iSixSigma

Screening DOE – question

Six Sigma – iSixSigma Forums General Forums Tools & Templates Screening DOE – question

This topic contains 12 replies, has 8 voices, and was last updated by  Venugopal G 9 years, 4 months ago.

Viewing 13 posts - 1 through 13 (of 13 total)
  • Author
    Posts
  • #53384

    van Helden
    Participant

    I am working on a new process where we have identified 6 critical factors. There is too little process knowledge to understand the impact of the six factors on the response Therefore we would like to run a screening DOE in order to find the significant few. I have experience with factorial designs, only I used 4 or less factors. To run this experiment there are two (or more..?) options.
    (1) 1/8 fractional factorial, which would have 8 runs. And replicate makes it 16 runs. For me this would be the minimum requirement. But at the same time, 16 runs is the maximum we can execute.
    (2) Plackett Burman, which would have 12 runs. But then if we would replicate that makes 24 runs which is practically hardly possible.

    I don’t have any experience with PB DOE’s. I did some research on it, however the only advantage of a PB design I could find so far, is the limited number of runs when testing 7 or more factors. Is there any other advantage of running a PB over a FF design. What would be the drawbacks of one over the other.
    Please advice

    Thanks
    J.

    0
    #189904

    Leader
    Participant

    For main effects only (and making the dangerous assumption that you don’t have any two-way interactions), go for a resolution III DOE, which is your 8-run solution. Why do a replicate for a screening design? Spend your money later on. The PB may get you a little more data, but I’m not totally sure it’s worth the extra four runs.

    0
    #189907

    Hody
    Member

    Resolution III confounds Main Effects with 2-Way interactions (which can be common, and thus a problem at Res III). If you spend extra runs for a Resolution IV (even without reps) you will insure no confounding until you may have main effects with 3-ways+( far less common)… I’d definitely recommend going for the Res IV !

    Good Luck

    0
    #189910

    Robert Butler
    Participant

    The whole idea of a screen design is to test to see if any of the variables of interest have a major impact on the product properties of interest. While you haven’t told us whether your variables are categorical or continuous your comments concerning earlier experience with designs would suggest the 6 variables are continuous. If this is the case then the most economical design would be the 8 point fractional factorial design with the addition of two center points. The center points will provide your replication and they will also provide information concerning the linear (or lack thereof) nature of the responses of your process to the changes in the variables. They won’t identify which variable(s) is/are responsible for a curvilinear effect( if indeed one is detected) but they will tell you whether or not one is present over the ranges of the variables of interest.

    You can use the results of this 10 point design to examine the main effects of your variables of interest and save your remaining 6 experiment for other things such as confirmation runs and/or the possible investigation of some select curvilinear/two way interactions – (the latter done by employing the methods of design augmentation).

    0
    #189912

    Hody
    Member

    I disagree. It’s overly cheap with a Res III DOE. If you think that you are examining main effects but you suggest a design that confounds main effects with 2-way interactions is a poor approach. It’s a better spend to avoid confounding rather than investigating curvilinear at this point trying to reduce factors. If you use Minitab for DOEs you’ll notice that Res IIIs are shown in RED (for good reason)…

    Cheers

    0
    #189913

    Robert Butler
    Participant

    The choice of a design is going to be driven by time and cost constraints. The original poster said, “16 runs is the maximum we can execute.” If we assume this is a statement of fact and not just exaggeration for effect then 16 runs is all he/she can afford/will be permitted to run. Under these circumstances a Res III design with 6 factors is as good as it gets. The Res III design will permit an investigation of a single 2 way interaction (of course it is confounded with another 2 way) but many times prior information can be used to choose variable assignment to the design so that the interaction of interest can at least be considered with as little confounding as possible.

    In the past, on this forum, we have had individuals whose post said, in effect, “I ran this Res III design because I was told that it would allow me to look at all of the variables of interest but when I try to build a model with intereaction terms the program gives me an error message.”

    I don’t know Minitab but I did know some of the people behind its development and I suspect they put Res III designs in red not to indicate they shouldn’t be used but just as an additional warning to make sure anyone considering their use really understand their limitations.

    0
    #189924

    Severino
    Participant

    I have to side with Robert on this one. Adding the center points to the 8 runs allows you to get a decent estimate of the experimental error very efficiently. Even if the design is Resolution III, so what? Let the process get started understanding main effects. Once you are up and running you can use operational tools to such as multi-vari inestigations and EVOP to gain better process knowledge. Once additional emprical understanding is gained, they can move on to higher resolution designs, robust parameters, RSMs, etc. etc.

    0
    #189975

    van Helden
    Participant

    First of all, thanks you all for your valuable comments. Secondly I would like to emphazise that we should not discuss the difference between a Res III and Res IV experiment. I do know that.

    My basic question was: what would be the advantage of a PB desing over a ResIII FF design?

    One of the questions/assumptions made was that the variables are continuous data – which is correct. We are discussing a chemical treatment plant. And that brings me to the following point.

    Going through your replies I found some interesting information. In my previous experience I used to replicate the whole design to give me more usefull information about the variation of the individual measurements. This was especially necessary since I experienced quit some noise/disturbances through the duration of the experiment. But some of you mentioned that you could only use center point(s) to estimate the experimental error. But then you would assume equal (lineair) error across your messurements….correct…?

