iSixSigma

DOE – Use Historian Data

Six Sigma – iSixSigma Forums General Forums Tools & Templates DOE – Use Historian Data

Viewing 6 posts - 1 through 6 (of 6 total)
  • Author
    Posts
  • #54215

    sacramento
    Participant

    Hi everbody.. I would like known about use historian data for DOE. Is possible do you use Historian Data for DOE? Yes or No.. If Yes why, If no Why? Thanks

    0
    #194257

    Robert Butler
    Participant

    The assumption of a DOE is that you have control over certain variables of interest and that you have randomized your design runs to neutralize the effects of any unknown/uncontrolled variables. The act of randomizing means that any effect due to unknown/uncontrolled variables gets dumped into the estimate of error and thus, any signal you get from the analysis can be legitimately attributed to one or more of the variables you examined in the design.

    With historical data you have no knowledge of what uncontrolled/unknown variables are confounded with the variables of interest. As a result there is no way to tell if the effect you are attributing to a variable of interest is due to that variable or to the fact that variable was confounded with one or more unknown/uncontrolled variables.

    Other problems with historical data:

    1. Co-varying independent variables of interest. This means you wouldn’t be able to separate the effects of two or more such variables even if you did have adequate knowledge concerning the confounding of other uncontrolled variables.

    2. It is very likely you will not find a correlation between your measurements and any variable known to currently impact the process for the simple reason that such variables will most likely be the variables you currently control. In other words, just as correlation does not imply causation, control of critical process variables frequently guarantee that causation does not guarantee correlation (which is as it should be given that you are doing what you should be doing to control the process).

    My recommendation – don’t waste your time. Instead I would recommend looking at univariate plots/regressions of individual variables of interest against historical output as a guide to your thinking during the define phase of your project.

    0
    #194263

    sacramento
    Participant

    Hi Robert.. thanks for your answer! In some cases is necessary to make experiment with historian data, because we donΒ΄t have the machine free for make the experiment, because we have hight cost of leave stopped machine. Thanks…

    0
    #194264

    Robert Butler
    Participant

    I’m not trying to start a fight here sacramento but that excuse is one of the oldest in the book and it is without merit. I’ve run DOE’s on any number of production lines where we couldn’t take things off line to run separate experiments. The method is quite simple – set up your design, get everyone’s buy in, randomize the order of the runs, and put them through the process as part of the ordinary production cycle. The big issue is that of coordination of effort at all levels – incoming raw material, production line practice, sampling, QC department preparation, etc. It also means that if you are the one heading up the effort you are going to really earn you paycheck.

    0
    #194266

    Chris Seider
    Participant

    Sacramento,

    My colleague RB is correct. If the problem was so important to address, you will be able to get time on your equipment to find a desired solution.

    There is nothing wrong with using past historical data to run multiple regressions, graphical analysis, and other hypothesis testing. If you find a potential answer using past historical data, try that solution out and dial that set of settings and see what results you get. The problem is without a controlled experiment, you don’t have the chance to prove real cause and effect and investigate interactions with a y=f(x) answer.

    0
    #194267

    sacramento
    Participant

    OK.. thanks yours anwers…I believe that you is correct, but I go to change my concept. If I have historian data I go to make multiple regression or others ( this tools is good but do not have same detail of the DOE).. thanks everbody..

    0
Viewing 6 posts - 1 through 6 (of 6 total)

You must be logged in to reply to this topic.