iSixSigma

Gage RR on Paired Data

Six Sigma – iSixSigma Forums Old Forums General Gage RR on Paired Data

Viewing 36 posts - 1 through 36 (of 36 total)
  • Author
    Posts
  • #29221

    GRRR
    Participant

    I am performing a Gage R&R to determine how good our scale is for determining weight loss during a wear test that we conduct.I am performing a Gage R&R on the test items before the test and a Gage R&R on the test items after the test. I intend to add the measurement system variation seen before the test and the measurement system variation seen after the test and compare this total measurement system error to the variation seen in the weight loss of the items. Is there a better way to perform a Gage R&R on this type of paired data? Should I do Gage R&R on the difference? IS there value in doing so?
    I would appreciate any guidance that you can give me. Thanks.

    0
    #74216

    aa
    Participant

    Hi, In my view you need to consider Gage RR and Paired t-test analysis in separately – if you’re trying to understand variability of your measurement system (namely gage). Here are two points for you to consider and see how you can apply to your activities:
    (a) Gage RR – what are you trying to establish.
    Repeatability, Reproducibility and Variation within the process. 
    The measurement system, where the repeatability or reproducibility is low (say 3%) and part-to-part, reports higher then 95% it seems you have a good gage – but can be improved, I usually get 99% for part-to-part and no of distinct categories = 40 and over.
    The other feature to consider is Number of Distinct Categories = higher the better. 
    My experience is with Minitab – therefore use of above terms.
    (b) With regards to paried t-tests – if you want to comparise the improvement post Gage RR work, yes before and after results can be utlised to draw an understanding.
    If you need any more help, post your email, I shall reply with further comments. If that is need.
     

    0
    #74228

    FJ
    Participant

    Gage R&R is normal done to assess the measurement system capability on a specific metric. You are looking at three seperate metrics here: 1) initial condition, 2) final condition, 3) delta. These are three seperate metrics and the capability of the measurement system for each should be treated differently since the width of the process data is different.
    Use your hypothesis testing (of course only after you know your “ruler” would give you good data – via Gage R&R studies) to look for differences in the two populations.
    -Fred

    0
    #74379

    DrSeuss
    Participant

    Hi GRRR,
    Maybe I am not understanding clearly what you are trying to do, but I would think your measurement GR&R should determine if your method for gathering data, i.e.  measurements is adequate.  It should not matter whether your data set is before or after.  I would think if you assess your measurement capability based on what type of resolution & error tolerance you need to see process differences, then your GR&R should be acceptable for both.  Then you could follow with a Paired-t test to show if there is a statistical difference on the before and after test samples.  Note, to use a Paired-t test, the samples are measured before and then after.  If the differences are statistically significant, then what ever you did to them had an effect.  Hope this helps.

    0
    #74385

    RR Kunes
    Member

    If you had Minitab or a similar software this task would be so easy you’d never posted.
    You can do the calculations manually, however, they are a bit time consuming.
    In six sigma terms you are looking for the components of Variation (COV) of this system.
    To truly get a handle on this you need two or more inspectors utilizing the same gages on different parts. This should be a blind study to minimize bias in the responses.
    You need to know the process repeatability and reproducability and determine how much ofthe variation is due to the operator the parts or  the inspectors. Even with CMM equipment I’ve seen huge variations in the way an operator sets up a piece. This huge variaition trickles down to needless adjustment of machinery.
     

    0
    #74415

    Mike Carnell
    Participant

    G,
    I used to use paired t test to evaluate measurement systems before we found this obscure document some guy had in his filing cabinet called GR&R. The first problem with using the paired test is you have to have 2. You can’t do more and you can’t do less. You can do multiple paired tests. The rational behind not doing that is well documented.
    The other issue with the paired test is the size shift you need the test to be sensitive to. You need to make sure you have the correct sample size.
    I am assuming the process changes the product in some significant way – since you are testing the GR&R before and after. If the gage is acceptable in terms of lineraity, stability, and bias across the range you use it, then it shouldn’t be any different before and after. If it is it is more than likely an proceedural issue than a gage issue.
    As far as your question about a better way of testing paired data. The GR&R is paired data. It can also be tripled, quadrupled, etc. That is the leverage of the ANOVA type analysis – you can go over 2.
    You asked for opinions I will give you mine which may or may not mean anything since I don’t really know the process. I would take a set of samples and use them to run the GR&R study before the test, process the material and rerun the study with the same parts. Use a the same operators before and after. If the reproducibility number blows up on you it could be one of two things. The gage R&R is different after the processing or the product is different. You truncate the data from after and analyze it by itself (not using the pre-process data) and you have the R&R on the post process. If the R&R hasn’t shifted then I would consider spending some time with the shift in material.
    I would separate the two issues. The gage issue and the process issue. You can’t determine anything about the process until you understand the measurement system capability.
    Good luck.

