iSixSigma

New Product DOE

Six Sigma – iSixSigma Forums Old Forums General New Product DOE

Viewing 20 posts - 1 through 20 (of 20 total)
  • Author
    Posts
  • #36706

    JAG
    Participant

    I am currently part of a team designing a new product and attempting, although unsuccessfully at the moment, to utilize DOE techniques to arrive at solutions earlier.  The question I have relates to a component that has 10 factors that contribute to the part performance.  When setting up the full-factorial DOE with one replication and two levels (which is too few), I come up with 128 diffent combinations required for testing.  I will take two to three months just to build the parts with all the diffent combinations – way too much time.  I could try a fraction factorial, but this is a critical component.  Does anyone know where I can get some case studies or has experience doing DOE’s on new designs where knowledge is limited when attempting to reduce factors?

    0
    #106580

    Mikel
    Member

    For starters, there is no way 10 factors are important. If you don’t have knowledge to reduce the number of factors, start with a screening design – a Taguchi L12 or a Placket Burman design and reduce the number of factors. After that, your DOE that is needed to look at effects and interactions will be clearer.

    0
    #106622

    Robert Butler
    Participant

         There is something amiss with your post.  A full factorial on 10 factors at 2 levels would be 1024 experiments. A full single replication of the entire experiment would be 2048 experiments.  Thus a design of 128 experiments for 10 factors at 2 levels would have to be a fractional factorial of some type.  Consequently, the statement
    ” I could try a fraction factorial, but this is a critical component.”
    doesn’t make sense. 
      Even if you had the resources to run a full replication of a 2**10 design it would be a huge waste of time and effort.  The whole idea of experimental design is minimum effort for maximum information and neither this nor a “reduced” design of 128 runs meets this criteria.
       Stan is correct, if you don’t have prior information and if you have investigated your problem and feel uncomfortable with the idea of further reducing the number of variables you wish to investigate then the first choice with respect to a design is a screen.
      There are a number of different choices for such an approach.  Stan has mentioned a couple and both are very good.  A screen is looking for the variables having a major impact on your process.  Central to the concept of “major” is the idea that if a variable is going to have an impact on a process it is most likely to exhibit that effect in the extremes of its settings.  Thus, 2 levels, while not necessarily ideal for a design whose aim is the fine tuning of a process, is usually adequate for an initial look.  If you are really concerned about curvilinear behavior at the screening level, a very efficient and inexpensive way to allow for this is to take a design such as a Plackett-Burman and, if it is possible to do so, run an experiment with all of your variables set at their middle level.  For the replicate, repeat this setting.  In the case of 10 factors, you could use the 12 point Plackett-Burman with an additional 2 runs (the center points) for a total of 14 experiments. 
      Based on the sentiments you expressed in your initial post, the idea of 14 experiments for a screen may strike you as folly.  Having worked on a number of problems like yours over the years I can assure you it isn’t. 

    0
    #106715

    Johnny Guilherme
    Participant

    Hi Jag. The 10 factors that you talk about-are these machine control parameters which contribute to the quality of the product you are making?? If so, why not run the process for a while and then see if you can cut down on some of these factors. Maybe you will find that some of them can be fixed and never have to be changed over time. This will make you DOE at lot simpler. There is someone I know in the US, his name is Dave Ingram. He might be able to help you. You can contact him at [email protected].
    Regards
    Johnny Guilherme

    0
    #106718

    Anonymous
    Guest

    Further to the previous posts – are you sure you have to use ‘hardware?’ We used cardboard modelling to ‘characterise’ the ‘ideal function’ of a film drive mechanism very successfully.
    In the past, I’ve also used simulation to model dynamic RAMs to study ‘stuck bits.’
    Have you come across ‘Working Model’ software? We couldn’t use it for on our film drive project because it couldn’t model rubber belts, but you might have more luck in your application. It can also be linked to most CAD systems. Once you understand the product all you need to do is to confirm a few cells in the experimental design!
    Good luck,
    Andy

    0
    #106743

    k.bhadrayya
    Participant

    Dear JAG
    Saw your problem satement. I have been using the placket burman methods quite often where number of variables were large and found it successful in screeninf the large no of varianles into few effective variabels as peratoes law. On screening few variables, they were used for optimization through DOE.
    I can guide your problem if you wish as I have the software to get a plan and analyse staistically to screen the variables.I have been guiding the research scholars through these methdos in their procss development projects.
    Only you need to spell out the variables and their lower and higher levels. If it is 10 variables you need to conduct only 12 expriements as per placket burman. I can give you the plan. Let me know. If you feel the levels are secret you can say -1 and +1 for lower and higher levels
    k.bhadrayya
     

    0
    #106745

    k.bhadrayya
    Participant

    Dear JAG
     
    Yes you can correspond to my mail: [email protected]
     
    k.bhadrayya

    0
    #106748

    howe
    Participant

    The key is SEQUENTIAL DOE methods.  Start with fractional that you can do quickly and can afford.  That can give surprising hints that you might not expect.  10 factors can be run in small first screening DOE.  Then later, after dumping the factors that don’t give strong signals, you can focus on the major key factors in next DOE’s. 
    The big full factorials usually fail if done first with 10 factors. First try two level screening designs, fractional.  Then trim factors to 2 or 3 and try 5 level RSM designs for optimization.  Get statistical help from local guru if you can.  Talk to local Industrial Engineering schools. 
    But before you start, try just graphing the historical data using Mulit-Vari plots to see how your factory variations seem to shape up.  Then your new product design can be focused to deal with those major natural variations (of course that effort in the factory could highlight variations that you can eliminate quickly! And that’s not a DOE). 
    Most of the DOE books found on Amazon or Google or this web site show small sequential DOE approach, with more up front planning and less “big solutions.”  Of course if your product is a CIRCUIT you might run that big 10 factor full factorial on a SIMULATOR for insight.  But most products are not that well simulated due to lack of precise physical models.  No easy answers.
    Mike

