1. This site uses cookies. By continuing to use this site, you are agreeing to our use of cookies. Learn More.
Dismiss Notice
You must be a registered member in order to post messages and view/download attached files in this forum.
Click here to register.

Understanding SPC

Discussion in 'SPC - Statistical Process Control' started by Guyselt, Oct 21, 2016.

Tags:
  1. Guyselt

    Guyselt New Member

    Joined:
    Oct 21, 2016
    Messages:
    2
    Likes Received:
    0
    Trophy Points:
    1
    Hi,


    I'm starting with SPC in my company. To my understanding to run SPC you need first to set up what I call the benchmarks (UCL and LCL) using a set of historical data. Once this defined you plot all new values against those and check for OOC results.

    My colleagues use to run statistical on the data and where looking also for OOC points. Doing like this you get moving UCL and LCL. They don't use historical data.


    What is correct practice ?


    Regards,

    Guy


    Please excuse my English, it is not my mother tong.
     
    Guyselt, Oct 21, 2016
    #1
  2. Bev D

    Bev D Moderator Staff Member

    Joined:
    Jul 30, 2015
    Messages:
    606
    Likes Received:
    664
    Trophy Points:
    92
    Location:
    Maine
    When the Control Chart is being used for 'on-line', real time, continual monitoring and control of processes the control limits are set from 'historical' data and the future points are plotted against those limits. The limits are not changed until a permanent improvement has been implemented. Limits not continually changed as new data is collected.

    If you are doing diagnostic ('off-line') studies for engineering purposes then you would have 'custom' limits for the period or periods of interest...
     
    Bev D, Oct 21, 2016
    #2
    MCW8888, RoxaneB, Jennifer Kirley and 1 other person like this.
  3. Guyselt

    Guyselt New Member

    Joined:
    Oct 21, 2016
    Messages:
    2
    Likes Received:
    0
    Trophy Points:
    1
    Hi Bev,
    Many thanks for your answer. If I understood you well, in any cases you need to have limits (calculated or “custom” ones).
    Now I need to convince my colleagues. Do you have papers, web sites, books excerpt that I could show them.


    Thanks
     
    Guyselt, Oct 24, 2016
    #3
  4. Bev D

    Bev D Moderator Staff Member

    Joined:
    Jul 30, 2015
    Messages:
    606
    Likes Received:
    664
    Trophy Points:
    92
    Location:
    Maine
    Well Shewhart said it in his seminal book "Economic Control of Quality of Manufactured Product". I'm sure it's in Wheeler's book on SPC and in Grant and Leavenworth's book and in Ott's Process Quality Control. Although these books spend a lot of time on many topics and so an explicit statement to not change the limits is difficult to find. It's a needle in a haystack fi you will.

    But it really just makes sense if you think about it. (and is ultimately more powerful than quoting chapter and verse, even the founder himself. your colleagues can no doubt quote any number of (hacks) sources that will say the opposite. The 'proof' of which way is correct lies in the intent and the math.

    The intent of a control chart used on the line to provide real time monitoring and process adjustment is that today's results are compared to the previous stable process. We calculate the initial limits based on some recent data, getting an estimate of the mean and the within subgroup variation. If a process changes - either in mean or variation - it will show up as an OOC point(s) on the chart. This is based on not changing the limits from the known stable period. If we were to change the limits with every new point we would be incorporating the process degradation into the limits and reducing the sensitivity of the chart to detect process changes. THAT is the reason to not continually recalculate the limits.

    Does that make sense?
     
    Bev D, Oct 24, 2016
    #4
    MCW8888 and tony s like this.
  5. RoxaneB

    RoxaneB Moderator Staff Member

    Joined:
    Jul 31, 2015
    Messages:
    926
    Likes Received:
    1,081
    Trophy Points:
    92
    Location:
    Ontario, Canada
    This is the key point, in my opinion.

    An OOC point could be "bad" or "good". If "good", my experience has been to question what the conditions were and are they repeatable? Once a permanent improvement is implemented, the control limits can be re-calculated/re-established.
     
