iSixSigma

Correct application of DOE?

Six Sigma – iSixSigma Forums Old Forums General Correct application of DOE?

Viewing 26 posts - 1 through 26 (of 26 total)
  • Author
    Posts
  • #44040

    IE
    Participant

    Hi guys, I’m working on a project with my door manufacturing line in which I’m trying to pinpoint why our daylight openings(DLO) are unpredictable, highly variable, and generally out of spec.  Because I want to identify and verify the cause of this problem, a DOE sounds like a logical tool to use.  There are so many potential causes of these problems with DLO’s that I think I’m going to start with a simple DOE and then, if none of my initial variables are significant I’ll take it one step further and look at more intricate variables.  Is it okay to do that?
    Furthermore, I’d like to take multiple DLO measurements.  For instance, I’d like to look at the vertical DLO in three places and the horizontal DLO in three places.  Is this out of the realm of DOE because I want to look at more than one output variable?
    Thanks, Andy

    0
    #140474

    GomezAdams
    Participant

    What you are looking to do is utilize a screening design once you have identified all your variables to include.
    Minitab/Statistica/…… others can assist you in its development.
    As far as including multiple readings, sure, you can do that. Make sure your measurement system is up to snuff first.

    0
    #140479

    IE
    Participant

    If I use a screening DOE, am I supposed to use my massive list of all the potentially significant variables or just my first “layer”(they’re easy to measure and intuitively the most significant)?
    I have another problem that is probably far more important: If I run a DOE, I need parts that are on the high and low end of the spectrum.  This would mean I’d need to painstakingly measure each part looking for high’s and low’s.  There has to be a better way or a better tool out there, right?
    Thanks, Andy

    0
    #140482

    Robert Butler
    Participant

    It sounds like you are confusing your X’s and your Y’s.  The X’s are the things you can vary at will to impact the end product properties.  Assuming your first “layer” of variables is the result of really thinking about variables impacting the process – the fact that they are easy to measure should not be an issue – then a screen with just these variables present is reasonable. 
      Once you have your system set up with a particular set of X’s you would go ahead and take your various measurements.  These measurements will be whatever they will be – high, low, centered, whatever. 
      If your X’s are such that you have to “painstakingly measure each part looking for high’s and low’s” then what you are saying is you don’t have any real control over these variables. 
      If that is the case then you need to stop and re-think the problem.

    0
    #140484

    IE
    Participant

    I realize now that I wasn’t clear.  We do have control over our X’s per what we specify from the supplier.  Of course we specify a nominal dimension with a +/- tolerance.  I assumed that in the experiment I would need to build doors with parts that are on the high end of the tolerance and parts that are on the low end of the tolerance.  Then I measure the DLO.  Is that correct?  If so, then it would be quite a chore to find samples of the parts that are exactly on the outer bounds of the tolerance.
    Perhaps a DOE isn’t applicable here, or perhaps I’m misunderstanding how a DOE should work.
    Thanks, Andy

    0
    #140486

    Anonymous
    Guest

    Andy,
    Do you have Pro Engineer at your site. If you do; you don’t necessarily have to build doors. You can use Off-line tolerance design to check your specifications.
    Andy

    0
    #140487

    Robert Butler
    Participant

      Maybe and maybe not. I’ve had to deal with this situation many times.  If you have enough range in the tolerance so that you can choose lengths that are “long” vs. lengths that are “short” and as long a members of either group don’t overlap then what you have is the problem of the “fuzzy X”.  The post below has some of the details.
    https://www.isixsigma.com/forum/showmessage.asp?messageID=96473
      Another option is that offered by Andy U and, again depending on what you have to work with, it may be a better choice than trying to do a DOE.

