Gage R&R Anova Method - Study Variance

R

Reach

Hi I have been studying about Gage R&R and did some experiments at my internship. and I ran into a question analyzing the results.

I see the general guideline for the AIAG method is that that percentage GR&R should be less than 10% or 30% to be marginal.

My question is, do i apply the same guideline when I'm using the ANOVA method (used MINITAB)? Do I see if my %StudyVariance is smaller than 10% or 30%? or is there some other specific guideline for the ANOVA method?

Thanks!!
 

Miner

Forum Moderator
Leader
Admin
There is no AIAG method. AIAG stands for the Automotive Industry Action Group who published the MSA manual for Ford, GM and Chrysler.

In the MSA manual, there are two optional methods: the Range method and the ANOVA method. Both methods will provide very similar results. The Range method uses simpler math, but the ANOVA method can detect a potential Operator x Part interaction.

The acceptance criteria is the same for both methods.

% Tolerance or P/T Ratio (gages used for inspection):

  • < 10% is ideal;
  • 10 - 20% is acceptable;
  • 20 - 30% is marginal but may acceptable if the characteristic measured is not critical/significant and better gaging is not economical or feasible;
  • > 30% unacceptable
% Study Variation or %GRR (gages used for process control or statistical studies):

  • < 10% is ideal;
  • 10 - 20% is acceptable;
  • 20 - 30% is marginal but may acceptable if the characteristic measured is not critical/significant and better gaging is not economical or feasible.
  • > 30% unacceptable
% Contribution (used to assess each component of measurement variation to prioritize measurement system improvements):

  • < 1% is ideal;
  • 1 - 9% is acceptable;
  • > 9% is unacceptable.
Number of Distinct Categories or ndc (for SPC):

  • > 10 Ideal
  • 5 -10 acceptable
  • < 5 unacceptable
 
Last edited:

bobdoering

Stop X-bar/R Madness!!
Trusted Information Resource
Number of Distinct Categories or ndc (for SPC):

  • > 10 Ideal
  • 5 -10 acceptable
  • < 5 unacceptable

Actually, for SPC it should be 10 minimum calculated as:

1.41[(USL-LSL)/GRR]

That provides 5 distinct catagories on either side of the mean. Anything less than that is borderline attribute. 5 gives you 2.5 on either side of the mean. Now that is a weak chart.
 

Miner

Forum Moderator
Leader
Admin
Actually, for SPC it should be 10 minimum calculated as:

1.41[(USL-LSL)/GRR]

That provides 5 distinct catagories on either side of the mean. Anything less than that is borderline attribute. 5 gives you 2.5 on either side of the mean. Now that is a weak chart.
I am providing the criteria as stated by AIAG, which is very conservative.

See this article by Donald Wheeler on "An Honest Guage R&R Study". As Wheeler states the use of standard deviations (which are not additive) instead of variances (which are additive) artificially inflates the measurement variation.

Wheeler has also done an article (Good data, bad data and process behavior charts) modeling the effect of varying levels of measurement variation on the effectiveness of SPC. This simulation showed that process control charts were relatively robust to measurement variation.

Given that, your advice is good regarding quality measurement for SPC. However, in some circumstances, it could drive someone to invest in a more expensive gage than may be necessary for most processes. Precision machining may very well be an exception to this. You would know better than I on that.
 

bobdoering

Stop X-bar/R Madness!!
Trusted Information Resource
I understand - oh so well - that it is the AIAG recommendation. And, I would say that for typical gaging ndc=5 may be fine. But, if you are trying to determine runs, other Western Electric rules - or any process activity - with a gage whose distinct categories indicate only 5 statistically valid "buckets", you have some pitiful resolution. I have seen SPC charts that jump in increments so large that they provide poor control limit analysis and overall poor control. It dilutes any predictive power you may have had. Point is, the people setting up the chart did not relate the statistical gage resolution to the resolution they thought they needed. Face it, they may have had 10 lines on their chart between the control limits, but with the gage they had, they only needed 5 lines. It would be a more accurate visual of their level of control.

