded like he was going for the "cross-tab" attribute study. He still isn't at his desk for our call...

We are testing the strength of gear teeth on a plastic gear. We use a digital torque wrench and a fixture to break the teeth. For the study, we tested 10 selected teeth (the 10 samples) on 9 gears from a batch (3 appraisers x 3 trials). So we are testing tooth #1 on 9 gears as 'sample 1', tooth #2 on 9 gears as 'sample 2', etc...

Of course no two gears are identical, but we feel comfortable that they are close enough for this study. We analysed the results using the standard average and range method with additional graphical treatments.

MSA gets much stickier when destructive testing is involved. The standard method you cite, known generally as the "long method" GR&R carries with it the assumption that all three operators will measure the same characteristic on the same parts, which is impossible when the feature has been destroyed in the initial instance. The AIAG 3rd Edition manual includes a broad disclaimer regarding MSA for "non-replicable measurement systems" basically saying that it's beyond the scope of the manual (see page 143). It's now even more important that you discuss this with the customer, as there's not much sense in engaging in testing that's likely to have unreliable results. In my past life as a job shop QM I made a point of speaking with the customer and getting them to specify what they want in such instances, as early in the development process as possible.

ScottyWM -

Your customer is asking for nothing more than the "standard" GR&R to be ran.

No, not if destructive testing is involved. For "standard" GR&R the measurements must, by definition, be replicable.


It's hard to express in words. But if you should look in any good DOE book, it will talk about this and will likely show diagrams that will illustrate completely. A GR&R is nothing more than a DOE. The effects are usually not of importance, only the variance components.

I think you're on the right track here, in that some sort of designed experiment is needed. I think it's important to remember that GR&R and MSA are not synonymous; GR&R is one type of MSA and is not appropriate in all instances. What's needed is a reliable statistical method that indicates that the measurement system is efficacious (capable of performing as intended).

Scotty -

Your description sounds like a nested design, with teeth nested inside of gear. Did you analyze it correctly?

Since it is destructive, you need to do a nested model for analysis. Some statistical software will do this, MINTAB V13 and higher is one. I use that now. If not available, can be done by hand. I can dig up references if need be.

If your customer wants a crossed design, I don't see how to do it in valid way.

If he really wants a attribute study to a break limit, than your sample sizes will likely need to be pretty large, I suspect.

I've seen the disclaimer, and agree that this is a destructive test. Using the process I described, we're saying that the batch parts are enough alike to qualify as a replication.

We've been doing this test for approximately 6 years with no complaints (same customer). It was originally qualified simply by looking at the stability and capability using x-bar r charts I believe. Now the customer has asked specifically for a "Gage R&R" on it. Then he requests specifically that we use the "Cross method". Makes me question his understanding. I would prefer to use different qualification methods, but we are trying to make the customer happy. And in fact, the standard Gage R&R came out to be pretty good. We feel it's a waste of time to further study this (simple) system when a GR&R (using described method - replicable samples would be even better) is good, we are miles away from a minimum specification, and we show good stability and capability.

Hope that he returns our call today...

I feel your pain. As with so many other issues we encounter, there's theory and then there's practice. In theoretical mode I would tell you that you can't assume replication. Variables average-and-range GR&R is not appropriate for destructive testing. There's a big catch-22 in the AIAG model, in that you're expected to do a capability study and also show that the measurement system used is efficacious. But one of the requirements for the MSA is use of parts that represent the operating range of the process, and that information can't be considered reliable without the MSA :frust: .

On the pragmatic side, you want your customers to be happy, and sometimes they're not happy until common sense has been burned beyond recognition. But it's important to remember that it's not necessary for the customer to be correct in the theoretical sense in order for him to be happy. So you gotta do what you gotta do, but not without first talking it over and hearing firsthand what the customer is looking for.

Okay... I think I've figured it out. I downloaded the 30 day Demo of Minitab 14. It has a "Crossed" method for GR&R and a "Nested" method. At first I tried the crossed since that is what the customer said to use, but couldn't duplicate his results. (He must be using an older version of Minitab because he had a 5.15 multiplier instead of 6. But it didn't match with either). The Crossed method appears to me to be a standard GR&R, don't know why it's called crossed. Then I tried the Nested method, and VOILA! Using the 5.15 multiplier with the nested GR&R, I was able to (very closely) duplicate his results. All of my personal GR&R reports use a multiplier of 6 as per MSA 3rd edition. Never even thought of changing that. The nested GR&R is for destructivbe testing where it's assumed the batch is homogeneous - which is what I was doing.

So we're GOOD! And I've learned something too!! Thanks y'all for your insight. I'll have to play with the Minitab and see what else I can learn.

Okay... I think I've figured it out. I downloaded the 30 day Demo of Minitab 14. It has a "Crossed" method for GR&R and a "Nested" method. At first I tried the crossed since that is what the customer said to use, but couldn't duplicate his results. (He must be using an older version of Minitab because he had a 5.15 multiplier instead of 6. But it didn't match with either). The Crossed method appears to me to be a standard GR&R, don't know why it's called crossed. Then I tried the Nested method, and VOILA! Using the 5.15 multiplier with the nested GR&R, I was able to (very closely) duplicate his results. All of my personal GR&R reports use a multiplier of 6 as per MSA 3rd edition. Never even thought of changing that. The nested GR&R is for destructivbe testing where it's assumed the batch is homogeneous - which is what I was doing.

So we're GOOD! And I've learned something too!! Thanks y'all for your insight. I'll have to play with the Minitab and see what else I can learn.
Congratulations, Scotty:
Did anyone address the issue of including "bad" parts? For GR&R to work for me, we had to include samples representing the WHOLE gamut of variation. We were given process tryout samples too close to identical; the measurement variation dwarfed the product variation (uh, oh). Worse than that, my Lab people were blamed for not holding the measurements tighter!

Did anyone address the issue of including "bad" parts? For GR&R to work for me, we had to include samples representing the WHOLE gamut of variation. If you can prove that your measurement system works at the specification limits (and in between, of course) what's gained by using parts outside the limits? Of course I'm talking about variables here; "bad" parts are always necessary for attributes studies (simply showing that an attributes gage will always recognize "good" parts isn't enough).

They never did expand on why they thought we needed parts outside the spec. limits. I'm convinced they thought it was an attribute test. The MiniTab GR&R's easily satisfied them, but it was similar to what we got with a standard R&R.

When it was all said and done, our company spent many MANY man-hours and $$ to chase a problem that was not even close to being ours. It turned out that our "Failed" unit was mishandled by the customer in an undocumented, non-normal, one-off assembly process for some test that they were doing. Once they found their involvement, they got real quiet right away. :bonk:

But I learned some new things, so all was not for naught!

You may want to review the following link.

http://www.aiag.org/forms/Non-Replicable_GRR_CaseStudy.pdf

This is a case study for performing a Gage R&R on product that cannot be retested, either because the test is destructive, or because the test changes the part in such a way that retests are always different from the original.