    0
    #74521

    Gabriel
    Participant

    Hi, GRRR
    I had a problem alike. We grease parts, and we want to know the mass of grease added. The measuring method is (using an electronic scale): Take one part, put in on the scale, press the “zero” button, take the part and grease it, put the part on the scale and read the weight of grease.
    The r&R study seems easy: Just make three operators make the measurement three times on 10 parts. The problem? The measurement can not be repeated, because “greasing” was part of the method. If you clean and re-grease a part, you are including the process variation in your measurement system variation.
    Then I thought to separate the two things: Weighting and greasing. Then you can perform an r&R for the individual weighting (first or second, probably the gage variation will be the same). It is possible, but I didn’t like the approach because: 1) The final result is “delta”, and we only want to know “delta”, never mind wich is the weight before or after (by the way, “weight before”, “weight after” and “delta” are not three different things, as someone posted, because they are mathematically linked by “delta”=W2-W1). So if you calculate the r&R of one weighting, you must take in account that the gage variation will be pressent twice in your measurement and therefore the combied variaton of the two weightings will be Scomb^2=S1^2+S2^2 (assuming that S1=S2, Scomb^2=2xS1^2). 2) This method does not take in account the “zeroing” step of the measuring process, which may have its own variation.
    The solution: We used 10 pieces of wire, each paired with one part, to “simulate” the grease. The differnt pieces of wire covered the expected range of the grease weight. So the measurement method for the r&R study was slightly changed as follows: Take one part, put in on the scale, press the “zero” button, take the part and put the paired wire, put the part on the scale and read the weight of the wire. Unlike the grease, you can put the same wire in the same piece as many times as you want, and belive me the scale will never know if it is grease or a wire. It’s easy, it’s fast and it is a better simulation of the full measurement method.
    In your particular case, you will use 10 already worn out parts and add some “thing” (like a piece wire) to simulate the weight of the lost material. You will weight first the part together with the “thing”, and then remove the “thing” and weight again. Remember you must keep the same “thing” with the same part all over the study (number each part and each “thing”). If your measurement requieres a VERY high accuracy, the density of the “thing” and of the worn out material should be the same, or calculate a density compensation (remember Arquimedes’ law).
    Hope this help. Tell me whether you liked the idea or not.

    0
    #74524

    Mike Carnell
    Participant

    Gabriel.
    I must have missed something. The a scale measures weight. It really doesn’t care if it is grease or metal or a combination (unless of course it hangs off the scale). Why would you need to simultate the grease with wire?
    Did you run the R&R without the wire and see if the result was any different than it was with the wire?

    0
    #74551

    Withheld
    Member

    My $.02: The Gage R&R checks the reliability and repeatability of the scale – period. It should yield the same results before your wear test as it does after your wear test. Let it stand on its own.

    0
    #74553

    Withheld
    Member

    I meant to add a comment about how I use Gage R&R. Since the study results in a % Tol Consumed (PTC), I reduce the spec limits by that %. My internal capability calculations (Cp / CpK, etc), therefore, occur *after* I have factored the PTC into the spec limits.
    I’ve not heard of anyone else doing this. Does anyone share my view of the merits? I’ve considered copyrighting the concept as the “Withheld Capability Calculation (WCC).” What do you think? ;-)

    0
    #74556

    Mike Carnell
    Participant

    Withheld,
    I use the same method as well but I got it from John Lupienski (Motorola, Elma, New York.
    Actually it should be used to do “Guard banding” as well. If you P/T ratio is 10%  (just using a number – ther isn’t anything special about 10%) then you should question any part that tests within 10% of the spec limit. If it is in spec and within the guard band it could in actuality be out of spec due to the measurement system. If it is outside the SL and within 10% then it could be good as well.
    Think about the rollup: before you even go into production take the spec width and divide it by 12 – that is your max Std Dev (if you don’t have resolution out this far and you intend to get to 6S you will never know when you get there). for a 6S project, the measurement system needs that level of resolution (which means one more decimal place), the calibration capability has to be better and the standard for the Cal Lab has to be better than that. It rolls up quickly.
    Dr. Shree Nanguneri and I presented a paper on this at the Rochester Quality Conference in 2001. We are working on a publication as well. Are you intersted in participating? Not that many people get this part of it.
     

