I recorded this video as a companion to my previous video on HOW to do a Gage R&R study (both the setup and the calculations): ruclips.net/video/aQligHSGvMA/видео.html Please check the video description for time tags if you're only interested in a specific topic If there are any other aspects of gage R&R you'd like to learn more about - drop me a comment =)
Thank you so much for this explanation. I'm just studying in case that I go back to the industry so I need to refresh my memory with all these concepts and formulas.
great video content and super easily to understand, I watched to understand a study that will have to do and you helped me so much with your explanations
Great to hear my videos can help you; and probably others too - I’d love it if you share it with people you think would also benefit from it. Thanks for your kinds words and taking the time to write them.
Hi Tom, I have ran the data in Mini tab for 0.3745 with tolerance of +-0.0002 which is very very tight tolernace and yielded Gage % tolerance of 36% which is not in acceptable range. I am super confused on how this data could be failing. For the provided data, The lowest dimensions that is checked is .37445 (.00015 above the low). The highest dimension that is tested is .37455 (.00015 off the high). The difference between the lowest dimension checked and the highest is .0001. I know we can still accept if gage r&r is above 30% with proper rationale, what justification can be used in this case?? This is for control shaft When we break it down like this, these parts are almost perfect, but I know all this is just data driven from Minitab. Indicator Mics was the gage used to measure diameter.
Hi Karthik, that's quite a question you have there - but even with all the data you provided, to answer it I'll need more numbers 😅 Some first thoughts: did you test randomly produced parts? If so, you seem to have a nice and stable process, having a Cpk of around 2 even including gage variance. If you hand-selected parts to suit your R&R study, you've probably taken a too small variance in parts, which results in falsely increased R&R percentage. You should take random sample parts. Is the 30% R&R vs observed variance (usual way that ANOVA shows result) or vs tolerance limits (my Excel template will show you both comparisons). If it's vs observed variance, the R&R vs tolerance would be around 15-20% in your case, which is not super, but can be acceptable for you (or at least not your organisation's first priority). If it's 30% vs tolerance range (and therefore 50+% vs observed in your case), you'll really want to do something with this measurement system. Is the variance coming mainly from repeatability (between trials by 1 operator) or reproducibility (between operators)? That will give you an indication of where to look for improvement. You might also want to do a new R&R with a larger (random!) sample, say 20+ parts, 3 trials, 3 operators. (if there is little to no difference between operators in your current study, maybe do 30 parts, 3 trials and only confirm repeatability).
Tom, I really enjoyed watching all your videos. You are talented. Keep it up and thanks for the great content, it would be great if you could add examples from industry👍🏻
Thanks, will do so when I can share things (not always allowed). The e-weight example in my latest video (ruclips.net/video/pGwB4olsSWY/видео.html) is directly from practice ;)
That’s just a rounding error, the 0,4 is actually 0,444. Not the best choice for a video example perhaps, but if you check the downloadable template - that has the same example as in this video, so you can see all the calculations.
Hello, what I know is before doing R&R study, I should check the precision and accuracy, but if I don’t check it, and the R&R study results is conform, does it mean the accuracy and precision are systematically ok?
Good question - this is an important topic in MSA (measurement systems analysis). R&R will check precision for you, which is to say: will everyone get the same result when testing the same product. For a good measurement system you also need accuracy, you'll need to calibrate your measurement system to test for that. ps: that also means that you should do calibration before/next to R&R (check accuracy), but you don't have to check the precision before R&R because that's what R&R is for 🙃
+/- 2,575 standard deviations from the mean captures 99% of the normal distribution. Going from 0,5% on the low end of the distribution, past the mean, up to 99,5% on the high end is twice that: so 5,15. In other words; 5,15 standard deviations is 'the same' as 99% of all values under the bell curve. Why did they pick 99%? I don't really know, it's probably based on previous tools using 99%. Some people like to use 6 in stead of 5,15 nowadays (that's 99,7% of the distribution), but the 10-30% guidelines for allowable measurement variation is linked with the 5,15 version, so I'd continue to use 5,15. ps: you can check the numbers in Excel with the formula =NORM.INV(0,995;0;1)-NORM.INV(0,005;0;1) The reason to use this number in the first place is because you're estimating the 'full range' of variation in your process from the observed values. You've only taken a sample from your process, so the range is not the highest minus the lowest value you've measured (there's a big chance you didn't randomly get a more extreme product). So you use statistics to determine the full variation range: based on the ranges observed in the measurements, combined with how many samples you took, you estimate the standard deviation in your process - that's the number you multiply by 5,15 to get 99% of the expected range of your process.
Thank you for the nice explanation. I have a use case that I would like to assess the Repeatability and Reproducibility for, but I'm not sure if it falls into the framework of Gage R&R completely. It is a system that is supposed to measure 3D location (XYZ) of any physical point with respect to a known reference point (it is a calibrated multi-camera system + a laser distance measurement). I can ask 3 operators to measure location of 10 physical points and repeat it 3 times. However, those 10 points are in different locations and have different XYZ coordinates. So, I cannot fully map it to the "parts" in your example. I would very much appreciate your help if you have any suggestions.
