Tool to tool matching

[#1]

Hello.  I was wondering best practices for proving tool to tool matching with many factors.

For instance I have 3 tools, 2 heads per tool (a left and a right).  I want to match 4 conditions for 10 parts per condition measured 10 times each. 

I would normally do a GRR if it only had 3 factors.  Should I just do fit model and use the variance components as metrics?

Should I just do a separate GRR for each condition then compare them?

The information I'm really trying to glean from this is if the gauges match and are capable at each condition (to some spec).

In reality, the part isn't measured at the same spot for the repeats of the measurement as it is moving.  And L&R won't line up exactly to the same spots either from tool to tool.  I think I can just nest those but I don't know if another method is preferred. 

I attached a mock table in case it's not clear what I'm asking. 

Even just recommended reading would be appreciated.

Thanks,

Vince

[#2]

Vince,

You've got a couple of options here... First is just to use the components of variance analysis (under variability study).  Another option with multiple matching criteria is to build a model in Fit Model with the conditions and see if any of them are significant at some level of alpha.  If they aren't significant, then the tools are effectively matched. 

Best,

M

[#3 · Accepted solution]

Vince: You may also want to take a look at Don Wheeler's EMP algorithm for analyzing data in this regard. I like Wheeler's method over the traditional Gage R & R approach primarily because there are many different graphical means by which Wheeler's method answers the various traditional questions around repeatability, resolution, reproducibility, interactions, etc. Here's a link in the JMP online documentation that illustrates the EMP approach as deployed in JMP:Measurement Systems Analysis

[#4]

I like this, but I the problem I come to again is just doing it by each nominal condition?  Let's say I'm trying to measure length of something.   I want to know how well it can control at 4 different lengths.  If I just plug everything into the MSA, it gives me an average of the whole, but I'm introducing false part to part variation.  So I can run it separately at every length, but then I get a classification for each condition.  Is there a way to get one classification of the tool for the whole range of conditions? So when I do the analysis of means it would give a UDL/LDL for each nominal condition?

[#5]

Thinking out loud here...always dangerous on a public forum...what if you calculated some kind of normalizing response for each level...like a z value for the level since it just sounds like you are trying to establish repeatability/reproducibility across the range of nominal values. Then the Average and Range charts in Wheeler's method wouldn't care about the pseudo false part to part variation?

[#6]

I like it.  That helps me get the information I'm looking for.   Thanks for the help.