Minimum number of parts to do an attribute gage R&R

[Jparfer]

Hi everyone!

Im going to do an attribute gage R&R but i have 2 questions:

1- How many parts has to be included in the studio as a minimun?
2- Must be the bad parts close to be good parts?

Thanks

 

[Curtis317]

If you are in the Auto industry and must follow the MSA, that will tell you. I think it is 50 pcs. However I have done Attribute GRR with 20 pcs. How close they are to good parts depends on the standard production parts variation . But I would make them as close to good as possible but still be bad. That way you are sure your gage can do what is intended. Which is to not allow bad parts to go to the customer.

 

[Jparfer]

If you are in the Auto industry and must follow the MSA, that will tell you. I think it is 50 pcs. However I have done Attribute GRR with 20 pcs. How close they are to good parts depends on the standard production parts variation . But I would make them as close to good as possible but still be bad. That way you are sure your gage can do what is intended. Which is to not allow bad parts to go to the customer.

In the MSA they do an example with 50 parts, but they dont said that u have to do with 50...

 

[Bev D]

If you are only tryign to 'check the box' for your Customer, then you probably can do any size you want as long as they accept your rationaliztion. Such as the parts areproduced in small volumes and defects are rare. I doubt that you will get away with "it takes too long, or 'it's too hard'

If you actually are trying to learn something about your system, I recommend 50 or more parts. I also recommend that the distribution of marginal and bad parts is equal to the current rate. to truly test your system. See my paper on MSA for details.

remember that it is necessary to have larger sample sizes with categorical data than with continuous data and teh lower the defect rate, the larger the sample size should be.

 

[Jparfer]

If you are only tryign to 'check the box' for your Customer, then you probably can do any size you want as long as they accept your rationaliztion. Such as the parts areproduced in small volumes and defects are rare. I doubt that you will get away with "it takes too long, or 'it's too hard'

If you actually are trying to learn something about your system, I recommend 50 or more parts. I also recommend that the distribution of marginal and bad parts is equal to the current rate. to truly test your system. See my paper on MSA for details.

remember that it is necessary to have larger sample sizes with categorical data than with continuous data and teh lower the defect rate, the larger the sample size should be.

I have read your Paper, but i cant read where is the minimum parts for an attribute R&R...

 

[Statistical Steven]

Hi everyone!

Im going to do an attribute gage R&R but i have 2 questions:

1- How many parts has to be included in the studio as a minimun?
2- Must be the bad parts close to be good parts?

Thanks

Depends on your acceptance criteria. If you need 95% confidence with 90% reliability for concordance, then 30 parts minimum if you observe 100% concordance (60 for 95/95 and 300 for 95/99).

 

[Bev D]

I have read your Paper, but i cant read where is the minimum parts for an attribute R&R...

That is because the minimum number of parts depends on the situation.

Statistical Steven is correct. there is NO universal minimum size.
As described in my paper, it all depends on what you are trying to do or understand and what defect rate you have and the distribution of that defect rate. There are statistical considerations as Steven pointed and non-statistical considerations.

An additional statistical consideration, as Steven alluded to, is the amount of disagreement you get. This is one of those iterative situations, you need the data to tell you how much data you will eventually need...the whole idea is to understand the limitations of your system, then improve it and test again. If you get 100% agreement there is no problem with a reasonably large enough sample (say 50 parts)

A non-statistical consideration is how many marginal units are in normal production and how much time the inspector or machine has to perform the inspection. If the defect rate is fairly low, there a decent proportion of parts that are marginal (both passign and failing) and the tiem to inspect is rather short a much larger sample size is necessarry to test teh actual systme capability. It's easier to determine pass/fail on marginal units when one isn't under time pressures so you get a better score, but in real life - where the capability of the system actually matters - it's very difficult to get it right. I have encountered this situation many times and have ended up testing 400-1000 parts on the line to mimic actual conditions.


in order to give you good advice we need to understand:

• what are trying to accomplish? (e.g. 'Check the box' to submit a study to your customer or do you want to learn about your system and improve it if needed)
• What is your measurement system?
• What type(s) of defect(s) are you trying to detect?
• what is the defect rate?
• What is the distribution of marginal units?
• What is the inspection environment? (e.g. quiet with plenty of time or inline high speed high volume?)

If you are really are just trying to 'check the box' (and many people are) then my answer would be to do 50 parts, 25 bad and 25 good. but others will have different answers because there is no single minimum size.

 

[Jparfer]

That is because the minimum number of parts depends on the situation.

Statistical Steven is correct. there is NO universal minimum size.
As described in my paper, it all depends on what you are trying to do or understand and what defect rate you have and the distribution of that defect rate. There are statistical considerations as Steven pointed and non-statistical considerations.

An additional statistical consideration, as Steven alluded to, is the amount of disagreement you get. This is one of those iterative situations, you need the data to tell you how much data you will eventually need...the whole idea is to understand the limitations of your system, then improve it and test again. If you get 100% agreement there is no problem with a reasonably large enough sample (say 50 parts)

A non-statistical consideration is how many marginal units are in normal production and how much time the inspector or machine has to perform the inspection. If the defect rate is fairly low, there a decent proportion of parts that are marginal (both passign and failing) and the tiem to inspect is rather short a much larger sample size is necessarry to test teh actual systme capability. It's easier to determine pass/fail on marginal units when one isn't under time pressures so you get a better score, but in real life - where the capability of the system actually matters - it's very difficult to get it right. I have encountered this situation many times and have ended up testing 400-1000 parts on the line to mimic actual conditions.


in order to give you good advice we need to understand:

• what are trying to accomplish? (e.g. 'Check the box' to submit a study to your customer or do you want to learn about your system and improve it if needed)
• What is your measurement system?
• What type(s) of defect(s) are you trying to detect?
• what is the defect rate?
• What is the distribution of marginal units?
• What is the inspection environment? (e.g. quiet with plenty of time or inline high speed high volume?)

If you are really are just trying to 'check the box' (and many people are) then my answer would be to do 50 parts, 25 bad and 25 good. but others will have different answers because there is no single minimum size.

I really want to check if a robot is doing right decissions (pass/fail). So i pretend to do a GR&R, with 30 parts (15 pass parts and 15 not pass [someone close to pass]).

I dont understand why 50 will be the minimum to do it.

The distribution i mean is a binomial one, because is a pass/fail decission.

It?s a quite enviroment.


Thanks!

 

[chiquinha]

thanks for your answers it helped me out too !