Dear Sirs,
Need your kindly feedback eagerly.
Is the following setence right?:-
Use SPC process monitors with one-sided specification. e.g. only USL or LSL is applicable, one-sided control limit may be used.
e.g. Control item A only have LSL. Plot Xbar-R chart, When I compute the control limit. Can I cancel the SPC rule for "One point above Xbar UCL" for this. i.e. When one point above Xbar UCL, I don't think it is a out of control situation & no need to shut down machine for adjudt machine and let the next point drop into the UCL.


Thanks and best regards
Guo Qing

Hi Guo,
If your concern for this thread is to find how to calculate the the process capability of the unilateral tolerances, I think it has been discussed a great deal on the cover, so you could Do a search for more info.
But I guess you are talking about how to carry out the SPC for the para with unilateral tolerances. Obviously I am not a stats guru, but I think this topic will be not easy to attract the guru attention. So here is my opinion for this subject.
First of all, the unilateral tolerances may cause the test data trend to be not normal distribution, and offten appear skewed one side. So we can not rely on the SPC theory, because you violate the assumptions of SPC theory. So you should evaluate the data before make any conclusion. And If it is normal data,you can use other stats tool to transform it to normal for further analysis.
Hope it could be help!
Jaco

The SPC is robust for almost any kind of distribution, but the rules don't.

To see about SPC with different kind of distributions, look for the attachement I posted in:
http://elsmar.com/Forums/showthread.php?t=8590

But I agree with Jaco, "the distribution affects SPC" and that's why I like non parametrics more.

I also agree with Guo about not using the rule out of control if the point is out in the side without spec, because the use of control rules must triger an action over the process, if no action is going to be made, forget about that rule or use it as a Christmas Tree ( and I tink that in your case <no USL> the bigger the better, so no actions will be taken).

For the other rules, be practical there is no "all fit shoe", look for the reasons why that rule say that (almost always is using probabilities of the Gausian distribution) and translate to your case (your distribution).

About control rules, the next attachement may enlight, convince you that you are on the right path or confuse more the things.

http://elsmar.com/Forums/showthread.php?t=6572
:agree1:

SPC does not rely on the normal distribution. Dr. Shewhart tested it against several skewed distributions using hat tokens. Dr. Wheeler has done similar testing with computer simulations. SPC is based upon the Tchebychev Inequality, which is nonparametric.

Now, to the original question. I would still run "regular" SPC even if the specification is one sided. This suggestion goes back to Taguchi principles - the farther you drift from the desired center of the distribution, the more costs incurred. Specifications generally don't do a very good job of capturing the losses, since they convert what is generally assumed to be a parabola to a step function.

Let's use the example of catching a commuter train (and let's assume it always leaves on time at 8AM). I want to be sure to get to the station before 8 AM. If my travel time to the station has high variance, I may need to arrive at 7:45 AM on the average to be sure I don't miss the train on unlucky days. It may help me to know that my arrival times vary between 7:30 and 8:00AM. Although I don't miss the train, I lose 15 minutes on the average every day waiting at the station. If I could better control my travel time, say within a 10 minute band, then I could arrive on the average at 7:55AM and only lose on the average 5 minutes waiting for the train.