    0
    #74559

    Withheld
    Member

    Thanks, Mike. My copyright comment was tongue-in-cheek. I can see it now – Mike Carnell, Dr. Nanguneri and Withheld! I started using the method we’re discussing back in 1994, largely because building it into my capability calculation forced me to perform the R&R. Another motivating factor was my belief that the great information obtained by the R&R seemed to merit use beyond a simple “good/bad” decision.
    My business is CNC machining and I have found the practice quite useful. I’m very sure it will be represented well by you and your team.
     

    0
    #74566

    Mike Carnell
    Participant

    Withheld,
    I don’t think it should be a tongue in cheek comment. It really isn’t well understood throughout industry and it is important in a variety of areas.
    It actually had a chance of getting some attention until things in the GR&R arena moved to the ANOVA method and they present two catagories: % Total Variation and %Study Variation. If you want percent tolerance you have to ask for it and most don’t. I unfortunately got into a situation and the percent tolerance got lost in the material because the academics didn’t understand the application. Now that I am back on my own we do it again.
    I think the roll up needs to be a part of DFSS. I don’t think the average designer understands the implications of placing a tolerance on a drawing beyond the impact of making parts.

    0
    #74567

    Gabriel
    Participant

    Mike:
    I can make the measurement with the grease instead of making it with the wire. In fact, that is what we do in the process (we measure mass of grease, remember?). The problem is that I can not repeat the measurement with the same part and the same greas because the measurement involves weighting the part before and after greasing. To assess r&R you are supposed to filter out part to part variation. That is made makeing the measurement ot the same part (for example, three operators “repeat” three measurement each on each part, so you measure 9 times each part, for a total of 10 parts) How do I measure the same grease over and over? Ok, I can measure the part with grease ten times, but is not the weight of the part + grease what I measure in the process. The full measurement process involves two weightings (before and after) and one zeroing operation in the middle. So the “scale r&R” (to call it somehow), appears twice, and you still have to add the variation of the zeroing operation. Remember to distinguish between “measurement instrument” and “measurement system”, in which the instrument is just one part, and operator, method, enviroment, etc are other parts of the same system.
    The answer to your second question in no, just because I don’t know how to make the r&R without the wire, using the grease. Unless I make an r&R of the scale alone.
    Hope now it is clearer.
    Gabriel

    0
    #74576

    Gabriel
    Participant

    Mike and Withheld:
    1) r&R has nothimg to do with “reliability” (as posted by Withheld). It is repeatability and reproducibility.
    2) Measurement: The “assignment of numbers to material things” (MSA manual – AIAG). The “material thing” here is not the weight of the part before the test, not the weight of the part after the test. It is the weight lost during the test. Yor parts weights, lets say, 10.00 Kg before the test and 9.96 Kg after the test, and the specification is (imagine) “the mass lost shal not be grater than 0.10 Kg”. And you can not wight that on the scale. You can make an r&R of the parts before the test ad obtain a value: When measuring about 10 Kg the r&R is 0.02 Kg. Is this value acceptable? Which is the specification you use to comare the r&R value with and obtain a %? Ok, the specification is “0 to 0.10 Kg” so an r&R of 0.02Kg is 20%, right? Wrong!. In this case, the scale is tot a direct reading gage. You want to measure “lost weight”. Again, you can not weight that because it is “lost”. But you have a system to measure that. The gage (scale, in this case) is just one part of the system. Measurement system: Colection of operations, procedures, gages and other equipment, software, and personnel used to assign a number to the characteristic being measured; the complete process used to obtain measurements (MSA manual – AIAG). It is usually called “gage r&R”, but if it involved only the gage, why to use different operators? In this context, 0.02 can be the scale r&R, but it is not the “measurement system” r&R which involves: a) Weight the part before the test, write down the value (with some rounding to the nearest mark invoved?), test the part, weight again (another rounding?), take the difference and now you have your measurement done. This can not be repeated on the same “material thing” = “lost weight”, so repeatability can not be assessed. This is the problem our friend has.
    3) Mike: Be careful with the resolution. You can have a good resolution with a poor accuracy. A gage showing a lot of digits is not very usefull if the measurement system (where this gage is just one part) has a measurement uncertainty of a lot of digits too. For example this happened to me: you can have a comparator (dial gage) with a resolution of 0.001mm. The accuracy of the instrument itself (according to the manufacturer specifications was 0.005mm meaning that the “true” value will be always somewhere 0.005 from the reading). Plus, combined with the stifness of the stand, the stifness of the part to be measured, the temerature effects, and the operator induced variation, gave us a total uncertainty of about 0.01mm. Not very good for a tolerance of 0.01mm, even when you have a 10% resolution!. Even worse if you try to control via SPC a 6 sigma process, where 99.97 of the parts are in a range of 0.005mm!. Of course the r&R of the system was awful! Usually the r&R % for a process that is charted is defined as r&R/(Tol/Cp), instead of r&R/Tol.
    Gabriel.