That does make it more complicated. Are deviations on any one of the axes independent of each other? Because in that case, you could analyze each of the axes independently: see if your operators all see the x-position the same way, then the y-position, then the z.
@@TomMentink Thank you for your reply. I actually tired to do that assuming the deviations are independent. However, I have difficulty estimating part to part variations, as in this case "parts" will refer to X coordinates of 10 physical points, which are not comparable as a single product.
@@sarehsaeidi7961 then the mathematics become more difficult - requiring a different calculation step: calculate the ‘average position’ and the deviation from that (you know: SQRT((x’-x)^2+(y’-y)^2+(z’-z)^2)). Just know that you’ll have to calculate several averages and ranges (the same that are in the R&R calculations): average per sample per operator, per sample for all operators, per trial for all parts - and then all kinds of differences between them to make the ranges that are the input for R&R. In the end, R&R uses ranges to compare the difference between the trials by one operator; to the difference between operators; to the difference between parts.
Great question - what you will often see is that you only test repeatability for automated systems: if the operator cannot really influence the outcome, there is no use to compare the results of different operators, but you still can check to what extend the machine generates the same results for the same part tested several times. If sample preparation of the way you present it to the automated system is important, it can be useful to have several operators in your R&R study. In any case: you somehow have to be able to present the measurement system with the same parts a couple of times - with a stone-alone system that is no problem, but automatic systems are often inline, so you'll have to trick the whole line a bit by inserting your 10-15 test parts into the line somewhere before the measurement device and taking them out before the process changes anything to them. The general R&R question remains the same: does this measurement device add more or less than 10% to the total observed variance? And also automated systems can generate a lot of variance if they're not performing properly.
@@TomMentink Thanks Tom! What if I have , let's say, three different (but identical) automated measurement systems. Can I treat each of the three different systems as three operators?
@@donnymac575 yes, certainly - that is the other logical reason to have a potential problem with reproducibility, so do test that if you have multiple measurement systems that should do the same thing. Be sure to send the same parts over all those 3 lines in this case (as you would with 3 operators).
Brushing up on my knowledge as it's been a few years since my greenbelt. Thanks for uploading!
Happy to hear you like my videos.
I recorded this video as a companion to my previous video on HOW to do a Gage R&R study (both the setup and the calculations): ruclips.net/video/aQligHSGvMA/видео.html
Please check the video description for time tags if you're only interested in a specific topic
If there are any other aspects of gage R&R you'd like to learn more about - drop me a comment =)
Very useful demonstration in a practical way. Thanks Tom!
Thanks, glad you liked it.
Thank you so much for this explanation. I'm just studying in case that I go back to the industry so I need to refresh my memory with all these concepts and formulas.
Glad to hear you liked it - and success getting back to the industry :)
Great video, in college class and was needing more explanation.
Great to hear that this helped you understand.
great video content and super easily to understand, I watched to understand a study that will have to do and you helped me so much with your explanations
Great to hear my videos can help you; and probably others too - I’d love it if you share it with people you think would also benefit from it.
Thanks for your kinds words and taking the time to write them.
Hi Tom,
I have ran the data in Mini tab for 0.3745 with tolerance of +-0.0002 which is very very tight tolernace and yielded Gage % tolerance of 36% which is not in acceptable range. I am super confused on how this data could be failing.
For the provided data, The lowest dimensions that is checked is .37445 (.00015 above the low). The highest dimension that is tested is .37455 (.00015 off the high). The difference between the lowest dimension checked and the highest is .0001.
I know we can still accept if gage r&r is above 30% with proper rationale, what justification can be used in this case?? This is for control shaft
When we break it down like this, these parts are almost perfect, but I know all this is just data driven from Minitab.
Indicator Mics was the gage used to measure diameter.
Hi Karthik, that's quite a question you have there - but even with all the data you provided, to answer it I'll need more numbers 😅
Some first thoughts: did you test randomly produced parts? If so, you seem to have a nice and stable process, having a Cpk of around 2 even including gage variance. If you hand-selected parts to suit your R&R study, you've probably taken a too small variance in parts, which results in falsely increased R&R percentage. You should take random sample parts.
Is the 30% R&R vs observed variance (usual way that ANOVA shows result) or vs tolerance limits (my Excel template will show you both comparisons). If it's vs observed variance, the R&R vs tolerance would be around 15-20% in your case, which is not super, but can be acceptable for you (or at least not your organisation's first priority). If it's 30% vs tolerance range (and therefore 50+% vs observed in your case), you'll really want to do something with this measurement system.
Is the variance coming mainly from repeatability (between trials by 1 operator) or reproducibility (between operators)? That will give you an indication of where to look for improvement.
You might also want to do a new R&R with a larger (random!) sample, say 20+ parts, 3 trials, 3 operators. (if there is little to no difference between operators in your current study, maybe do 30 parts, 3 trials and only confirm repeatability).
Tom, I really enjoyed watching all your videos.
You are talented.