    0
    #106750

    howe
    Participant

    Placket-Burman is good advice, but it does spread the interactions over the main effects, and some processes/products suffer most from strong interaction effects.  So the other fractionals, used sequentially, perhaps with help of an expert locally, can discover more about the interactions.  But in any case, the small sequential DOE approach is best.  12 runs is better than hundreds.  Big DOE’s almost never work for many reasons, or don’t get done at all. 
    Mike

    0
    #106751

    Willis Major
    Member

    Using the Taguchi approach, you could perform a screening DOE to reduce the number of factors by using a L12 Design Matrix in which you could put in your ten factors @ 2 levels each (choose them to cover total expected level range). This matrix will give you 12 combinations to run. Taking this one step further, you could replicate the 12 combinations giving you a total of 24 cominations in total. From the analysis of the 24 combinations you could choose a few significant facts of your component (narrowing down from ten factors to three/four significant factors) to use in a Full Factorial DOE  or a three level DOE that is more managable with fewer factors.

    0
    #106754

    Kim Niles
    Participant

    Dear JAG:
     
    One more thing to suggest that I don’t see mentioned in other posts is to prioritize your 10 factors and fix in place as many as seems reasonable to the team before you start your DOEs.  This is assuming your goal is to produce evidence of a stable process for selling that product.   
     
    I like to adopt a RPN prioritizing system common in Reliability FMEA programs where I rate each factor from 1-10 for ease of obtaining data, estimated importance of obtaining that data, and expected significance relative to the factor ranges available.  When those three team based best guess numbers are obtained, I multiply them together to get one priority number and then sort the list by that number. 
     
    Once the list is prioritized, test only the top 2-3 factors while holding everything else as constant as possible and you will often find you are already to a point to justify selling the product as you only need to test until your experimental error rates are low enough to justify a reasonably controlled process (given assumptions of long-term variation control).  Of course as Taguchi, Deming, and others points out, long term process understanding and improvement efforts are justified to continue forever. 
     
    Good luck with it.
     
    Sincerely,
    Kim Niles
    https://www.isixsigma.com/library/bio/kniles.asphttp://www.kimniles.com/  

    0
    #106763

    CN12
    Participant

    You can use Taguchi approach.  But instead of Studying two levels, try three levels and run L18 for 8 factors.  You can test remaining factors later.
    Regards.

    0
    #106766

    Mikel
    Member

    Why?

    0
    #106770

    Vahok
    Member

    My first thought is WOW! Now are the 10 factors you have selected really all that important  (?) or do you just think they are important based on previous process knowledge. Is this a mechanical design? I have run many designs on mechanical designs that at first the process owners felt that there were many more important factors that there really were. I would suggest that you team with the people who are living that process and try to get some sort of reduction in the factors and then perform some screening designs to look for some very strong factors. Possibly break this into several smaller designs and look for the strong factors then possibly run more tests or take the limited information and do some trial runs and see if you are at least headed in the right direction.
     
    Good luck

    0
    #106781

    CN12
    Participant

    With L18, one will get lot more information about 8 factors.
    Regards
    CN12
     

    0
    #106787

    Brett Allen
    Participant

    Hi Jag
    A suggestion –  have you considered using some corelation studies to define which factors of your DOE have a direct and significant effect on the response. This may help eliminate some of the lesser contributors.

    0
    #106795

    Mikel
    Member

    bad advice

    0
    #106809

    Chris Seider
    Participant

    Stan,
    I agreed with your advice 2 posts ago about not wasting time with  a 3-level DOE.  However, I suspect they don’t understand why you said no and it’s a waste of time in your 2 posts.
    I’ll put into motion a string of posts with the statement that a 3 level DOE, even if fractional, would be a waste for initial understanding of potential key factors–as I think you were hinting about.  Keeping in mind 2 things…A.  If no process understanding exists, then doing a 3rd level instead of just 2 increases the cost of the experiment since the point of the first experiment is to NARROW down which factors to do more work with.  B.  Budget and credibility will be used up without finding too many details from the first pass of fractional experiments which are just narrowing down the factors.

    0
    #106819

    Obiwan
    Participant

    Jag
    Johnny’s post has merit as do many of the others.  At a previous employer, we ran DOE’s on all new products…via their processing parameters.  Our DOE philosophy was to run the process for a day, allowing the process engineers to “play” with the process. 
    From this learning, we were able to do a Plackett Burman screening design on 7 factors (usually, sometimes 11)…7 factors took 8 runs (and 11 factors took 12 runs).  The downside to a Plackett Burman (or Taguchi) is that it ONLY analyzes the main effects and points out the MOST signficant factors.  This usually narrowed the list of factors down to about 2 or 3 factors.
    We then ran a full on full factorial with replicates to fully analyze our process.  And, if it were a critical process, we would run some Central Composite or Response Surface designs to really dial the process parameters in a tight space.
    We did this very successfully with over 400 DOE’s.
    Hope this helps
    Obiwan

    0
    #106858

    Mikel
    Member

    I wanted to see if CN12 actualy knew something, he doesn’t.
    I string on 3 (or more) level designs and when and why to ust them would be good.

    0
Viewing 20 posts - 1 through 20 (of 20 total)

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