Sample Size for GRR - Is it possible?

[FirstTimeBuild]

Our customer wants to validate a new single CMM fixture through MSA/GRR. Is it possible to perform a 5 parts 2 trials , 2 appraisers study using the ANOVA method? Is the sample size too small? I cannot find a good resource to definitively recommend a higher sample size > 5. The population would be 10 and the 5 parts sampled from this population. The part has over 120 features and the tolerances range from .0001" to .004" I would appreciate if someone could guide me as to the minimum sample size/population for the ANOVA method. Thank you.

 

[Miner]

Ideally, the total number of measurements should be in the 90-100 range. You planned 5*2*2 = 20 which is inadequate. The traditional 10*3*3 = 90 is adequate. The most important aspect is the number of parts and the number of trials. 2 Appraisers is sufficient.

Note: 5 parts is better information than no information. The problem is the size of the confidence intervals around that information. The smaller the sample size, the wider those intervals.

 

[Bev D]

The issue here is that an R&R estimates the measurement error - which is the measurement standard deviation. Standard deviations are best estimated with many small subgroups. The subgroup is the repeated measures of each part. In estimating the SD the increase from 2 repeats to 3 is negligible. You are better off increasing the number of parts. I typically recommend 30 parts but have g one to 20 or 15 when the measurement is time consuming and/expensive. If really expensive I’ve gone down to 5 parts but this has mostly been for bio;o GI alas that come in huge batches and are only produced a few times a year.

You might find the articles by Donald Wheeler helpful at SPC.com.

 

[John Predmore]

@FirstTimeBuild mentioned the part has 120 features. I hope you have chosen one or two critical dimensions, probably the dimensions furthest from the datums, to amplify positional error of the CMM. Measurement of these few dimensions is representative of the entire capability of a largely automatic measuring machine. Then you can focus on a few dimensions of many more parts.

 

[Ron Rompen]

Just to advance an alternative, have you considered doing a Type 1 gauge study? If the measurement process (including the locating of the datum features) is completely automated, then (in theory at least) the operator influence on the measurement system is negligible.
2 or 3 parts, measured 50 times each will give you more information that 5 parts 2 times or 3 times.