    0
    #140490

    Anonymous
    Guest

    Robert,
    Once again .. congratulations on an excellent post (fuzzy x.)
    Fuzzy X’s are a major problem in some industries – for example, the chemical industry where there is little certainty concerning the ‘activity’ of raw materials. In my opinion has led to Georege Box and others being over defensive with respect to ‘psuedo-interactions’ – those interactions that arise due to fuzzy x’s.
    Pehaps this is why Taguchi has not been not quite so defensive: anyway, it will certainly provide an interesting topic for ‘discussion.’
    Best regards,
    Andy

    0
    #140493

    IE
    Participant

    Unfortunately I don’t have anything as sophisticated as what Andy U has suggested.
    Robert, I read your post about “fuzzy x” and I have a basic question.  Which should come first, designing the experiment or picking my “high” and “low” parts?
    Thanks, Andy

    0
    #140494

    Wayne Leao
    Member

    Hey, I think before any advanced quality tool’s application, we should ask ourself 3 times that did all primary quality tools applicated. For example, Pareto, BrainStoming, Fishbone? Per Deming’ and the practise, almost 80% of a process could be decreased. So, study the prcess from now on, you will find some actions without any prepartion than 6Sigma could improve it too.

    0
    #140496

    Robert Butler
    Participant

    IE,
      I’d recommend designing the experiment first and if your X’s are going to be “fuzzy” I would recommend sticking to main effects designs.  As I mentioned in the cited post, the biggest problem with noise in the X’s is the loss of information concerning some interactions. 
      To that point I find Andy U’s comment about pseudo-interactions interesting.  I wasn’t aware of any controversy surrounding the issue of interactions and noise in the X variables I’m well aware of the Box Taguchi issues and, in truth I have issues, and have had words,  with both gentlemen).  Every time I’ve had to face the problem the issue was the recognition that the price paid for the no longer perfectly orthogonal X matrix was the loss of information about some of the interactions.

    0
    #140498

    IE
    Participant

    So, in short, its up to me to find parts that are as close to the upper and lower limits as possible…but if I can’t find perfect parts its nothing to lose sleep over because I can still get good results.
    Some of that other stuff you’ve mentioned is a little over my head, but it looks like I need to crack open a couple of my old textbooks.
    Thanks, Andy

    0
    #140500

    Mu Joe
    Participant

    IE,
    I’m not sure if I quite understand your problem. You said in your first post that you had a quality problem with your DLO’s. (I’m not in the door manufacturing business, but I preseume that means the holes where your glass panels would go in your doors.) You said they are unpredictable and generally out of spec. What I take that to mean is the dimensions come out the wrong size so your glass panels don’t fit properly. Stop me if my assumptions are wrong.
    So what a DOE will do for you if done properly is help you identify those inputs (X’s) in your manufacturing process that are most responsible for your out of spec DLO’s (your Y). I think you were interested in both height and width dimensions (two possibly independent Y’s). and yes, you can collect data for both Y’s at the same time as you adjust the inputs.
    What throws me off is your comment about having to look for doors that were at the upper end of the spec and at the lower end of the spec. Why would you do that? The DLO spec is your Y not an X (if my assumptions above are correct). You need to choose the X’s (inputs in your manufacturing process) that contribute to the out of spec Y. So you need to run some doors through your process with your factor X’s set at “high” and “low” settings, which just means at opposite ends of their operational limits. The output (Y) of your experiment will be the dimensions of your DLOs and you will analyze that data to determine which inputs had the biggest impact on your Y. Let’s say for example that you thought one of the factors was the RPM of a saw blade. You might set your factor levels at 500 RPM for the low level and at 1,000 RPM for the high level. You set up your experiment with your other factors identified and run the appropriate number of units at the various hi/lo settings then analyze all of the output data to see which X’s had the biggest impact. 
    Sorry to get so elementary in my discussion, but it seemed to me that you were about to go about this problem the wrong way. Unless of course, I have completely misunderstood the situation.
    Mu Joe

