## Forum Replies Created

- AuthorPosts
- June 21, 2018 at 8:58 am #202716

Rip StaufferParticipant@ripstaur**Include @ripstaur in your post and this person will**

be notified via email.No, Mike; I’m not trying to imply that you don’t know the difference

between spec limits and control limits; had I believed that, I wouldn’t

have asked the question that way. I just wanted to know how you are using

those controllers (I’d still like to know).I also realize that most people don’t use cigarette lighters any more

(although they often still use cigarette lighter sockets). That was from

the Japanese Control Chart example we discussed earlier. The actual part

doesn’t matter to this discussion, anyway. They were, and are, achieving

the same levels of quality in producing most of their components (many of

which are still being produced for cars in 2018). I was responding to your comment that I “might want to work through that with some numbers;” the numbers are there, in the Wheeler example.Again, I am well aware that attributes charts are not useful once you get

close to a Cpk of 1; I don’t know what I’ve said that makes you think I

believe otherwise, but I must have said something, because you keep

returning to that point.My point about the Tokai Rika chart was that they achieved that kind of

quality using SPC; they just used the control chart (which they were using,

anyway, to monitor the process). If they had had an automated way to get the special cause signals, they could have done the same thing (but it

still would have been SPC). At the time, they just had a paper control

chart.0June 21, 2018 at 8:43 am #202715

Rip StaufferParticipant@ripstaur**Include @ripstaur in your post and this person will**

be notified via email.I’m not sure how they arrive at that; of course, operational definitions are important. Here’s how I did it: If you use =NORM.S.DIST(-6,TRUE) in Excel, you get 0.000000000987 for the percent of the curve to the left of 6 sigma; doubling that for the other tail yields 0.000000001973, which translates to about 2ppb. That is what I have seen in unshifted tables for years (usually expressed as “0.002 ppm”). I’d be interested to see how you and your friends calculated it.

0June 20, 2018 at 10:12 am #202704

Rip StaufferParticipant@ripstaur**Include @ripstaur in your post and this person will**

be notified via email.That sounds about right, Mike. I’m very glad to hear that the shift has become a footnote, at least here. I talked to Mikel Harry back in about 2011…at the time, I was on a couple of ANSI TAGs, including one for statistical methods. When I joined, they were putting the final touches on a DMAIC standard, in which they explained the metric and the shift, and said they were using it “by convention.” Unfortunately, I joined too late to get that statement removed. I still run across training materials that contain those tables and the shift. I won’t use them, but I still see them and talk to people who don’t know better but think that maybe I “just don’t understand it well enough.”

Mikel was happy to talk to me, but he did express some disbelief that anyone was still interested.

The Monte Carlos you are talking about are the very same simulations he was talking about.0October 12, 2010 at 8:52 pm #190864

Rip StaufferParticipant@ripstaur**Include @ripstaur in your post and this person will**

be notified via email.In his original book, Walter Shewhart concluded that normality was neither sufficient nor required for a state of statistical control. In point of fact, most control charts actually do not “require” normality. Normality is nice, but what’s important is that the data be reasonably unimodal and symmetrical. For some of the WE Zone tests, a bell shape is also useful. This is why the Binomial and Poisson distributions are so useful in quality applications…once the counts are high enough, they model unimodal, symmetrical, bell shaped distributions that work well in control charts.

You cannot assess stability with a histogram; for reasons given by others and for a more fundamental reason; a histogram is a snapshot in time. Stability by definition must be assessed over time. That’s why you need a control chart. It plots the data over time, using local dispsersion measures to estimate the within-process variation so that you can tell when excursions from natural process variation happen.

Actually, if you must test for normality with process data, you’d better check to see if the process is stable first. If the process is not stable, you can’t say anything about the shape or the distribution.0 - AuthorPosts