will u please give me some details about ppk.

:bigwave: :bigwave: :bigwave: will u please give me some details about ppk.

Is some body here to ans. me.

Welcome to the Cove, Nitin Sharma. I moved your posts here, where they will probably get more attention. There are lots of threads that could possibly fit your needs. You might want to check the links at the bottom of the page, or try using the search function.

:bigwave: :bigwave: :bigwave:

Is some body here to ans. me.

Please be patient. Sometimes it takes a day or two to get an reply, while sometimes it only takes an hour or so.

There are lots of existing discussion threads about Ppk (http://Elsmar.com/Forums/search.php?do=process&titleonly=1&query=ppk).

After looking through some existing discussion threads, if you have specific questions by all means ask.

PPK is usually calculated as the minimum value of both PPU and PPL. PPU is the ratio of [(USL-processMean)/(3*Stdev)]. You can calculate this for your subgroups, or the entire process.

I'm by no means an expert in stats, but there are many different ways of analyzing data for Capability, I'd suggest, if you're looking to do a capability analysis, researching at a local public library or such, there are many good books available, etc.

Good luck :)

Hello,
While calculating Ppk, Sigma is calculated (normal maths cal., existing as a function in exel), while during Cpk Calculation Sigma is calculated through RBar / d2.

Hello,
While calculating Ppk, Sigma is calculated (normal maths cal., existing as a function in exel), while during Cpk Calculation Sigma is calculated through RBar / d2.

Very correct and welcome to the Cove !!!!!
This is indeed the difference between the 2 indicators.

Best regards,

Antoine

Hello,
While calculating Ppk, Sigma is calculated (normal maths cal., existing as a function in exel), while during Cpk Calculation Sigma is calculated through RBar / d2.

Just for completeness, I'd like to add that Cpk can also be calculated using Sbar / c4. While the Rbar approach is more common and easier to calculate by hand, the Sbar approach is actually more accurate. Since all the work is done by computers now, I haven't figured out why there hasn't been a shift from R-bar to S-bar.

Tim F

The ppk/pp indicator is called by Donald Wheeler as performance indicators and cpk/cp indicator is called capability indicators.

The use of ppk and pp is recommended for non stable processes altho some guys don't recommend any indicator because of the fluctuation of the indicator. IMHO "if you don't meassure, you don't administrate" and the ppk indicator is a way to do so.

ppk should be lower than Cpk, and according to "Quality Systems Requirements" (AIAG, 1995)

"Ongoing process performance requirements are difined by customer. If no such requirements have been established, the followinf default values apply:
....
For chronically unstable processes with output meeting specs and a predictable pattern, a Ppk value >= 1.67 should be achived."

The ppk/pp indicator is called by Donald Wheeler as performance indicators and cpk/cp indicator is called capability indicators...

ppk should be lower than Cpk, and according to "Quality Systems Requirements" (AIAG, 1995)
....
For chronically unstable processes with output meeting specs and a predictable pattern, a Ppk value >= 1.67 should be achived."

here you have fallen into the AIAG Ppk definition gap (or is it a paradox, or jsut the twilight zone?):
There is a definition within AIAG athat states that short term capability can be assessed by taking 30 units made under the same conditions (same material batch, equipment setup, operator, etc). then the 'short term' CAPABILITY can be determiend using the Ppk formula which calculates the standard deviation from all of the indiviual values*.
AIAG then goes on to recommend a Ppk of >1.67 for chronically unstable processes. the Ppk in question is the one just described. The intention here is the 'short term' variation be so small that relatively small changes from the chronically unstable process will allow enough time to be detected before bad product is manufactured. Of course, the other paradox here is that Ppk and Cpk indices have no predictive value for 'unstable' processes. And indeed, technically, a stable process is a prerequisite for calculating and reporting capability indices...




*This would be advantageous in the development stages when 25 or 30 subgroups spanning multiple material lots, setups, operators etc. would be oppressive. (The original definition of a short term capability index, Cpk uses the standard deviation based on within sample standard deviation - the intention being that thsi number would assess capability if you were to eliminate the between subgroup variation.)

One last thought is that neither of these calculations are applicable to precision machining. Capability is tolerance divide by process spread. For the uniform distribution, capability=(USL-LSL)/(UCL-LCL). Notice that there is no relation to the standard deviation there. They need standard deviation for the normal distribution to find the 'ends' of the tails. But, the uniform distribution has no tails, and it ends abruptly at the control limits.

Cp is the capability index. It measures how well your data might fit between the upper and lower spec limits.