    0
    #74577

    Mike Carnell
    Participant

    Gabriel,
    The point I was making was the scale doesn’t care what it is measuring. AS long as your parts are not creating some issue by the way they are positioned (hanging off the scale) you don’t have to use actual parts at all. The scale only cares about the mass and the study only cares about repeating a measurement, within and between people.
    As long as I use something (even mass standards) I can assess R&R acurately.
    If I believe that the grease is changing in some way I can create grease samples and do a study there but I need to know the scale will repeat before I do that.
    AS far as greasing parts that is process variation and doesn’t have anything to do with the R&R on the scale.
    Why wouldn’t I do R&R with Mass standards (you do can spread them across the range of interest). This will also let me check linearity and bias across the range of use. Now I understand what the scale will do and I can separate it from anything else the process is inducing because I understand 4 out of 5 of the characteristics of the scale all in one study. Stability will need a time factor (the fifth characteristic).

    0
    #74582

    Withheld
    Member

    You’re correct, Gabriel. I tripped over my “R” words. Mea Culpa. That is one of the better arguments for going by, “Withheld.” I post on the fly for the most part. While I hold the substance of my comments near and dear, I would not want to post typos and related tongue-stomping things under my name.
    Regarding comment #2, I tend to disagree with you. Your deep thinking notwithstanding, it seems to me you’ve complicated the issue. You are using the same gage (the scale) to measure before and after. The repeatability and reproducibility does not change simply because you are measuring (weighing) the same part twice. If you’re using two scales, of course, you’d need two studies.

    0
    #74597

    Gabriel
    Participant

    A final word: Measuring is not reading an instrument (not allways, at least). Let’s do this mental puzzle.
    We are manufacturer of sheets of paper (in coils). Our main customer doesn’t care about the size or even the thickness of the material, but specific gravity “SG” (or density) (weight per unit of volume in Kg/m3) is a key characteristic for them. They have a specification for this: 1 +/- 0.2 Kg/m3. How do we measure specific gravity? We cut a triangle, and we measure the thickness “t” with a dial gage (r&R=0.02mm), the three sides L1, L2 and L3 with a ruler (r&R=1mm) and the weight “w” with a scale (r&R=0.05grams). The triangle is, nominally, 100cm x 100cm x 141,42cm, but anyway we always measure the size because the cutting is not very repetitive. And the thickness is nominally 0.2mm, but we measure it each time. Finally, after converting everything to Kg, and meters we apply the following formula:
    SG = weight / volume, where weight = w and volume = t x surface. Finally, for the triangle L1, L2, L3:
    surface = 0.5 x L1 x (L2^2 – L1^2 / (1+L3/L2)^2)^0.5.
    Don’t worry. The formula is loaded in a spreadsheet, so we only have to enter the values for t, L1, L2, L3 and w and Excel give us the specific gravity.
    Our customer requires an r&R of no more 30% of the tolerance for the specific gravity measurement. Do we meet the requirement?

    0
    #74599

    Mike Carnell
    Participant

    Gabriel,
    It looks to me as if you have a process confused with an R&R. It does not matter what you do before or after you place a part on a scale. The scale will weigh whatever is sitting on it period. It does not care about the sample prep, etc. It weighs things. If you want to check R&R you weigh them multiple times. If you are doing destructive testing you can run a nested design but I can’t imagine a scale that destroys a sample.
    I understand the difference between resolution and accuracy. A lack of resolution can drive a lack of accuracy.