    0
    #140501

    Anonymous
    Guest

    Robert,
    Seems to me one of the great advantages of Miniab is ease of data generation. Using Minitab it’s easy to investigate the effect of fuzzy x’s; and I hope someone will try.
    Sometime ago I decided to design a simple two factor full factorial with two ‘contrived ‘ main effects, using random data. I then added another random component to the  x values to study the effect of control variable noise and I found it produced an AB interaction. (My design was similar to a study I had recently completed on the chuck of a 3-D laser imaging system that printed a quadraphile helix on a miniature, microwave antenna.)
    On my first attempt, my replications were put into ‘blocks,’ so I wondered what would happen if instead I put them in an outer array …  with a view to using an average instead of individuals –  to lessen any fuzziness in the x values. (At the time, I asked the forum if anyone had an interest in fuzzy regression, and I got some sharp replies – one notably from Mike Carnell.)
    I then found the AB interaction ‘disappeared’ leaving my contrived, original A and B main effects. This was important because in the ealier experiment on the chuck we could not explain the ‘interaction,’ as the fixture appeared to be linear. (As you know interactions tend to be more difficult to set than linear effects.)
    Later, I began to think about why George Box and others were so defensive in their experimental designs and wondered if there was something specific about the chemical industry and if some of these interactions might well be due to fuzzy x’s.
    It occured to me people often measure ‘exact’ quantities of raw materials without giving any consideration to the ‘chemical activity’ as measured by IR spectroscopy, which I suspect amounts to a fuzzy x.
    Anyway, I hope someone will be interested to investigate this independently because it might explain why so many of us have had good results with Taguchi Methods, and without being too defensive towards interactions. (Justified after running confirmation runs.)
    Cheers,
    Andy

    0
    #140502

    Mu Joe
    Participant

    IE,
    Sorry, I just read through the whole thread again, and what it looks like is you “assemble” doors from parts that are supplied to you. So, when you say you are looking for parts near each end of the spec, you’re talking about door components not finished doors.
    If that is the case, then I agree that maybe a DOE is premature. Exhaust your other tools first. It could be as simple as requiring a tighter spec on your components, or there could be a procedural issue in your assemply process. Do you know what your out of spec proportion is for your input componenets? Does anyone ever check that?
    Mu Joe

    0
    #140503

    Jered Horn
    Participant

    IE,
    You’ve started a good discussion on DOE…
    However, I’d like to point out that DOE is probably NOT the easiest tool to use in your situation.
    I’ve heard of a class you can take that teaches you some of the easier tools (along with DOE and other “hard” stuff).  Six Sigma, or something like that.
    What I’m referring to is…Process Map, FMEA, Cause & Effect Diagram, C-T Matrix, Pareto Analysis, Hypothesis Testing, ANOVA, etc.
    In general, going directly from problem statement to DOE is not advisable…IMO.

    0
    #140504

    IE
    Participant

    Yes, I would have to go find parts near each end of the spec to complete the DOE…that is, unless I let some fuzzy x’s into the experiment(which I’m not against).
    It really may be as simple as requiring a tighter spec on the parts/components, but how can I figure this out and be sure of it?  To me, it seems like a great application of DOE.  Our panels have been variable and unpredictable for so long that everyone just accepts it as a fact of life.  Everytime our quality team is called in to find a part out of spec, they find nothing.
    Is there a better tool that I should be using?  In essence, I am starting from square one.  I’m certainly open to new ideas.
    Thanks, Andy

    0
    #140505

    Mikel
    Member

    IE,
    I would bet you have a measurement problem.
    Go check your MSA on all the critical piece parts, sub-assemblies, and the final assembly.
    You’ll find your problem there.
    After fixing the MSA, you’ll want to consider DOE’s for the purposes of getting the specs right – and you’ll relax as many as you tighten 

    0
    #140573

    IE
    Participant

    We haven’t had any past MSA’s done on anything in the plant.  I don’t even know where to begin with that.  I have Lean Six Sigma Handbook and its got a little information on MSA, but I feel pretty confident in the way I’m measuring my parts and my DLO’s.
    I want to do a DOE just for that reason(tosee if the specs should be changed!)
    Thanks, Andy