Cpk is the centering capability index. It measures how well your data is centered between the upper and lower spec limits.

Use Cp, Cpk when you have a sample. Cp, Cpk use sigma estimator.

Use Pp, Ppk when you have the total population. Pp, Ppk use
standard deviation.

If Cp, Pp are fairly different, then you may have an
unstable process. Run a control chart on your data to analyze stability.

Can you explain this ? (Or give a reference to book or paper)

When taking samples from an uniform distribution (rolling a dice for example), the sub group averages will approximate a normal distribution (central limit theorem), the sub group will have a standard deviation.

Can you explain this ? (Or give a reference to book or paper)

When taking samples from an uniform distribution (rolling a dice for example), the sub group averages will approximate a normal distribution (central limit theorem), the sub group will have a standard deviation.

Hi Kales Veggie,

Sample averages taken from a stable process (the distribution of individuals does not have to be normally distributed) will always approximate a normal distribution. As you already stated: "The Central Limit Theorem." It is the principle on which the control chart theory is based. This principle is explained in various SPC texts.

I'm sure our fellow Statistical Cove Experts will comment as well.

Stijloor.

Stijloor,

Thanks for your reply. My question is to Bobdoehring, who states:

"For the uniform distribution, capability=(USL-LSL)/(UCL-LCL). Notice that there is no relation to the standard deviation there."

Just wondering why there is no relation to standard deviation and how UCL/LCL are calculated for a uniform distribution.

Is there some kind of article available which explains in detail how this calculation can be done?

The CpK is often called as the short term (potential) capability.
The PpK is the long term (overall) capability.

CpK capability considers the variation within subgroups, as the PpK consider the overall standard deviation.

If the within and overall variation are close, you will get more or less the same CpK and PpK as a result. But if the difference is significant then CpK will vary from PpK.

The CpK is sensitive to the order of the data, you can strongly get different CpK if the data behavior has a "time-scale trend", while the PpK will always give the same result.

Take care of that

Is there some kind of article available which explains in detail how this calculation can be done?

Hello Vyasan,

Welcome to The Cove Forums! :bigwave::bigwave:

Take a look at Post #9 (http://elsmar.com/Forums/showpost.php?p=185425&postcount=9) in this thread. It has an attachment that may be of help to you.

Stijloor.

Is there some kind of article available which explains in detail how this calculation can be done?

I utilize the formulas in the attachment if I am forced to provide a process capability index for a uniform distribution. The formulas and a decent explanation are available in:

"Process Control and Evaluation in the Presence of Systematic Assignable Cause"

by: Ashok Sarkar and Surajit Pal
Quality Engineering, Vol. 10, No. 2, December 1997, pp. 383-388
1997, Marcel Dekker, Inc. and ASQ

you can purchase an electronic copy of this article from the ASQ website; $5 for members, $10 nonmember. (this why I maintain my membership...)

Stijloor,

Thanks for your reply. My question is to Bobdoehring, who states:

"For the uniform distribution, capability=(USL-LSL)/(UCL-LCL). Notice that there is no relation to the standard deviation there."

Just wondering why there is no relation to standard deviation and how UCL/LCL are calculated for a uniform distribution.

When analyzing a process that is normal, it has a particularly difficult problem: where does it end? The curve continues well beyond the mean, but at its further reaches it represents very little probability. So, most people will consider 3 sigma from the mean "the endpoint". +/-3 sigma, then represents the distribution of the process.

For a uniform, or rectangular distribution, we know exactly where the distribution ends - right at the control limits. Sure, in real life there may may small tails for the variation of when the operator takes their readings, but we account for that by not setting the control limits at 100% of the specifications, but rather 75% of the specifications. That yields a capability of 1.33 based on the above equation.

Much of the detail is in this thread:

http://elsmar.com/Forums/showthread.php?p=187696#post187696

I utilize the formulas in the attachment if I am forced to provide a process capability index for a uniform distribution. The formulas and a decent explanation are available in:

"Process Control and Evaluation in the Presence of Systematic Assignable Cause"

by: Ashok Sarkar and Surajit Pal
Quality Engineering, Vol. 10, No. 2, December 1997, pp. 383-388
1997, Marcel Dekker, Inc. and ASQ

you can purchase an electronic copy of this article from the ASQ website; $5 for members, $10 nonmember. (this why I maintain my membership...)

From your exerpt, they sure complicate a simple calculation. It may be handy if you are calculating capability on data that you find to be uniform, but if you have a process that was controlled as a uniform process (USL-LSL)/(UCL-LCL) should suffice. It is the simple math of a rectangle.