Keep it up and thanks for the great content, it would be great if you could add examples from industry👍🏻
Thanks, will do so when I can share things (not always allowed). The e-weight example in my latest video (ruclips.net/video/pGwB4olsSWY/видео.html) is directly from practice ;)
Hey Tom,
Sorry for silly question. referring to 4:18. I tried to divide sm/st*100 =0.16/0.4*100= 40%. How did it become 36%
That’s just a rounding error, the 0,4 is actually 0,444.
Not the best choice for a video example perhaps, but if you check the downloadable template - that has the same example as in this video, so you can see all the calculations.
Hello, what I know is before doing R&R study, I should check the precision and accuracy, but if I don’t check it, and the R&R study results is conform, does it mean the accuracy and precision are systematically ok?
Good question - this is an important topic in MSA (measurement systems analysis). R&R will check precision for you, which is to say: will everyone get the same result when testing the same product.
For a good measurement system you also need accuracy, you'll need to calibrate your measurement system to test for that.
ps: that also means that you should do calibration before/next to R&R (check accuracy), but you don't have to check the precision before R&R because that's what R&R is for 🙃
I've rewatched the video, and I still can't understand where did 5,15 value come from. Could you explain?
+/- 2,575 standard deviations from the mean captures 99% of the normal distribution. Going from 0,5% on the low end of the distribution, past the mean, up to 99,5% on the high end is twice that: so 5,15. In other words; 5,15 standard deviations is 'the same' as 99% of all values under the bell curve.
Why did they pick 99%? I don't really know, it's probably based on previous tools using 99%. Some people like to use 6 in stead of 5,15 nowadays (that's 99,7% of the distribution), but the 10-30% guidelines for allowable measurement variation is linked with the 5,15 version, so I'd continue to use 5,15.
ps: you can check the numbers in Excel with the formula =NORM.INV(0,995;0;1)-NORM.INV(0,005;0;1)
The reason to use this number in the first place is because you're estimating the 'full range' of variation in your process from the observed values. You've only taken a sample from your process, so the range is not the highest minus the lowest value you've measured (there's a big chance you didn't randomly get a more extreme product). So you use statistics to determine the full variation range: based on the ranges observed in the measurements, combined with how many samples you took, you estimate the standard deviation in your process - that's the number you multiply by 5,15 to get 99% of the expected range of your process.
@@TomMentink Wow... I didn't expect you to reply.. Thank you so much!!! 🙌🙌
Thank you for the nice explanation. I have a use case that I would like to assess the Repeatability and Reproducibility for, but I'm not sure if it falls into the framework of Gage R&R completely. It is a system that is supposed to measure 3D location (XYZ) of any physical point with respect to a known reference point (it is a calibrated multi-camera system + a laser distance measurement). I can ask 3 operators to measure location of 10 physical points and repeat it 3 times. However, those 10 points are in different locations and have different XYZ coordinates. So, I cannot fully map it to the "parts" in your example. I would very much appreciate your help if you have any suggestions.
That does make it more complicated.
Are deviations on any one of the axes independent of each other? Because in that case, you could analyze each of the axes independently: see if your operators all see the x-position the same way, then the y-position, then the z.
@@TomMentink Thank you for your reply. I actually tired to do that assuming the deviations are independent. However, I have difficulty estimating part to part variations, as in this case "parts" will refer to X coordinates of 10 physical points, which are not comparable as a single product.
@@sarehsaeidi7961 then the mathematics become more difficult - requiring a different calculation step: calculate the ‘average position’ and the deviation from that (you know: SQRT((x’-x)^2+(y’-y)^2+(z’-z)^2)). Just know that you’ll have to calculate several averages and ranges (the same that are in the R&R calculations): average per sample per operator, per sample for all operators, per trial for all parts - and then all kinds of differences between them to make the ranges that are the input for R&R.
In the end, R&R uses ranges to compare the difference between the trials by one operator; to the difference between operators; to the difference between parts.
How do you perform Gage r&r for automated systems, with no operators?
Great question - what you will often see is that you only test repeatability for automated systems: if the operator cannot really influence the outcome, there is no use to compare the results of different operators, but you still can check to what extend the machine generates the same results for the same part tested several times.
If sample preparation of the way you present it to the automated system is important, it can be useful to have several operators in your R&R study.
In any case: you somehow have to be able to present the measurement system with the same parts a couple of times - with a stone-alone system that is no problem, but automatic systems are often inline, so you'll have to trick the whole line a bit by inserting your 10-15 test parts into the line somewhere before the measurement device and taking them out before the process changes anything to them.
The general R&R question remains the same: does this measurement device add more or less than 10% to the total observed variance? And also automated systems can generate a lot of variance if they're not performing properly.
@@TomMentink Thanks Tom! What if I have , let's say, three different (but identical) automated measurement systems. Can I treat each of the three different systems as three operators?
@@donnymac575 yes, certainly - that is the other logical reason to have a potential problem with reproducibility, so do test that if you have multiple measurement systems that should do the same thing. Be sure to send the same parts over all those 3 lines in this case (as you would with 3 operators).
@@TomMentink Great! Thank you for the guidance.