    RoxaneB, Oct 25, 2016
    #5
    Chris Glover likes this.
  6. tony s

    tony s Well-Known Member

    Joined:
    Sep 10, 2015
    Messages:
    1,350
    Likes Received:
    1,055
    Trophy Points:
    112
    Location:
    Laguna Philippines
    The spread of the control limits depend on the variation in the process. If your historical data was based on a stable but wider variation (e.g. range or standard deviation), your control chart will be less sensitive to detect OOCs. If the historical data was based on a stable but with narrower variation, then your control chart will have a lot of OOCs.
     
    tony s, Nov 17, 2016
    #6
  7. ncwalker

    ncwalker Well-Known Member

    Joined:
    Sep 21, 2015
    Messages:
    261
    Likes Received:
    168
    Trophy Points:
    42
    Location:
    North Carolina
    Honest question: And this is why though I like control charts, I am not necessarily a zealot about them.

    If the point of them is to monitor a process, then to me we are looking for two things:
    1) Is the process drifting? Which would typically result in a setup response - retargeting something. (Definition of too much drift could be - too much change OR proximity to spec limit).
    2) Is the variability increasing? Which would typically result in a maintenance type response. Something is loose, or out of balance, or a feedback sensor is stuck.

    In my head, the whole purpose of the control chart (remember, they are OLD, circa 1920 or so) was to allow us to do this with statistics AND (at least to me) the whole subgroup and using the range as an estimator for sigma part of it is because of the tediousness (not difficulty) of calculating sigma. That's not the case anymore. I can do it in a computer with trivial ease.

    So I still don't get why we don't just plot mean with +/-3 sigma bands (or whatever sigma you choose) and monitor this against the spec limits (replacing the Xbar chart) and then just plot sigma for the Rbar chart. Where sigma is a rolling number calculated from the last 30 measurements.

    To me, this would show both drift and increases in variability in a better way than control limits would.
     
    ncwalker, Nov 17, 2016
    #7
  8. Bev D

    Bev D Moderator Staff Member

    Joined:
    Jul 30, 2015
    Messages:
    606
    Likes Received:
    664
    Trophy Points:
    92
    Location:
    Maine
    Just because something is old doesn't mean it's wrong or bad. On the other hand just because something is new and easy, doesn't make it good or right

    You're correct that the within subgroup range was used originally because it was easy to do 'in your head' prior to calculators and computers.
    With today's software and some of our more complex processes than in Shewhart's day, the Xbar S chart is perfectly acceptable instead of an Xbar R chart.
    However using the rolling standard deviation of the last 30 measurements is NOT a good approach. It accommodates expanding variation and violates the very power of how control charts actually work. The foundation is that the within subgroup variation is directly related to the between subgroup variation for a stable, homogenous subgroup scheme. When the process changes (becomes non-homogenous) this relationship will no longer hold. (this is a fundamental relationship of statistics from t-tests to ANOVA to confidence intervals to SPC, to ANOM...) So if our limits are calculated from the average WITHIN subgroup standard deviations from 10-30 subgroups, set at +/-3*(average within subgroup SD/c4*sqrt(n) . Set them and plot future values against those limits.

    I recommend the following to gain a better understanding of the theory before trying to answer your good questions on your own. These will no doubt also help the OP...
    The right and wrong ways of computing limits
    When do I recalculate my limits
    Foundations of Shewhart's Charts
     
    Bev D, Nov 17, 2016
    #8
  9. Serious Man

    Serious Man Active Member

    Joined:
    Oct 24, 2017
    Messages:
    79
    Likes Received:
    28
    Trophy Points:
    17
    I have different approach to starting SPC. I do not care specification, process capability value, distribution type, etc. I am interested only in whether there were determined process characteristics affecting product characteristics. I call it "Do not ask, what's a width of the road you want to drive, when you don't have a steering wheel."
    I was really surprised how fetishized Cpk can be, while lack of means to change product characteristic and then Cpk was not determined as a problem.
     
    Serious Man, Oct 25, 2017
    #9