In any event, they did illustrate that the standard deviation does not apply in the uniform distribution, so that was handy - and it looks like a good read. I will add it to the collection.

Of course, the book I usually reference for these calculations is CorrectSPC: When 'normal' is not typical, (2007), which can be found at the website listed in my avatar. :cool:

From your exerpt, they sure complicate a simple calculation. It may be handy if you are calculating capability on data that you find to be uniform, but if you have a process that was controlled as a uniform process (USL-LSL)/(UCL-LCL) should suffice.

Yeah I agree that they complicate it adn I don't do it exactly that way. BUT the article is published in a respected peer reviewed journal, so theoretical math at some level is a requirement for the authors. And knowing that the the theory is solid helps when debating the approach with 2nd and 3rd party auditors who are "check the box' types and with others who don't trust their own eyes when looking at plots of data that have an obvious message.

Yeah I agree that they complicate it adn I don't do it exactly that way. BUT the article is published in a respected peer reviewed journal, so theoretical math at some level is a requirement for the authors. And knowing that the the theory is solid helps when debating the approach with 2nd and 3rd party auditors who are "check the box' types and with others who don't trust their own eyes when looking at plots of data that have an obvious message.

That is so painfully true! :cool:

Thanks, Bev, for recommending this article.


"Process Control and Evaluation in the Presence of Systematic Assignable Cause"

by: Ashok Sarkar and Surajit Pal
Quality Engineering, Vol. 10, No. 2, December 1997, pp. 383-388
1997, Marcel Dekker, Inc. and ASQ


I am surprised this snuck past the normalcentrics at ASQ, although it was a while back. Anyway, I looked at their sample - which is the key to reviewing any journal article. When we discussed journal articles in my doctorate classes, the biggest error we found was in the sample. The statistics were calculated correctly, but the sample was suspect at best. The sample they used was diameter of bar stock. They properly discerned the sawtooth curve (congrats to them), but their error is in their measurement. The took one measurement for the diameter for each part. That creates a degree of measurement error that can mask the true variation. If you compare their Chart 1 to the black line in the attached chart, it this error may become clearer. This measurement error is random, and normal. That is what provides the tails on their distribution. Had they used X-hi/lo-R methodology, those tails would have be significantly diminished. Therefore, calculating the variation directly from their data, then trying to develop an accurate model of the variation may give inaccurate results, especially with curve fitting, etc.

Their calculation for Pp, 2h is really equal to UCL-LCL. So, we have some agreement there. Trying to find a shift of the mean in relationship to the specification limits only assumes that the control limits were not correctly centered to begin with.

I do agree with their conclusion, although I prefer to define "systemic unavoidable assignable causes" as common causes. I prefer to think that for a cause to be worth dumping into the assignable bucket, there should be a manner to remove the cause. But, that is my opinion.:cool:

yes it's not a perfect analysis but they are/were academics not practioners. The fact is that they did an appropriate analysis of the data they had and it is one of the only articles I've seen besides Wheeler's "sloped limits" article which wasn't peer reviewed.

I don't mind their definition of a systemic cause as it fits the age old commonly accepted definition of "assignable cause" yet still alludes to it's 'but you can't get rid of it' certainty. We need to remember that the terms assignable and common cause are human ioperational definitions and not laws of physics.

and I've found that Quality Engineering and the Journal of Quality Technology (which once banned articles on Cpk/Ppk because they were psuedo statistics) are not edited or controlledin any way by ASQ....

yes it's not a perfect analysis but they are/were academics not practitioners. The fact is that they did an appropriate analysis of the data they had and it is one of the only articles I've seen besides Wheeler's "sloped limits" article which wasn't peer reviewed.


I agree - I have been looking for articles on this topic, and it is the closest one I have come across. I certainly thank you for bringing it up. Even the tome "Measuring Process Capability" by Bothe, that has a massive collection of capability indices, does not quite get it right when dealing with tool wear. He identifies the sawtooth curve, but then pollutes his analysis with normal curves and also misses the measurement error from not accounting for roundness.

Although it is fun to banter about the academics of this issue, the real tragedy is people who are trying to implement SPC in precision machining, and are frustrated because they are not getting the right info. That said, I am sure trying my best to get the word out - for their good more than mine. After all, I am in assembly for the automotive industry right now. The only interaction I have with precision machining is with my suppliers. But, having been in precision machining for years, I know what they are going through and sure hope to help them use the correct SPC so that they can yield the full benefit of their charting efforts.

Luckily, ASQ chapters are very open to this concept. I have found great success sharing the concept during their meetings. :cool: