NDC (Number of Distinct Categories) - Long Gage R&R Study in MSA 3rd Edition

[Bev D]

Well, I would start by asking you 2 questions:
why do you think it's so bad?
what training have you had in Gage R&R?

I ask the second question because all to often the most training someone has in gage R&R (particularly when these kinds of questions are asked) is a simple: follow this protocol to run the test (3 operators, 10 parts, and 3 readings of each part by each operator) and then enter the data in these cells and look in these cells for the resutls and compare with the answer you 'should' get...

 

[Miner]

I couldn't get more than 1 ndc. See below data I used. I get % GRR 92.15 AND NDC 0. Please help...

For what purpose are you using this gage? It is marginally acceptable for use as an inspection gage at a P/T Ratio of 23.6%

If you are using this gage for process control (i.e., SPC) it is unacceptable at %GRR 92.15 and ndc = 0. There are several possibilities for the cause:

1. Does the variation of the parts reflect the actual process variation? If this was a restricted sample with less variation than the actual process, it will make the %GRR and ndc numbers worse than they really are.
2. If the product variation is correct, then your gage does not have adequate resolution to see the product variation or to see any differences between operators. This invalidates any interpretation of the MSA graphs.

 

[Bev D]

I plotted the data (see attached) and I would not hesitate to use this gage for either SPC or product acceptance. There is NO product variation in the 10 parts selected (which is why rpatel has an NDC of 0; actually it's less than zero...) and the measurement error Sd = .0008 vs. a tolerance range of .020

It's amazing the things you see when you plot your data, look at it and think about it...

SPC will clearly work although the results will be a bit chunky - you will still see trends and shifts...

As for product acceptance I would want to know how much real variation there is in the process. if it's fairly capable - I wouldn't change anything. If it is not caapble and has a lot of results at teh spec limits, I might try to improve the measurement system or invoke a multiple measurement of parts that are 'close' to the limits for awhile. However I would really focus on improving the process capability...I would also ask how 'real' the spec limits are - were they truly engineered?

Other than those fine tuning questions I would use the gage as is.

 

[bobdoering]

Gage R&R is not a tool - it is a tool box. It tries to do many things all at one time. Many have cited - to some degree of accuracy - that it doesn't do any specific thing very well. It can suffer further from people that attempt to plug and chug, but not really analyze the results. Bev has done a good job of looking beyond the "results" page - which I highly recommend. However, in spite of its critics, it does have some very good guidance, when used with some thought.

NDC is a valuable tool, because it is the tool to fight against those people that believe that if the resolution of the tool is in the right range it is good enough. I have dealt with many operators - and tool makers - that say if the spec is +/- .001 and the tool reads to .001 it is good enough. That doesn't even meet the rule of 10:1. So, it is far from the truth - but how do you show it?

MSA tries to show it by dividing the part variation by the GRR. However, it assumes that you have presented to the test the full range of variation from your process. You may have heard people mention that you must do this to make the GRR valid. True, it helps - but quite frankly you cannot always do that. When you are doing a GRR for a PPAP, you only have one lot. You have a world of variation out there waiting for you in the future - or you live a blessed life. Even if you have a wide presentation of variation, does 10 parts represent a process capability study? Let's keep this tool in perspective!

So, what can you do?

The GRR tool does provide clues as to what the operators will measure when they get a group of parts. It will collect that level of error - and that error is the error of interest. Rather than attempting to analyze that error to the part variation presented, you may want to analyze it to the expected part variation.

Careful now...we are leaving the MSA ndc calculation...the purists may want to close their eyes here:

Assumption 1: the process distribution is normal
Assumption 2: the 10 pc sample represent the distribution

At this point, both assumptions are weak. It would be handy to have a capability study. Now, here is one of the many chicken-and-egg dilemmas: you do not want to measure a capability study without know the gage is acceptable! But, if you did a capability study and took your 10 samples from its range, you would have a better developed GRR study. Ever think of that?

Anyway, the assumptions being the case, you would have control limits (based on X(MI)-R assuming parts were in order) of 2.4607 and 2.4629. The range is .0022. Divide that by GRR of .001 and you have an NDC of 2. You need an NDC of 10 for SPC (10 distinct categories - 5 above the mean, 5 below the mean) to be of any value. Otherwise, you will not pick up runs, etc., within the expected performance band of the control limits.

Not doing SPC? Lets look at the specification. You have .020 tolerance, divided by .001 GRR gives you an NDC of 20. Not bad there. So, the gage will read issues within the tolerance, but is not acceptable for control purposes.

That would be my call.

 

[sowmya]

I have done R&R for a connector received for incoming inspection. We could find the parts only in the upper tolerance and there is no much differentiation in readings. I have searched the entire lot and could not find the part with different readings for study. What to do?. NDC is 3 only

 

[Miner]

Your post states that you are using this gage for receiving inspection. The appropriate metric for this is the P/T Ratio (% Tolerance), not ndc.

ndc, as provided by virtually all software packages is intended as a metric to assess suitability for process control. bobdoering recommends a modified version of ndc where the the process spread in the equation is replaced by the tolerance equivalent. This is certainly valid, but you would have to manually calculate this, and you will have to note and explain the difference between this and the standard use of ndc.

 

[bobdoering]

As far as what should be a "good" value for NDC, if you are doing SPC, the value should be at least 10 - calculated using the control limits, not PV. You want a resolution of 10 increments that you can trust - 5 above and 5 below the mean (or 10 across the region if you are doing non-normal charting). If you had an NDC of 5 calculated to the spec, you would be able to trust less than 2 increments above the mean and 2 increments below the mean - which is useless in my estimation.

As I was rooting around the AIAG MSA 3rd edition, there was an interesting find on page 47 (Chapter 1 Section E). They have an excellent example (Figure 6) of why you need a ndc of 10 calculated by (UCL-LCL) rather than PV (the specified NDC variable in the AIAG calculations) if you are doing SPC - but they never work through the issue with the math. I did, though - so if you are using the gage for SPC, calculate the number of discriminate categories as 1.41[(USL-LSL)/GRR], and its value should be greater than 10 for good charting.

 

[equesnel]

I have developed a Excel spreadsheet after the AIAG manual calculations, but it doesn't work for this scenario as it would require a SQRT of Negative Number?!?

However, it works for me in most cases.

I would like others to review it and add comments (if possible)!!

Thanks.
Eric

 

[mech2008]

Have a question. What theory to support NDC is greater than 5 is acceptable for automobiles industry? Why not 3, 4, 6, 7... or 10? Thanks.

 

[Bev D]

I am not aware of any theory that would support a specific NDC. IT appears to be a 'rule of thumb' at best or rather arbitrary at worst.

The first issue with stating a NDC goal is that the assumption would have to be that the product variation in the data set is actually representative of the actual process. Too often we take 'stratified samples' that either have very little of the natural variation existing in the process OR we deliberately over sample the tails and the middle to get the RANGE of variation. this will either under inflate or over inflate your product standard deviation and result in a mis-estimated NDC.

The second issue is that the useful NDC number is dependent on the gage's use (SPC or product acceptance or problem solving or design). Often a poor NDC can be mitigated by sample size or by taking multiple measurements of the same device.