Measurement System Analysis - Nested Gage R&R
Recently hit road blocks in trying to complete an MSA for a Instron, peel testing device.  The % Tolerance  and distinct number category is not being met (<30% Tolerance, and > 5 categories).  I've attempted multiple runs, carefully choosing samples, and still the variation is too large.  I can't even get the parts to be distinct, no matter how much I change the seal temp, pressure, and time between parts.  The parts are simply not distinct from each other.  In addition, within the same parts, the repeatability is high enough that %Tolerance fails.  I can't definitively say that the equipment is at fault, or the process simply can't make seals consistently.   

So.....I've been thinking. The purpose of MSA's are to determine the inherent variability in the measuring process caused by the equipment and/or operator.  The actual capability of the machine producing the samples doesn't  matter.  Why do I have to use samples from the production floor?  Why can't I create my own to make sure that each part have different peel strength, so that the distinct category will be met.  For example, why can't I simply use 5 different types of glue, and seal tyvek strips to a poly-film to create my own samples.  Are there vendors that sell test samples to qualify peel testing equipment?


11 Replies
Daniel, if you send me the data I maybe in a better position to respond

Sent from my iPhone
John Vandenbemden
Principal consultant Q-Met.Tech
Mobile: 859-240-1739
Dear Daniel Callahan,

I have run a small simulation on what I think you are talking about and it may be due to the inability to create the different homogeneous sample subgroups.

"I've attempted multiple runs, carefully choosing samples, and still the variation is too large.  I can't even get the parts to be distinct, no matter how much I change the seal temp, pressure, and time between parts.  The parts are simply not distinct from each other. "

You can run an Anova and confirm that the categories come from different homogeneous sample subgroups. When I ran a GR&R the number of distinct categories was less than 1.

You might want to run a simple 2^2 DOE to create different homogeneous groups. The idea is not to create perfect samples but to create the 4 or 5 different homogeneous groups. An alternative may be to use some batch rejected parts that have been rejected for different reasons or even combine the different forms of rejection.

Note: the part to part variation will be accounted for in the nested calculation and the different homogeneous groups should be similar enough to each other (recall... between (different groups) and within variation(same group))

When I simulated the ability to create 3 homogeneous groups it increased the ndc to 3 with 27 measurements total. I was able to increase the ndc to 4 -5 when the number of total measurements reached about 45.

Hopefully, the numbers required are practical enough if we use the rejected parts.

A simple correlation may be done by running an Anova to test whether the homogeneous groups are sufficiently different and compare it with the GR&R ndc results.

Let me know whether my simulation is similar to your results if you do try this instead  of creating test samples which may not in some circles be representative of the process under consideration.

Yours sincerely,
Here are my results from a 45 Run, Nested Gage R&R
I will review this and let you know what I find!
The results overlap which causes the distinct categories to be 1 and not recognized as 5.  When running an R&R it is normally conducted with 10 different measurement ranges (parts) in order to achieve to acquire at least 5 categories.  I ran a histogram of the results which depicted this situation.  

The suggestion you present on creating a different type of study can be completed using a randomized block design or simple ANOVA comparison.  I not sure what your actual output question is.  Are you looking for the best glue.  If you have only one factor a simple on-way ANOVA would work.  If it is a little more complex a randomized block design or factorial experiment can be used.
Dear Daniel Callahan ,

Your current data shows an ndc of 1. I used a boxplot and noticed that part labelled 2 has a high variation and parts labelled 4 & 5 are similar...

Dropping parts labelled 2 & 5 results in an ndc of 2.


Give it a try and I hope you will be able to make sense of what John and I are saying.

Yours sincerely,
Dear Daniel Callahan ,

In case you were wondering what my simulation results looked like... for a higher ndc... 


Hope this makes things clearer,

Thanks for the thoughtful responses.  My difficulty is in creating the "10 different measurement ranges".  I ran a DOE to try to create seal variation to allow me to get these distinct parts, and as my MSA results show, most of these parts overlap.  I can't create samples that are well suited for this peel test MSA.  I wish I could just purchase parts from a vendor to test....
Dear Daniel Callahan,

Do you think that what we have done is able to demonstrate the due diligence we have undertaken to show that we have carried out the feasibility of using our parts?

Would any reasonable person think differently and be able to prove that it can be done...

I think you have your justification you need through this process to use other than your parts (I believe you did get a higher ndc > 1) ... but it hasn't stopped you from reviewing from time to time :)

Hope that answers your question,

You have 2 issues:
One is production samples versus lab samples; the other is your MSA design. I'll tackle that first.

It does not appear the experiment is sufficiently nested.
You can never ever separate the variation in the samples from the measurement system variation with destructive testing because you cannot recheck the same part the way you can with a gage. You need to construct a different experiment to remove or identify as many of the sample related variables to minimize the test variation and sample combined.
How many samples can you make from one seal?
How many seals can you make at one time - ie is this a continuous process or is this a single or multi-cavity tool that makes the seals?
How many seals can you make from one lot of material (ie - substrate, adhesive, coating all the same lot)
There may be others, but I don't know enough to speculate further than this.
For those of you who have participated in round-robin testing between test labs for any given test, the solution to this MSA is very similar.
For the first part - why can you not just use lab parts?
This is where so many practitioners go astray. While it is true, you can improve your MSA results under more tightly controlled conditions, you may end up missing some significant factors affecting variation within or between samples that you might otherwise detect when you use real production samples.

MSA is to determine if your measurement system is adequate to make decisions for one of two questions: Can you get a reliable judgement of good or bad with minimal risk of a wrong judgement based on the measurement? Is the measurement system suitable to detect process changes when used for statistical evaluation - ie SPC and or Capability Analyses?

It appears you may have reason to question the test method and/or the process making the seal. I would recommend you perform a lab sample MSA to evaluate the measuring process relative to the results you obtained so far. Again, you will still need to develop a more rigorous experiment to analyze the components of variation.

If you need help with either of these, let me know.
Dear Daniel Callahan,

Thank you for the good refresher on materials, adhesion, substrates and homogeneous groups for GR&R.

Your seal is
"why can't I simply use 5 different types of glue, and seal tyvek strips to a poly-film to create my own samples"

An interesting piece of information is the different surface energies of materials at

An interesting option as opposed to changing glues...maybe to create samples with the different types of pressure sensitive tapes available not necessarily only for tyvek to setup the homogeneous groups... Your values are quite small? (2-3.6) close at the moment...(Instron load cell range/accuracy check).

If you do try it, please let us know if
1. homogeneous groups can be established (check the adhesive substrate and tape matrix)
2. the different Pressure Sensitive Adhesive tapes allow some sort of interesting bench marking with the glue used

I remember specifications calling for 3 M tape for the cross hatch peel test... but this seems to be interesting if it can be used to help homogeneous groupings for peel tests...

Thanks for the feedback,