Informational Control Chart Interpretation - General "Rules"

If you use either CLT or Tchebychev inequality, the claim is for Shewhart charts you can chart them for any random distribution directly because the "signals" will appear.

You can -and may have to as a last resort - transform the data for capability analysis (not charting), since that is distribution specific and has no relationship to the CLT or Tchebychev inequality. Even Shewhart infers transformation masks valuable information, and to avoid it unless you need to do it.

Remember, both CLT and Tchebychev inequality require that the data is from a distribution whose variation is random. More importantly, the Western Electric rules demand it. They are looking for order in a condition that is supposed to be random.

Steve,

I understand what you are saying - and I don't disagree with you, however I am in an environment where "buisness decisions" are made with a lot of weight attributed to CPK information...

For example - If I want to ask for a tolerance change - I will always be refused unless I can first show that my CP is not capable of meeting the current tolerance.

CPK is an area where I hope to see improvement over time. At this time 1.33 is the golden rule. I would like to be able to make improvement in this area by showing that a CPK value of X with a risk assesment value of Y dictates appropriate inspection frequency.
If a non-conformance would be a very minor impact than a CPK of nine does not make sense (unless it already just happens to be at this level [which I doubt]) then a 1.33 makes sense but If I have a CPK of nine - even if a non-conformance is more critical - I should be able to reduce inspection based on the consistancy of the process.

It seems to me that the 2 issues should be kept seperate. If I need a CPK of 2.0 then I set my control limits to 6 standard deviations and make process improvements to get me there. In order to have a meaningful CPK calculation with non normal data I THINK (though I am not sure) that I should run all data transformations and find the best P value to calculate my CPK from.

Then I set my control chart to 3 standard deviations and continue with manufacturing...

By the way I am no longer in Ohio - Now located in New Jersey - not sure if that is close to you or not?


To all,

Can someone explain to me what the significance of the AD value is in an Anderson Darling test for normality. I get the P value - not so sure about the AD value?

Thanks!

Coleman Donnelly said:

Then I set my control chart to 3 standard deviations and continue with manufacturing...

Click to expand...

Control limits are derived from the data, not "set"...unless you are not using Shewhart charts and are dealing with X hi/lo-R charts which are based on non-random variation or pre-control charts.

Your tolerance should be approximately +/-4 sigma of a sample run if you want your process to meet +/- 3 sigma over time, because you need to account for lot to lot variation, such as material lot variation, etc. (see total variance equation)

Also, AIAG is a good reference for handling non-normal variation, and how Cpk does not apply to it. See AIAG PPAP 4th Edition Section 2.2.11.5. If your "business' is using it incorrectly (as in "rubber stamp"), then there is little justification for any subsequent action.

I think you may also be confused about the differences in capability studies and control charting. It may be helpful for you to post an example or two so we can deal with a practical situation....

Unfortunately, I do not have data available to post as we are currently developing the process and have not reached a point were we have ran 30 consecutive pcs to establish process stability.

I think I do understand the difference - Control chart is the tool you give your operator so he can properly control the process.

A capability study is used to determine if your process is able to meet the specification requirement on the drawing.

cp and cpk are the result of your process being controlled over time.

...or maybe I am confused?

I am trying to be set up so that when we do move forward we are on the right path

Bob,

I believe that I will be using both pre-controlled and traditional Schewart control charts on this project as I have parts/procesies that contain the dominant tool wear factor and some that do not.

I am getting hung up trying to determine if after the control charts are in place and I am using them properly as Dr. Schewart and (Dr?) Bob have advised me, is it then appropriate (or even necessary) to run test for normality using multiple types of data transformation techniques and find the best fit normal collection of data to use for calculating a cpk (which is entirely a separate process from the control chart question asked originally)?

Do you need to report your Cpk values to a Customer or are you doign this for yourself?

In general Cpk is a weak indicator of process capability and often not worth the effort if you get into transformations and Normality tests...

Process capability index is the leverage that we the developers have to ask for more tolerance on a design print.

If we have a tolerance of +/- 0.00005" and "we know" we are not going to be able to hold that tolerance then R&D says to us "How do you know you can't hold it?" our best leverage comes from the cpk value being too low.

Coleman Donnelly said:

I am getting hung up trying to determine if after the control charts are in place and I am using them properly as Dr. Schewart and (Dr?) Bob have advised me, is it then appropriate (or even necessary) to run test for normality using multiple types of data transformation techniques and find the best fit normal collection of data to use for calculating a cpk (which is entirely a separate process from the control chart question asked originally)?

Click to expand...

My suggestion is to get some run chart data, avoid measurement error, and let's take a look at them. The key for Shewhart is random variation - from the process, not the sampling, measurement and/or gage error.

I have attached a couple slides I have added to my presentation over time to help describe traditional SPC. Perhaps they will help.

And, BTW, although I collect degrees, I don't have a PhD. Started a doctorate program a while back, but work kept overwhelming it. I did finish the statistical core classes though, and they have come in handy.

Ok - we will be collecting run chart data - i will probably only get about 10 pcs before shutting down for the night... at what point is there enough run chart data to be useful?

bobdoering said:

My suggestion is to get some run chart data, avoid measurement error, and let's take a look at them. The key for Shewhart is random variation - from the process, not the sampling, measurement and/or gage error.

I have attached a couple slides I have added to my presentation over time to help describe traditional SPC. Perhaps they will help.

And, BTW, although I collect degrees, I don't have a PhD. Started a doctorate program a while back, but work kept overwhelming it. I did finish the statistical core classes though, and they have come in handy.

Click to expand...

Unfortunately, the sampling, measurement and/or gage error is part of the process! It's one of the sources of process variation. There are various methods for determining how much variation your measurement method introduces, and that variation may be one of your targets for reduction, if the process variation is unacceptable. But measurement variation is an inherent part of the the process and needs to be considered when doing statistical process control.