    0
    #74601

    Withheld
    Member

    Gabriel said, “Our customer requires an r&R of no more 30% of the tolerance for the specific gravity measurement. Do we meet the requirement?”
    For the *measurement* or for the *instrument* for measuring? My understanding of the gage R&R process differs from yours, Gabriel. I take a measuring instrument (whatever it is), two or three operators (whomever they are) and a few sample widgets within the measuring range of the instrument. The operators each take a series of measurements, the math (built into a speadsheet) is done and the result is returned. Except for factoring the P.T.C. into the spec limits, that’s the Gage R&R for me. All the talk about methods and specs and other variables (to me, anyway) seems beyond the scope of the gage R&R.
    I have long since delegated the performance of R&R’s. I admit it’s possible they have changed and I am behind the times. If this is the case, please tell me where I’ve gone wrong.

    0
    #74606

    Mike Carnell
    Participant

    Withheld,
    I have a similar interpretation as you do. When someone uses the term GR&R it is a gage characteristic. If they use the term MSA (Measurement System Analysis) it includes 5 types of analysis – repeatability & reproducibility (precision) and stability, linearity, & bias (accuracy).
    If one of those catagories is unacceptable it may lead back to the sample preparation – I treat that as a process unto itsself. There is a difference between knowing and understanding what the gage is capable of and if you are screwing up a good gage with a poor sample preparation process.
    If I have to report on R&R for specific gravity then before I complicate it with sample prep (particularly if I already know a part of that process isn’t working well) I had better know how well it does at the measurement. It is basic compartmentalization. Just because a customer specifies something be reported in aparticular fashion does not mean I need to analize it that way. Ultimately I have to understand it, understand the customers requirement and if the customer isn’t getting a measurement that really serves their purpose explain to them what they need to change. If they are correct then I need to be concerned with compliance. It is part of the differentiation between “Total Customer Satisfaction” and “Total Customer Success.”

    0
    #74608

    GRRR
    Participant

    Very good discussion! Thank you all, especially Gabriel, Mike and Withheld. Gabriel, my thought process was pretty close to yours. I want to be convinced one way or the other.
    The problem I discussed was about a wear test of fabric. There are “before” and “after” measurements and the difference is just one of the measures of wear.
    Since the weight loss when I measure the fabric after it is worn out, can be extremely small,  I want to know for sure, if the variation in the difference (Wear, and not absolute mass) that I see, is  because of the process (part to part difference) variation or because of my Gauge.
    ‘Coz, if my Gauge is giving me an illusion, I am in trouble.
    With some cautious skepticism, I believe, it is not a typical Gauge R&R problem, rather a “paired characteristic” measurement situation. The Gauge might be ok for individual measurements-either before or after. But the result of my process (which is what I want to measure, using this gauge) is wear and not weight in absolute terms. My primary interest is not to know whether the gauge is suitable for measuring the weight of a fabric piece.
    And I am not convinced yet, that the % GRR variation would not eat up some of the “accuracy” of our wear measurement…..(loosely said, imagine it as “part-to-part variation” of the “worn material”, assuming it were possible for my team to collect and meaure the products of wear)
    I agree with what Mike and Withheld has mentioned about MSA, but want you all to just give some critical thought, to what I have just mentioned.
    Either way, I deeply appreciate your participation in the discussion.
    GRR
     

    0
    #74640

    Mike Carnell
    Participant

    gRrR,
    Sorry about losing track of your  question.
    I don ‘t believe you can trust the results of the t test for weight loss until you do an R&R particularly since you are trying to detect a small shift. If you think about the paired test poor repeatability will kill you.
    The weight loss does seem to lend itsself to the paired t test. That was why in one of my earlier answers I said to separate the two issues.
    Good luck.

    0
    #74644

    Withheld
    Member

    I’ve learned that the only way to keep a conversation on the first page is to reply to the first message. This comment is intended to reply to grrr’s latest message.
    Regarding “convinced one way or the other,” I can best summarize my opinion by saying that the Gage R&R should stand alone and serve the sole purpose of evaluating the scale. Subsequent tests such as hypothesis testing are looking for something other than what you learned from the gage R&R and have nothing to to with that aspect of your work.
    For a generic example, suppose I have a gasoline budget and have evaluated the various influeneces on this expense. The efficiency of my car (tire pressure, tune up, etc) contribute to the expense just like the miles traveled and cost of gas do. Tuning my car, just like the gage R&R is the right thing to do no matter how many miles I travel or what the price of gas is.
    Tuning you car influences the price of gas no more or less than the gage R&R influences the wear on your fabric. Although both have bearing on a successful evaluation/conclusion, neither have any ability to influence the subsequent variables.
    Forgive me if I’ve provided a poor example. My heart was in the right place!

    0
    #74684

    Gabriel
    Participant

    gRrR:
    As I told you, I understand your problem because I had almost the same problem. The main thing here is if a full measurment system can be evaluated with r&R just as we do with a single gage. I think it is valid. My recomendation, try what I said: Take 10 parts and pair each of them with some weight that simulate the worn out material (the 10 extra weights should be in the range of the weights of worn out material). Then have three operators perform the simulated test (three times on each part-weight). They shuld weight each part with its extra weight, remove the extra weight (simulating the wear test), weight again and calculate in that way the weight of the extra weight (worn out material). Let them inform you only this result (not the weight before or after) and perform the r&R calculation with it. It worked for us: It was accepted by the customer, it was accepted by the third party auditor and, most important, we are satisfied with it. Just try it and see what happens. You will lose nothing by trying.
    Good luck

    0
    #74801

    Bahram
    Participant

    Hi AA,
    How do you get 40 distinct categories?  The highest that I have seen is 9. Is that mean there is less variability?

    0
    #74899

    Bahram
    Participant

    Mike,
    I have been delighted by your discussion and approach, but if the intention of GR&R is to identify the intensity of variation from the measurement/measurement system and improve it, don’t you think that the measurement steps/procedure must be well understood and completely analyzed?

    0
    #74901

    Mike Carnell
    Participant

    Bahram,
    Glad you joined in. Gabriel and I are still exchanging views on email.
    R&R is a precision metric. You can do it with or with out sample preparation as part of the measurement. If my total R&R is unacceptable I don’t have any choice but to take the process apart and determine if it is my sample prep or the gage. If I leave the two confounded I am going to be guessing at what I need to fix.
    If I am using a scale I can run a R&R on the scale in about an hour using mass standards (in the weight range of my part). It is a very clean and easy measurement process. That makes it a very inexpensive piece of information.
    Greg Brue did this in a wall paper factory where we were seeing some issues with color consistency. Before we started screwing around with anything else he ran a fast (he does everything fast) R&R on just the scale – no confounded with anything else. The scale was screwed up. Problem fixed.
    If I had a measurement device that was more difficult to do that with I can come the other way and do it with the sample prep included and deduce that if the total is process is OK then I don’t need to take it apart. This has more risk. I am deducing that the samples I prepared for the study represent the totoal amount of variation that my LT sample prep will have. That is a risk in its self. It is a risk/benifit/cost call. What is the information worth to you.
    When you leave the two confounded you do not know the gage R&R. You know the process R&R.
    Just my opinion I could be wrong.

    0
    #74903

    Bahram
    Participant

    Mike,
    Thank you for your profound response.

    “When you leave the two confounded you do not know the gage R&R. You know the process R&R.”

    0
    #74904

    Mike Carnell
    Participant

    Bahram,
    Thanks. My name and profound do not occur in the same sentence very frquently.
    Good luck.

    0
    #74909

    Gabriel
    Participant

    Wait a second! Mike, Bahram: There is no confounding between the process variation and the measurement variation in the method I proposed! What you are including is “scale variation” and “measurement process” variation (which is not “process variation”). Note that for the r&R you do not need to know the “real value” of the samples you are measuring, so using a gauge mass will not give you any adventage. In the method I propose, you have ten parts and EACH part is measured three times by each operators. If the measuring system was perfect, it would give you a repeatability of 0 (no variation in the three measurementd made by each operator on each piece) and a reproducibility of 0 (no variation in the average value obtained on the same part by each operator). ABSOLUTELY NO PROCESS VARIATION IS INCLUDED IN THE STUDY.