    Thanks
    Joost

    0
    #189977

    bbusa
    Participant

    J

    Sorry for joining this discussion late

    My first point is : With too little process knowledge how did you identify the 6 critical factors

    My opinion : dump the PB and go for a simple FF design – first without the center points. I don’t think the PB has any advantage over the Res III design in your case

    bbusa

    0
    #189980

    Robert Butler
    Participant

    The value of a PB is that the number of runs in one of these designs is some multiple of 4 whereas in a standard factorial the number of runs is a power of 2. In those cases where the number of runs in a PB is equal to the number of runs in a factorial the design is referred to as a geometric PB and it is essentially the same as a two level fractional factorial. With 6 factors the basic design would be 8 points whether you use PB or a Res III factorial.

    As for replication – the question you need to ask is this – what is the point of your experimental design? Are you trying to identify variables that are significant in the presence of noise or are you trying to generate a great estimate of precision?

    You said ” I experienced quit some noise/disturbances through the duration of the experiment.” Given this statement I would conclude you have a noisy process and that your primary interest is identifying process factors that will have a significant influence on the outcome regardless of the systemic noise. If this is the case then randomizing the run order of the experiments and running a simple replication on two center points will be the most economical choice. If this makes you too nervous then try the following – plan for a full replication of the Res III with two center points (20 runs) run the first 10 and analyze the results of these 10 runs – based on what I’ve seen over the years you will find so many interesting things that you won’t want to waste any time with a full replication and your choices for additional experiments will be driven by other interests.

    If you are interested in identifying process variables that imact process variability then running full replicates and looking at two or three point estimates of variation at each of the design points is a waste of time and money. As mentioned previously, the method of choice for this effort is the Box-Meyer. Box-Meyer “requires” a Res IV design to make sure you are not confounding two-way and main effects, however, you can run this on Res III – you just have to remember what you are doing and the assumptions you are making.

    0
    #190269

    Venugopal G
    Participant

    joppie wrote:

    I am working on a new process where we have identified 6 critical factors. There is too little process knowledge to understand the impact of the six factors on the response Therefore we would like to run a screening DOE in order to find the significant few. I have experience with factorial designs, only I used 4 or less factors. To run this experiment there are two (or more..?) options.
    (1) 1/8 fractional factorial, which would have 8 runs. And replicate makes it 16 runs. For me this would be the minimum requirement. But at the same time, 16 runs is the maximum we can execute.
    (2) Plackett Burman, which would have 12 runs. But then if we would replicate that makes 24 runs which is practically hardly possible.

    I don’t have any experience with PB DOE’s. I did some research on it, however the only advantage of a PB design I could find so far, is the limited number of runs when testing 7 or more factors. Is there any other advantage of running a PB over a FF design. What would be the drawbacks of one over the other.
    Please advice

    Thanks
    J.

    Hi,

    I have experience in both classical DOE and shainin DOEs.
    When there are more than 3 factors,a shainin DOE tool called as ” Variable Search” is the best one. It minimizes the total no. of runs based on the results of run1, run 2 etc. What I want to say is the succesive runs are based on the results of previous run results. You may end up also in just doing 3 runs, that is how the tool will guide you.3 is just a hypothetical number. But the problem is shainin techniques material is not a freeware, you will not find much on net. If you are interested I can guide you step by step.

    Venu.

    0
    #190271

    Mikel
    Member

    Shainin is public domain if you don’t accept their nonsense that Dorian created this stuff. He didn’t. You will note they only have service marks, not trademarks.

    You can get all of the info you need on either variable or component search from World Class Quality by Keki Bhote or from any competent BB materials.

    A Taguchi L8 or a fractional 2^(6-3) is a better way to go anyway, You will get more information. The only people who believe Variable Search is better are the Shainin cult. What major results can they show for their work? GM? Chrysler? They were in both places for way over 20 years. I believe that speaks volumes.

    0
    #190273

    Venugopal G
    Participant

    Stan wrote:

    Shainin is public domain if you don’t accept their nonsense that Dorian created this stuff. He didn’t. You will note they only have service marks, not trademarks.

    You can get all of the info you need on either variable or component search from World Class Quality by Keki Bhote or from any competent BB materials.

    A Taguchi L8 or a fractional 2^(6-3) is a better way to go anyway, You will get more information. The only people who believe Variable Search is better are the Shainin cult. What major results can they show for their work? GM? Chrysler? They were in both places for way over 20 years. I believe that speaks volumes.

    Hi,

    I am not a supporter of Shainin/Taguchi/Classical. I am not working as a marketing guy for either of these. I dont mind using a tool if it helps irrespective of where it is from as long as it is proved. It has been proved as a good tool since I applied it quite a few times. I also appled Classical/taguchi DOE, but variable search proved as a good tool especially when there are more factors. The advantage with Variable search is you will end up in less no# of runs , coz it is a proactive tool. Proactive in the sense, the preceeding runs are based on succeeding runs.
    Joppie I recommended this tool based on your six factors. If you had less than 3 factors, I would have recommended still the full factorial DOE. Since you have experience in factorial designs try to gain some experience in variable search also. Judge yourself on which is better (even later also, when ever you get a chance).

    Venu.

    0
Viewing 13 posts - 1 through 13 (of 13 total)

You must be logged in to reply to this topic.