That said, if (for example - with enough exaggeration to make a point) your control limits are 10% of your specification, your PV represents that 10% of the specification and you have a true normal (or similar mean-central distribution) then ndc=5 could well be tolerable. But, if your distribution needs to be controlled via adjustments for any reason (not just precision machining), then I would recommend buying a more suitable gaging system to get statistically valid 10:1 within the control limits. You need some decent resolution to let "the voice of the process" (to use a Dr. Wheeler term) speak in clear terms - instead of mumbling.

I am also aware of the concern for the calculation for the GRR that Dr. Wheeler has brought up (I have read both articles). What I have not seen is a corrected ndc calculation from Dr. Wheeler from that analysis. Maybe I missed it. That could be a good improvement on the analysis. The concept of statistically significant resolution is critical to measurement system decisions. If there is an improvement, I am anxious to see it. Even so, I would go with a "corrected" (but still based on USL-LSL) ndc>=10 unless atribute (or near attribute) is "good enough".
 

Miner

Forum Moderator
Leader
Admin
Wheeler recommended the Intraclass Correlation, see "An Honest Guage R&R Study" section IV, Interpreting the Results. This is used to categorize the gage as a 1st through 4th class monitor. The table in Honest Step 11 then defines the capabilities of each class of monitor.
 
Last edited:

bobdoering

Stop X-bar/R Madness!!
Trusted Information Resource
Wheeler recommended the Intraclass Correlation, see "An Honest Guage R&R Study" section IV, Interpreting the Results. This is used to categorize the gage as a 1st through 4th class monitor. The table in Honest Step 11 then defines the capabilities of each class of monitor.

OK, to utilize Wheeler's approach (which starts out by completely tossing out AIAG's approach - if the OP is willing to go there), then it appears:

The value of 10% of (USL-LSL) must fall between the Smallest Effective Measurement Increment is (0.2 PE) and the Largest Effective Measurement Increment (2 PE) - where PE is The Probable Error of a single measurement. That would give you 10 valid "buckets" between your control limits.

I interpret that he has replaced ndc with the area between the Smallest Effective Measurement Increment and the Largest Effective Measurement Increment.

Note: Do you like his Gage R&R logo? Scary, huh?
 
Last edited:

Bev D

Heretical Statistician
Leader
Super Moderator
Note: Do you like his Gage R&R logo? Scary, huh?

It's only scary becuase he's left out the specific identifyier to AIAG...as you all know I am not a user or fan of the 'traditional' approach perpetuated by AIAG...
 

bobdoering

Stop X-bar/R Madness!!
Trusted Information Resource
It's only scary because he's left out the specific identifier to AIAG...as you all know I am not a user or fan of the 'traditional' approach perpetuated by AIAG...

That is understandable, and Wheeler has done a thorough job in describing his issue with the AIAG approach. I admit, at this point I have not had a chance to do a thorough comparison of the techniques.

But, the worst approach is not looking at the "statistical resolution" of a gage versus its intended output. People will do gage R&R, it will "pass", but it is not related at all to the resolution needed (10:1 minimum to the area between the control limits) to properly utilize SPC. It is a real missing link, especially for AIAG technique (and not too clear in Wheeler's writing, either).

Like most quality tools, gage analysis is not plug, chug and walk away. It needs some thought.
 
Last edited:
Q

quincy

Do not use % study variation to assess whether or not the gage is acceptable. The study variation is an estimate of the width you need to capture 99% of your process measurements (using a multiplier of 5.15). Focus on 2 things. First, does the % R&R fall between 10% and 30% (better yet, is it below 10%)? Second, use the "Number of Distinct Categories" feature in Minitab to see if there is enough resolution in your measurement device to pick up the variation between parts. For example, if you get a value of 1, this tells you the measurement system is poor (it cannot differentiate between parts). A value of 4 or higher indicates that the measurement system is acceptable. A value of 3 says that the measurment device can segregate the parts into 3 unique groups - low, middle, high. Hope this helps.

Regards,

Q
 
Top Bottom