    0
    #74988

    Mike Carnell
    Participant

    Gabriel,
    You don’t need to know the weight or any other measurement to ever do a R&R study. It just has to repeat the same value. If I can do a fast R&R with mass standards, I get linearity, bias, repeatabilty and reproducibility all in one shot.
    Maybe I have missed something in your method. If I do anything that changes a sample between trials I have induced more variation than just the scale create in the measurement it takes. There isn’t a way to unwrap the two.

    0
    #74997

    Mike Carnell
    Participant

    Gabriel,
    First, I agree you guys did some creative stuff to get the process variation out of the R&R. You did not remove the confounding. The part you introduced to remove the variation from the process turns the scale into a process (which you mentioned in your original post). It now has the process of zeroing and weighing confounded. You also add the potential variation from someone forgetting to zero the scale.
    If the R&R on the zeroed weights is poor and I have to do something about it, the first step (for me anyhow) would be to run it without the zeroing so I could understand the whether there is a difference between the two.
    Just my opinion.

    0
    #75001

    Gabriel
    Participant

    Mike:
    I agree about both paragraphs of your message. If you understood that you change samples between trials, yes, you missed something (which can be my fault for a poor explanation). You take ten samples, which represent the worn out parts, and attach to each sample an thing such a piece of wire (extra weight) that represents the worn out material (simulating the material that has gone in the test, but you do not perform any test during the study). Now, you weight a part with the extra weight, then remove the extra weight (simulating that you are performing the test), then weight the part alone again and take the difference (the difference, which should be equal to the extra weight, is the measurement itself, i.e. the assignment of numbers to material things, in this case the weight of the lost material due to the wear test in the “true life” or due to removing the extra weight in this study). Repeat this process with the different parts and operators. Now, during the whole study, you keep each part paired with the same extra weight, so you can repeat the process described before with the same part-extra weight pair as mny times as you want (for example, three operators three times each) without introducing any part to part variation. Is it valid to use mass standards? Not direcly, if you use one standard alone in each measuremnt.You must notice that the measurement in this case is the difference of two readings, and not a direct reading in the scale. And the r&R of the difference is not the same than the r&R of a reading alone. If we call “d” the difference between the measurement beofre (b) and after (a), then d=a-b, and an an approximation could be r&R(d)^2=r&R(b)^2+r&R(a)^2 and, assuming that r&R(b)=r&R(a) (what is very likely) this leads to r&R(d)^2=2.r&R(m)^2, or r&R(d)=1.41.r&R(m), where r&R(m) is the r&R obtained using one mass standards in each measurement, and is asumed to be equal to r&R(b) and r&R(a). In this way, you could use the mass standards as you propose. I still prefer my method because it is more similar to the measurning method that will be used in real life to determine weight of worn out material.
    But you gave me an idea. Take ten samples that represent the worn out parts. Then, instead of using any extra weight in the range of the worn out material, such as a piece of wire, use mass standards in this range (very light mass standars) and apply my method. This method is about as fast as using the standards alone and, as you said, you can get bias, linearity, repeatability and reproducibility in one shot but for the difference between readings, that is what your measurement is after all, and not for the absolute individual readings (what might tell you something about the quality of the scale but is not useful to compare with the tolerance which is given for the diferential weight). Everybody happy?

    0
    #75003

    Gabriel
    Participant

    Mike:
    How will you compare the “full measurement system” r&R with the “scale alone” r&R? The scale alone will be at least 1.41 better just because only one reading is involved, when in the “full system” you have two (please read the previous message I have just posted).
    Anyway, I would do first the “full system” r&R. If it is Ok, that is the end. If it is not Ok, then further invetigation will be needed to see where the problem is (if it is the sclae, the “zeroing process”, the operator or whatever).
    Regards
    Gabriel

    0
    #75005

    Gabriel
    Participant

    Mike:
    By the way: the process of zeroing or taking the difference is not a process introduced in the study. It’s a part of the measurement process you use in the “real life” test, then it is a part of the measurment system, then it should be taken into account in the r&R study. It is useless to have an instrument with very low variation if the measuring process is poor and introduces so much variation that the measurement variation itself “eats” all the tolerance. That’s why the in the r&R study the measurements are performed by the real users (usually operators, not metrologits), using the real measurement process (including a zeroing with a master, if any) and in a realistic enviromet (i.e. factory floor, not metrology room).

    0
Viewing 36 posts - 1 through 36 (of 36 total)

The forum ‘General’ is closed to new topics and replies.