Re: Six Sigma and Shewhart

[John <jsdwd@samart.co.th>]

on 6/10/00 4:33 am, Dr A D Burns at mmedia@ozemail.com.au wrote:

> In effect, this is what the "I" of a "PID" controller does. As long as a
> process is encountering random variation, there is little reset windup and
> little effect from the "I" component of the controller. When trends above
> the mean, or runs occur, the integrating effect of the controller
> progressively comes into play. It is not so effective with a trend around
> the mean. Controllers also generally work on very short time intervals which
> I guess makes them susceptible to auto-correlated data.
> 

Interesting input.  A few points:

1.  Above it says, "As long as the a process is encontering random
variation.."  If I might make two points.  First the process doesn't
'encounter' variation, it produces it.  Second, what 'rules' are used to
test for randomness...Of course, we don't know if it's 'random' or not
(indeed randomness may be best thought of as a hypothetical construct), so a
set of rules is needed.....back to Shewhart, etc.

2.  Again, "When trends aove the mean, or runs occur..."  How are 'run' and
'trend' operationally defined?  The so-called sensitivity of the chart is
dependent on these defintions and the sampling protocol.
 
3.  "It is not so effective with a trend around the mean...."  Are you sure
you are not talking about the target value?  I have frequently found that
individuals working in industries that use process controllers quite a bit
(e.g. refining, paper making, etc.) that folks use these terms
interchangeably and, of course, they are not interchangeable.

My experience is, like Myron's I guess, that usually process controllers
have a 'dead band' around the target in which no action is initiated by the
controller.  it is important to be clear when one is talking about the mean
and when one is talking about the target value.

Which, of course, brings up another issue which is that if a process is
being continually 'corrected' or 'controlled' then how would one go about
ascertaining the mean?  Is there one in any usual sense of the word.  Does
the concept of 'constant cause system' have much meaning in this
application?  How would one go about defining it, estimating it, or trying
to construct limits for it?

Heero Hacquebord may have some ideas here or Donald Wheeler.  I know Heero
has worked a lot with the Paper Industry...  Back when I was working with
Chevron refinery in Richmond, I encountered the name of a statistician from
Canada who had done some work with the application of control charts and
other statistical methods to feedback controllers and other process control
devices.  Unfortunately the one brain cell I have currently functioning is
not recalling his name....

A final quote, if I may once again borrow from above, "...which, I guess
makes them suscetible to auto-correlated data.."

Yeah, I guess.

Thanks for a good post on a difficult technical topic.