Advanced MSA of Continuous Data Part 4: How to Sample Parts

Another fine article from Minitab on Gauging Gage Part 3: How to Sample Parts.

This article addresses a personal pet peeve of mine. I have seen many posts recommending these same bad practices. As with the three previous posts, Minitab has used the power of Monte Carlo simulation to illustrate why these practices are so bad.

Always a chicken and the egg problem. The conversation seems to start after there is hsitory with the part, but that is not likely the case. Usually, you have a new part, new process...and you need a gage. Make a few parts, but any way to estimate the variation you will see over the life of the process? Oh, and when the designer measured parts to verify his design....what gage did he use? Would it "pass gage R&R"? If not, how much "better" of a gage do you need to make a part than the designer used to validate the design? Aspirin, anyone??

Hello,
thank you for sharing your knowledge
i have a question please: is the RandR study enough to accept measurment instruments in automotive mechanical spare parts industry?

Is the answer to sample randomly or exact?

The MSA calculation assumes the sample spread you submit represents the variation of the process you expect to see over time. You need to look at your sample variation and assure yourself that it is a reasonable representation. You can avoid some of that stress by using a historical variation of the process if it is available. The highly recommended historical variation treatment is found on page 121 of the 4th edition MSA. If you do not have historical variation, use a Ppk target value of 1.33 - since your gage should be adequate over the range of a capable process. By doing that, your actual sample spread needs to be as wide as possible, so that the response can "see" the difference between the samples. Samples that are too close to identical will fail any gage R&R. So, long story short, random may not be good enough.

If a process is producing products with very little variation to be able to naturally produce samples spanning the allowable variation and is in no danger of drifting towards the upper or lower limits, but the measurement system has more than enough resolution to detect products near, or over those limits and the study fails due to not enough part to part variation inflating the measurement

Is it allowable to manipulate or create samples than span the range of allowable variation?

Include samples above and below the limits to show the system can detect out of spec parts?

Accept the "failed" study and justify why the measurement system is still accurate enough even though the R&R might be above the recommended criteria?

Or should the equipment be replaced for something more accurate than needed?

Thanks for the reply.

It depends on the application of the measurement device. If you are strictly using it for inspection purposes, it does not matter how you select the parts as the part variation is not part of the %Tolerance calculation and %Study Variation does not apply.

However, if you are using the device for SPC, part selection is vital. I agree with @bobdoering on the following:

bobdoering said:

The MSA calculation assumes the sample spread you submit represents the variation of the process you expect to see over time. You need to look at your sample variation and assure yourself that it is a reasonable representation. You can avoid some of that stress by using a historical variation of the process if it is available. The highly recommended historical variation treatment is found on page 121 of the 4th edition MSA.

Click to expand...

If your process is so highly capable such that the part variation is too small to be acceptable by the standard R&R study, the answer is not to play games with the part selection, but to read up on the writings of Dr. Wheeler on Class 1 - 4 monitors and which extended SPC rules to use with each class. These may be found on the SPC Press and Quality Digest websites. Dr. Wheeler has done extensive research on "Honest Gauge Studies" and the true ability of a measurement device to detect shifts in a process behavior (control) chart. If applicable, I also recommend @bobdoering 's posts on precision machining and the special application of SPC (i.e., Hi/Lo Range chart) to that niche.

My point is if your sample is too tight, and it does not represent the variation of the process, it will fail the gage R&R when the gage may actually be fine with a more representative sample.

If your process really is tight all the time, and you are using the gage to determine the process variation (and not product acceptance), then you will need a much higher resolution to be able to watch the process vary and get the feedback needed to see that the process is not acting as it was and may have a special cause. At least that is the point of using process variation in the assessment. However, it ignores that the process may have plenty of room to vary, and is expected to do so especially as process components wear in, and may have no significant impact on the output. With no history (as in the beginning of a process, with new gages) estimating the overall process as a capable Ppk is a reasonable place to start.

adding to the great posts above, when conducting an MSA G R&R or a production demonstration run/run at rate some well meaning folks consciously or unconsciously will "stack the deck" aka, "selecting select parts, or installing especially proficient operators or inspectors hoping to influence the outcome of these efforts in their favor.