    0
    #140700

    Tierradentro
    Participant

    Andy,
    Have you every used Statistical Engineering methods taught at the Shainin Institute?  To me, the first step always is to eliminate measurement error.  2nd, is to use methods which ‘divide’ the sources of variation so that you end up with a FEW suspects… not a lot.  In this manner you end up doing ONE (1) effective experiment vs. many.  The problem I have with DOE is it seems everyone wants to start there and that’s costly and the wrong way to resolve problems.
    I don’t know much about door assembly, but first, find a measurement which is repeatable.  Then, pair up doors that meet spec vs ones that don’t.  There is a method called Paired Comparison you can use to compare good vs. bad.  It’s very effective but you have to be thorough in your investigation.  Your pattern of Ys will give you the Xs that are suspect.  This is used if you cannot ‘assemble/disassemble’ the door without destroying it.
    If one relies soley on DOEs, you may end up doing many of them … again, a costly adventure without assurance of resolving the problem.  I see this being done constantly by Black Belts as well as Master Black Belts unfortunately.

    0
    #140703

    jediblackbelt
    Participant

    John –
    Gotta love Tukey Tail count type of experiments.  Can’t say enough about them to everyone else.  How something so simple can save so much time in a project and then lead down to fix the problem.
    IE – You may also want to look into multi-vari charting against your production/quality defects to see if you have a problem across the board or if you can narrow it down to different inputs from shifts or operators or suppliers.
    Might save you a lot of time and money to do more investigation first versus jumping into DOE.  But several people have said it already – you may feel confident in  how you are measuring things, but do you know everyone else is just as good?  Start with MSA, I have found about 75% of all my projects usually end here with a correction of past bad practices when measuring the defects/defectives.
    Good Luck

    0
    #140706

    Tierradentro
    Participant

    JediBlackBelt,
    Great supplement !  Thanks for your input.  Usually, Multi-Vari is where I always start as well (after ISOPLOT) but didn’t know the situation re: the door assembly if it’s hand-assembled or automated.  Good point though.
    re: Tukey, I like Shainin’s 6pack test (B vs C).  So simple to use compared to all the various statistical garbage that people in manufacturing really do not need to be bothered with.  He did a great job improving on Tukey’s end count didn’t he?
    good luck and thanks for your input.
    John

    0
    #140709

    ww
    Member

    Good morning all;
    I enjoyed greatly the conversation and have gleaned a great deal from it.  Please correct me if I’m wrong but there seems an underlying current in all of the discussion, an out of control process.  I must agree with Mr. Horn.  This subject smacks of process control or actually a lack of it with regard to output.  To get a better bead on the cause(s) its FMEA time!!  I love these things.  If conducted with the right mix of contributors, they can really cut to the failure modes in the current process.   Then the control plan (if there is one) can be modified to one that provides a constant stream of conforming product to finished goods.  You may get the best bang for the buck using some Mfg. Process Control tools.  Just an opinion, I could be wrong.  I wish you well.

    0
    #140717

    IE
    Participant

    Wow guys, I really appreciate all the responses.  It appears as if I’ve got ahead of myself, as always.
    I wish I understood more of this stuff you guys are talking about.  I’ve been in the “real world” for a year, but I haven’t been applying the tools I learned in college.  Apparently I just didn’t get how and when to apply of all these tools.
    I think I need to start from square one, establish a baseline.  Then move on to identifying potential causes with an Ishikawa diagram.  Then verify which ones truly are causing the problem using the correct tool.
    I think its safe to say that because my daylight openings are so variable, my process isn’t in control.  Perhaps I can look at this after establishing a baseline.
    John: Part of our doors are hand assembled while other parts are automated.  Which tools ‘divide’ the sources of variation???
    Thanks again for all the comments, Andy

    0
    #140800

    Neutral Observer
    Participant

    Not every problem can be solved with Six Sigma methodologies. Why don’t you do a Kepner Tregoe to find the root cause of the problem and then try to solve it with Six Sigma tools if you need to after you know what the problem is.

    0
    #140810

    PlasticsEngineer
    Participant

    I agree with others that MSE (MSA) is first step.  Process and Product Map, at least a basic one, before anything else.  We generally use COV before a DOE to determine focus.  Make sure your sampling plan is well thought out, or you will just end up with more questions or misleading results.

    0
Viewing 26 posts - 1 through 26 (of 26 total)

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